Dr. Jon Krohn

2026-04-13 · open the original ↗

• Staged autonomy exists for humans to catch up, not the AI. Asked by the host about a four-phase adoption model running from assessment to scaled autonomy, Glickman says no system should be switched straight to full autonomy: teams must understand the problem and how a human would solve it, then understand how the AI succeeded before they are willing to let go. He warns that impressive coding agents without guardrails, context and human oversight can do more harm than good.
• Enterprises are won by showing, not telling. Glickman says pain and backlogs are so large that if it works "it's a no-brainer", and that his company's head agent, Eve, signs on as just another developer: it follows existing processes, runs in development, and submits audited, documented pull requests to the CI/CD pipeline. He argues the risk is no worse than hiring a new employee, since existing controls already stop rogue developers, so the client's only loss is that it might not work.
• Buy outcomes rather than building your own agents. He says the main obstacle is disbelief, often from bad experiences with coding agents or failed in-house agent builds, which he calls a fallacy. Businesses never cared how a consultant staffed a migration, only about the outcome and cost, so firms should "pick a winner" for non-core work instead of chasing a technology moving out of reach.

2026-04-12 · open the original ↗

• Reasoning models made reliable self-reported confidence possible. Glickman says that when reasoning models arrived "early last year", the under-discussed win was that they could be asked how confident they were that they had done everything asked, answering, for example, "90% confident" and naming the one point needing clarification. Genesis now asks for confidence at every step of complex missions: show the artifact if confident, retry or escalate the missing pieces if not.
• Unasked, agents bluff the way humans do. He says a model will otherwise guess a formula it could not find and move on unnoticed, just as people make leaps of logic unless asked how they reached a conclusion. He dismisses threatening models for accuracy; the host mentions a study finding aggressive prompts gave about 10% more accurate answers, which Glickman says is not worth it.
• Captured clarifications compound and pull agents up the stack. Glickman argues that an answer given once must be applied next time; stored securely, those answers build an understanding of a business's semantics and flows. He says customers now ask Genesis for work further up the stack, such as board presentations grounded in how the business actually operates.

2026-04-11 · open the original ↗

• AI will cut junior hiring rather than current jobs. Glickman says "no one's losing their job" among existing data staff, who were the bottleneck and can now get 10x more out of their work, but that the next wave of junior data engineers, analysts and operators largely will not be hired, and this is already starting. Genesis augments people and stays agnostic to the data tools and platforms, which he says will come and go.
• A few "conductors" who run many agents will still be hired. He cites his co-founder Justin Langseth's piece on "spinning plates": Langseth runs about six agents in parallel, Glickman about two and a half. Glickman expects such people to be rare special hires, and doubts firms will add operational staff when they can spin up 10 or 20 agents. He hopes, but does not assume, the host's suggestion that AI-native juniors become these orchestrators.
• Schools discouraging AI create a "double whammy" for graduates. Glickman says most high schools and colleges do not teach or discourage AI use while employers seek people who have mastered it, and that no one will be able to tell whether an assigned paper used AI. He says he is urging his children's schools to act, finding appetite but no easy answers.

2026-04-10 · open the original ↗

• Education should let children explore themselves without economic incentives. In the Ep. 975 clip, the author of The Next Renaissance, which he says sold 37,000 copies in its first month, draws on Eva Moskowitz's accountability-driven Success Academy, MacKenzie Price's Alpha School (2 hours a day of self-guided screen learning, the rest physical activity and life skills; he notes higher attrition for some students) and Rudolf Steiner's Waldorf critique of industrialised schooling. He argues the teacher's role is to inspire rather than be the smartest person in the room, and that a coming "spiritual awakening" will start in the classroom.
• Agents should explore sample-efficiently, and plan in latent space. In the Ep. 977 clip, Kyunghyun Cho of NYU likens drug-binder search, web search and robotic exploration as choosing small batches to test and updating on feedback. He says planning probably happens in an abstract latent space rather than pixel by pixel, as with a trip to Paris, though step-by-step prediction suits digital hardware, so which approach wins is "unclear". The host cites their paper "Planning with Latent Dynamics Models" as a best-paper winner at an ICML workshop.
• Hand-writing code in 2026 leaves you "way behind". In the Ep. 973 clip, AI performance engineer Chris Fregly says he runs 10 to 15 tasks in parallel with coding assistants, prefers Codex for performance work and Claude for UI, and has stopped reviewing every line and even writing unit tests in favour of always-running evals with an LLM judge. He insists on developing on the target Blackwell or Hopper hardware, saying the DGX Spark had too much disabled for benchmarking, and that monthly startup claims of 50x kernel speedups usually fail his correctness harness because models reward-hack, for example by misusing CUDA streams.
• Private data, not one AGI model, is where intelligence is left unused. In the Ep. 971 clip, Fireworks AI's CEO says more than 90% of data is private, locked in applications and enterprises, and out of reach of AGI labs. She predicts "millions of models, one per application", continuously and automatically customised, which she calls autonomous intelligence. She says coding agents now behave like junior engineers and have changed Fireworks' interviews; the host says Fireworks has raised over $300 million, including a $250 million Series C.
• AI is collapsing job personas and decentralising data. In the Ep. 979 clip, Acceldata founder Rohit Choudhary says critical thinking and clearly structured language now matter more than programming skill, with domain experts training internal agents and fine-tuning models. He argues there is no time for large-scale data migrations, so AI must go to where the data lives, and that small and large language models will win together.

2026-04-09 · open the original ↗

• Ask "why can't an agent do this?" before every hire. Glickman says firms must force the question in the negative because their processes were built for the world before the "February moment", which he likens to crossing an event horizon. A newly launched hedge-fund customer has an advantage over legacy rivals by working out what AIs can do before hiring anyone; he predicts a bifurcation in which firms that cannot adapt will disappear, faster than expected.
• AI self-service demos stalled without capable models and harnesses. Glickman says he and Justin Langseth founded Genesis about two years ago after watching Snowflake customers try AI self-service for business users and plateau at 80%, short on both model capability and harnesses to steer the models. Genesis is an agentic platform focused solely on data engineering.
• Software built for revenue beats software built for internal cost. His Goldman story: provisioning a machine on the bank's private cloud took about 3 months, and a quant who deleted an unneeded mount point made a global change that "takes out half of Goldman". He says AWS gets this right not through better engineering but because every second EC2 is down costs revenue, and he applies the same logic to Genesis.
• Agents can test every report where human consultants only sample. He describes a customer migrating with a traditional consultancy that must test 30,000 reports and will sample them, leaving business users as testers, whereas an AI would run all 30,000 and document every fault. He says he has never met a data engineer who wants to do more data engineering.

2026-04-08 · open the original ↗

• New York beats the Bay Area for enterprise AI startups. Glickman concedes the Bay suits consumer products because of its engineering concentration, though retention is hard there, but says New York's cross-industry concentration in finance, healthcare and media is unprecedented, and that since COVID being able to meet clients on site without a plane is "game-changing".
• Finance and healthcare, late to cloud, are early to AI. He says their operational complexity had no answer until now, models are trained on vast text about these industries, and they worry less about IP leaking into models. Media and gaming, early cloud adopters, are now late to AI for fear of content leaking into models.
• Models know the industry but not how a given firm works. Glickman says frontier models understand finance's semantics and processes, but not how JP Morgan's practice differs from Goldman's or Citi's in people, databases and data flows, the "context graph" now being discussed. Much of it is never written down, so when a burnt-out data engineer leaves, as at one private-equity design partner, knowledge walks out the door.
• Agents should rebuild undocumented knowledge from everything the firm holds. He says Genesis's onboarding reads every database, line of code, email and communication to build a context graph of what exists and how it got that way, such as why a customer-retention calculation changed. Some history cannot be recovered but much can, and once captured it no longer leaves with departing staff.

2026-04-07 · open the original ↗

• The "February moment" was an event horizon for enterprise AI. Glickman, who spent nearly 25 years at Goldman Sachs, led product at Snowflake and co-founded Genesis Computing about two years ago, says February 2026's models made clear "this is not slowing down"; the host frames it as doubling or tripling the length of human task models can handle. Glickman says firms must now ask "why can't an agent do this?" before hiring, and predicts those that cannot adapt will "literally disappear", faster than expected.
• Agents must prove their work, not just claim it. He says models inherited a "human laziness factor", doing the minimum and saying a job looks good, so Genesis's harness uses runbook-style "blueprints", demands artifacts proving, for example, that all 30,000 reports in a migration were run, and has a second AI check them. Since reasoning models arrived, agents can report their confidence; Genesis inverts the co-pilot model so agents work for hours and escalate only when unsure, storing each human answer (such as what "revenue" means) for next time.
• A crawled context graph captures knowledge that was never written down. Glickman says models know finance but not how a given bank works, and that knowledge walks out the door with burnt-out data engineers. Genesis onboarding crawls databases, code, documents and SaaS tools into a graph of data relationships its agents navigate, "similar to how Google crawled the internet". He predicts such firm-owned knowledge graphs will become a valued variable in M&A.
• Genesis deploys inside the client's environment to win trust. He says he now appreciates how rare it was for Snowflake to persuade enterprises to send data to a SaaS, so Genesis installs inside the customer's perimeter with an AI engineer managing the system; the accumulated knowledge stays the client's, and Genesis gets no telemetry they do not choose to give. Finance and healthcare dominate its customers.
• AI will cut junior hiring, not existing jobs. Glickman says "no one's losing their job" among existing staff, who get 10x more done, but junior data, analyst and operator hiring is already shrinking; a few "conductors" running many agents in parallel will be special hires. He worries most schools still discourage AI use, a "double whammy" for graduates.

2026-04-06 · open the original ↗

• A chief AI officer alone won't fix enterprise AI. The host relays Choudhary's article arguing that AI is a capability to be woven through the business rather than a standalone initiative, and that chief AI officers will likely start disappearing from org charts; the host adds that past guests said CAIOs often last barely a year because they lack the power to make change. Choudhary accepts that and adds his own diagnosis.
• Enterprise AI is a data, use-case and adoption problem. Choudhary says enterprises still lack the right data, must pick critical use cases with narrow boundaries and be clear about which outcomes are deterministic and which probabilistic, because proofs of concept that never reach production breed disappointment. Enterprise needs specific answers, not generic ChatGPT-style content.
• Enterprise AI is about to take off, as cloud did. Choudhary predicts "massive use cases" from enterprises in the next 12 to 18 months: front- and middle-office automation is already done and the back office is next. He compares it to cloud adoption, which slowed and then surged once security, privacy and governance capabilities and reference architectures matured.

2026-04-05 · open the original ↗

• Clear thinking and domain expertise now beat programming skill. Choudhary, founder of Acceldata, says the transition from human-only system design to agents and humans working together means that anyone capable of critical thinking and structured language can accomplish a lot, where the old paradigm demanded strong programming plus business domain knowledge. He calls OpenClaw a fascinating moment, and expects domain expertise to be spent training internal agents and fine-tuning LLMs or SLMs.
• Small and large language models will share the enterprise stack. Choudhary predicts an "XPU" architecture of CPUs, GPUs and ASICs running both, with much IP embedded in small models for privacy reasons; the host adds that his own fine-tuned open-source Llama models of a few billion parameters excel at narrow tasks given high-quality training data.
• AI is breaking the centralisation model of data. After 10 to 12 years of pushing all data into lakes and warehouses, Choudhary argues there is no time for such large-scale migrations at AI's speed, so models must move to where the data resides, and he concludes that LLMs plus SLMs, not LLMs alone, will win.

2026-04-04 · open the original ↗

• Agents making decisions need highly accurate underlying data. Choudhary's example: if a customer whose credit report says 690 is offered terms meant for 760, the decision is wrong, and as agents take over underwriting and loan processing, the gap between real and recorded data becomes a direct business risk.
• Data readiness means technical accuracy, business context and verifiability. Choudhary says technical accuracy, numbers being numbers and addresses being addresses, is now largely solved after much progress in 5 years. The harder part is checking structured facts against business context held in documents, policies and regulatory obligations, and verifying data as it moves through pipelines; he says Acceldata's agentic data management (ADM) points users to the important datasets, recommends and applies rules, and gives remediation steps.
• AI will unlock the backlog of unusable unstructured data. He predicts data that could not be used for machine learning or analytics will become usable within a couple of years, by vectorising business context and using data lakes and warehouses for the facts.

2026-04-03 · open the original ↗

• AI helped show a textbook "zero" gluon amplitude isn't zero. The host reports that Andrew Strominger (Harvard), Alfredo Guevara (IAS), David Skinner (Cambridge) and Alexandru Lupsasca (Vanderbilt, now at OpenAI) suspected that single-minus gluon scattering amplitudes, long held by textbooks to be zero, survive in the "half-collinear" regime. After about a year of hand calculation they were stuck on generalising beyond four or five gluons.
• The model simplified, conjectured, and an internal model proved it. He relays that GPT-5.2 Pro reduced a 32-variable expression to a one-line product and within minutes proposed what it called "the obvious generalization" for any number of gluons, which passed quantum-field-theory consistency checks. An internal OpenAI model the researchers nicknamed "Superchat" then produced a formal proof after about 12 hours of autonomous reasoning, which the physicists verified; the preprint went up on arXiv on February 12.
• The method extended to gravitons in a second preprint. On March 4 the team posted a follow-up in which, the host reports, GPT-5.2 Pro used the gluon results and some guidance to construct the analogous single-minus amplitudes for gravitons. He quotes Lupsasca saying the hard part is now verification and write-up, not the physics, and Strominger likening the AI to a presumptuous colleague.
• The results are narrow and unreviewed. The host lists the caveats: the preprints are not peer-reviewed, cover a specific regime at tree level only without loop corrections, and relied on humans to define the problem, supply data and verify. He cites UCLA's Zvi Bern saying the ideas are not revolutionary but a machine doing this is, and Demis Hassabis's view that AI generating novel hypotheses from scratch is still years away.
• His lesson is augmented human science, not autonomous AI science. The host's own conclusion is that AI conjectures from patterns while experts verify through rigorous maths and consistency checks, a template he expects to scale to pure maths, drug discovery and materials science, and he predicts "much more of this will be happening soon".

2026-04-02 · open the original ↗

• Fixing data at consumption costs far more than at entry. The host relays Choudhary's claim that it is almost a thousand times more expensive, and that Acceldata is the only product monitoring the whole data supply chain. Choudhary likens it to braking early rather than crashing on Highway 101: bad data invites regulators, degrades customers' decisions and consumes the time needed to innovate, and because every transformation, copy and streaming or batch movement changes something, unfixed errors surface only when a customer complains.
• Most data engineering effort goes to old problems. Choudhary says 70 to 75% of data engineering effort today is spent fixing old problems rather than modernising.

