tech

2026-05-25 · open the original ↗

• You can buy two of intelligence, cost and latency, not all three. The host takes this triangle from the guests' book Architected Intelligence, along with its techniques for "cheating" it: dynamic routing, streaming, semantic caching and ensembles.
• Streaming is the most effective cheat; ensembles are the most underused. The guest answering cites, without being sure of the figure, a statistic that every added 100 milliseconds loses about 10% of users. Caching is now easy because OpenAI and Anthropic often apply it automatically, with semantic caching as an optional extra layer. The guest sees ensemble methods gaining interest since Andrej Karpathy released his "Council of LLMs" a few months earlier.
• On-the-fly routing by difficulty is the hardest cheat. The guest notes that adaptive thinking in ChatGPT and Anthropic's products drew complaints: "having it try to guess which to do bothers everyone." The advice is to leave on-the-fly routing to the labs, and to route only where a deterministic rule for a given problem set exists.
• Good UX design can hide LLM latency entirely. The guest likens it to Instagram uploading a photo while the user types the caption. Designers can start AI jobs early and walk the user through other steps so the result appears instantly. Slow LLMs, they argue, mean engineers must relearn latency tricks that faster internet let them forget.

2026-05-24 · open the original ↗

• Hallucinations usually signal missing context, not a faulty model. The guests' book argues that hallucinations are gaps in "canon", the information a system accepts as authoritative. The guest answering says that with 2026 models, a model told you live on the third floor won't say the second. A hallucination therefore flags that the context is missing or that retrieval failed to fetch it. The host adds an example of his own: an HR agent given an undated two-year-old guide whose policy has since reversed, which would answer confidently and wrongly.
• A "canon manager" should own what agents treat as true. The authors coined the title to consolidate roles that organizations name differently. The anti-pattern they describe is critical agent context buried in a GitHub repo that only an engineer can change, when the engineer is not the subject matter expert. The canon should be moved in front of the experts who are accountable for the agent's behaviour.
• AI tools that cut friction make company wikis viable. The guest says no one used the old company wiki, but a well-maintained one now unlocks "a million things". The biggest gain for their company and others has been AI that spots conflicts in the knowledge base and proposes a diff for an expert to accept or tweak.

2026-05-23 · open the original ↗

• AI treats whatever data it is given as truth. So, Mumford argues, AI engineering falls back on data engineering. The book quotes Anchorman: whatever you put on the teleprompter, Ron Burgundy will read. Mumford likens this to data scientists who are hired, find the company's data in a mess, and spend two years doing data engineering. The hope is to pass that lesson on to AI engineering.
• AI engineering's data problem is unstructured documents, not missing numbers. Data science asked where the numbers were. Mumford says AI engineers in 2026 ask where the PDFs, images and Google Docs are, which are authoritative and which are junk, and how to chunk, embed and search them.

2026-05-22 · open the original ↗

• Shareholder primacy is a recent legal invention, not a law anyone ever passed. Eric Ries, author of the forthcoming "The Uncorruptable," traces it to a small group of academics and judges in the 1960s-70s who reinterpreted "any lawful activity" charters as requiring shareholder-value maximization — a norm he says has never been subject to legislation or referendum, yet is treated as binding by courts, as when Twitter's board was compelled to sue to complete Elon Musk's takeover.
• Founders keep losing companies they built even while succeeding by every normal measure. Ries cites Twilio's Jeff Lawson, ousted 199 days after his dual-class control sunset expired despite revenue growth since IPO, and Sol Price, locked out of FedMart in 1975 by a new board after 20 years building it; FedMart went bankrupt within seven years of adopting standard shareholder-value practices, while Price's next venture became Costco.
• Independent board directors don't solve investor control — they just make it look balanced. Ries argues independent directors have no equity stake in a company's mission but do have a financial incentive to be seen as pro-investor, since investors recommend them for future board seats, making the classic two-VC/two-founder/one-independent board effectively investor-controlled; he cites data that companies using a separate nonprofit-trustee structure instead are six times more likely to survive to year fifty.
• A nonprofit trustee board can preserve value the market would have destroyed. Ries recounts Novo Nordisk's foundation trustees blocking a lucrative 2000s merger because it wasn't necessary for the company's survival — the rival company that merged instead had its R&D cancelled after being acquired by Merck — allowing Novo Nordisk to keep funding an unproven GLP-1 program for over a decade to a later $600 billion valuation; he points to Anthropic's similar long-term benefit trust as enabling its refusal of a $200 million contract.

