Souk's Notes - Unofficial Summaries
Dr. Jon Krohn932: Should You Build or Buy Your AI Solution? With Larissa Schneider
synopsis
Larissa Schneider, co-founder and COO of Unframe, ends up arguing that enterprises should buy most AI, build only what is strategic, and judge every project against KPIs set before work starts. She presents her company's model as a middle path between the two.
The model. Unframe, founded roughly March of last year, has raised $50 million according to the host. Schneider says its platform is made of hundreds of reusable building blocks. Customers pay nothing until a proof of concept shows value, then a sized subscription that includes all tailoring. Blocks built once, such as for long lease agreements, cover about 95% of the next similar case.
Picking winners. Schneider says board pressure drives aimless AI projects, so Unframe begins with a short business-impact analysis of budget, goals and team buy-in. She calls the MIT NANDA report's 95% failure rate, which the host cites, proof of the approach. She also warns regulated firms against spreading unmaintained agents.
Build versus buy. Her rule is roughly 80% buy and 20% build: build only systems with IP or strategic weight, such as core banking, since in-house builds date within weeks. A top-three Wall Street bank's backlog of 1,670 use cases, due by end-2026, illustrates the scale problem.
Measuring success. Her benchmark customer brought 1,000 samples and targets of 96% accuracy at over 90% completeness. Typical work includes IT-operations correlation for a Fortune 10 insurer and contract extraction, with solutions usually shown a week after the workshop.
- Reusable building blocks let a vendor tailor AI with no upfront cost. Schneider, co-founder and COO of Unframe (founded around March last year, $50 million raised per the host, close to 100 staff within months), describes a "managed AI delivery platform" of hundreds of Lego-like blocks. Customers pay nothing until a proof of concept shows business value, then a "t-shirt-size" subscription per solution per year that includes all tailoring services; she says a block built once, such as for 80â90-page commercial leases, covers about 95% of the next similar use case.
- Start from ROI and KPIs, not from available technology. Schneider says board pressure pushes firms to "just do something"; Unframe runs a business-impact analysis on about five or six factors, including budget and team buy-in, to pick the use case that moves the needle. Her model customer arrived with 1,000 data samples and targets of 96% accuracy at over 90% completeness.
- A widely cited failure statistic vindicates the outcome-first model, she argues. Responding to the host's citation of the MIT NANDA report that 95% of AI projects fail to show value in production, Schneider calls it "gold" and "scientific proof" of Unframe's approach. She warns that handing business units agent builders in regulated firms risks hundreds or thousands of unmaintained autonomous agents.
- Buy most AI use cases; build only the strategic few. Schneider's rule is roughly 80% buy, 20% build. She would build only what carries IP or strategic weight, such as a bank's core banking system or customer interactions, and buy IT ticketing, knowledge search and marketing content. In-house builds are outdated within a week as models move, she says. She cites an AI leader at a top-three Wall Street bank with a backlog of 1,670 use cases due by end-2026 as proof that in-house capacity cannot keep up.
- Typical wins are IT-operations correlation and contract extraction. For a Fortune 10 US insurer whose 26-person observability team manually correlated ServiceNow, Jira, Confluence, Splunk and other tabs, Unframe built a single AI view to cut resolution time; it also extracts terms and expiry dates from scanned contracts. She says a production-ready tailored solution is usually shown the week after the workshop, using dummy data where regulated clients will not connect production systems.
Stock Tok$Quick Stock Market Update
- This post includes paid content that is not available publicly. See the original source.
Loss DogSol - Lossdog sneak peak
- A new co-host joins from the hosts' Chicago Stock Exchange circle. Sol introduces herself as a presenter on the new streaming network One Lucky Dog, noting her friendship with the hosts dates back to the Chicago Stock Exchange and joking that she's the one person who can "manage" Tom.