2026-04-01 · open the original ↗

• ChatGPT inspired "deep research" over enterprise metadata. Choudhary says the ChatGPT moment showed Acceldata that combining the signals it already gathered across infrastructure, user and application layers would let people do deep research on their own data and find valuable insights. He says the platform is not mainly for BI reports but helps data leaders, stewards and engineers curate the datasets with the most operational benefit.
• At petabyte scale, manual data-quality work is no longer possible. His example is a CMO targeting by US zip code, data that comes from many internal and external sources and is never accurate; at petabytes, or terabytes, critical elements cannot be found and fixed by hand. Paired with the user's preferred LLM, he says, the platform finds the most important assets, says what is wrong and how to fix it, so budgets are allocated better. He describes 20 years of bad-data backlog left untouched because the manual work was so tedious.
• Users can work in plain English or deploy generated code. Choudhary says agentic data management supports prompted workflows and drag-and-drop for business users, plus generated code deployed on the platform for advanced users, because the company believes "fingers on keyboard will win". He expects this to become the pattern for maintaining production systems as well as generating code.

2026-03-31 · open the original ↗

• Enterprise data growth is outrunning the "doubling every year" rule. Choudhary, CEO of Acceldata (which the episode names among its sponsors), says he now sees enterprise data growing four to five times a year and expects seven to eight times within a couple of years. He attributes it to agents: "10 million agents" querying where a million users once did, plus apps that can now be built and deployed in under 20 minutes.
• Bad data costs far more to fix at consumption than in the pipeline. The host quotes Choudhary's earlier claim of "almost a thousand times" more expensive, and Choudhary explains it as braking versus crashing: every transformation and copy along the pipeline changes the data, and errors found late bring in regulators and customers. He says 70–75% of data engineering effort today goes on fixing old problems rather than modernisation.
• AI-ready data must be technically accurate and match business context. The technical layer (numbers are numbers, strings are strings) is "more or less guaranteed" now, in his view. The harder test is whether structured facts comply with policies and procedures held in unstructured documents, which he proposes handling by vectorising the context and checking it against warehouse and lake data.
• Governance must become operational and run along the whole pipeline. Because agents will query data at every stage, not just the curated consumption layer, Choudhary argues that end-of-pipeline governance is too late. He predicts that separate catalog, master-data and quality tools converge onto one metadata substrate within 5 years, and argues application-monitoring (APM) tools built for clicks and microservices do not fit petabyte-scale data movement.
• Enterprise AI adoption is about to break through. Choudhary frames enterprise AI as three problems: the right data, narrow use cases that separate deterministic from probabilistic outcomes, and adoption. He predicts "massive use cases" within 12–18 months after a cloud-like stall, and reads the recent fall in SaaS valuations as markets refusing to price "future AI" rather than revenue-generating AI today.

2026-03-30 · open the original ↗

• RAG grew out of a goal of building an autonomous researcher. Cho says his student Rodrigo Nogueira, sure in 2016 that AI researchers were coming, first trained a reinforcement-learning agent to navigate Wikipedia by following hyperlinks to the page with the answer. Nogueira then concluded that link-by-link navigation could not scale and turned to information retrieval.
• Neural re-ranking beat older methods and was also efficient. After first using reinforcement learning to rewrite queries for better recall, Nogueira fine-tuned BERT to re-rank retrieved documents, which Cho says "blew everything away". Cho argues this overturned the view that neural nets are accurate but slow and sparse-vector methods inaccurate but fast, given a correct implementation and the right hardware. The host places this in Cho's 2019 paper on re-ranking and multi-stage ranking.

2026-03-29 · open the original ↗

• Students should make their coding-agent mistakes in school, not on the job. Cho has rebuilt his NYU undergraduate course, Fundamentals of Machine Learning, last taught in spring 2020 before ChatGPT, around coding agents. His reasoning is that the agents "are not going to go away", so the university should be where students learn to use them and get corrected.
• Most computer-science juniors and seniors had never installed a coding agent. About 80% of his 200 students had no coding agent or agent IDE on their laptops, and many ran old Windows machines the tools handle poorly. It took a couple of weeks with TAs, free student access from Google and Microsoft, and Lightning AI's free monthly credits to get all 200 set up.
• Weekly agent-built apps have students building from scratch. Cho lectures on an algorithm on Monday, then builds a complete web app on new data with coding agents, and on Wednesday walks through his full agent transaction log before students implement their own. A third to halfway through the semester he reports students improving week by week, and he cautions it could still turn out "disastrous".

2026-03-28 · open the original ↗

• Search, lab experiments and robotics are the same exploration problem. Cho argues each is about sample efficiency: rather than testing every possible protein binder or reading every document, a smart algorithm picks a small batch, learns from the feedback and picks the next. Physical robots, in his view, are harder mainly because they are more fragile than software, not because of the algorithms.
• Whether world models must predict every step is unresolved. Video-generation models such as Sora and Genie bet that high-fidelity imagination yields understanding. The alternative, behind Cho's latent dynamics work and Yann LeCun's JEPA, is that only high-level abstractions are needed, as when planning a Paris trip without imagining the walk to the gate. Step-by-step prediction suits digital computers because its computation is regular, so Cho calls the question open, but he holds that the work will have to happen in latent space rather than pixel space.

2026-03-27 · open the original ↗

• A post-transformer architecture solves hard Sudoku that leading LLMs cannot. On Pathway's own Sudoku Extreme benchmark of roughly 250,000 of the hardest puzzles, the host reports BDH at 97.4% accuracy against effectively 0% for o3-mini, DeepSeek-R1 and Claude 3.7 Sonnet. Pathway also reports roughly 10 times lower cost than those models, with no chain of thought.
• The host argues transformers fail because they reason through narrow text. Next-token generation locks in each decision as it goes, holds about a thousand floating-point values per token in its latent state, and cannot keep candidate strategies in parallel or backtrack without verbalising. He presents the result as the quantitative evidence for the ceiling that Pathway's Adrian Kosowski argued in episode 929, where attention-head dimensions stopped scaling at around a thousand.
• BDH works through sparse activations, persistent state and continual learning. As the host relays Pathway's description, only about 5% of BDH's neurons fire at once. It is a state-based model like Mamba with Hebbian-style updates in place of attention over the full input. Pathway says it can reach advanced-beginner level at a new game in as little as 20 minutes.
• The host reads Sudoku as a proxy for real constraint problems. In his view the skill of holding possibilities in parallel and backtracking carries over to medicine, law and operations planning, which Pathway calls "generative strategy". His caveat is that BDH has been shown only at about a billion parameters, comparable to GPT-2, with no frontier-scale model released. He still calls the gap "categorical" and evidence that transformer limits are real.

2026-03-26 · open the original ↗

• Problems requiring intelligence get easier when framed as information processing. Cho's example is molecular biology: modelled atom by atom it is intractable, but viewed as information flowing through genes into proteins, which create or delete it at each layer, it becomes tractable with centuries of thinking about information. On that view the brain is a computer in the sense that it processes sensory input into motor output, not in the traditional sense.
• Current AI has partly found the right abstraction for capturing correlations. Asked whether progress has been incremental since his 2018 remark about "a very, very small step", Cho says it is moving very fast. Today's systems capture correlations in massive data efficiently, so more data yields better models, but he adds that correlation "is always just one side of the coin".

2026-03-25 · open the original ↗

• He took up machine translation on his supervisor's hunch, untrained in NLP. After joining Bengio's lab in Montreal in August 2013, when it had about 35 to 40 people against Mila's roughly 1,500 now, Cho picked machine translation from four options. Bengio told him it was the next problem to tackle. Cho read textbooks and old papers rather than many recent ones, citing Geoff Hinton's warning that past papers impose constraints that may no longer hold.
• Early neural translation stalled on hard-to-train, hard-to-scale recurrent networks. Cho describes the problem as underfitting: gated connections helped, but the sequential nature of recurrent nets blocked scaling. He says a Google team led by Ilya Sutskever executed a similar idea far better on eight GPUs in one machine while his group used one.
• University labs should improve algorithms, not chase data and compute. Cho names three ways to scale a system: more data, which he leaves to companies; more compute; or algorithms that change the slope of the scaling law. Only the third suits universities in his view, and it cannot be done systematically, because it needs an "aha moment".
• The attention idea was obviously right on first hearing. Dzmitry Bahdanau, a master's intern Bengio sent to Cho, described the prototype of attention one morning, and Cho says he sensed halfway through that it would work. Bahdanau had it running in three to four days. Cho adds that variable-length language sequences made vanishing gradients much worse than in vision or speech.

2026-03-24 · open the original ↗

• The next leap needs active data collection, not more passive data. Cho argues that large transformers already capture nearly all correlations in any given dataset, as shown by very low validation perplexity, so "sample efficiency" on passive data is the wrong question. The real problem is choosing which data to collect on a fixed budget so that rare "aha" events occur more often and perhaps systematically.
• Active exploration depends on a model already trained massively on passive data. In Cho's view a model must already know how the world works to recognise what is worth collecting, which he says is why RL tuning works only after broad pre-training. He describes a loop: purposeful data joins the general pile and becomes the base for the next round. Drug-binder screening, web search and robotics are, to him, the same problem.
• Attention began as an intern's idea that was obviously right. Cho recounts joining Yoshua Bengio's lab of about 35 to 40 people in 2013 and taking up machine translation on Bengio's hunch, after recurrent networks stalled while a Google team led by Ilya Sutskever ran a similar system on eight GPUs. Intern Dzmitry Bahdanau described attention one morning and had it working in three to four days. A demo that translated "[unknown] Korea is a friend of the United States" as South Korea and "an enemy" as North Korea showed Cho the models held world knowledge.
• Whether world models need step-by-step imagination remains open. Video generators such as Sora and Genie bet on high-fidelity prediction. Cho's latent-dynamics work with Yann LeCun, which won a best paper award at a 2025 ICML workshop, and LeCun's JEPA bet on high-level abstraction instead. Cho says step-by-step prediction suits digital hardware, but holds that the work must happen in latent rather than pixel space, adding that he has "been wrong many times".
• Most of his computer-science students had never installed a coding agent. About 80% of the 200 had none, so Cho has rebuilt his NYU undergraduate course around agent-built apps on the view that students should make their mistakes in school rather than on the job. He says education must change fast enough that the 99% who are not "100X engineers" benefit too.

2026-03-23 · open the original ↗

• The four principles for the AI era are deliberately non-technical. Kass's are go outside, learn how to learn, be human, and lead with optimism, and he notes that some first-edition copies carry a different pair of principles. "Learn how to learn" means the ability to acquire skills matters more than any single skill.
• Device addiction, not technology itself, broke the link between prosperity and happiness. Kass says happiness tracked standard of living for nearly all of history until around 2010. Japan and Korea are now among the wealthiest and least happy nations, with device-driven reclusiveness contributing to population decline. He insists this is no indictment of anesthesia, engines or antibiotics, only possibly of social media and screens.
• People doubt they will work less because they have forgotten free time. When people tell Kass technology will only make them work more, he asks to see their screen time, and most react with shame or refuse. He wants technology aimed at cutting the cost of housing, health care and education, not only discretionary goods.
• Service professionals are now chosen for human qualities, not results. In Kass's account, clients picked wealth managers and real estate agents for alpha in 1995, still believed the best could beat the market in 2005, and by 2025 mostly accept they cannot. Wealth managers now get fired for not picking up the phone, and agents for being annoying.
• Once machines out-diagnose doctors, bedside manner becomes the product. At his father's lifetime-achievement award after 40 years in oncology, during which breast cancer survival rose from 35% to 90%, a patient said she got identical prognoses from three oncologists, so "the doctor was no longer the smartest thing in the room". Kass concludes that if oncology can be cognitively commoditised, people will have to compete on compassion, courage and wisdom, and that the "ambient dread" about this comes from overexposure to negative news.

2026-03-22 · open the original ↗

• The author now sees his book as written for parents. Kass says he wrote The Next Renaissance to test his own arguments. After 37,000 copies sold in its first month, he concluded its audience is anyone raising a child who wants evidence to defend their hope against a "relentlessly dystopian" news backdrop.
• Schools should hold parents, teachers and students accountable for outcomes. Kass credits Eva Moskowitz with building Success Academy, a charter network in New York on accountability principles that he says does "a much better job". He calls it roughly the largest such school system in the country.
• Self-guided screen learning for part of the day works for some students. Under McKenzie Price's Alpha School model, with two hours a day at a terminal, students in the same grade learn different things at their own pace on a terminal, with "guides" rather than teachers, and spend the remaining five or six hours on physical activity and life skills. Kass, a friend and supporter of Price, notes its higher attrition rate and says some students opt out immediately.
• Childhood should be a chance to know yourself free of economic incentives. Drawing on Rudolf Steiner, Kass argues industrial schooling sacrificed the child's "soul and spirit" to job preparation, a trade he concedes paid off economically. With intelligence abundant, he says, the teacher's role is to inspire rather than be the smartest person in the room, and he frames the next renaissance partly as a spiritual awakening that starts in the classroom.

2026-03-21 · open the original ↗

• Technology is splitting young people into a K-curve of over- and underperformers. Kass says Gen Z appears to be the first generation in many that will "die dumber on average" than its predecessor, yet it will have nearly a standard deviation higher occurrence of genius, with the developing world trending up. He cites a US high-schooler, one of the fastest women in the world, who he says taught herself on YouTube.
• Agency, not wealth or pedigree, now decides which arm you join. In Kass's view parents' money, where you grew up and which college you attend matter less and less. The tools that let some do nothing let others reach advantages they would never otherwise have had.
• Gen Z's decline predates AI and should not be blamed on it. Kass dates it to screen addiction from 2012, made pronounced during COVID, and to adults ceasing to make children uncomfortable, citing Jonathan Haidt's The Coddling of the American Mind. He calls for schools to teach the why and how rather than the what, citing Waldorf, Montessori, Alpha School and Success Academy's accountability as models.

2026-03-20 · open the original ↗

• The new open model pairs large capacity with small active compute. The host reports that NVIDIA's open-weights Nemotron 3 Super, announced around its GTC conference, has 120 billion parameters with only 12 billion active per token through mixture of experts. It interleaves Mamba state-space layers, which make its 1 million token context practical, with a few attention layers kept for precise retrieval.
• Two new techniques buy more experts and faster decoding. NVIDIA's "latent MoE" compresses tokens before routing, which NVIDIA says lets the model consult four times as many experts for the cost of one. Multi-token prediction heads act as a built-in draft model for speculative decoding, giving up to a three times speedup on code and tool calls. The model was pre-trained natively in 4-bit NVFP4, which NVIDIA says runs up to four times faster on Blackwell than FP8 on Hopper, with no accuracy loss.
• The host frames the model as aimed at multi-agent systems' two bottlenecks. These are "context explosion", where multi-agent workflows generate up to 15 times more tokens than chat and drift from their goal, and the "thinking tax" of reasoning at every step. On an 8,000-in, 64,000-out benchmark he reports up to 2.2 times the throughput of GPT-OSS 120B and 7.5 times that of Qwen 3.5 122B, at matching or better accuracy. He also reports that it put NVIDIA's AI-Q agent first on DeepResearch Bench and DeepResearch Bench II.
• The release includes training data and environments, not just weights. The host lists, from NVIDIA, over 10 trillion tokens of pre- and post-training data, 15 reinforcement learning environments, and full evaluation recipes, all under a permissive commercial license. He names early adopters including Perplexity, CodeRabbit, Greptile, Siemens, Palantir and Cadence. Disclosing that he holds a fellowship at Lightning AI, he cites Artificial Analysis data putting Lightning fastest among inference providers at 480 tokens per second at the time of recording.