2026-05-22 · open the original ↗

• Recent computer science grads now fare worse than the average new graduate. The host cites the New York Fed: 7% unemployment for computer science grads aged 22 to 27 and nearly 8% for computer engineering. That compares with about 6% for all recent grads and about 4% for the US workforce. He adds that US computer science enrollment fell 11% in 2025 and programming enrollment 26%, and that Handshake postings are down roughly 50% from their 2022 peak. Revelio Labs data suggest entry-level software and data analytics postings fell by as much as 67%. He contrasts this with White House adviser Kevin Hassett's claim of no sign in the data that AI is costing jobs.
• Some evidence says AI is shutting the young out of entry-level jobs. A 2025 study by Stanford's Erik Brynjolfsson and colleagues found employment among 22 to 25-year-olds in AI-exposed jobs down 13% since 2022, while older and less-exposed workers held steady or rose. The host says the Dallas Fed replicated the pattern this year and found it driven by non-hiring, not layoffs.
• Other studies find no AI effect, blaming rates and over-hiring instead. Two Google economists found posting declines in AI-exposed occupations as steep for seniors as juniors, and starting before ChatGPT. A Federal Reserve study of over a million firms found "precisely estimated null effects". Skeptics point to high interest rates, the end of the tech boom and 2021 over-hiring. Stanford's Eric Roberts likens the moment to the dot-com bust, after which hiring recovered by 2004.
• The host's view: macro explains much, but entry-level exclusion is real. He says the granular evidence that young AI-exposed workers are uniquely shut out is hard to dismiss. At his own consulting firm, Claude Code and OpenAI's Codex make code so cheap and fast that he is not sure why it would hire any entry-level developer in the foreseeable future.
• Grads should differentiate on design, domain, portfolio, agents and network. The host's five tips: compete on system design rather than raw coding output; pair AI fluency with a domain such as healthcare or law; build a public portfolio, since recruiters distrust CVs and applicant-tracking systems filter most out; learn agentic tooling such as RAG, evaluation and multi-agent orchestration, citing PwC's finding of a 56% wage premium for AI skills; and rely on referrals over job boards. He says the advice applies at any career stage.

2026-05-22 · open the original ↗

• Project Nomad now supports AMD GPU acceleration for its AI features. Version 1.32.0 of Crosstalk's free, open-source offline server runs Ollama on AMD GPUs with ROCm support, where only Nvidia was supported before. The team tested it on a Radeon 780M iGPU and an RX 6800.
• Knowledge-base ingestion was redesigned and now defaults to manual. Users pick "always" or file-by-file ingestion in the setup wizard or on first AI chat; manual is the default because a CPU-only install can peg the CPU for days. Only text is indexed, so ZIMs heavy with images or PDFs show errors, and deleting an entry removes only the ingested data, not the file.
• Indexing a ZIM can take 5 to 10 times its size on disk. The host says ZIMs are compressed, so a 10 GB file could become 50–100 GB once indexed. Nomad now warns before ingesting anything over 50 GB or more than 10% of remaining free space.
• Maps and content sources were extended for international users. Maps now show coordinates and take named, coloured pins, with pin notes slipped to a 1.32.1 due within days. Maps can be downloaded by country at a chosen zoom (15 is street level; Nicaragua was about 125 MB), and the content explorer adds Kiwix mirrors and custom ZIM libraries.
• AI handling got more robust and is now fully optional. Nomad auto-recovers Ollama when Docker changes drop its Nvidia passthrough, and switching models mid-chat now unloads the old one so it doesn't spill out of VRAM into slow system RAM. Leaving AI unselected in the setup wizard skips every AI step.

2026-05-21 · open the original ↗

• OpenAI is investing $2 million in tokens, not cash, into every spring and summer 2026 YC company. Garry Tan explains the money comes via an uncapped SAFE, effectively pricing the investment at the company's future Series A valuation but committing it at the earliest stage, and can be spent on Codex or the OpenAI API.
• The deal targets founders "token maxing" through software-factory patterns. Tan describes dispatching many agents to burn large volumes of tokens as a way to reach what would have been 18 months of engineering and Series-A-level product polish in a fraction of the time, and says YC observed its best current companies already spending heavily on tokens this early in their life cycle.
• YC extended a special application window since the summer batch deadline had already passed. Tan says the offer, developed together with Sam Altman, is being run as an experiment for just this spring and summer batch, with applications due by Memorial Day weekend (May 25) and decisions promised by June 5.

2026-05-21 · open the original ↗

• Zillow Offers lost to a winner's curse once its features drifted. The host gives the loss as almost a billion dollars. Miller describes the trap: an offer set too high is accepted and Zillow loses, while one too low is refused. The pilot went well at first, but COVID and other changes caused model drift. Miller treats any adversarial position as a shift in the feature space the model never planned for.
• Modern AI products face far more adversarial environments. Miller would hate to build a customer-service bot that hands out discounts to settle complaints, since Reddit posts teach people to trick it. Such projects need observability that flags out-of-norm behaviour immediately. At the start of COVID, Miller says, many data science teams threw out their models and forecast on the fly because the present no longer resembled the past.