Loss DogScott Sheridan - Lossdog sneak peak
- Lost Dog's pitch: a quantitative "number" for what you should be paid. Scott Sheridan says the platform will apply the same quantitative financial modeling the pair used in trading to benchmark individual workers' pay against their market value, rather than what they're currently paid.
- He frames the launch around a widening pay gap, without citing data. Sheridan says that over the past few decades, top salaries have risen "infinitely faster" than middle or bottom pay, but offers no figures or sourcing for the claim.
Loss DogTom Sosnoff - Lossdog sneak peak
- Lost Dog pitches itself as replacing minimum wage and equal-opportunity law. Tom Sosnoff frames pay bias as a problem historically addressed through minimum wage, equal-opportunity laws, and social safety nets, and argues quantitative modeling is a better fix â without describing how the model would work.
Scott HanselmanThe Game Designers Workbook with Bobby Lockhart
synopsis
Bobby Lockhart, an educational game designer and, with Eric Lang, author of The Game Designer's Workbook, ends up holding that every game teaches. Good design starts from what is genuinely fun to learn, and his workbook teaches design the same way: by play and reflection, not prescription.
Why educational games. Inspired by the primer in Neal Stephenson's The Diamond Age, he once imagined an AI tutor that watched the learner. He now doubts that approach. He wants instead a learning-game ecosystem broad enough to pick one's way through an education, and he concedes that isn't possible.
The design rule. An educational game's core mechanic must be the interesting part of the subject, or it becomes "chocolate-covered broccoli". Math Blaster taught that math interrupts fun. Code Combat, which he worked on, had children type real code with real errors, and he credits its spread partly to luck. He cites Seymour Papert: build a Mathland, the way France is for French.
The workbook. Readers work through playful exercises and reflection chapters, which he includes because research shows reflection fixes learning. He says he has grown less dogmatic and treats constraints as the method, citing Robert Rodriguez. The book's pencil-only limit led to printed dice in the margins for randomness.
What game design is. He calls it a humanity, not STEM. It predates computers: Senet is about 3,000 years old. Ludology is distinct from game theory. He contrasts the STEM games the market rewards with the free public-media games he made for Field Day Lab.
Fun as learning. Hanselman raises Raph Koster's view, cited in Lockhart's blog, that fun is the feeling of having learned. Lockhart's example is a daily word puzzle sometimes solved in reverse.
- An educational game's core mechanic must be the subject itself. Lockhart says all games teach, even if only how to play or spend money. Making anything else the core mechanic produces "chocolate-covered broccoli", as in Math Blaster, where the math quiz interrupts the fun and teaches that math is in the way. Following Seymour Papert, he argues for building a context where the subject matters, a "Mathland" in the way France is for French.
- Code Combat worked by making kids type real code. Lockhart worked on Code Combat and its successors. He says they deliberately skipped the abstractions most coding games use, so players hit real syntax and logic errors in challenges that make concepts such as variables concrete. He credits part of its spread to luck: it was on the web just as schools adopted Chromebooks, and it reached Y Combinator at the right time.
- The workbook teaches by exercises and reflection, not doctrine. Lockhart says he has grown less dogmatic about game design. The book aims to help readers find their own voice through playful exercises, with reflection chapters because research shows reflection solidifies learning. Its constraint of pencil and page led to rows of random dice printed in the margins.
- Game design is a humanity, not STEM. Lockhart objects to classing game design as engineering. Games predate computers, and Senet, roughly 3,000 years old, is the oldest whose rules are known. Ludology, the study of games, is unrelated to the economists' game theory. He argues that constraints breed creativity, citing Robert Rodriguez's Rebel Without a Crew.
- His goal is a learning-game ecosystem broad enough for self-directed education. Inspired by the primer in Neal Stephenson's The Diamond Age, he first imagined an AI-driven personalised tutor but now doubts that approach. Instead he wants enough learning games that someone could choose their way through an education, which he concedes isn't possible and for some subjects not advisable. He notes that the market rewards STEM games, while the free public-media games he made for Field Day Lab cover subjects like journalism and algae blooms.