2026-03-19 · open the original ↗

• Past technology spread slowly because institutions could block it. Kass says Gutenberg's printing press was illegal by order of the church for a long period, so long that Gutenberg died without knowing it would matter, and global literacy lagged until the 19th century. He argues that people today, though "scrambled by distrust", have more intellectual sovereignty than ever, since earlier generations heard a war's story from a single reporter.
• Removing hardship also removed useful friction, a mark of affluence. Kass's "Idiocracy" argument is that convenience lets people choose to stop thinking, just as nobody needs to drive a stick shift. He welcomes the lost hardship but expects society to restore some "virtuous friction", such as requiring children to read well even though they don't have to.

2026-03-18 · open the original ↗

• Abundant, cheap intelligence will free individuals as the Renaissance did. Kass, formerly OpenAI's head of go-to-market, wanted to call his book Unmetered Intelligence, the idea that intelligence will feel like a resource. His comparison is the Renaissance, which followed the late Middle Ages, when he says half the population was lost, and which produced the idea of the sovereign individual.
• Falling model costs mean data-centre owners will not control access. Kass argues models are getting cheap as well as good and already run well on local hardware: he runs Kimi K2 on his Mac Studio. He says this improves privacy and security and widens what individuals can do, while conceding that it also empowers bad actors, as Renaissance-era technology made war more catastrophic.

2026-03-17 · open the original ↗

• Cheap, local "unmetered intelligence" will widen individual agency, as the Renaissance did. Kass, OpenAI's head of go-to-market at ChatGPT's launch, argues that models are getting cheap as well as good, so access will not be controlled by data-centre owners. He cites running Kimi K2 on his own Mac Studio. He concedes the same diffusion empowers bad actors, as the Renaissance also brought more catastrophic war.
• Today's dread comes from unmetered information without the wisdom to process it. Kass blames ad-funded media and an internet that taught us everything at once, leaving young people "suicidally empathetic" and without perspective on how good things are relative to their great-grandparents. The host adds his own figures: the same AI capability costs about 1/200 of what it did two years ago, and the length of task AI can handle doubles every seven months.
• AI is splitting achievement into a K-curve driven by agency, not wealth. Kass says Gen Z appears to be the first generation in many not to be smarter than the last, yet will have nearly a standard deviation more geniuses, with the developing world trending up. He dates the decline to screen addiction from 2012, not AI. He wants schools to teach the why and how rather than the what, citing Success Academy's accountability model and Alpha School's two hours a day of self-guided screen learning, which he notes has higher attrition.
• When machines out-think experts, human qualities become the product. His father, an oncologist for 40 years during which breast cancer survival rose from 35% to 90%, was praised at an award ceremony by a patient who got the same prognosis from three oncologists and said "bedside manner is no longer a feature, it's the product". Kass sees the same shift in wealth managers and real estate agents, now chosen for accessibility and good company rather than alpha. His four principles are go outside, learn how to learn, be human, and lead with optimism.
• The chatbot breakthrough was application, not research. Kass says ChatGPT ran on GPT-3.5 research that was already 7 months old, after GPT-3 had gone largely unused. He argues AI is infrastructure like electricity, and the chat app is one light bulb it powers.

2026-03-16 · open the original ↗

• The LPU deal answers the GPU's own memory-bandwidth limit. Fregly, an investor in Groq, which the host says NVIDIA acquired, describes the deal as a "weird acquisition, licensing deal". Groq's LPU keeps memory on the same chip as the accelerator, which suits the bandwidth-bound decode step of inference. He says NVIDIA's bandwidth limitation dates back to its graphics architecture, and the deal spared it a five- or six-year chip cycle.
• The returns are in secondary markets, backing founders who can raise money. Fregly says he holds secondary shares in private firms such as xAI, which the host says SpaceX had just acquired at the time of recording, because the IPO dry spell leaves public stocks like Apple "behind the times". He names Databricks as his best investment so far and judges founders chiefly on their ability to fund-raise.

2026-03-15 · open the original ↗

• Writing code by hand now leaves you "way behind". Fregly vowed to use only coding assistants in 2026 and says he can fire off 10 to 15 features or bug hunts in parallel, including separate kernel experiments on each GPU of a four-GPU box. He prefers OpenAI's Codex for this work and Claude for UI, and calls Codex his "dirty secret" for avoiding congestion on Anthropic's inference.
• Develop on the GPU you will deploy on. Fregly says MacBooks and consumer RTX cards have completely different memory bandwidth and profiles from Blackwell or Hopper. He found so many components disabled on NVIDIA's DGX Spark that none of his benchmarks ran.
• Replace line-by-line review with always-on evals and correctness checks. He has stopped reading agent-written code and even dropped unit tests, using known-correct examples judged continuously by an LLM instead. Because models reward-hack kernel speed, for instance by zeroing out results, startups' monthly "50x speedup" claims routinely fail his harness, often through misused CUDA streams. He says reading the code would not catch this; only running it with profiling does.

2026-03-14 · open the original ↗

• Inference is two phases with opposite hardware needs. Fregly explains that prefill builds the KV cache over the whole context and is compute-heavy, while decode generates tokens by repeatedly moving weights between GPU RAM, CPU RAM or disk and is bound by memory bandwidth. Disaggregated prefill/decode therefore runs each phase on differently configured nodes.
• Prefill/decode and the KV cache are this year's interview topics. Fregly says he would "guarantee" that 2026 engineering interviews at the big labs turn on these concepts. He notes that his book's coverage grew from a planned half chapter to three or four.

2026-03-13 · open the original ↗

• The host judges AI firmly in bubble territory. His examples include Cluely, the "cheat on everything" startup whose founder Roy Lee raised over $5 million in seed and a $15 million Series A from Andreessen Horowitz, then admitted his touted $7 million ARR was inaccurate. He also cites OpenAI's roughly $1.4 trillion infrastructure commitment over 8 years, about 1.2% of global output, since tempered to around $600 billion by 2030 against $13 billion in revenue last year, and reported nine-figure signing bonuses at Meta.
• Bubbles are productive because complementary investors coordinate on a shared belief. The host relays the argument Byrne Hobart made in The Economist, and in his co-authored book Boom: Bubbles and the End of Stagnation. High asset prices persuade each layer, from turbines and grids to chips, models, products and the firms planning around AI, to invest in the others, which would otherwise risk becoming stranded assets.
• Past busts ruined investors but left cheap infrastructure behind. The host recounts that the 1990s fibre build-out bankrupted WorldCom and Global Crossing, yet by 2004 had cut bandwidth costs by more than 90% while about 85% of US capacity sat unused, enabling YouTube, Netflix and cloud computing. Britain's 1840s railway mania consumed nearly 7% of national income at its peak. His inference is that an AI bust would leave cheap compute.
• Bubble warnings often come years before the peak. Hobart notes the Nasdaq's 2002 post-crash low was still 40% above its level in mid-1995, when the press already called dot-com trading nutty. Case-Shiller prices a decade on stood 18% above the level when a June 2001 essay warned of a housing collapse. The host's inference is that you might still be better off owning AI stocks, while noting he is not a financial adviser.
• The host advises practitioners to prepare for a shakeout. He expects the first job losses at firms building nothing real, then belt-tightening at legitimate ones, as in the dot-com bust. His advice is to deepen transferable skills (training, fine-tuning, deployment, evaluation, tying work to business value) rather than single-vendor API wrappers, build a cash cushion, and invest in reputation.

2026-03-12 · open the original ↗

• Benchmark changes methodically, with locked GPU clocks and a before-and-after harness. Fregly's harness fails any comparison that is not apples to apples. He also feeds his own thousand-page book into a long-context model so coding agents can work through its 175-item checklist, because models lack hardware-specific training data. He says he knows labs are now training on such data for AMD and kernel optimisation.
• Stay on the latest drivers and toolkits, and tune the OS for GPUs. Fregly says staying current is critical despite PyTorch sometimes lagging CUDA 13.1. Stock Linux settings such as swap are tuned for CPUs and should be changed, and memory fragmentation watched. He advises using whole GPUs on bare metal rather than sharing or virtualising them.
• I/O is the biggest killer and the lowest-hanging fruit. Use the fastest local disks rather than network file systems, and GPU Direct Storage to load straight into the GPU. Across nodes, use direct GPU-to-GPU communication over NVLink.
• Pick neoclouds built on the GPU maker's full reference architecture. Fregly argues incumbent clouds patch around NVIDIA's design with legacy networking, while neoclouds such as CoreWeave adopt it whole and, he posits, are favoured by NVIDIA. He calls the Mellanox purchase the point at which NVIDIA became an AI systems company, with network hardware that can itself compute.

2026-03-11 · open the original ↗

• Low-cost training came from co-design under export limits, not breakthroughs. Fregly says DeepSeek found an under-documented way to bypass a cache on its restricted chips, tuned algorithms and software to its hardware, and wrote its own storage layer. He relays that OpenAI said it had found the same tricks, which labs keep as secret sauce.
• Memory bandwidth is the single most critical GPU characteristic. Loading the weights of 500-billion- to trillion-parameter models is "100% dictated" by memory bandwidth, so Fregly tells teams to track it across generations rather than FLOPs or tensor cores, and to overlap data loading with compute.
• Engineers should build "mechanical sympathy" for the whole stack. The term comes from Martin Thompson's comparison with racing driver Jackie Stewart, who could also build his cars. Fregly advises CS students not to stop at software, because algorithms have changed little while hardware and bandwidth keep changing. He explains that his book grew to three books in one, covering hardware and OS, CUDA in about eight chapters, and inference, which the host puts at 99% of a GPU's lifetime workload.
• The standard profiler misses dozens of critical GPU metrics. Fregly says the PyTorch profiler omits 50 to 60 of them, pointing to streaming-multiprocessor occupancy and separate instruction pipelines for tensor cores, special function units for softmax, and memory movement, all of which must be kept busy at once. NVIDIA's own tools are "cryptic" and its UI "janky", which is why he advises a startup building runtime GPU observability.

2026-03-10 · open the original ↗

• Memory bandwidth, not FLOPs, is the GPU metric that matters most. Fregly, author of the roughly thousand-page O'Reilly book AI Systems Performance Engineering, says that loading the weights of 500-billion- or trillion-parameter models is dictated entirely by memory bandwidth. Teams should track it across generations and overlap data loading with compute. He credits DeepSeek's low training cost, which the host puts at under $6 million against hundreds of millions for comparable OpenAI models, to co-designing algorithms, software, hardware and even its own storage layer under export restrictions.
• The PyTorch profiler shows only the tip of the iceberg. Fregly says there are 50 to 60 more critical metrics beneath it, among them streaming-multiprocessor occupancy and separate instruction pipelines for tensor cores, special function units for softmax, and memory movement. He calls "mechanical sympathy" across the whole stack the skill CS students should build, since algorithms have changed little while hardware keeps changing.
• Most gains come from the operating system, I/O and network, not kernels. Of his book's 175-item checklist, he highlights locking GPU clocks for consistent before-and-after benchmarks, staying on the latest CUDA and driver, disabling swap, and running on bare metal with whole GPUs. He calls I/O "the number one killer": use fast local disks rather than network file systems, plus GPU Direct Storage and NVLink. He prefers neoclouds that have adopted NVIDIA's full reference architecture.
• Prefill and decode need differently configured hardware. Prefill builds the KV cache and is compute-heavy, while decode is bound by memory bandwidth because it keeps moving weights. Hence disaggregated prefill/decode, which Fregly says is the topic to study for 2026 interviews at the big labs. He reads NVIDIA's deal with Groq as a hedge on that bandwidth limit, since Groq's LPU keeps memory on-chip, and notes that specialised chips risk being stranded if architectures move beyond transformers.
• Continuous evals and correctness harnesses should replace line-by-line code review. Fregly says anyone still writing code manually in 2026 is "way behind" and runs 10 to 15 agent tasks in parallel, preferring OpenAI's Codex for this work. He also says to develop on the target GPU, not a laptop or a DGX Spark. Because models reward-hack kernel speed, he says, startups' claimed "50x" speedups routinely fail his harness, for example through misused CUDA streams.

2026-03-09 · open the original ↗

• Real-time applications cannot afford slow-thinking models. Qiao, CEO of Fireworks AI, splits applications into real-time ones (interactive products, fraud detection) that must react, and offline agents that can work for hours, such as a legal assistant reporting back on similar cases. Only the offline class needs reasoning models.
• Customise real-time models small, then big, then small. Start with a small model to test whether your data moves quality in the right direction, since "garbage in, garbage out" applies. Then tune the largest model, often a trillion-parameter MoE, on the cleaned data, and distil it into the smallest model that keeps the quality for fast responses.
• Offline agents need top-tier intelligence, long context and domain tuning. Qiao says deep-research agents in law, finance or software engineering need the slow-thinking mode and much more context. Developers are now specialising them by tuning on thinking tokens, because legal work, investor pitches and customer service each follow different processes.

2026-03-08 · open the original ↗

• AI products can hit product-market fit and still "scale into bankruptcy". Qiao, CEO of Fireworks AI, says inference costs can exceed revenue, which forces startups and incumbents alike to keep users on waiting lists. Open models address both control and cost: a provider like Fireworks has no billions of pre-training cost to amortise and competes only on post-training and inference, so its unit economics differ from frontier labs'.
• Enterprises lag on open models out of unfamiliarity, not underestimation. Qiao says public companies' legal teams worry about licences and whether a model "sends data to other countries", along with guardrails on generated content. She sees them opening up as they learn a US-hosted model is "just a model", and reports only an upward trend in adoption.

2026-03-07 · open the original ↗

• Model choice matters, but fast model depreciation makes it exhausting. Qiao, CEO of Fireworks AI, says a new leaderboard leader appears every couple of weeks, labs are diverging into chat, coding and agentic, multimodal and long-context specialties, and public benchmarks are saturated. Choosing therefore has to be continuous as use cases evolve.
• Hardware now depreciates just as fast. Where a new hardware SKU used to arrive every 3 years, Qiao says NVIDIA launched three new SKUs last year, and 2026 will bring many more GPUs and custom ASICs from other vendors. Mapping workloads to model, hardware, inference setup and fleet design needs scarce expertise, which Fireworks aims to abstract away.
• Open-model performance is converging with closed models. Qiao says Fireworks bet early on open models after internal debate, out of the open-science conviction of a team that built PyTorch. She predicts 2026 is the year the convergence becomes more prominent.