2026-05-20 · open the original ↗

• Product taste is judgment, distinct from personal preference. Mark separates disliking a film from judging a product, and says the Sysinternals and ZoomIt design choices are Mark's own preferences, which suit others because they rest on intuitiveness and ease of use. Hanselman's example of missing taste is Windows voice access, whose bar takes over the top corners where users throw the mouse for close and the system menu; Hanselman says its designer made voice access the best rather than Windows.
• Their preceptor experiment showed most developers lacked it. A small group of Microsoft developers built ZoomIt's video trimmer without a visual spec, and both hosts judged most submissions clunky in both UX and expected behaviour; Mark's own version was the intuitive one.
• Taste comes from exposure and cannot be bootcamped. Hanselman calls it "having seen stuff and then filed it away"; Mark calls it experience, which "directed, methodical exposure" can accelerate only to some extent. Neither thinks a boot camp yields a senior engineer, and Mark says it takes sustained, repeated exposure to variants of the same thing.
• Software can't be fully specified, so engineers must still read code. Mark rejects a friend's claim that nobody will need to look at code, describing "sculpting" by iterating on AI output, as with the trimmer and a reference-checker app with dozens of behaviour questions; Mark notes that Carnegie Mellon's Software Engineering Institute, formed in the '80s, never codified programming. Hanselman adds that whatever is left unsaid gets filled by "the default part of the switch statement", and that AI does well on common apps but goes wacky at edge cases with little prior art.
• Mark still sees value in a degree. Mark calls it structured training, a gauntlet that serves as shorthand for evaluating someone, and says CS matters beyond simple web-plus-database apps. Hanselman worries that schools are not moving fast enough to teach thinking alongside AI, and Mark predicts "a lot of schools turning out ChatGPT degrees".

2026-05-20 · open the original ↗

• The client was paying twice for poor 5G at two nearby offices. A Southern California catering company had a Verizon 5G home internet connection in each of two offices about six units apart, getting 40–60 Mbps with the modems buried among stainless-steel kitchen equipment. Charter Spectrum offered wired service billed separately per site at well over $150, and that only as a first-year promotional price.
• Crosstalk's design put one 5G router on the roof feeding both offices. It uses a Crosstalk Mobile SIM, the firm's own product, in a Peplink Max BR1 Pro 5G inside an IP67 Antenna Max enclosure magnet-mounted on the roof. It is powered by PoE++ over a single cable, clear of the kitchen, with outdoor Cat 6 run across the roof into the second office.
• A cheap TP-Link Omada switch and Wi-Fi 7 APs finished the build. The host found a $75 unmanaged DS106GPP with one PoE++ port for the router and three PoE+ ports for EAP772 access points, which adopt by QR code in the Omada app with no on-site controller. The team hadn't realised the APs came without PoE injectors, and bought a TRENDnet one locally for $30 for the second office.
• Both offices now share one LAN and one connection at mid-300 Mbps. The host puts that at about eight times their previous roughly 40 Mbps, with a single bill. He and a Crosstalk engineer add that Crosstalk Mobile SIMs have exceeded 700 Mbps elsewhere, in Pasadena.
• A neighbour's parking-lot link uses APs not built for point-to-point. The host spotted a pair of Ubiquiti U7 Pro Outdoor omnidirectional APs, presumably meshing across about 60–70 ft of clear line of sight. He thinks it probably works at that distance, but would not have chosen them, and worries about their RP-SMA connectors being left uncapped in the weather.

2026-05-20 · open the original ↗

• With near-instant prototypes, an AI project must prove value on day one. The guest says enterprises hold on to ideas too long because traditional software rewarded sticking with a partial product. Their example is the text-to-SQL chatbot that, they say, every company tried, including theirs early on at Pattern: a day yields an impressive demo, so companies commit and keep going. The guest cites AWS, which had just announced that six engineers rebuilt its AI infrastructure over two or three months in a move from Bedrock to "Project Mantle", as the resolve to start over.
• Once building stops being the bottleneck, definition and feedback decide outcomes. In the define-build-feedback cycle, slow building used to hide weak definition and feedback. The organizational fix the guests describe is narrow lanes: two or three people with a clear target and freedom to decide, who fail more readily because they are not failing in front of a large audience.

2026-05-19 · open the original ↗

• Zepto's founding pivot came from data, not strategy: one dark-store neighborhood outperformed the rest 3-4x. Aadit Palicha says he and Kaivalya Vohra started at 17 delivering groceries via a WhatsApp group in Mumbai during COVID, under the name Kirana Cart, but only switched to running their own mini-warehouses — the first literally inside a co-founder's apartment — after seeing that neighborhood's order volume dwarf areas still served by doorstep delivery from existing stores.
• They designed backward from an extreme ideal experience rather than forward from what was operationally feasible. Palicha credits a discussion of Airbnb's "seven-star" thinking for the approach behind 10-minute delivery, and argues that nailing customer experience — not cost-cutting — created the demand and throughput per warehouse that then made costs low, concluding "customer delight is where financial value starts."
• He delayed dropping out of Stanford for about eight to nine months until reaching early product-market-fit signals. Palicha says he waited until the business hit roughly 10,000 orders a day and a run rate of 60-70 crore, plus secured a term sheet, before committing full-time — advising founders to have proof a business is working before taking that kind of risk, not before.
• AI operations work has let Zepto's supply chain and corporate teams run with far fewer people. Palicha says machine-learning forecasting now runs dark-store replenishment for millions of units a day without humans in the loop, replacing forecasts that used to take teams days, while GenAI-driven ad-keyword tools boosted advertising revenue and internal automation let the company cut nearly all SaaS and outsourced-service spending with a corporate team of about 500 engineers plus 150 in data, product and design.