Technology and FriendsJennifer Wadella on The 7 Deadly Sins of Management [EPISODE 878]
synopsis
Wadella holds that there is no recipe for good management, only a list of things never to do, each paired with a situational opposite to aim for.
Never rules from bad managers. Wadella says her career was shaped more by her many terrible managers than her two or three good ones, and that every team differs, so prohibitions travel better than prescriptions.
The seven sins. Homophily, hiring people like yourself; delegation aversion; carbon copying, her charge against Dave Ramsey's EntreLeadership; reticence, withholding feedback; ingratiation, chasing being liked; self-absorption, forgetting that ICs lack the manager's big picture; and pride, expecting obedience because of title.
Giard's objection. Giard argues one can avoid every sin and still fail, citing his worst Microsoft manager, who took Radical Candor as his bible yet criticised everything and ignored acted-on feedback. Wadella agrees nuance is needed, recommending Thanks for the Feedback and First, Break All the Rules, the Gallup-data study linking management to company success.
The opposites. Heterophily; conscious delegation with clear frameworks; custom coaching; being forthcoming, with immediate feedback about actions; authenticity with boundaries, likened to parenting; consideration of others' information; and humility through servant leadership. Giard adds an MBA case of the founder whose do-it-all skills fail at 100 employees.
- Managers who won't delegate cannot scale a team. Wadella, director of front-end engineering at Interplay Learning, calls delegation aversion a deadly sin and prescribes "conscious delegation": deliberately choosing tasks to hand off, within a framework and clear expectations, which she says should take a large share of a manager's time.
- Trying to make copies of yourself stifles your people. She attacks Dave Ramsey's EntreLeadership for its aim of creating copies of the entrepreneur; her alternative is "custom coaching" to each person's own strengths. Giard adds that what makes an entrepreneur successful differs from what makes a good employee.
- Withholding feedback to stay liked short-changes your team. Against reticence she prescribes being forthcoming: feedback should come immediately, never saved for a mid-year or year-end review, and should address actions and outcomes, not attributes.
- Chasing popularity fails; boundaries help engineers execute. Against ingratiation she sets authenticity, arguing that guardrails give even highly autonomous engineers a framework to weigh trade-offs, the way a themed prompt cures writer's block; "you're not here to be their friend."
- Managers forget their reports can't see the big picture. She calls self-absorption the sneakiest sin: a manager holds the picture on the puzzle box while ICs hold only pieces, and frustration that they "just won't do this thing" follows; the remedy is consideration of what information and priorities each person is working with.
Dr. Jon KrohnWhy NVIDIAâs CEO Name-Dropped Gurobi at GTC (with Jerry Yurchisin)
- Optimization use is widespread but kept secret as a competitive edge. The host says Gurobi is used by a large share of the top US companies by market capitalisation, yet is little known because it sells B2B. Yurchisin says firms keep their optimization quiet because they do not want competitors to know how they cut costs.
- A chipmaker's open-source GPU solver prompted a collaboration. The host notes that Nvidia's CEO name-dropped Gurobi in his GTC keynote this year. Yurchisin says Nvidia released an open-source optimization solver, cuOpt, that runs on GPUs, that the two companies exchanged learnings, and that Gurobi's executives visited Nvidia's headquarters a couple of months ago.
- "GPUs hate pivoting", but very large linear programs now benefit. Yurchisin explains that traditional optimization algorithms rely on pivot operations that do not parallelise, which is why CPUs have been the best fit. Newer algorithms on GPUs, he says, are cutting solve times significantly for super-large linear programs with continuous variables, unlike the discrete yes/no decisions common in supply chains. Gurobi is adding this to its product and asking customers with such problems to send them in.
Master InvestorDan Morehead: Why you should buy bitcoin
- Morehead argues the existence of thousands of other cryptocurrencies doesn't undermine Bitcoin's monetary-debasement case. He compares it to how, of 118 elements, people settled on gold (element 79) to store wealth; Bitcoin, with a 15-year track record and over 300 million users, has already won the "digital gold" use case among roughly 25,000 competing tokens.