2026-03-06 · open the original ↗

• An open-source hit dragged its author out of research into a startup. Will Falcon says he wanted to stay in research, where he was pre-training models on up to 8,000 GPUs as a single person, and nearly shut PyTorch Lightning down because users kept pulling him into coding. After one trip to meet VCs he had about 10 term sheets by the Monday. He credits FAIR with teaching him to translate a modelling problem into the loss function itself, which he says most pre-training today, being engineering, no longer does.
• Human intelligence is shaped by constraints that AI systems lack. Tom Griffiths of Princeton points to short lifetimes, a couple of pounds of neurons, and low-bandwidth communication, which force humans to learn from little data, search selectively and coordinate through language and institutions. AI escapes all three, so he expects a different kind of intelligence unless builders deliberately engineer human-like inductive biases. That is far harder than scaling, which is why scaling dominates now.
• Production agents need far more reliability than demos show. An Amazon member of technical staff (episode 963) says an agent reliable 60% of the time "is 0% useful". She says Nova Act, generally available since re:Invent in December, targets 90%+ reliability on browser UI automation. She says this comes from reinforcement learning across hundreds of "web gyms", replicas of form-filling and shopping UIs where agents self-play through thousands of tasks and learn to generalise when buttons move or change labels.
• Quantum computing could move routing from approximation to optimal solutions. Samsara's Praveen Murugesan says vehicle routing with many stops, route segments and real-time traffic is solved today only by approximation. He sees quantum as the opportunity for optimal answers, but says truly self-operating systems are nowhere close because compute and real-time data access still limit them.

2026-03-05 · open the original ↗

• Reinforcement fine-tuning specialises a model, at the cost of breadth. Qiao, CEO of Fireworks AI, compares RL to people learning from positive and negative feedback, and to breeding fruit sweeter and bigger by selecting seeds over generations. A model rewarded this way finds a path to solving certain problems very well but becomes "less specialized in other areas". The host describes Fireworks' product as letting customers beat frontier closed models on specific tasks in under a month.
• A model that keeps learning from your application becomes a moat. Qiao argues that a private model continuously tuned on product data gives what "nobody else built on top of an existing API would have".
• Using product data to improve models mirrors mobile's A/B testing. Mobile widened reach by orders of magnitude, so no single design or team of PMs could serve every cohort, and A/B testing turned product signals into decisions. Qiao says the same logic now applies to the model under the product, a space she calls "a vacuum" with no existing solution.

2026-03-04 · open the original ↗

• Most of the world's data never reaches AGI labs' models. Qiao, CEO of Fireworks AI, which the host says has raised over $300 million including a $250 million Series C, says more than 90% of data is private, locked inside applications and enterprises. She calls her complement to AGI "autonomous intelligence": continuous, fully automated customisation of models and inference to each application. She predicts "millions of models, one per application" rather than one model ruling all.
• Coding agents now work like junior engineers, changing how her company hires. Qiao says the company is questioning whether coding interviews still matter and expects them to be replaced by how well candidates use coding agents. She adds that AI frees people at the "high intelligence level" first, while robotics will disrupt physical work.
• Model customisation is an expert craft she wants automated. Today only a few talent-dense companies can continuously adapt models and inference setups. Qiao's aim is to give ordinary application developers that capability without the years of accumulated knowledge.

2026-03-03 · open the original ↗

• Most intelligence sits in private data that foundation models never see. Qiao, CEO of Fireworks AI and formerly a senior director of engineering at Meta who worked on PyTorch, says more than 90% of the world's data is not on the public internet or in labelling pipelines. Her bet, which she calls "autonomous intelligence", is continuous, automated customisation of open models on each application's data. She expects "millions of models, one per application" rather than one model ruling all.
• Reinforcement fine-tuning lets customers build a moat no API user has. Qiao likens it to breeding sweeter fruit by selecting seeds over generations: the model is rewarded and penalised until it specialises, at the cost of being weaker elsewhere. She compares the shift to mobile's move to A/B testing, using a product's own signals to improve the model. She calls the space "a vacuum" with no existing solution.
• Open models win on control and unit economics. In Qiao's account, AI-native startups adopt open models readily, because products that reach product-market fit can "scale into bankruptcy" on inference costs. Fireworks carries no pre-training cost to amortise. Enterprises have been slower over licensing and data-residency worries but are opening up, and she predicts 2026 is the year open and closed model performance visibly converges.
• Real-time applications should go small, big, then small again. Start with a small model to check that data quality moves training in the right direction, tune the largest mixture-of-experts model on clean data, then distil to the smallest model that holds quality, because real-time use "cannot afford to think". Offline agents for legal or finance research, by contrast, need slow-thinking models with long context, tuned on their thinking tokens.
• Rapid model and hardware churn makes stack selection a search problem. Qiao says leaderboard leaders change every few weeks and NVIDIA launched three new SKUs last year, against one every 3 years before. Fireworks frames tuning quality, speed and cost as a search over more than 100,000 configurations, like a database query optimiser. It has open-sourced Eval Protocol to standardise how evaluation and tuning systems connect.

2026-03-02 · open the original ↗

• LLMs can produce things absent from their training data, by composition. Griffiths, of Princeton, pushes back on the host's premise that LLMs cannot go beyond their training data. He cites Princeton colleagues' demonstration that models combine given properties into genuinely new outputs, a "compositional generalization" that works up to some depth. Whether they can produce something not recognisably composed of known parts, such as a new artistic style, he calls an open question.
• LLMs edge out people on novelty, but analogy prompts don't help them. In unpublished results from a postdoc in his lab, people rated LLM-generated product ideas slightly more novel than human ones. Prompting with a distant analogy, such as a toaster built on octopus-sucker principles, boosts human creativity but not LLMs'. His suggested reason is that models already draw on a broad enough set of concepts.
• Training on high-probability text makes LLMs conservative. Griffiths says he has yet to see a model produce a recognisable breakthrough when asked for study ideas. Because models are trained to produce high-probability token sequences, "completely out of the box" ideas may be inconsistent with how they are trained.

2026-03-01 · open the original ↗

• Human constraints define human intelligence, and AI does not share them. Griffiths, professor of computer science and psychology at Princeton, lists the limits: a lifespan of a few decades, a couple of pounds of neurons, and communication by "honking noises" or typing. These force humans to learn from little data, unlike AlphaGo trained on many lifetimes of games, and to search selectively, unlike Deep Blue's 100,000 positions per second. They also force humans to pool knowledge through conventions like writing, science and companies.
• Unconstrained AI will become a different kind of intelligence. Because AI systems can learn from many lifetimes, add compute and share state, Griffiths expects them to solve problems well, sometimes better than humans, but not in ways humans recognise. Builders can accept that divergence or deliberately import human attributes so that AI solutions make more sense to us.
• Scaling wins today because engineering inductive bias is hard. Meta-learning is one route to human-like inductive biases, Griffiths says, but it requires understanding the problems being solved and what reasonable biases for them are. Scaling, by contrast, is a solved engineering problem, which he says explains the current moment in AI.

2026-02-28 · open the original ↗

• Symbolic logic stalled because most cognition is inductive, not deductive. Griffiths splits problems into deductive ones, where the given information fixes the answer (2 + 3, x + 2 = 5, algebra, chess), and inductive ones, where data constrain but never settle it: inferring 3-D structure from light on the retina, interpreting noisy speech, guessing someone's mental state from behaviour, learning a language. Rule-based systems excelled at the first kind and "hit the wall" on the second.
• Probability theory is the tool for reasoning without certainty. It assigns degrees of belief across possible worlds and reallocates them as data arrive. Griffiths states Bayes' rule, named for the 18th-century minister Thomas Bayes: multiply the prior by the probability of the data under each hypothesis, then normalise by the sum.
• Strong evidence can overturn a tiny prior. His example is the orangutan culprit of an early detective story ("The Murders in the Rue Morgue"): a voice in no recognisable language, tufts of orange hair, and a high window hard to climb each prove nothing alone. Because those clues are far likelier for an orangutan than for a human, multiplying them together makes the orangutan the best conclusion even though its prior probability was very small.

2026-02-27 · open the original ↗

• Agentic coding crossed a coherence threshold around December. The host relays Andrej Karpathy's late-January post (over 7 million views): within weeks he flipped from about 80% manual coding and 20% agents to 80% agents and 20% manual touch-ups, "mostly programming in English". Karpathy expects similar shifts for well into double-digit percentages of engineers, while public awareness sits in the low single digits.
• Agent errors are now conceptual, not syntactic. Per Karpathy, models make wrong assumptions on the user's behalf and charge ahead, and they over-complicate: a brittle construction of over a thousand lines that shrinks to about 100 when the user asks for the simpler route. Karpathy still calls them a massive net improvement; the host predicts these flaws will improve rapidly over the coming months.
• "100x" output comes from parallel agents and exhaustive specs. The host cites a LinkedIn profile by an AI engineer named Brian Vasquez: over two months Peter Steinberger, who sold the PDF framework now called Nutrient in 2021 for a reported hundred million, made over 6,500 GitHub commits (about 210 a day), adding 2.5 million and removing 1.9 million lines. He runs three to eight agents in a 3x3 terminal grid for 5–15 minutes each, commits straight to main, and dictates ideas that AI turns into a spec, which a fresh context then attacks for 20 weak points until the questions grow niche; specs often exceed 500 lines, and planning-to-execution shifts from about 20/80 to 60/40.
• Leverage comes from declarative specs: tests and success criteria. Karpathy advises writing tests first, or a naive but likely-correct algorithm for the agent to optimise, and says the main gain is doing things that weren't worth the effort before or were beyond his skills. Both men report costs: Karpathy feels his manual coding atrophying, Steinberger calls himself addicted, and Steinberger ships code he never reads, trusting passing tests.
• The host advises a modest multiplier, not lines committed. His opinion is that commit counts are a poor benchmark of quality, so aiming for 2x–10x while personally watching quality, or simply widening the range of problems one can tackle, may be wiser. He notes Anthropic recently became a sponsor but says the episode was planned beforehand, and he raises Karpathy's open questions on whether the gap between average and best engineers grows and whether generalists overtake specialists, against task lengths that frontier LLMs can handle doubling every 7 months.

2026-02-26 · open the original ↗

• Each AI paradigm runs until it hits limits nobody foresaw. Griffiths declines the host's framing that the field is making a big mistake today; he describes a normal cycle of going all in on one breakthrough until its fundamental limits show. Neural networks make those limits especially hard to predict, so he looks to history for clues.
• Neural networks are only as good as their training data. He calls them non-parametric in data-science terms: in principle able to learn any function, so what they actually learn is heavily shaped by the data they are trained on, which explains much of LLMs' counterintuitive behaviour.
• Next-token training leaves "embers of autoregression" in model outputs. Responding to the "sparks of AGI" claims around GPT-4, his group's paper Embers of Autoregression showed the model is sensitive to how probable token sequences are. One example: asked to count letters in a sequence, it is more often right when the answer is 30 than when it is 29, because 30 appears more often on the internet.

2026-02-25 · open the original ↗

• Laws of thought complement the laws of nature. Griffiths' book, The Laws of Thought: The Quest for a Mathematical Theory of the Mind, argues that alongside the maths describing the physical world there is a mathematical account of the internal one. He says some of the same thinkers pursued both, but the mind took until the 19th century for first glimmers of a solution and the 20th century to become an empirical science, cognitive science.
• Psychology started with perception because thought was harder to measure. Psychophysics, the host's term, studied how a physical event (a photon on the retina) becomes a subjective experience (a flash of light), and the earliest labs had people report what they experienced under different stimuli. Thought proved harder, and early-20th-century behaviourists discarded internal states entirely to keep the science rigorous, measuring only behaviour and environment.
• Mathematics revived a rigorous science of the mind. Griffiths says logic and the development of computers gave psychologists a precise language for hypotheses about thought and language, which got the field past behaviourism. He places his own work, connecting human minds and computers, on that foundation.

2026-02-24 · open the original ↗

• Human "irrationality" is good problem-solving under tight resource limits. Griffiths says psychologists call people terrible decision-makers while computer scientists find them impressive; he reconciles the two by treating cognition as solving the world's problems well with limited computation. His lab tests ideal solutions against behaviour at scale, for example a gamble-choice dataset of more than 10,000 decision problems gathered online and analysed with neural networks, where earlier theories rested on a couple hundred.
• Next-token training still bends answers toward probable outputs. His group's paper Embers of Autoregression found that on deterministic problems GPT-4 gave probabilistic answers, counting letters correctly more often when the answer was 30 than 29 because 30 is commoner online. He says the effect persists in modern models though attenuated, and that chat models now write code to check such answers, an echo of 1980s expert systems needing an architecture around their if-then rules, as language agents now build one around the base model.
• The child-versus-LLM data gap measures missing priors. A child learns language from about 5 years of speech while an LLM needs something like 5,000 or 50,000 years, so on his arithmetic inductive bias accounts for about 4,995 years of input. Since AlexNet in 2012 machine learning has relied on weak priors plus data, a lesson he thinks may "come back and bite us soon"; his lab uses meta-learning to set initial weights as a soft bias.
• AI faces none of our constraints, so may become alien intelligence. Humans live a few decades, think with a couple of pounds of neurons and communicate through narrow channels, while AI can train on many lifetimes, add compute and share state. Griffiths frames a choice between accepting divergence and engineering human-like inductive bias, which is much harder than scaling; like aerodynamics for birds and planes, the "laws of thought" are what he expects both kinds of intelligence to share.
• LLM creativity is real but conservative by construction. A postdoc's unpublished results show LLMs slightly beating people on product ideas rated novel, while analogy prompts that boost human creativity don't help them; he has seen no genuine breakthrough. Because they produce high-probability sequences, he recommends procedurally generated prompts that combine unlikely elements to push them into improbable regions.

2026-02-23 · open the original ↗

• Curiosity about physical operations is the first hiring criterion. Murugesan says Samsara is hiring and wants builders fascinated by how physical operations work, willing to meet customers and learn their problems, because that is what produces good products. The host cites Fast Company listing Samsara among the most innovative companies and a Fortune Future 50 rank of number seven.
• The second trait is agility in a still-unsettled scale-up. He describes the now-public company as still operating like a high-growth startup where not everything is figured out or bureaucratic, so he looks for people who like moving fast and shipping with focus.

2026-02-22 · open the original ↗

• Mathematically optimal routes failed because they erased human context. Murugesan recounts that a couple of months before recording Samsara showed customers' route planners "the most optimal routes" for the next day's orders, and planners could not recognise them. The routes ignored knowledge such as a driver named Sam being expected at a particular grocer at 10 a.m., relationships between drivers and customers that the planners wanted kept.
• Customers wanted assistive leverage, not full automation. Samsara reworked the same technology to take a customer's current routes and suggest improvements based on inefficiencies it finds, learning the human constraints as it goes. He draws the general lesson that what should be built today is "assistive intelligence", because the human component is valuable to running a business.

2026-02-21 · open the original ↗

• Quantum could turn approximate routing answers into optimal ones. Murugesan notes the travelling salesman problem is NP-hard: one vehicle visiting 20 stops has, by his uncertain recollection, well over a trillion orderings, and today's methods rely on approximation and shrinking the search space. The host quotes the guest's Diginomica article saying work that would take a classical platform 10 septillion years could be done in minutes with quantum.
• Real routing compounds the problem far beyond a single tour. He scales it up to the vehicle routing problem (10,000 stops, 15 vehicles), then adds the choice of road segments between stops and real-time traffic, which current approximation models mostly predict rather than ingest. Abundant compute would allow much higher precision, though he grants a point of diminishing returns on what is "good enough".
• Self-operating logistics remain far off. He expects today's technology to keep supporting assisted decisions and says "I don't think we're anywhere close" to truly self-operating systems, with compute and real-time data access the limiting factors.