2026-05-19 · open the original ↗

• Design every process so a human, an AI or both can run it. The book's "user agnosticism tenet" is illustrated by a sales chatbot. A human can join at any point and pick up where the bot left off, so automating the pipeline does not lock the human out and a later rollback does not force a massive refactor.
• When building is cheap, definition and feedback become the bottleneck. Prototypes are near-instant, so the guests say AI projects must prove value on day one or be scrapped. Their example is the text-to-SQL chatbot every company tried, whose day-one demos lured teams into persisting. Miller's fix is narrow lanes of two or three people who can fail repeatedly without a large audience. A guest cites AWS's just-announced rebuild of its AI stack by six engineers over two or three months as the resolve needed.
• Build workflows before agents, and remove AI where a rule will do. If you aren't happy with a workflow where you set the path, the guests ask, why trust an agent improvising one? They see teams progress from skills (markdown files subject matter experts can iterate on) to workflows to agents. Many "agent builders" are really workflow builders, they say, and deterministic steps cut cost and risk. One guest swapped an AI step for a decision tree of canned statements and user complaints vanished overnight.
• Hallucinations mostly mean the canon is missing or wrongly retrieved. Mumford invokes Anchorman (Ron Burgundy will read whatever is on the teleprompter) to argue that AI engineering is now mostly data engineering over unstructured PDFs and docs. The guests propose a "canon manager" who owns authoritative context. That context should move out of engineers' GitHub repos to accountable subject matter experts, with AI tools proposing diffs to resolve conflicts.
• Fine-tune rarely, and treat speed as the only durable moat. A guest says more people fine-tune than should; the case is a narrow, stable task such as classification at millions or billions of data points. Miller weighs a gain of 93 to 95% accuracy against the projects forgone, and the host cites BloombergGPT's eight-figure cost and the few weeks before GPT-4 eclipsed it. Since everyone has the same frontier models, Mumford argues advantage comes from cycling define-build-feedback faster, with taste setting how much each cycle improves.

2026-05-18 · open the original ↗

• Visual checks at every pipeline step catch errors early and cheaply. Manz calls himself a "visualization person" who wants to see the data at each stage. He says he has often loaded a dataset with problems such as missing values that a quick visual check would have caught. Missing them cost him an afternoon down the wrong path, or baked the flaw into a result.
• Scientific visualization tools should live where the data work happens. In biomedical fields, Manz says, new experimental techniques yield abstract data with no familiar representation, so tools such as the genome browser had to be invented. Microscopy tools tend to be one-off applications outside the analysis environment. He argues for composable tools at hand for cheap checks, such as overlaying a segmentation on a tissue image to test the thresholding.

2026-05-17 · open the original ↗

• Coding agents are already a form of recursive language model. Manz describes recursive language models, from research released last year, as a model with an environment it can offload context to rather than stuffing everything into the prompt. He says this counters "context rot", where models forget or go off track as context grows. For today's coding agents that environment is the file system, with tools to fetch context when needed.
• Marimo pair extends the agent's environment into live Python memory. With Marimo pair, Manz says, context can be values held in memory as well as files, so the agent need not pull all the data's context into its window. The notebook becomes a structured representation of the agent's working memory. Ask a plain coding agent the same data question repeatedly and it might find 10 different ways through the file system; a notebook yields a similar scaffold each time, with each cell recording a step to the answer.

2026-05-16 · open the original ↗

• Data work now starts in the coding agent, which drives the notebook. Manz describes asking Claude Code, or open code or Codex, to start a Marimo pair session, then declaring at a high level which datasets and libraries to use. The skill teaches the agent to start or connect to Marimo. The agent then edits the notebook open in a browser while the human can still edit cells by hand.
• Shared kernel state lets the agent see what the human sees. Because the agent can run code in the kernel, read cells and take screenshots of them, a user can circle something interesting and ask about it. Manz says notebook state that used to be trapped becomes context beyond the file system. Streaming the human's edits back to the agent is not built yet; he mentions exploring MCP's "channels" feature.
• Agents make exploratory analyses cheap enough to try. Just as agentic coding made software ideas feel cheap, Manz says he can now ask for, say, five validations he would otherwise skip for not remembering an API. He then lets the agent make plots and return when he can decide.

2026-05-16 · open the original ↗

• YC says the India event drew over 25,000 applications. A speaker tells the crowd that India's AI ecosystem will be defined by the people in the room, not by the prior generation.

2026-05-15 · open the original ↗

• The IBM speaker claims Java runs best on Z. The speaker says IBM has invested in Java on Z for over 25 years and optimizes machine-level instructions for it (sorting, shorter garbage collection), calling Java on Z "more performant" and "more efficient" than on any other platform, "full stop". The same passage appears in the 2026-05-11 Hanselminutes interview with Skyla Loomis.
• Modernization is an architecture question, not a language swap. The speaker argues that Z's value is full-stack optimization rather than the language, and asks whether a Java app moved off the platform would keep its performance, resiliency and latency once it is "disconnected from the data".