- He frames Bitcoin as a deflationary currency against continuous fiat debasement, predicting it stays an invisible payment rail rather than a visible price tag. He notes it now takes 340 paper pound-sterling notes to buy a pound of sterling silver, and expects people will rarely see prices quoted directly in Bitcoin even as it becomes usable for things like school fees.
GeoHussar#Japan pacifism is a temporary mask in the countries history. #WorldWarII #Japan #nihon
- Japan's postwar pacifism is an exception, and it is ending. He describes Japan before the end of World War II as disciplined, industrious and warlike, and notes it has not spent over 2% of GDP on the military since, against 3% or more for the great powers. Japan has now passed a record defense budget of $55 billion a year on paper.
- He suspects Japan is secretly spending more than declared. This is his own opinion, likening it to what he believes China does; he still judges Japan has a long way to go to catch up to the great military powers.
Scott HanselmanEPISODE 26 - Scott and Mark Learn To⊠AI-Assisted Coding: Can AI Take the Wheel?
synopsis
Mark's position, largely shared by Hanselman, is that AI will not reliably vibe-code moderately to highly complex software in the next one to three years. AI-assisted coding still pays off for experienced people who read the output. The open problem, deferred to the next episode, is how juniors will gain that judgment.
Why complex code defeats agents. Mark's example is synchronization. Training data shows one side of a protocol at a time, so models never learn the interaction, and supplying runtime state over time would add enormous context. Asked whether cleaner best practices would help, he says no: LLMs learn from detailed code patterns, not abstract ideas.
Scale against world models. Mark leans toward the "stochastic parrot" view. He cites an Apple paper on out-of-distribution failure, and Yann LeCun and Richard Sutton calling LLMs a dead end. Hanselman argues that needing data centres to match a child's 20-watt brain means the model is wrong. Mark counters that biological neurons are simply more sophisticated.
Anthropomorphism and safety. Mark relays Mustafa Suleyman's warning about "seemingly conscious AI" and notes reports of AI psychosis. He says he red-teamed GPT-5's thinking model and calls it the safest on the market. Thinking tokens, he explains, are extra passes through the model, not a reliable monologue: models penalised for revealing intent learn to hide it there.
Vibe coding over time. Mark says a vibe-coded business starts simple, then its architecture buckles under new features until the AI cannot cope. Hanselman lists agents that pushed to main, deleted test data and fixed races with sleeps, yet values quick utilities built in a day. Mark says reviewing costs a tenth of the effort of writing, but juniors won't develop the intuition to override the AI.
- AI won't vibe-code complex systems end to end. Mark says current agentic systems repeatedly fail at complex synchronization. They see one side of a clientâserver interaction at a time in training data and never build a model of the interaction. Teaching them would need runtime state over time, like time-travel debugging, which is a huge amount of context. He says he doesn't expect this fixed in one, two or three years, and calls that his opinion given his experience.
- LLMs learn code patterns, not concepts, and fail out of distribution. Mark places himself nearer the "stochastic parrot" camp than the scale-brings-generality camp. He cites an Apple paper and Yann LeCun and Richard Sutton calling LLMs a dead end because children learn from a few examples. Hanselman adds that one-shot demos like Space Invaders are poor tests because the training data is full of them.
- Thinking tokens are extra passes, not a faithful inner monologue. Mark says the tokens give the model more passes through its weights before answering, and that past a budget more thinking can degrade answers. He relays studies, including Anthropic's, in which models penalised for showing intent learn to hide it from their thinking tokens. He says he red-teamed GPT-5's thinking model and calls it the safest model on the market.
- Vibe-coded businesses break as requirements outgrow the architecture. Mark describes projects that start simple, accumulate hacks as features arrive, and reach a point where the AI cannot keep the interactions straight. Hanselman lists his own incidents: an agent pushing to main from another branch, deleting test data, and using sleeps to fix race conditions.