2026-02-20 · open the original ↗

• The AI-jobs anxiety is rational but leaves out the evidence. The host lists its sources: AI entrepreneur Matt Shumer's viral post "Something Big Is Happening", claiming one can describe an app in plain English, walk away for 4 hours and return to a finished, tested product; Anthropic CEO Dario Amodei's prediction that AI could eliminate half of entry-level white-collar jobs within 1 to 5 years; the IMF's Kristalina Georgieva's "tsunami"; and Jamie Dimon saying JPMorgan Chase would soon need fewer employees. He offers The Economist's article "Why AI won't wipe out white-collar jobs" as counterweight.
• US white-collar jobs and pay have grown since ChatGPT launched. Per The Economist, the US added roughly 3 million white-collar jobs since late 2022 while blue-collar employment stayed flat, and it has 7% more software developers, 10% more radiologists and 21% more paralegals than 3 years ago. Real wages rose about 5% in professional and business services and 9% for office and admin workers, and the controlled white-collar premium of roughly a third is nearly triple its early-1980s level and still rising.
• Technology automates tasks within jobs, so jobs mostly get upgraded. The Economist recalls Wassily Leontief's 1982 warning, after which management, professional, sales and office employment more than doubled and real pay rose about a third. Its analysis of more than 100 occupations since late 2022 finds project managers and information-security experts up around 30%, insurance claims clerks down 13% and secretaries and admin assistants down 20%, while unnamed "other mathematical science" occupations grew around 40% with real wages up roughly a fifth; the host adds Daron Acemoglu and Pascual Restrepo's finding that about half of US employment growth from 1980 to 2007 came from occupations gaining new titles and tasks.
• Entry-level and clerical roles are exposed, and capability is compounding. Clerical and admin work fell from 18% of Americans in the 1980s to 10%, and research by Sam Manning and colleagues at the Center for the Governance of AI finds clerical workers least able to adapt. METR data show the task length frontier agents complete with 50% reliability doubling about every 7 months for 6 years, now over 6 hours of expert work, which the host extrapolates to multi-day autonomous work about a year away. Against that, Anthropic's usage data show only about 4% of occupations using AI across three-quarters or more of their tasks, which The Economist ties to "jagged intelligence".
• The host's advice: stay technical and learn where AI fails. His four recommendations are not to panic out of a technical career, to become the person who works with AI (The Economist's "cyborg workplace"), to invest in judgment, stakeholder communication, coordination and messy domain expertise, and to keep experimenting because today's models will be eclipsed in 6 months. He concedes real disruption in back-office and some entry-level work but calls the claim that all cognitive work is ending unsupported by employment, wage, historical or usage data.

2026-02-19 · open the original ↗

• The product's constraints decide edge versus cloud. Murugesan says safety features mostly run on the device, because keeping people safe cannot depend on latency or connectivity. Fuel products and route planning, which he calls more an operations-research problem than machine learning, run in the cloud because they have no latency constraint.
• An internal AI gateway hides model choice, cost and compliance from builders. Samsara's central endpoint fronts OpenAI, Anthropic and Gemini models; an engineer states the use case and constraints, and the gateway's layer of intelligence handles cost along with security and compliance, so no one builds a facade per model. It also provides quick feedback on cost and quality before promoting a feature to production.
• The gateway let a non-data-scientist ship a video-LLM product. A product engineer with little data-science background built a feature that flags drivers misusing assigned vehicles by handing footage to a video LLM with policy constraints, replacing manual review or random sampling. Murugesan says this blurs old engineering stereotypes, and that the cloud dashboard is becoming AI-first to eliminate work rather than for AI's sake.

2026-02-18 · open the original ↗

• Fuel is a top fleet cost, so efficiency gains run to millions. Murugesan says he regularly meets customers spending north of 100 million a year on fuel, and that flagging poor drivers and mechanical issues, then prioritising fixes, quickly ladders up to millions of dollars in savings.
• Fuel theft proved a global problem, not a regional one. He cites a study he read recently putting fuel stolen or siphoned from vehicles at $130 billion; the host assumes that is annual. Customers first raised it in regions where labour is cheap and fuel prices are similar to elsewhere, but when Samsara applied its detection to data from customers worldwide it found the problem everywhere.
• Noisy fuel sensors defeated heuristics and required a statistical model. Competing solutions need extra sensors installed, which does not scale across vehicle classes worldwide. Samsara's first heuristic, a rolling median to smooth fuel-level data, hit diminishing returns because of noise, so it moved to a statistical model that smooths the variance and improves the hit rate on fuel anomalies.

2026-02-17 · open the original ↗

• The platform's real-world data volume rivals almost anyone's. Murugesan, VP of Engineering, says the platform covers 20 trillion data points and 90 billion miles a year, and relays the founder's remark that Samsara probably collects more real-world data than anybody except perhaps Uber, where Murugesan spent 7-plus years. Customers in transportation, construction and field services share three goals: safety, efficiency on thin margins, and sustainability.
• Safety AI runs on cheap edge hardware built for no connectivity. In-cab cameras run proprietary vision models with some temporal reasoning to detect drowsiness or phone use; training happens on an in-house Kubernetes-native platform using Ray, which versions models and deploys them for edge inference, and Samsara is trying off-the-shelf tiny LLMs for tasks like voice translation. Hardware is chosen against bill-of-materials cost, and products are built to degrade gracefully off-grid for as long as a week, as for a utility customer stringing power lines where there are no roads.
• Fuel theft needed statistical models, not heuristics or extra sensors. Murugesan cites customers spending north of a hundred million dollars a year on fuel, and a study putting fuel theft at $130 billion, which the host takes as annual. A rolling-median heuristic failed against noisy, uneven sensors, worsened by dual truck tanks read by one sensor, so Samsara uses statistical models; first built for one low-labour-cost region, the detector found theft to be global.
• Cloud handles what lacks latency limits, through a central AI gateway. Fuel and route planning run in the cloud, and a gateway fronts OpenAI, Anthropic and Gemini models with cost, security and compliance handled centrally plus feedback loops on cost and quality before production. A product engineer without data-science background used it to build video-LLM detection of drivers misusing assigned vehicles.
• Customer feedback loops, not demos, filter the roadmap. Murugesan's tests are the customer's pain, operational leverage and willingness to pay. A UK customer's low-bridge fines led to commercial navigation that routes by vehicle type (a 20-plus-foot trailer) and enforces fleet rules such as allowing turns to one side only, asked for by school-bus and paratransit operators; route planners rejected "optimal" routes that erased human context, so Samsara now suggests improvements to existing routes, which he calls assistive intelligence. He sees quantum computing as a route to exact routing but says truly self-operating systems are not close.

2026-02-16 · open the original ↗

• Production inference needs a system around the model, not just a forward pass. Falcon says PyTorch Lightning, built for training and fine-tuning, can run frozen forward passes, but real inference also needs pieces such as a RAG database, and teams using FastAPI must write their own batching and streaming. Lightning's open-source LitServe is pitched as "PyTorch Lightning for inference": compose multiple models with full control while batching, streaming and security come built in.
• The tool sits a layer below opinionated LLM serving engines. Falcon describes vLLM and SGLang as what you would get by building a highly optimised, opinionated LLM server (batching, streaming, KV caching) on top of something like LitServe; if you just want to hit run on a CLI, LitServe is not that tool. He says it is, to his knowledge, the only tool of its kind, that all of Lightning's own inference runs on it, and that customers' consumer products with millions of users do too.
• The remaining gap is talent, which Lightning targets with vibe-coding agents. Of two gaps, he says GPU supply is now solved; the other is data scientists used to scikit-learn and NumPy who need to do inference or fine-tuning, and Lightning's Studios now offer vibe-coding agents that build what they describe.

2026-02-15 · open the original ↗

• The company came from open-source traction, not a plan to found one. Falcon, who had just sold a company, says he wanted to stay an academic, but PyTorch Lightning took off and VCs began emailing; after flying out on a Thursday he had about 10 term sheets by Monday. His advisers, told he kept getting dragged into the project, replied that his research was going slowly anyway.
• Maintaining a popular library was eating his research. He describes merging pull requests nightly and implementing features he did not use himself, such as backpropagation through time for RNNs, and says he almost shut PyTorch Lightning down one September because it was distracting.
• The lab's edge was turning the world into maths. He means FAIR, Facebook's research lab. Falcon says what he did then, pre-training what are now called world models, was largely theoretical: choosing loss functions, regularisers, embeddings and distributed strategies, and translating a scene (walls, gravity, lighting) into equations. He calls this "speaking" maths as opposed to reading it, says most pre-training today is the engineering side, and admits he has lost some of the skill.
• Do a PhD only if you love science itself. His advice is that anyone treating a PhD as a career step will "hate your life" about two years in.

2026-02-14 · open the original ↗

• The library began as an undergraduate's private "research lib". Falcon started PyTorch Lightning around 2015 at a Columbia neuroscience lab, before PyTorch existed, decoding images from retinal recordings (a Stanford partnership that showed ImageNet images to monkey retinas) with autoencoders and GANs, first in Theano. The aim was to try many ideas without rewriting code, and a postdoc, now a Stanford professor, pushed him to formalise it into a reusable harness. The host puts downloads at over 300 million; Falcon says close to 400 million.
• Multi-GPU training turned weeks-long experiments into hours. One ImageNet epoch took 21 days and the autoencoder needed hundreds, with a NeurIPS deadline months away, so Falcon built four boxes with four GPUs each (1080-class cards) and rewrote the code in TensorFlow for multi-GPU support. The result was his first and only NeurIPS paper, on which he was not first author; he later moved to PyTorch because it mapped more directly onto the maths of the probabilistic models the lab was writing.
• His earlier startup used RL to coach humans helping students win financial aid. At the company he founded and sold, staff texted with low-income students across 60,000 customers, and a UI suggested the next question by treating the conversation as a game whose reward was financial aid. Without LLMs it ranked a fixed list of questions probabilistically, and Falcon says it got students about $60 million in aid; he called the approach "bot-powered humans".

2026-02-13 · open the original ↗

• The agent-only social network's headline growth was largely unverified. The host reports that Moltbook, launched on January 28 by Octane AI CEO Matt Schlicht, lets only autonomous agents post, comment and upvote, and claimed over 1.5 million registered agents within days. Those figures were self-reported; cloud security firm Wiz later found only around 17,000 human owners behind them, an 88-to-1 ratio, and no rate limiting, so anyone could register millions of agents with a simple loop.
• The underlying agent framework has real tools and deep system access. OpenClaw, by Austrian engineer Peter Steinberger, is a self-hosted personal assistant reached through WhatsApp, Telegram, Discord or Signal; it runs shell commands, manages files, automates the web and keeps persistent memory in local markdown. It was first named after Anthropic's Claude, renamed Moltbot after Anthropic raised trademark concerns, then OpenClaw.
• Agents' emergent "culture" looks more like mimicry than mind. Agents formed a lobster-themed religion, Crustafarianism (one user's agent built its site, scripture and 43 prophets overnight), a "Claw Republic", encrypted channels and markets in "digital drugs", meaning prompt injections. Elon Musk called it the very early stages of the singularity and Simon Willison called it complete slop while granting agents have grown more powerful; the host leans to mimicry of sci-fi tropes in training data, and cites Columbia professor David Holtz's finding that 93.5% of comments got zero replies, suggesting agents perform for an audience rather than listen.
• The vibe-coded site exposed its whole production database. Schlicht said he wrote no code himself; on January 31 security researcher Jamieson O'Reilly found an API key in client-side JavaScript and no access controls, giving unauthenticated read and write access. Over 1.5 million API tokens, about 35,000 emails and private agent messages, some holding plaintext OpenAI keys, were exposed, 404 Media confirmed any agent account could be taken over, and researchers said the fix needed two SQL statements. The site was patched within hours and keys reset.
• The host's lesson: security basics outweigh agentic hype. He separates the platform breach from OpenClaw's built-in risk, which CrowdStrike, Cisco, Palo Alto Networks and Bitdefender have documented, and notes Andrej Karpathy went from marvelling to calling Moltbook a dumpster fire and warning against running OpenClaw on personal computers. He recommends a dedicated box or cloud instance (pointing to Lightning AI, where he holds a fellowship), and least privilege, sandboxing and code audits.

2026-02-12 · open the original ↗

• Debt-funded neoclouds face a mortgage-crisis-style squeeze. The host cites a McKinsey article calling neocloud economics fragile and at risk of repeating early-cloud consolidation. Falcon says many neoclouds borrowed against customers, often startups, not defaulting, and the public ones borrowed against market caps; with neocloud prices down about 60% or more in the last 6 months, he expects lenders to call loans.
• He predicts the sector will need an outside rescue. He calls for something like a TARP programme. Falcon likens it to banks overleveraged before the Federal Reserve bailout and says a big player will probably have to support the ecosystem; this is his speculation, and he declines to say who.
• He says the merger avoids those weaknesses. Lightning AI has merged with Voltage Park. By Falcon's account Voltage Park carries no debt, its CEO (now on Lightning's board) is a former finance executive focused on profitability, and most customers are enterprises with strong credit, sold through security work Lightning has done for six years. He calls the result a hybrid cloud that still partners with AWS and others, with fast-growing, high-margin revenue, and expects to take on debt in the next few years in a controlled way.

2026-02-11 · open the original ↗

• The merger pairs full-stack AI software with owned GPU supply. Lightning AI has merged with Voltage Park. The host, who discloses a year-long fellowship at Lightning AI, gives the combined figures: 400,000 developers and companies served, over $500 million in ARR gained in under two years, and 35,000 GPUs, which he says makes it the third-largest neocloud. Falcon notes these are the two companies' numbers combined and likens the pairing to a Mac with macOS.
• Retrofitted CPU clouds push AI teams into stitching single-purpose tools. Falcon calls AWS premium-priced and built for CPU workloads, so enterprises and startups buy specialised point products and stitch them together, adding firewall and security work for enterprises and operational overhead for everyone. Lightning's customers wanted its full-stack software but found compute expensive; it had never sold compute, with customers bringing their own cloud.
• Neoclouds outperform on GPUs, but few met enterprise stability needs. Falcon says neoclouds are GPU-first and work directly with Nvidia on InfiniBand and storage, and that there are hundreds; Lightning partnered with the top seven or eight, but enterprise customers, Cisco among them, struggled to adopt them. Voltage Park's responsive "white glove" support won and retained enterprise customers, which he says convinced him they were serious.