2026-05-15 · open the original ↗

• AI compute demand is outrunning a supply chain with several simultaneous bottlenecks. The host dismisses "tokenmaxxing", maximizing tokens consumed in Claude Code or Codex to look good on internal usage dashboards, as a vanity metric. He cites the symptoms of scarcity: Anthropic's weekly Claude Code rate limits from last August, which it blamed on power users running it 24/7 and said it was "very compute constrained", OpenAI pausing Sora sign-ups, GitHub halting new Copilot subscribers, and unreliable on-demand H100s without reserved capacity. Fixing one bottleneck, he says, does not unstick the system.
• Packaging and memory, not GPU fabrication, are the chip choke points. He reports Nvidia data-centre GPU lead times of 36 to 52 weeks, Blackwell largely sold out through mid-2026, and H100 spot rents up about 30% since November. TSMC's CEO C.C. Wei has said CoWoS packaging is sold out through 2026; Nvidia reportedly holds about 60% of it through 2027, and new fabs take two to three years. HBM demand has roughly quintupled since 2023 (an H100 needs 80 GB, a B200 192 GB), and SK Hynix, Samsung and Micron are largely sold out into 2026. Nvidia reportedly cut RTX 50 output 30% to 40% as a result, and new HBM fabs take 18 to 24 months.
• Agentic AI needs far more CPUs per GPU. Morgan Stanley analysis cited by the host puts chatbot systems at about one CPU per 12 GPUs and agentic systems at one-to-one. Intel's CFO David Zinsner said on the Q1 call that its server CPU shortfall "starts with a B". The host says server CPU prices have risen 10 to 20% in a couple of months, a big part of why Intel's market cap more than doubled in 6 months.
• Electricity is the most intractable bottleneck. Hyperscalers report build-outs gated by grid interconnects (18 to 36 months) and transformers (another 18 to 24). Gartner projects power shortages will restrict 40% of AI data centres by 2027. The host counts at least 12 states with moratorium bills as of March, a Maine statewide moratorium passed in April and vetoed, and more than 50 local moratoria. He also cites rate rises of 8 to 15% near clusters in Virginia, Texas and Georgia.
• Suppliers aren't matching hyperscaler spending, so relief is years away. The host says even an aggressive supplier ramp-up wouldn't bring meaningful relief until 2028 or later. The five largest hyperscalers are on track for about $725 billion of 2026 capex, roughly three-quarters for AI, against about $120 billion in 2022. He says Anthropic committed over $100 billion to AWS for up to 5 GW, took about 5 GW more from Google, and took all of the "SpaceX Colossus 1" site, which let it double Claude Code limits. Suppliers have raised capex only about half as much as hyperscalers' tripling, wary of overbuilding. The host forecasts 2 to 3 years of physical constraint shaping which AI firms survive, while efficiency gains (Google's TurboQuant, Amazon's Trainium cluster of over 500,000 chips, mixture of experts) keep cutting application costs.

2026-05-15 · open the original ↗

• A skill is a loose contract between a runtime and a model. Marimo pair is a skill, a folder of markdown files that teaches an agent to start or connect to Marimo and run Python inside its runtime. Manz says the agent discovers the APIs hidden inside Marimo rather than relying on a strict contract. If the APIs change between releases the model adapts, and upgrading Marimo simply adds capabilities, without the usual worry over semantic versioning.
• Skills took off once models got good around last fall. Manz dates the shift in coding agents to around November with Anthropic's Opus. He first said 4.6, then accepted the host's correction that 4.6 came out in February and settled on 4.5. When people returned in January, skills became a high-impact way to share team processes that once sat in documents people had to read.
• Repetition and failures should update the skill. Building Marimo pair taught Manz a habit: if he is repeating himself to the agent, that belongs in a skill. When something goes wrong he updates the skill rather than "yelling at the model".

2026-05-14 · open the original ↗

• IBM deliberately keeps the physical mainframe visible. The speaker describes "hug your mainframe day" social-media programs, on which people post photos with their data-centre mainframes on LinkedIn, and says real design thought goes into the machine's door so that it makes an impactful visual at launches. The same passage appears in the 2026-05-11 Hanselminutes interview with Skyla Loomis.

2026-05-13 · open the original ↗

• Foundry agents ship with default guardrails that block or log. Working on the agent from episode 215, Giard shows two built-in guardrail versions, which he says do much the same thing, with version 2 doing a little more. They cover categories such as hate speech, self-harm and sexual content, on user input (prompts), on output (responses), or both. Each one either blocks the content outright with a refusal message, or only logs that it happened.
• A custom guardrail can be built once and reused across agents and models. Under the project's guardrails section, the options Giard lists are a jailbreak check that applies to user input only, a protected-material check for code or text that applies to output, which he says is "so that you're not revealing secrets", and categories such as hate, self-harm, sexual and violence that apply to both. Each risk has a severity level that sets how severe content must be before the guardrail kicks in. He attaches the guardrail to his one agent and to both models he has deployed, then names it.
• The demo hate-speech guardrail blocked a threatening prompt. After he assigned the guardrail to the agent, the test prompt "I will kill you" got a refusal saying the request was blocked by guardrails.
• Block lists forbid specific terms in prompts or responses. Foundry includes a built-in profanity block list, which Giard says rejects prompts containing swear words and filters them out of responses. A custom list can be loaded from a CSV file for long lists of terms, or entered one term at a time. Each term is matched either exactly, ignoring case, or by a regular expression.
• A block list takes effect only through a block-list guardrail. Giard builds a block list containing "rosebud", then creates a guardrail of type block list that uses both it and the built-in profanity list, on input and output. Annotations are not supported for this type, so it can only block. Once it was applied to the agent and its models, a prompt asking about "rosebud" was refused because the block list forbids the term.