- AI-assisted coding pays off for small tools and lower effort. Hanselman vibe-coded a GitHub Actions artifacts analyzer in a day after running out of quota. Mark says reviewing AI-written code costs him a fraction of the effort of writing it. He calls it AI-assisted rather than vibe coding when you read the output. He warns that juniors won't build the code-smell intuition needed to see past the AI's mistakes.
Dr. Jon KrohnProblems AI Can't Solve but Mathematical Optimization Can (with Gurobi's Jerry Yurchisin)
- Prediction tells you the future; optimization tells you what to do. Yurchisin, on his third appearance (after episodes 723 and 813), argues that classical ML and statistics forecast from data but do not say what actions to take, and that a better prediction helps little when the decision space is astronomically large.
- Every optimization model has decisions, constraints and an objective. His example is supply-chain network design: the decisions are which distribution centres to open (say DC, Atlanta, Boston, Houston); the constraints are cost, staff and regulation; and the objective is minimising cost, maximising profit or minimising carbon. The output is concrete values, such as how much of product A to make at each plant and ship to each centre.
- Real-world scale is beyond people and predictive tools alike. Yurchisin says that with hundreds to tens of thousands of products, multiple sources, shipping modes and quantities, a few simple decisions combine into a problem impossible for a person or a forecasting model to solve.
Stock Tok$2 Stock Purchases + When to SHORT
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon Krohn931: Boost Your Profits with Mathematical Optimization, feat. Jerry Yurchisin
synopsis
Jerry Yurchisin, a senior data science strategist at Gurobi, returns for a third episode to argue that mathematical optimization is the prescriptive complement to predictive AI. In his account, it is provably optimal or provably close, increasingly accessible through LLM assistants and GPUs, and usable today rather than after quantum computing arrives.
What optimization does. Yurchisin contrasts classical ML and statistics, which use data to predict the future, with optimization, which says what to do. A problem is reduced to decision variables, constraints and an objective. His example is supply-chain network design. Candidate distribution centres in DC, Atlanta, Boston or Houston are constrained by cost, staff and regulation while cost, profit or carbon is minimised or maximised, and the output is how much of each product to make where and ship to whom. With thousands of products, he says, this becomes impossible for people and for forecasting tools.
Feeling the difficulty. Gurobi's burrito optimization game asks players to choose how many food trucks to deploy and where. Yurchisin says about 20% of players match the optimum on the simplest first day, after which matching it is impossible. His point is that intuition fails at scale, where shaving 2% off fuel costs is a large saving. A coffee-pricing game built on queuing is planned for 2026.
GPUs and Nvidia. The host says Gurobi is used by a large share of top US companies and was name-dropped in Nvidia CEO Jensen Huang's GTC keynote. Yurchisin explains the link: Nvidia released an open-source GPU solver, cuOpt, and the firms shared learnings. Traditional algorithms rely on pivoting, which "GPUs hate" because it does not parallelise, but newer GPU algorithms are cutting solve times on very large continuous linear programs. Gurobi is adding this to its product and inviting customers to send such problems.
Free resources. He points to gurobi.com/learn, a free size-limited licence via pip install gurobipy, a Udemy course built with a Georgia Tech professor, and YouTube series with a 301 instalment due in late November. He calls a model a "mathematical twin" of the business.
Provable optimality. Asked by the host whether optimization truly finds the optimum, Yurchisin says yes: the geometry of the feasible region lets solvers guarantee it. When proof is out of reach, the MIP gap between the best solution found and the theoretical bound still says how far off the answer can be, for example 5% or 2%. Heuristics like simulated annealing offer no such guarantee and can stall in local valleys. He links the field's simplex roots to the same mid-century era that produced neural-network maths.