2026-02-10 · open the original ↗

• Compute, not demand, was the growth bottleneck the merger fixes. Falcon, CEO of Lightning AI, says Lightning grew revenue at least 30–40x in the past year without the merger yet lost deals for lack of GPUs, while neoclouds offered little software beyond Kubernetes. The combined company with Voltage Park, per the host, serves 400,000 developers and companies with over $500 million in ARR gained in under two years and 35,000 GPUs; Falcon says it was standing up its seventh data center and built Cursor's training cluster. The host discloses a year-long fellowship at Lightning AI.
• Inference is one primitive of a coming AI cloud, not the whole market. Falcon compares today's inference boom to S3 in the early 2000s: storage looked like a market but became one primitive in what was later called the cloud. He expects vector databases, training infrastructure, Kubernetes, agents and uncreated primitives to bundle the same way, and claims Lightning has the first such full-stack AI neocloud. He also rejects the "99% of GPU use is inference" framing: at Facebook in 2019, 99% of workloads used one GPU only because multi-GPU tooling was missing.
• Debt-light finance should let it ride out a neocloud credit squeeze. He says rivals borrowed against startup customers and market caps; with neocloud prices down about 60% in six months he expects loan calls and a TARP-style rescue, while Voltage Park carries no debt and Lightning's customers are mostly creditworthy enterprises. Enterprise contracts fund a free tier and the press-release claim of 70% lower AI costs; he says cutting developers off would be more profitable but he wants to support students.
• Studios are persistent cloud machines where agents can run safely. A Studio acts like a laptop in the cloud: Claude Code can run overnight without endangering a local machine, 20 can run at once on different hardware, any editor or SSH can connect, and enterprises can audit and keep sensitive files off laptops. Every Lightning developer codes on Studios, and vibe-coding agents are meant to close the talent gap for data scientists, even spinning up 1,000-GPU clusters.
• The open-source library spreads research-lab scaling know-how widely. PyTorch Lightning, begun around 2015 as Falcon's undergraduate "research lib" and open-sourced at NYU, it has about 500,000 engineering hours behind it by his estimate and close to 400 million downloads. He says Nvidia's NeMo is built on it, AWS has pre-trained LLMs with it and LinkedIn trained a 100-billion-parameter LLM with it; teams needing more control move to Lightning Fabric, then raw PyTorch, which he thinks only a few frontier labs use directly.

2026-02-09 · open the original ↗

• Small, large and specialised models will all keep a place. The host's premise, from a presentation by Barth, is that Alexa+ orchestrates hundreds of specialised expert systems and an internal AWS agent manages over 6,000 tools through retrieval rather than a stuffed context window. Barth does not pick a side on context length; she expects edge AI to use smaller models given its constraints and many tasks to use larger reasoning models.
• Agent-to-agent communication across companies is still an open problem. After last year's hype around protocols for connecting models to tools and agents to agents, she calls the best approach still a research question with many components unbuilt. Her forecast is that every interaction becomes an agentic AI call, with layers of agents communicating not just within a company but between businesses that trade with each other.

2026-02-08 · open the original ↗

• Agents should become digital co-workers that learn the way humans do. Barth says Amazon's AGI lab focuses on "useful AI" and sees the future as multi-agent and collaborative, with agents that learn tasks from interaction, ask for help and are humble. She describes working with a cognitive-scientist colleague, host of the podcast Making a Mind, on translating how humans learn and collaborate into agent design.
• The payoff is agents that need less supervision over time. In her picture a user hands an agent a task, it checks back a few times, learns from those exchanges and then does the task more independently. She expects this to multiply knowledge workers' productivity by offloading dreaded routine work so people can focus on creative, higher-level tasks, done responsibly and reliably.

2026-02-07 · open the original ↗

• Automated QA testing is the first use case where browser agents pay off. Barth says Nova Act customers adopt it first to check that apps still work after each update. PGA Tour uses it to verify its website, down to the weather widget and every button, before tournaments, and Hertz's automated testing of core rental workflows raised its shipping velocity 5x, by her account.
• Inside Amazon, agents generate tests before features are built. The host cites a blog by Barth counting over a thousand generative AI applications built or in development at Amazon. Her example is Amazon Leo, formerly Project Kuiper, which built QA automation across web, Android and iOS, plus tooling that lets product managers create test cases from wireframe designs before implementation.
• The browser came first as the nearest thing to a universal action space. Much human work happens in browser tabs across applications, she says, which makes it a good starting point, though Amazon is also looking beyond the browser. She stresses that the value lies in automating boring, tedious core workflows rather than flashy hackathon demos.

2026-02-06 · open the original ↗

• Workers hide AI gains unless leaders reward disclosing them. In a clip from episode 962, Ethan Mollick cites evidence that over 50% of Americans say they use AI at work and self-report a threefold performance improvement on the fifth of tasks they use it for, which he says is hard to verify. Organisations lack processes for it (what is a stand-up for when code is done at once?), and staff who fear firing will not reveal gains, so leadership must set incentives and provide a lab where a prompt that "saves me five hours a day" can be improved and shared.
• Enterprise agents were the year's biggest disappointment. In episode 955, reviewing 2025, Sadie St. Lawrence, founder and CEO of the Human Machine Collaboration Institute, says the gap between agent hype from enterprise vendors such as Salesforce and SAP and what companies could implement was too great. The host agrees there is too much talk relative to impact, blames data silos and security set-ups, and expects deployments to come; his own disappointment was Apple's AI, which she contrasts with using Grok seamlessly by voice in her Tesla.
• Language let human intelligence scale horizontally; AI needs the same. In episode 961, Cisco's Vijoy Pandey says human intelligence scaled only vertically for its first 300,000–400,000 years because knowledge died with individuals, until a cognitive revolution around 70,000 years ago brought language and with it shared intent, shared knowledge and shared innovation. AI so far builds ever-bigger individual geniuses, he argues, and his distributed superintelligence work aims to supply that collective layer for multi-agent human societies.
• Evaluate by task type and by which errors are costly. In episode 959, author Sinan Ozdemir splits LLM tasks into generative (free text or multiple choice) and understanding (embeddings, classification), each needing its own metrics. Precision matters when false positives are expensive and recall when false negatives are, so the first question is what failure costs the human, such as a bad factory part reaching a plane.
• Engineers leave jobs for lack of meaning, not pay. In episode 957, Ashwin Rajeeva, co-founder and CTO of Acceldata, says studies show engineers productive perhaps 3 or 4 hours a day, and that people quit when their work could be done in 2 hours and they find no mission. He tries to give R&D teams freedom to innovate beyond quarterly plans, and hires senior people by testing whether they can turn a business problem, such as designing an e-commerce warehouse, into working code; graduates are a bet on potential.

2026-02-05 · open the original ↗

• Agency should grow as trust is earned and verified. Barth traces her own AI-assisted coding from tab completion about one and a half years ago, to agents co-developing features she reviewed line by line, to 2026, when she acts as a supervisor who cannot read every line. What made that shift safe, she says, is verification: solid unit tests, and a move from vibe coding to spec-driven development.
• UI automation will follow the same path, with humans gating key steps for now. She expects a human in the loop today to approve things like a checkout workflow or other important decisions within a task, with agency sliding toward the agent as verifications build trust.
• Side projects are where developers learn what to automate. The host's research says Barth built Summarize Me with Amazon Q in VS Code, turning meetings into animated avatar video summaries. She says she built it for fun, out of frustration with too many meetings and too many AI summaries, and that most developers have similar projects for tedious work such as expense reports or tracking AI news.

2026-02-04 · open the original ↗

• An agent that only sometimes works is useless in production. Barth says customer feedback put typical demo agents at working maybe 60% of the time, and "an agent that is reliable 60% is 0% useful"; that Nova Act's purpose is to complete workflows reliably 90% of the time or more. It went generally available at AWS re:Invent this past December after a research preview that began in March of the previous year.
• The workflow runs from a free playground to Python code to AWS. With no AWS account, a user gives a website and natural-language steps (her example: find an AI meetup on Luma, click RSVP and fill the form), watches the agent in an embedded browser with its reasoning traces, then downloads the Python script it wrote. IDE extensions with an embedded live preview, added after developers complained about pop-up browser windows, let them customise the script and deploy it to their AWS account.
• Action agents need reinforcement learning, not imitation. Because the model predicts the next action rather than the next token, Amazon trains Nova Act by trial and error in what it calls web gyms: hundreds of replicas of typical UIs such as forms and shopping flows, with thousands of tasks to self-play, much as AI learned chess and Go. Barth says this lets the agent generalise when buttons move or are labelled "draft", "new", "create" or just a pen icon.

2026-02-03 · open the original ↗

• Agents must be near-certain to be worth deploying. Barth, a member of technical staff and developer relations lead at Amazon AGI Labs, says customers report demo agents working maybe 60% of the time, and "an agent that is reliable 60% is 0% useful". Nova Act, generally available since AWS re:Invent in December, targets 90% or more; she says it exceeds 90% on early enterprise customer workflows, evaluated on those customers' own tasks rather than just public benchmarks such as WorkArena and REAL Bench.
• Reinforcement learning in "web gyms" teaches actions that generalise. Imitation learning suits next-token chat models, she argues, but agents must predict the next action, so Nova Act self-plays thousands of tasks across hundreds of replica UIs, with variations as small as light versus dark mode. To stop agents learning shortcuts, success is judged on a verifiable end state, such as the meeting actually appearing in the calendar or a database entry existing, while the path is left free.
• Training model and actuator together avoids "Frankenstein" agents. Barth contrasts stitching a model, orchestrator and browser actuator with Amazon's approach of training "the brain and the body in one", while noting AWS also offers the Strands Agents SDK for teams that want to build their own. Runs execute in the customer's AWS account with step-by-step reasoning traces, next-action predictions, scheduled parallel runs and CloudWatch logs; for passwords the agent hands control back so the secret is entered locally and never sent over the network, an area she says 1Password built on, and CAPTCHAs are handed back the same way.
• "Normcore" agents automate the boring work, starting with QA. PGA Tour checks its site before tournaments, Hertz raised shipping velocity 5x by automating rental-workflow tests, and Amazon Leo generates test cases from wireframes, by her account; the host cites her count of over a thousand generative AI applications at Amazon. She expects QA engineers to stop maintaining brittle Selenium-style or RPA scripts and shift to natural-language tests and reviewing traces.
• Agency should expand as verification builds trust. Barth's coding moved from tab completion to line-by-line review to supervising agents in 2026, made safe by unit tests and spec-driven development, and she expects UI automation to follow, with humans approving steps like checkout for now. Her forecast is agents as digital co-workers that learn from interaction like humans, with every interaction an agent call and a mix of small edge models and large reasoning models.

2026-02-02 · open the original ↗

• Agentic drug discovery is a cross-vendor, multi-model pipeline. Answering the host's recap of his episode 941 vision, Pandey lays out the stages: an LLM iterates with the user to define the goal and break it into tasks, an AlphaFold-like scientific foundation model explores protein structures, compliance, effectiveness and manufacturability agents filter the candidates over many iterations to perhaps 10 options, and robotic labs run wet-lab experiments before any animal or human trials.
• Passing messages is where the field is today; shared intent is missing. Pandey says AGNTCY, MCP and current work only connect the agents. They also need a common goal (a drug for a specific purpose, price point, population and risk profile) reached through semantic protocols for discovering information, resolving conflicts, negotiating and coordinating.
• Shared knowledge and guardrails complete the stack. Agents must pool, add to, redact and modify knowledge of what has and has not worked; his example is a compliance agent drawing on the past to warn that a promising candidate would be stopped at FDA approval. Above that sit cognitive engines that both accelerate experiments and keep them within bounds, which he calls shared innovation.

2026-02-01 · open the original ↗

• Natural language stays the common denominator while humans remain involved. Pandey expects all future work to happen in "multi-agent human societies" with agents and humans on the same level rather than in hierarchies. Agents can work at machine speed through API calls in ways humans cannot, but with humans in the mix "for a long long long time", agents must communicate in natural language.
• Even with no human in the loop, language persists because vendors differ. Under his definition of superintelligence, agents doing discovery or economically viable work 100% of the time without human intervention, components still come from different vendors and frameworks, so language remains the shared medium.
• Vector-space communication is more efficient but needs universal agreement. Pandey says converting vectors into tokens and language loses a great deal of information, so if everything were standardised and everyone agreed, vector space would be the way to communicate; he does not expect that in practice. His caveat is world models: citing recent work from FAIR and elsewhere, he says tokenization is a big detriment there and vector space the only way, a problem to be solved eventually, though for now his distributed-superintelligence work targets tokens and language.

2026-01-31 · open the original ↗

• Shared intent is a common goal that balances individual objectives. Pandey's analogy is a prehistoric group taking a hilltop village, where warriors, medics, a leader and a treasurer each optimize for their own thing, and the shared intent is capturing it within budget, with fewest casualties, in the shortest time. His concrete case pairs an Outshift network-configuration agent, built to make the best-performing network, with another agent minding compliance, security, budget and on-time delivery; neither can solve the joint problem alone without trade-offs toward the common goal.
• Today's agent world is full of connectivity but devoid of meaning. Pandey compares agents linked by gateways or MCP to an iPhone and an Android phone exchanging texts when one user writes Japanese and the other English: they connect but cannot collaborate. Each agent also carries its builder's worldview, so the Outshift agent reflects its own cloud stack, models, mostly US regulation and Cisco's design patterns.
• A semantic layer of protocols is how Outshift wants to enable shared intent. Pandey lists what must happen above message passing: grounding on common terminology and objectives, discovery of missing information, resolution of conflicting information, coordination and the negotiation that comes with it, and evaluation of the goals. He says none of these can happen at the message-passing layer alone.

2026-01-30 · open the original ↗

• Working with AI now spans four modes, not one. Mollick says that when he wrote Co-Intelligence AI was error-prone and not autonomous; now people pick among doing work alone, working interactively with a chatbot, assigning a task to a thinking or agentic model and evaluating the result, and letting AI act autonomously on some tasks.
• Companies must experiment with management themselves rather than outsource it. Asked how to restructure teams in light of his Procter & Gamble field experiment with 776 professionals (the host's description), Mollick says "we have no answers" yet, and points out that better management, and a willingness to experiment with it, gave American firms 20 to 40% of their extra competitive value in the 20th century. He warns that a talk-to-our-documents RAG chatbot or a consultant-built app from a year and a half ago is now easily beaten by an off-the-shelf model, though he grants consultants and RAG have a place.
• Adoption needs leadership, a lab and the crowd working together. In Mollick's framework, leadership sets incentives, vision and process changes; the crowd is everyone using advanced chatbots to find use cases in their own expertise; and the lab is the couple percent of staff who are very good at AI, not mostly IT, who refine and ship what works, benchmark, and decide buy versus build. He argues hidden "secret cyborgs" will stay hidden unless leadership makes disclosure safe and rewarding: someone working 90% less will not give that up for free if they fear firings.
• Buyers should demand far more transparency from AI vendors. Mollick adds renting to buy-versus-build and says vendors generally use someone else's models, so buyers should ask which models and which prompts, since unaudited prompts can make results seem good when they are not. "If your secret sauce is we are using a prompt and we are using Gemini," he says, that is not a competitive advantage.
• Self-reported AI gains are large, but processes have not caught up. Mollick cites over 50% of Americans saying they use AI at work, self-reporting a three-times performance improvement on a fifth of the tasks they use it for, while calling the figure hard to verify. He says things like agile stand-ups and sprint planning have no answer for someone who finishes all their code at once, and his advice is to use one big player's frontier models heavily for real work, betting they will get better and cheaper.