2026-05-13 · open the original ↗

• Fraud must be caught during the transaction, not after. The speaker says that a delay of even 30 to 60 seconds forces a bank to let the transaction through, and that fraud let through is basically a loss, which many banks treat as a cost of doing business at about 8% of revenue a year.
• On-platform AI makes scoring every transaction feasible. Running inference where the transaction happens removes network latency, which the speaker says gives sub-millisecond or single-digit-millisecond responses and allows scoring on 100% of transactions. The same passage appears in the 2026-05-11 Hanselminutes interview with Skyla Loomis of IBM Z.

2026-05-13 · open the original ↗

• Every field concentrates in one place, and the historical answer is always to go there. Paul Graham compares Silicon Valley today to Paris for painting in 1870 or Göttingen for math in 1900, arguing the talent pool at a center improves in two dimensions — better people and more of them — which he says produces disproportionately valuable serendipitous meetings that he cannot fully explain but has observed repeatedly in biographies of accomplished people.
• Silicon Valley investors decide faster because competition punishes hesitation, and it doesn't hurt their returns. Graham argues the more confident an investor is about an opportunity, the less time they can afford to wait, illustrated by a Boston VC firm that faxed Dropbox a blank-valuation term sheet only after learning Sequoia was interested — too late, since Dropbox went with Sequoia and became the first YC company to go public, in 2018.
• Comparing yourself to visible "big fish" resets ambition from impossible to merely hard. Graham says moving to a big pond lets founders measure themselves against known peers like Brian Chesky or Sam Altman and conclude they aren't a different species, just people who worked exceptionally hard — turning a fogged-in, seemingly unreachable summit into "a high but definite threshold."
• Founders should go to Silicon Valley, ideally via YC, and then return home — even though data shows returners do worse on paper. Graham argues going and coming back is the best way to help a hub like Stockholm, bringing back money, imported culture and a stronger startup, but acknowledges YC data shows founders who return home are only about half as likely to become unicorns as those who stay — a gap he attributes partly to selection bias and to Bay Area startups raising at higher valuations, not necessarily performing better.

2026-05-13 · open the original ↗

• Marimo pair is a skill that lets agents use notebooks as a tool. Manz describes it as an agent skill that teaches an agent to do anything a person can do in a Marimo notebook, and more. Marimo is MIT-licensed and fully open source.
• Agents lacked access to in-memory state, so notebook work meant babysitting. Manz defines an agent as a model plus tools in a loop. Coding agents are good with files and commands, but data work depends on exploring data loaded into memory, which agents could not see. Using notebooks with agents felt like prompting chatbots a couple of years ago, relaying errors by hand. In software, by contrast, agents run long self-correcting loops of edits and tests.
• Opening Marimo to the agent brings those self-correcting loops to data work. Marimo pair lets the agent work with code and its values in memory, and ask about anything the user can see on screen, by letting it write and run Python inside the runtime.

2026-05-12 · open the original ↗

• Mainframe development's friction was the tooling, not COBOL. Asked where AI fits with COBOL, which the questioner calls "shockingly a growing business", the other speaker blames decades-old, complex applications that lack documentation and tests, along with the developer tools. The speaker says z/OS applications can now use VS Code, Jenkins, automated-test CI/CD, Ansible and Artifactory, the same experience as distributed platforms. The clip is from the 2026-05-11 Hanselminutes interview with Skyla Loomis of IBM Z.

2026-05-12 · open the original ↗

• Peplink is raising list prices on select hardware. The increase takes effect June 1, making May 31 the last day to order at current prices; it covers select hardware models, not service plans, and the full list of affected models was not expected until mid-May.
• Crosstalk, which resells Peplink, offers a May purchase incentive. Buyers of Peplink hardware through Crosstalk in May can choose 25% off a two-hour support block or 50% off the first month of a Crosstalk Mobile unlimited 5G SIM, and the host offers to lock in current prices for pipeline projects emailed in before May 31.