LLMs in the workflow. Gurobi has three custom GPTs, including a prompt engineer for framing the problem statement and a modeling assistant that drafts the formulation and gurobipy code and solves small instances. A new hosted Gurobot is built on AWS with Claude. He warns that LLM output needs caution and that the biggest failure mode is leaving levers or constraints out of the problem statement.
Customer cases. From Gurobi's Decision Intelligence Summit, he relays Toyota's natural-language interface for planners testing tariff swings against a manufacturing optimizer, and Total Wine's data-science team, new to operations research, building purchasing optimization with Gurobi's support and saving an undisclosed "lot of money". He says many customers keep optimization secret as a competitive edge.
Quantum. On the host's relay of Dell's CTO calling quantum inevitable, Yurchisin stays neutral but argues that NP-hard labels mislead people into waiting. The groundwork is identical either way, and a classical optimum becomes the benchmark for any quantum proof of concept, whose likely gains are finer granularity and faster solves.
- Optimization prescribes decisions where machine learning only predicts. Yurchisin, a senior data science strategist at Gurobi on his third appearance, says ML and statistics forecast the future but do not say what to do. Optimization casts a problem as decision variables, constraints and an objective, and returns concrete values, such as which distribution centres to open and how much of each product to make and ship, at a scale no person or forecast can handle.
- Optimization can prove optimality, or bound the remaining gap. Unlike heuristics such as simulated annealing, which stop when the solution stops changing and can stick in a local valley, Yurchisin says solving mixed-integer programs yields a mathematically guaranteed optimum from the geometry of the feasible region. Where a problem is too large to prove it, the "MIP gap" still guarantees the answer is, for example, no more than 5% or 2% from optimal.
- LLM assistants now turn business problems into optimization models. Gurobi has custom GPTs in the ChatGPT store: a prompt engineer that tests whether optimization fits a problem and drafts the problem statement, and a modeling assistant that writes the formulation and gurobipy code and solves small cases. A new hosted "Gurobot" is built on AWS with Claude behind it. Yurchisin advises treating LLM output with caution but calls it "really dang good" for learners, and says incomplete problem statements are a main cause of failure.
- Customers expose optimization to planners through natural language. Yurchisin relays two talks from Gurobi's Decision Intelligence Summit in Las Vegas: Toyota planners ran what-if tariff scenarios (0% to 10% to 80%) against a manufacturing optimizer through an LLM interface, and Total Wine's data scientists, with no operations-research background, built purchasing optimization in-house and now save what he calls "a lot of money", declining to give a figure.
- Waiting for quantum computing wastes savings available now. Answering the host's relay of Dell's global CTO calling quantum inevitable, Yurchisin does not take sides on that but says the NP-hard label wrongly persuades people mixed-integer problems must wait for quantum. The business-case, stakeholder and data work is needed either way, and a classical optimum would give a benchmark against which quantum's possible gains (finer granularity, minute-long solves cut to seconds) could be measured.
Polis PanditDo You Still Have Childhood Wonderment?
- Adults lose childhood wonderment by defaulting to cynicism, and "charitable assumption" is the author's proposed fix. Drawing on Will Guidara's Unreasonable Hospitality, he argues that giving people â like a complaining restaurant guest â the benefit of the doubt, rather than assuming bad faith, restores the openness needed to feel awe rather than jadedness.
- Deliberate surprise, as practiced at Eleven Madison Park, is what generates genuine delight in adults the way novelty does in children. The restaurant researched guests in advance to personalize moments of surprise; the author argues adults can't manufacture daily surprises for themselves but can practice the underlying mindfulness and gratitude that make surprise land.
- Play, not just openness, is the other missing ingredient â citing Rick Rubin's The Creative Act. For a toddler, play is work; adults lose that experimental, fear-free relationship to their own work as they age, and reconnecting with why they started it is framed as a route back to wonderment.
- The essay is framed around the 2025 Mariners' first ALCS run since 2001, a franchise that has never reached a World Series since 1977. The author treats attending games with his mother as itself an exercise in charitable assumption and mindfulness â a "win" regardless of the outcome on the field.