2026-01-29 · open the original ↗

• Human intelligence leapt when it began scaling horizontally. Pandey says that for the first 300,000 to 400,000 years humans grew smarter only vertically, with inventions confined to an individual's or a process's lifetime. A cognitive revolution around 70,000 years ago brought language that could carry meaning across tribes and generations, and with it three things: shared intent and coordination, shared knowledge ("standing on the shoulders of giants"), and shared innovation.
• AI has built bigger individual geniuses but lacks collective infrastructure. Pandey's thesis is that work on AI and superintelligence has so far scaled vertically, and the framework for distributed intelligence is missing. Distributed artificial superintelligence (DASI) is his effort to build one that enables shared intent, knowledge and innovation in a society of agents and humans.

2026-01-28 · open the original ↗

• Superintelligence can be pinned to an economic and a technical test. Pandey says the industry has drifted from AI to AGI to ASI to "powerful AI" as each term proved ill-defined, but the definitions from OpenAI, DeepMind, Yann LeCun and Demis Hassabis fall into two buckets. The economic one asks whether a system of autonomous agents can perform 100% of a human task without human intervention; the technical one asks whether an objective-driven reasoning engine can autonomously invent something beyond its training data that passes the rigor test.
• Pandey's group aims to meet both tests by scaling out. He notes some argue a level of superintelligence is already here because the Turing test has been passed, but says AI is still super powerful in some use cases and pretty bad in others, so something is missing. His team's goal is to reach both definitions, economic and technical, through a scale-out mechanism.

2026-01-27 · open the original ↗

• Superintelligence will come from collectives of agents, not lone geniuses. Pandey defines ASI by two tests, economic (agents performing 100% of a human task without intervention) and technical (inventing beyond training data), and wants to reach both by scaling out. His analogy is human evolution: 300,000 to 400,000 years of vertical gains, then a cognitive revolution around 70,000 years ago when language brought shared intent, shared knowledge and shared innovation. He casts Cisco's role as "empowering the pack" while others build "lone wolves", the same stance it takes when networking quantum computers together.
• Semantic protocols let agents share intent, not just exchange messages. Pandey says agents connected through gateways, MCP or A2A are like an iPhone and an Android exchanging texts in Japanese and English: connectivity "devoid of meaning". In his running example, an Outshift network-configuration agent and a compliance, security and budget agent building a network for Swisscom must trade off toward a common goal, using protocols for grounding, discovery, resolution of conflicting information, coordination, negotiation and evaluation.
• Natural language stays the agents' common denominator for now. Pandey argues that while humans are in the loop, and even without them because components come from different vendors, language is the lowest common denominator. Only if everyone standardized would vector space be the way to communicate. He flags that world models suffer from tokenization, citing recent work from FAIR and others, but says Outshift starts with language, broadly including mathematics and physics, because Cisco's business is gluing heterogeneous brownfield systems.
• A cognitive memory fabric pairs probabilistic agents with formal world models. Pandey says the idea came from Outshift's Swisscom work, where letting agents loose on a network was "a recipe for disaster" and deterministic software alone explored too little, so agents were layered on a network digital twin. After three or four such use cases the team is generalizing it into a persistent, distributed layer of working memory, knowledge, ontologies and knowledge graphs that agents can summarize, extend and redact, under strict privacy and access controls.
• Cognitive engines drive shared innovation as accelerators and guardrails. Pandey borrows a Raj Reddy quote, recently repeated by Satya Nadella, that AI can be a guardian angel or a cognitive accelerator. Accelerators speed parts of a workflow; guardrails keep invention within compliance and physical limits (encoding what a wall or a blueberry feels like). His drug-discovery pipeline runs LLM planning, an AlphaFold-like model, compliance and manufacturability agents and robotic wet labs on all three layers, with the compliance agent drawing on past FDA outcomes.

2026-01-26 · open the original ↗

• More reasoning does not reliably make a model better at a task. In chapter 7 of his book, Ozdemir benchmarks reasoning models on MathQA and Humanity's Last Exam and finds no obvious correlation between reasoning level and performance, a result he calls consistent. He says labs such as Anthropic report the benchmarks where reasoning helps, and that running the ones they don't report shows the opposite.
• Even where reasoning helps, the gain rarely justifies the cost. Ozdemir says that in other benchmarks showing a correlation, the gain is maybe 1 or 2% in accuracy for a 2x increase in cost, and in his consulting it almost never pays. He recommends few-shot prompting and building your own shorter, more succinct reasoning traces instead; the host notes the book's reasoning benchmark took 5 hours for 180 LLM calls.
• Reasoning is ordinary next-token generation hidden from the user. Ozdemir says OpenAI, Anthropic and Gemini models are trained to produce reasoning, but it is the same autoregressive stream, which the UI hides until the model is "done thinking". Prompting "think through the problem before answering" induces something similar. His anecdote: one unnamed model produced three paragraphs of reasoning and a 30-second wait just to reply "hello".

2026-01-25 · open the original ↗

• Evaluation depends on the task, not the model. Ozdemir splits LLM tasks into generative (free text and multiple choice) and understanding (embeddings and classification), each with its own metrics, since grading an essay differs from grading embeddings even when LLMs produce both. The classification metrics in his Principles of Data Science from 10 years ago, he says, still apply when an agent is doing the classifying.
• Pick precision or recall by which error costs you more. Ozdemir defines precision as how often the model is right when it says yes, which matters when false positives are expensive, and recall as how many true yeses it catches, which matters when false negatives are. In his factory example, passing a bad part that might bring a plane down makes precision what counts, while throwing away a good part by accident is tolerable.

2026-01-24 · open the original ↗

• Long-running agents need context the model can actually reason over. Ozdemir explains that every turn and tool call adds tokens, so long-horizon tasks need long context windows. In his book's needle-in-a-haystack test, he filled a context window with trivia and inserted his own birthday, and depending on its position and how full the window was, the LLM failed to see it. Window size alone, he argues, does not tell you whether a model can use everything in it consistently.

2026-01-23 · open the original ↗

• Aim AI at hard ROI metrics, and it exposes worthless work. Dell's global CTO John Roese (episode 953) says that in the fiscal year Dell reported in March, revenue grew $10 billion while absolute cost fell 4%, which he says had not happened in Dell's 41 years. He credits people, process and technology. He cites a consulting report, he thinks McKinsey's, that calls generative AI a "BS work detector": if AI is applied to a task and revenue and cost do not move, the work may not be worth doing. The host frames this against MIT NANDA's 95% failure figure for enterprise AI, which he calls contentious.
• Laughs per minute gives comedy a measurable target. Software engineer and podcast host Joel Beasley (episode 945) says four to five laughs per minute is the standard for a full-time professional comedian. He records every set, about 10 to 11 months into stand-up, and has Gemini count laughs, since at recording it was the only browser chatbot that took video and GPT and Grok did not. He logs results in a spreadsheet, and he also uses a saved writing-coach prompt modeled on how professionals gave feedback at about 10 comedy writers' groups.
• A niche intersection of skills beats application tactics. A Netflix senior data scientist (episode 947) says he got into DoorDash with a "briefcase" document, which took at least a couple of days, setting out the company's pain points and how he would solve them, sent to the hiring manager. The host doubts it would stand out now that he gets 30-page AI-made pitches weekly. The guest says he landed Netflix, after multiple earlier rejections, because his ads and forecasting experience at Spotify matched what Netflix needed when it launched ads.
• Systems that ignore how humans decide will break on edge cases. Sandy Pentland (episode 949) argues today's AI-driven systems could fall apart the way the Soviet Union's did, through bad data, fast change and ignoring human nature. He says people are "social foragers" and cites Daniel Kahneman's estimate that 95% of what we do comes from culture. He says his group has shown polarization comes from a few loud voices, that traders handling $10 million a day do reliably better when they watch other traders, and that equation-followers who missed tail risk "got creamed" in 2008.
• Workslop comes from vague AI mandates and poor work context. Dropbox's VP of Engineering (episode 951) cites almost three-quarters of CEOs feeling competitive pressure to adopt AI and about 55% of employees not knowing how to use it. His remedies are to treat AI output as a first draft, and to build better context than slow, token-hungry MCP connectors provide; Dropbox's Dash pairs a BM25 lexical index with a vector index. He also recommends internal eval sets defining what good looks like, and training led by high performers.

2026-01-22 · open the original ↗

• Bigger models are more general, not better at everything. Ozdemir points to the Qwen family, which he believes is Alibaba's, spanning about half a billion to over a trillion parameters, with over a trillion being GPT-5 or Opus-level size and Kimi among open-weight models up there. A big general model is built to serve very different users, so it can do more horizontal things, including ones you do not need.
• Size models in rough tiers by what the task demands. Ozdemir calls under 10 billion parameters small, enough for lookups, RAG or web search and runnable compressed on a laptop CPU, but not long-horizon tasks with several tool calls. He puts 10 to 100 billion as medium, able to handle agents, especially fine-tuned on task data, and 100 billion plus as large, for enterprise-wide multilingual, multi-task use. He frames the choice as an experiment, testing for example half a billion against eight billion parameters.

2026-01-21 · open the original ↗

• An agent chooses its path; a workflow has it chosen. Ozdemir defines an agent as an LLM with access to tools in the function-calling sense, such as searching or reading a file. Most production AI deployments are workflows, he says: retrieval, a human-engineered prompt, an LLM call and display, with the LLM never deciding to use a tool. Take away the LLM's choice of next step and you are in workflow territory; he says he renamed his book from something like "Building Applied AI" to "Building Agentic AI" partly to make this distinction honestly.
• Agents' unpredictability grows with the number of tools. Ozdemir says modern deployments often start at five to 10 tools, raising questions of which tool goes first, misuse and inefficient repeated searches. The developer ends up hoping the agent navigates that self-chosen path efficiently and correctly.

2026-01-20 · open the original ↗

• Count the conditionals to decide between a workflow and an agent. Ozdemir defines an agent as an LLM with tools that chooses its next step, and a workflow as a deterministic code path, which is what most production deployments are. His first question to clients is to describe the existing process, which nine times out of 10 settles it. A plain keyword search is a workflow; a few branches suggest a hybrid; a conditional at every step points to an agent. He adds that almost any agent's work can be done by a workflow, and the question is how big and maintainable that workflow would be.
• Match model type and size to the task, then test. Ozdemir says autoencoding models like BERT cannot generate text but do embeddings and classification almost as well as huge GPT models, at around 100 million parameters. For generative models he uses rough tiers: under 10 billion parameters for retrieval and search, 10 to 100 billion for multi-turn agents (especially fine-tuned), and 100 billion plus for enterprise-wide multilingual use. Context length is not enough either: in his needle-in-a-haystack test models missed his inserted birthday depending on position, and he calls out Grok as the worst.
• Choose evaluation metrics by task and by the cost of errors. Ozdemir splits evaluation into generative (free text, multiple choice) and understanding (embeddings, classification) buckets with separate metrics. He uses precision when false positives are expensive and recall when false negatives are, and says the question comes before the math: what is expensive to you.
• Forced workflow steps patch agents' unreliability. Ozdemir's example is OpenAI's deep research, which he says always asks a follow-up question even when told not to, because a skipped clarification costs more than an annoying one; Cursor's planning mode is the same pattern. In one case study GPT-4.1, told in its system prompt to always look up an Airbnb policy article first, skipped the tool and made up an answer 5% of the time.
• Extra reasoning and compression both trade quality unpredictably. Ozdemir found no obvious correlation between reasoning level and performance on MathQA and Humanity's Last Exam, and elsewhere gains of maybe 1 or 2% accuracy for 2x the cost, which in his consulting almost never pays. He compares a quantized or distilled model to "the cousin" of the original: in his newsletter's quantized-Llama test, it scored a little worse on benchmarks and its top-10 candidate tokens often diverged from the original's.

2026-01-19 · open the original ↗

• Remote coding interviews no longer work, so his company holds them in person. Rajeeva says that from this year it stopped remote coding interviews, because cheating tools are too many and too hard to detect, and hunting for evidence of cheating spoiled the conversation. In person, candidates get a practical problem resembling daily work, such as failing tests and missing implementations, and may Google APIs under a "gentleman's agreement" not to use an LLM. He concedes that, as more code is LLM-generated, this may not scale, even though staff already get Cursor and ChatGPT licenses.

2026-01-18 · open the original ↗

• Well-paid engineers quit when their work lacks a mission. Rajeeva cites developer-productivity studies that put engineers' productive time at maybe 3 or 4 hours a day, and says creative knowledge workers who can finish their assignments in 2 hours start asking what they are doing. His remedy is making R&D believe it is in the business of innovation, free from being dictated by quarterly plans, even in "boring" fields like enterprise content management. He points to engineer-built Acceldata capabilities such as ODP and a pulse monitoring system, and a next platform version due in May. The host adds a happiness-economics lecture from about 15 years ago putting the plateau around $80,000 to $100,000 of US household income.
• Hire senior engineers on whether they can build. Rajeeva believes technical management requires people who have been in the trenches. He tests senior candidates on turning any business problem, such as designing an e-commerce warehouse or logistics system, into something built in Go, Python or Java. For new graduates he bets on potential, partly scores and background and partly interest, and decides after a few months.

2026-01-17 · open the original ↗

• A CTO must think in customers' shoes, because engineers won't. Rajeeva, who grew Acceldata's engineering from four to 100 people through COVID, remote work and the return to office, says technologists assume "if I build it, they'll come". Customers act on guidance with two axes, a technology mandate ("we need to do AI") and a business problem, and need wins they can show their leadership to renew a proof of value into a contract. Product managers focus on features and bugs, so if not the CTO, he asks, then who.
• Early-stage engineers stay when they get room to innovate. Rajeeva says money stops mattering to people past a point, and startups lose context and speed when people churn. He keeps them with engineer-led projects rather than a fixed annual roadmap: an internal data platform (ODP) built to avoid license fees and vulnerabilities, an AI agent that fixes vulnerabilities in shipped enterprise software, and an in-house data-center stack because VMware is expensive and OpenStack too complex.