2026-05-12 · open the original ↗

• A reactive notebook guarantees the document matches the actual state. Manz says Marimo, unlike Jupyter, tracks dependencies between cells and re-runs stale ones automatically. That removes the overhead of tracking state and of relying on deleted variables. He jokes that it is "disguised Excel" with Python instead of Visual Basic. Marimo is MIT-licensed open source, was acquired by CoreWeave last fall, and is meant to double down on open source.
• Agents must see in-memory state to iterate on data work. Manz says coding agents excel with files and commands but couldn't see notebook state, so users babysat them by relaying errors. Marimo pair is a skill that lets Claude Code, Codex or open code start or connect to a Marimo session, run Python in the kernel, read cells and screenshot them, while the human edits alongside. The result is a reproducible notebook that stakeholders can audit, not logs or loose scripts.
• Skills work best as declarative guides, not step-by-step scripts. Manz says the agent discovers APIs hidden inside Marimo, so the contract is loose and API changes between releases don't break it. His first imperative version of the skill was followed too dogmatically, mishandling exceptions, so it became more declarative: teach the tools and their nuance, and let the model work out how. He dates the rise of skills to around November and Claude Opus 4.5.
• Coding agents are recursive language models, and the notebook is their memory. Manz describes recursive language models as models that offload context to an environment to avoid "context rot", which for today's coding agents is the file system. Marimo pair extends that to Python values in memory. The notebook's cells then form a structured record of the agent's working memory.
• Widgets lower the activation energy of visualization, like a catalyst. Manz built AnyWidget, a standard and Python implementation for interactive widgets across Jupyter, Colab and Marimo. Visualization tools usually need an app download and data export, he argues, while an algorithm is just an import away. A selection made in a widget becomes a variable that reactively triggers downstream cells, so an outlier can be clicked rather than filtered by hand. His "progressive" design lets a widget start as plain JavaScript in a notebook and move into a library, with tooling such as TypeScript added only as it grows.

2026-05-12 · open the original ↗

• Talking to computers is close to practical. One speaker uses Scotty in Star Trek 4 talking to a mouse to say "we're almost there." The other describes a voice agent, "still in testing", that controls a set of coding agents: it fires off prompts, wakes up to notify, and can be run through headphones while walking around the yard.

2026-05-11 · open the original ↗

• Loomis says today's mainframe is modern, scale-up compute. Loomis, general manager of IBM Z software, compares it to a car: invented in the 1960s, but nothing like a Model T today. Loomis claims eight nines of availability "inherently in the box" (roughly 300 milliseconds a year, by Loomis's estimate), and says the platform runs z/OS or Linux, is built to run dense and hot, and does heavy online transaction processing as well as batch.
• Z is a fit-for-purpose "race car", cheaper the more it carries. Loomis advises against it for small or experimental workloads where a client has no Z already, and for it on core transactional work where an outage risks brand trust, and where data gravity makes copying data elsewhere a security, cost and compliance burden. Boxes ship fully loaded, and clients pay only for the capacity they enable (say 60%), run that at 80–97%, and turn on the "dark capacity" to scale. Loomis's example is Citibank moving its MongoDB from x86 to Linux on Z, reportedly bringing over 50% power savings and meeting cyber-resiliency requirements it could not meet on x86.
• AI belongs on the mainframe because latency decides fraud outcomes. Loomis says many banks accept losing about 8% of revenue a year to fraud, and that inference next to the transaction gives sub-millisecond or single-digit-millisecond responses, enough to score 100% of transactions. IBM added the Telum chip (traditional ML) in z16 about four years ago, and Telum 2 plus PCIe Spyre cards for small-to-medium generative models in z17, announced last year. Hanselman cites 128 GB and 32 cores per card and up to 48 cards; IBM is also bringing agentic AI to operators so they need not use green screens.
• AI helps COBOL developers when grounded in static analysis. Loomis says mainframe development was held back by tangled, undocumented, decades-old applications and by old tools, not by the language; z/OS now supports VS Code, Jenkins, CI/CD, Ansible and Artifactory, and some clients deploy to a CICS/DB2 banking app 20 times a day, with cultural change the hardest part. IBM feeds static and dynamic analysis of how transactions touch DB2 and IMS into the models as context, which Loomis calls its differentiator across code bases of "tens or hundreds of millions of lines". Hanselman's framing is that the LLM orchestrates tools while the development life cycle stays the same; Loomis agrees that good engineering and architecture matter more as code arrives faster.
• Modernization is architecture, not a language swap. Loomis says Java has run on Z for over 25 years with machine-level optimizations for sorting and garbage collection, and claims it "runs better on the Z platform than any other platform, full stop." Loomis asks whether an app moved off the platform keeps its performance, resiliency and latency away from its data, and adds that "COBOL is not that hard of a language."

2026-05-11 · open the original ↗

• Engineers should answer governance worries by asking what security needs to see. Gupta's questions for a zero-trust world are: how to get visibility into what agents are doing, how to govern what they can and can't do, and how to recover when they inevitably err. She says many organizations are deadlocked between security listing risks and engineers needing to build. Engineers can break that by asking what visibility, policies and resilience plans would make security comfortable.
• Enterprises lack AI literacy training on a par with privacy training. Al-Dhubaib says annual privacy and cyber training instils a sense of obligation about consequences. He sees no widespread equivalent that teaches workers the larger blast radius of the AI tools they now hold.
• AI-enabled knowledge work is cross-functional by default. Al-Dhubaib points to Rubrik's dedicated AI attorney, who began by asking about risk exposure and then upskilled through specialist conferences. He argues data scientists and engineers likewise need to learn the language of risk management.