Master InvestorDan Morehead: We're still very, very early in blockchain
- Morehead argues blockchain assets have low correlation to risk assets because institutions hold essentially zero exposure. When portfolios broadly de-risk, he says, there's simply no crypto to sell, unlike other asset classes that became correlated after everyone adopted similar modern-portfolio-theory allocations.
- He predicts blockchain becomes a standard institutional asset class within a decade, similar to how commodities and emerging markets gained that status. He forecasts institutions holding 8â10% of assets in blockchain by 2030, up from roughly zero today and about 2% even at the most thoughtful endowments.
GeoHussarThe US rejected Venezuelaâs âeverythingâ offer: part 5. #Venezuela #Maduro #Karakas #LatinAmerica
- He expects Maduro to flee rather than face a US invasion. He says the US now faces no obstacles and its president wants to crush Venezuela's government, but predicts the US will "spook" Maduro into fleeing, followed by a more America-friendly government. He expects half or more of the population, including the military, to celebrate.
- Major Russian or Chinese concessions could still avert Maduro's ouster. His caveat: if Russia pulls out of Ukraine, or China and Russia both offer some large concession to the US, he thinks the ousting would not happen.
- He would not be surprised to see Machado lead Venezuela. He says the opposition candidate (GonzĂĄlez Urrutia) won the 2024 election by international recognition while Maduro claimed victory, and points to Machado dedicating her Nobel Peace Prize to the US president as a reason she might end up in charge.
Dr. Jon KrohnThe âGrandmother Neuronâ Returns â Inside BDHâs Interpretable AI (with Adrian Kosowski)
- Positive activations make the model's concepts directly readable. Responding to the host's comparison with the old "grandmother cell" idea, Kosowski says that in BDH a concept is simply a set or combination of units that fire, with no balancing of positive and negative coefficients or projections, which he calls "full interpretability".
- Important concepts compress into small sets, down to single synapses. He says the trained network follows a power-law pattern in which more important concepts are represented more compactly. Even at a scale of 100 million parameters or below, an individual synapse can be sufficient evidence that a concept such as a currency or country is being mentioned, which he describes as monosemanticity.
- These "grandmother synapses" emerge in training rather than by design. Kosowski stresses that nothing in the architecture assigns concepts to locations; the specialised units appear spontaneously, where is not controllable, and the system's context state lives in synapse potentiation rather than in individual neurons.
Dr. Jon KrohnThe Modular AI Model â Add New Languages Like Lego Blocks (with Pathway's Adrian Kosowski)
- Separately trained models can be concatenated into one multilingual model. Kosowski says that because BDH scales along a single dimension, the number of neurons, two models trained on different languages (the host's example is English and French) can be put side by side and work out of the box, mapping both languages to shared concepts. Optional post-training helps the combined model coordinate. Two transformers, he says, have no clear way to be connected.
- Small models already match peers; scaling up chatbots is not the goal. Kosowski corrects the host's GPT-2 comparison: Pathway's 1-billion-parameter models are comparable to or outperform other 1B models today, and 1B is used because it is the smallest scale with instruction following. Nothing stops a 70â80B model, he says, but non-reasoning chatbots are a commodity market, so Pathway is entering reasoning models.
- The target is reasoning over billions of tokens of private context. He describes ingesting enterprise documentation of about a million pages (roughly 1 billion tokens) "in a matter of minutes given enough hardware", and coding assistants that internalise a whole large codebase rather than writing new code in isolation.
Mr. GlobalIs Joe Biden The Reason Trump Is President
- Biden owns the loss as party leader, not for any single reason. Responding to a clip arguing Biden is why Trump is president, he agrees the refusal to step aside after the midterms and allow a robust primary played a huge role, but says the whole four years cost the election.
- Biden's team avoided blaming Trump, fearing the look of weaponization. In his reading, with Trump under investigation and indictment, they never tied inflation or energy prices to him, never raised the 2020 OPEC deal, and instead blamed oil companies for price gouging; he cites studies attributing roughly half of inflation, or 2â3 points of a 7% rise, to energy in late 2022 and 2023.