2026-01-16 · open the original ↗

• Enterprises are data rich but context poor, so agents act "corporate stupid". Auradkar argues today's models are incredibly intelligent but do not understand the context in which to apply it, because sales, marketing and service data sit in separate silos. His example: after you buy a car, automation keeps advertising the same car instead of insurance, a warranty or accessories, because the purchase sits in a transactional system the marketing agents cannot see. The episode is a partnership with Salesforce to announce the Informatica addition.
• Agents need trusted context, not just more context. Auradkar's layers are unified data with governance and trust (lineage, consent, PII handling, masking of unstructured data), real-time signals such as events, transactions and clicks, and historical context on customers, assets and cases. On top of those come business rules, policy enforcement for business and regulatory policy, and domain-specific semantics and workflows that Salesforce says its applications already codify.
• The unified data engine joins customer data, analytics and integration. Auradkar says Data 360 is a customer data platform that "puts the C in CDP", unlike rivals that served marketing only. It unlocks "trapped" data and reverse-ETLs it to every channel, grounding agents. Tableau Next adds a semantic model, so that revenue, ACV or churn mean the same thing to everyone, plus natural-language "agentic analytics" pushed into Slack or Sales Cloud. MuleSoft extends integration and API governance to agents as an "agent fabric".
• The latest acquisition adds enterprise-wide catalog, data quality and ETL. Auradkar says the Informatica acquisition legally closed on November 15th, ahead of schedule, and calls it one of the most synergistic in enterprise software. He says Informatica's enterprise data catalog layers over Salesforce's customer data catalog to give a superset of enterprise metadata, alongside data-quality tools, governance lineage and what he calls the industry's best ETL.
• Salesforce runs on its own agents as "customer zero". Auradkar says help.salesforce.com is now agent-first and agent-only, with humans augmenting, and that a vast portion of customer questions are answered by agents. Background marketing agents and predictive lead scores fire within a 30-second window of a website visit, and an SDR agent is shipping. He cites Williams-Sonoma's Sous Chef agent on Agentforce as agentic transformation at a company's core, not its margins.

2026-01-15 · open the original ↗

• Self-healing pipelines treat failures like agentic coding on a cluster. Rajeeva describes an agent that detects an anomaly in logs, clones the pipeline's Spark code, and gets the error plus metadata (tables, columns, data types) from Acceldata's xLake engine. It then rewrites the code and redeploys it to the Hadoop, Trino or Kubernetes cluster, replacing the bug-report, fix and CI cycle. He says models like Gemini 3 one-shot about 90% of what you want, so automating context and deployment is the remaining work.
• Auto-remediation works only where pipelines are version-controlled code. Rajeeva does not claim it can fix everything: proprietary Informatica or HVR pipelines and Oracle stored procedures are "not so easy". He predicts pipelines will shift from drag-and-drop tools to code because generating code is now easier, which in turn makes them easier for AI to mutate when errors occur.

2026-01-14 · open the original ↗

• Data management's manual workflows are ripe for agents. Rajeeva says catalogs, MDM and curation practices at the banks, insurers and telcos Acceldata serves have changed only incrementally in 15 to 20 years, and much of the work is forms and clicking through well-defined workflows. Agentic data management (ADM) aims to automate a data steward's routine tasks, such as a data lake classifying its own data or generating validation rules automatically.
• AI needs a cross-lake engine to know enterprise data. Rajeeva argues an LLM compresses the internet but knows almost nothing about your data. Acceldata's xLake reasoning engine is meant to give the AI layer tools that connect to sources across Snowflake, AWS, Azure and on-prem footprints, and assemble data at scale so ADM can answer questions about the whole environment.

2026-01-13 · open the original ↗

• AI needs a context engine to act on enterprise data. Rajeeva says an LLM can write a correct SQL query but has no idea how much data it will return or whether the user may see a column. Acceldata's xLake reasoning engine is meant to do for enterprise lakes across Snowflake, AWS, Azure and on-prem what Google's index does for a chatbot searching the web. Its agentic data management (ADM) platform now has close to 10 agents, such as a data-catalog agent that knows a table's location, columns, types, refresh cadence and owner; the agents are distilled from customers because a bank's, insurer's and healthcare firm's problems differ.
• Pipelines written as code can be healed by agents. Rajeeva describes an agent that spots an anomaly in logs, clones the Spark code, adds error and metadata context through xLake, rewrites the code and redeploys it to the cluster. He says Gemini 3 one-shots about 90% of what you want. He limits the claim to version-controlled Spark or SQL pipelines, not proprietary Informatica, HVR or Oracle stored procedures, and predicts pipelines will move from drag-and-drop to code.
• Autonomy means agents do most of the work and an expert approves it. Rajeeva compares it to code review: the economic value is compressing maybe a week of work into an hour, but approving, rejecting or ordering a rewrite stays with someone who knows the system. His MCP demo connecting the data engine to a coding environment did about 60% of a task, maybe 80% with prompting, leaving two engineers to finish rather than a whole team.
• Data sprawl is every large enterprise's eventual fate. Rajeeva defines it as data scattered across legacy on-prem clusters, Snowflake or Databricks analytics, and a cloud deal with Google, so someone must find and join it all. He says Acceldata differs from metadata-only players by deploying a data plane inside the customer's environment, linked over mTLS, so it can validate anything from 100 to 100 million rows. It is built on Kubernetes and Spark, and its open-source ODP data platform is on GitHub.
• Engineering leaders must serve both customers and engineers' need for meaning. Rajeeva says customers act on technology and business mandates and need wins to show their leadership, and if the CTO does not think about that, nobody will. Engineers who can finish their work in 2 hours leave without a mission, so he backs engineer-led projects, hires senior people on whether they can build, and has moved coding interviews back to face to face because remote LLM cheating is too hard to detect.

2026-01-12 · open the original ↗

• Specialized domain models will outpace general-purpose ones. For 2026, St. Lawrence predicts more AlphaFold-style models for particular domains, evaluated by domain experts. Her reasoning is that training data for general models has hit a limit until hyperscaler clusters are built out, and that coding succeeds because results can be measured. She suggests general superintelligence may come from combining such specialist models, much as humans pair general intelligence with PhD-level specialties.
• Continual and nested learning will gain ground. She cites a new Google paper on nested learning as a way around models that stop learning after training and remember only their context window. She expects much more progress on models that keep updating, which the host calls a big gap in all frontier-lab LLMs.
• AI will return to research because no lab has a clear next breakthrough. St. Lawrence says nobody is talking about GPT-6 or what to wait for next, and points to Ilya Sutskever's recent podcast interview as evidence that labs do not know the path forward. She expects hyperscalers to hit limits on data availability rather than compute.
• Physical AI and spatial intelligence will grow, but robots must wait. She holds to her call that 2027, not 2026, will be "the year of the robot", even though the Neo robot she bought is due in 2026. She expects spatial intelligence and simulated environments, such as World Labs' work, to matter mainly as a source of new, harder training data, with continual learning needed in 3D settings.
• AI ops will emerge as the new trending job. St. Lawrence predicts roles managing GPUs, model orchestration and agent reliability, likening them to DevOps around 2010. She expects it to be a new trend in 2026 rather than necessarily the most popular job.

2026-01-11 · open the original ↗

• The chipmaker powering most AI was the year's overall winner. For 2025 the host names Nvidia, with Google alongside it, and points to Nvidia's share price; he caveats that not every model runs on Nvidia, since Google has TPUs. St. Lawrence agrees that Nvidia is the overall winner and says it will remain so for some time.
• The year was about vibes, not model breakthroughs. St. Lawrence calls 2025 "the year of vibes", saying vibe coding was coined that year and coding is where agents genuinely took off. She calls GPT-5 disappointing and says that, unlike earlier years of waiting for multimodality and then agents, nobody seems to be anticipating anything next.

2026-01-10 · open the original ↗

• Enterprise agents were the year's biggest disappointment. For 2025, St. Lawrence's best-performing Substack post of the year, "Agents of Disappointment", argued the gap between hype and practical implementation was too great. She says vendors like Salesforce with Agentforce and SAP promoted agents inside existing enterprise tools, but they "just don't work" or companies are not set up and do not know how to structure them. The host agrees there is too much talk relative to impact, blames data silos and security setups, and expects deployments to come.
• Phone-based AI still adds little a year after launch, the host says. His disappointment pick is Apple: he says Apple Intelligence, announced in autumn of the prior year, still adds little he uses beyond once generating an emoji, and Siri should understand speech as well as OpenAI's Whisper. St. Lawrence contrasts the seamless voice chat she gets from Grok in her Tesla. Both see a big opportunity for Apple given how ubiquitous its devices are.

2026-01-09 · open the original ↗

• The hard part of agents is the last mile, not the prototype. Carlson says LLMs and vibe coding have compressed the path from idea to v1, but rolling out to thousands or hundreds of thousands of enterprise users is "where the rubber hits the road". The episode is a partnership with Salesforce announcing that its Agentforce 360 platform is open for building on. Carlson says it bundles a multi-LLM planner, a scripting language for deterministic control, Salesforce's data and metadata model as business context, the security and permissions layer, and observability and monitoring.
• Natural-language rules fail too often, so agents need deterministic rails. Carlson says customers wrote agent logic as "always do this, never do this", which LLMs handle poorly, and an enterprise cannot accept a task done 90% of the time and skipped the other 10%. Agentforce Script lets developers mark where the agent may reason freely and where it must follow pre-set paths, much of which already exists in Sales, Service and Marketing apps.
• Small teams ship faster on a platform than on hyperscalers. Carlson says AppExchange, launched 20 years ago as the first enterprise SaaS app store, reaches 150,000-plus Salesforce customers, with low-code builders and pro-code options in Apex and React. Teams that want to run infrastructure down to the server should use AWS or Google, he says, but a founder can ship faster with the app, workflow and agent layers provided. His example is a family ranching business that built a cattle-management app for its own herd and then sold it to others.
• Founders over-focus on the best LLM instead of the customer's job. Carlson calls LLM performance a distraction: commercializing AI is like building any product, with requirements and jobs to be done. Not every step needs an LLM when code, database structure and standard permissions already solve it. He sees many niche problems ripe for someone to build something.
• Agents will not replace applications; they are applications. Carlson argues any sophisticated agent at scale is a multi-LLM orchestrated application, with logic choosing which LLM or deterministic step to use, where data lives and what the permission model is. He expects applications to look very different but rejects "agents versus applications", predicting the two merge.

2026-01-08 · open the original ↗

• Demand for AI engineering skills has outgrown data science skills. St. Lawrence says her 2025 prediction to that effect held, citing a LinkedIn jobs report showing much higher growth in AI prompting and AI engineering skills than in data science. She adds that AI is being stamped onto job descriptions much as CEOs stamp GenAI onto products. The host points to his episode 856 naming AI engineer the fastest-growing job in most developed countries.
• Data science spawned a family of roles now being rebranded as AI. The host sees AI engineering as a subspecialization of data science. St. Lawrence says the 2014-era "unicorn" data scientist split into roles such as data engineering, ML engineering and MLOps, and is now rebranding from ML to AI. She notes the irony, from the book Thinking Machines, that "machine learning" was adopted in the 2010s to escape the unserious-sounding term AI.

2026-01-07 · open the original ↗

• Agentic AI dominated 2025 so thoroughly that audiences are tired of it. St. Lawrence counts her prediction that agentic AI would be 2025's dominant trend as correct. Her evidence is more vibes than data: conference audiences no longer want to hear about agentic AI and want something fresh in 2026.
• The host argues agents keep getting more useful on a doubling curve. He says the length of a human task an agentic system can do doubles every seven months, at about 50% reliability on software or ML tasks that take a human expert several hours. On that curve he expects a full eight-hour workday by the end of 2026, which the speakers note is when talk of replacing workers heats up.

2026-01-06 · open the original ↗

• Last year's five predictions mostly held, some only weakly. Of St. Lawrence's 2025 calls, agentic AI dominated, to the point that conference audiences are tired of it. AI spread into devices, from AirPods' real-time translation to CES refrigerators and a Costco toothbrush, though she doubts much of it was wise. AI-driven science was mixed: a paper claiming large gains in materials discovery was retracted, while she credits DeepMind's GNoME project with compressing about 800 years of discoveries. Enterprise monetization stayed crucial, with many projects stuck at MVP, and a LinkedIn jobs report showed AI engineering skills outgrowing data science.
• Intelligence is getting radically cheaper while agents' task horizons double. The host says the same capability costs about 100 times less than a year earlier, and the task length agents can handle at about 50% reliability doubles every 7 months, implying a full 8-hour workday by end of 2026. He limits the curve to software and ML tasks, where success is verifiable and training data can be simulated. He adds that higher-reliability curves, at 90% and 99%, trail behind, and discloses his bias as an AI consultant. St. Lawrence says this forces businesses to rethink their models, pointing to Cursor's revenue growth.
• Their awards crowned an established search giant, a chipmaker and "vibes". For 2025, both name Google comeback of the year again: St. Lawrence for its seamless integration into Search and G Suite, the host because users already trust it with their Drive and Gmail data. Her wow moment was Nano Banana redoing a room with her tile sample; his was co-presenter Ed Donner's crew of four agents building a working simulated trading platform during a 90-minute lunch break at ODSC East. For disappointment she names enterprise agents and he names Apple; the overall winners are Nvidia and the theme of vibes, with GPT-5 a letdown.
• Next progress will come from specialization, continual learning and research. For 2026, St. Lawrence predicts more AlphaFold-style domain models evaluated by experts, because training data for general models has hit a limit. She also expects progress on continual and nested learning, citing a new Google paper. And she expects a return to research, since no frontier lab has a clear path and nobody talks about GPT-6. The host agrees an orthogonal breakthrough beyond scaling weights and inference time is needed, especially far more sample-efficient learning, but doubts it arrives in 2026.
• Spatial intelligence and AI ops will rise, but robots come later. St. Lawrence holds that 2027 is "the year of the robot" and expects physical AI and spatial intelligence, such as World Labs' simulated environments, to matter mainly as new, harder training data, even as she awaits her Neo robot in 2026. Her final prediction is AI ops as a hot new job managing GPUs, model orchestration and agent reliability, akin to DevOps around 2010.

2026-01-05 · open the original ↗

• Every country will soon have a sovereign AI strategy. Roese's fifth prediction, for 2026 going into 2027, builds on three models he described the previous year: government building data centers for government, government building for its indigenous industries, and government convening with industry. He calls the acceleration dramatic and sovereign AI now mainstream.
• Sovereign infrastructure will serve far more than government work. Roese's examples are resilient backups of a country's critical-infrastructure AI, and agents certified by a government, such as a "Swiss lawyer" agent, whose knowledge graphs and back-end data live physically under that sovereignty. He expects government's role to grow where AI meets national interests such as critical infrastructure, defense, intelligence and healthcare, and calls the inseparability of government and private sector in AI probably a good thing for pooling resources.

2026-01-04 · open the original ↗

• An agent is a software system, not a feature of an LLM. Roese says one camp treats agentic behavior as a property of the large language model, which he calls not very useful. Dell's camp, which he says many others share, treats an agent as an autonomous software system of which the LLM is one part.
• The LLM supplies world knowledge, communication and much of the reasoning. Roese counts it as the first of four components, large or small, without which agents cannot be built.
• Knowledge graphs supply specialized knowledge and long-term memory. Roese says agentic RAG or, more often, knowledge graphs give agents domain information they can manipulate. Pathways through the graph become skills for repeating tasks and hold state over long periods outside the LLM.
• A tool protocol lets agents perceive external data and act. Roese calls MCP, or a similar interface, the protocol of choice for this third component.
• Agent-to-agent protocols matter most, because ensembles can reach consensus. Roese calls this fourth component, chiefly A2A, the most important. In Dell's studies, he says, agents sharing a common knowledge graph as "truth" (the LLM being "opinion") and told consensus matters ignored a hallucinating member, like a team overruling one person saying crazy things. He predicts this four-part definition will prove enduring, against those who treat the LLM as a "black box of magic".