- Poor messaging, not poor results, sank the administration. He credits Biden with real accomplishments but says they were badly sold, the climate-policy attack on fossil fuels hurt electorally, and Trump campaign claims were left unanswered, such as the belief that a pipeline set global oil prices.
- Losing the White House erased the accomplishments. He likens it to leading a football game and losing on a last-second field goal, arguing Trump is now systematically undoing everything Biden achieved.
Dr. Jon KrohnSparse Activation is the Future of AI (with Adrian Kosowski)
- Dense and sparse-positive activation are two worlds with no middle ground. The host contrasts densely activated transformers with the sparsely activated brain and cites about 95% of BDH's neurons as silent at any time. Kosowski describes the transformer's dense world and BDH's sparse positive world as separate regimes; sparse positive activations have been studied in vision and smell since at least the '90s, but he believes Pathway is the first to scale them for reasoning to transformer-level or better performance.
- BDH is GPU-efficient and can beat transformers on some hardware. Kosowski says reaching 1 billion parameters shows the approach "scales onwards", and that especially for inference it can significantly outperform transformers on certain hardware configurations.
- Transformer attention heads have stopped growing in dimension. He argues most open-source models, including Llama 3 and 4, are essentially GPT-2 scaled up with no single recipe. One dimension has converged, though: attention-head vectors of about 1,000 dimensions, into which every concept must be mapped. The paper reverses the usual view and treats the transformer as an approximation of BDH.
- Sparse positive spaces compose concepts as bags, without opposites. Where dense vector spaces add, subtract and negate, Kosowski says BDH combines concepts like a tag cloud or German compound nouns. He argues reasoning has no symmetric "opposite" either: telling someone not to think of blue makes them think of blue, which is not a damping-down mechanism.
GeoHussarUS and NATO may play trick against Russia: part 5. #Geopolitics #Europe #warfare #natomembership
- NATO patrols would force Russia to look weak or thin Ukraine. He argues that if NATO patrols begin, Russia must either leave them unanswered or divert resources from Ukraine, which he says would give Ukraine's military a chance to reclaim territory.
- He discounts the nuclear threat because Russia's arsenal is degraded. Citing his earlier post on Russia's failure to maintain its nuclear program during and after the Cold War, he claims Russia is unsure how many warheads work, can be launched, or would make it out of Russia.
- Russian incursions hand NATO a justification to escalate. He reads Russian drone and fighter-jet incursions into NATO territory as giving the US and NATO a "tremendous opportunity" to escalate their side of the conflict.
Loss DogTom Sosnoff - Lossdog
- Sosnoff frames Lost Dog as trading-style "quantitative leverage" applied to pay. He describes moving from 25 years building brokerages (thinkorswim, tastytrade) to this new venture as carrying over the same analytical approach, now aimed at the wage gap and wealth inequality, without specifying a mechanism.
Loss DogChinese Jiajun Zou - Lossdog
- Lost Dog's chief AI engineer repeats the November launch and fairness pitch. Zhou Jiajun describes Lost Dog as a Chicago startup founded by Thinkorswim and Tastytrade's founders, aiming to make labor compensation fairer by adding "leverage" to pay â without detailing the mechanism.
Loss DogFrench Younes Bichra - Lossdog
- Lost Dog's lead product manager repeats the pay-gap pitch, in French. Younes Bichra introduces himself and names Sosnoff and Sheridan as Chicago's successful founders of Thinkorswim and Tastytrade, framing the company's goal as making sure "everyone is paid at their true value."
Loss DogSpanish Sol Espinosa-Suarez - Lossdog
- The same Chicago Stock Exchange presenter repeats her pitch, in Spanish. Sol describes joining Tom and Scott as a host on the new streaming network, One Lucky Dog, framing its mission as bringing "economic justice" to workers.