2026-05-10 · open the original ↗

• Agent Rewind can undo system changes, not delivered messages. Gupta says the product draws on Rubrik's cyber-recovery base. It can restore a dropped production table, repair a Salesforce change that broke the link between opportunities and accounts, and reverse identity changes that granted excess access. An email that has already reached a customer cannot be recalled.
• Irreversible actions need runtime blocking; ambiguous changes need rewind. For high-risk actions such as trading or emailing customers, Gupta says the right defence is defining them up front so Sage, Rubrik's policy engine, blocks them at runtime. System changes are hard to judge at runtime because most are legitimate and telling them apart needs more context, so rewinding after the fact fits them better.
• Enterprises need four or five overlapping layers of defence. Gupta likens the layers to Swiss cheese slices, stacked so that together they leave no holes.

2026-05-09 · open the original ↗

• Non-deterministic agents must be governed by AI, not rule engines. Gupta frames Rubrik's approach around three pillars: visibility, governance and control, and remediation. She says over 80% of organizations feel they lack visibility into which agents are running and what they do. Because agents are non-deterministic, a simple rule engine can't monitor them.
• A fine-tuned small model wins on accuracy, cost and latency together. Sage, Rubrik's Semantic AI Governance Engine announced at RSAC in March, is built on a fine-tuned SLM. The host describes Rubrik's published head-to-head as five times faster than GPT 5.2 and more accurate at detecting policy violations. Gupta says an enforcement layer can't wait 5 seconds per decision when agents chain many actions. The capability came from Rubrik's acquisition of Predibase about a year ago, bought to add model and fine-tuning expertise to its data and identity knowledge.
• Rubrik Agent Cloud pairs agent discovery with natural-language policies and rewind. Gupta says it gives visibility across Microsoft Copilot, AWS Bedrock and Claude Code agents. Policies such as "no financial advice" are written in natural language and enforced by Sage at runtime, and Agent Rewind undoes unintended actions. Al-Dhubaib likens the discovery side to e-discovery for agents: a dashboard revealing hundreds of agents, 50 of them with read-write access to four databases.

2026-05-08 · open the original ↗

• Many AI tools hide code; this one puts files first. One speaker praises an unnamed tool as feeling like "a software development life cycle tool", not vibes, because it puts the file viewer and diffs up front. The other, speaking for its makers, says the end product includes the designs as well as the code, so the work should stay "local first, open formats first" on disk rather than in a cloud database behind "a thin view".

2026-05-08 · open the original ↗

• Garry Tan rebuilt the same blog platform three times, and the cost collapsed by roughly 20,000x. He says the 2008 version (Posterous) took $4 million, six or seven people and about a year and a half; a 2013 rewrite took roughly $100,000, two people and three months; the 2026 version, built with a Claude Code Max subscription, cost about $200 and five days while adding full agentic retrieval and research capability the earlier versions never had.
• "Boiling the ocean" — paying for exhaustive completeness — is now cheap enough to be standard practice. Tan argues that because an agent can cross-reference twenty sources as easily as one, builders should stop economizing on research or code depth the way a human effort-budget would force them to, calling this "token maxing" and arguing it will spread from coding into every kind of knowledge work.
• A "thin harness, fat skills" philosophy puts judgment in markdown, not code. Tan describes building G Stack around markdown skills — a "CEO" review for ambition, a plan-and-engineering review, a developer review, and browser-based QA — while keeping the underlying agent harness minimal, arguing that anything requiring human-style judgment belongs in plain-language markdown rather than brittle deterministic code, and that skipping 80-90% test coverage produces output "10x worse than human-written code."
• Measuring logical lines of code, Tan found he was producing about 400 times his 2013 output, directing roughly 15 agents at once, even though the code he personally typed dropped about 70% from his 2013 baseline — a distinction he says got lost when his "100x" claim drew online criticism. He frames the larger stakes as builders needing to control their own prompts and tools rather than depend on centrally controlled AI products, calling it "the defining question" of the moment.

2026-05-08 · open the original ↗

• Mythos's hacking ability emerged from general gains, not targeted training. The host recalls Dario Amodei calling GPT-2 too dangerous to release in 2019, which drew mockery, and says that this time he may be right. Amodei, now Anthropic's CEO, declared on April 7th that Claude Mythos preview was too powerful for wide release. Anthropic says it did not train the model to find vulnerabilities; the ability came from broad improvements in code understanding, reasoning and agentic autonomy. It reports 94% on SWE-bench, against top public models nearing 80%.
• Anthropic's figures show a generational leap in exploit capability. Against 147 Firefox vulnerabilities, Opus 4.6 produced working shell exploits twice in several hundred attempts and Mythos 181 times. On roughly 7,000 OSS-Fuzz entry points, Sonnet 4.6 and Opus 4.6 each reached tier five (full control-flow hijack) once, against 10 fully patched targets for Mythos. Using a simple scaffold of a container, Claude Code and a prompt, Anthropic says Mythos found thousands of zero-days across every major OS and browser. Mozilla patched 271 Firefox vulnerabilities in one release, against 73 high-severity fixes in all of 2025. Contractors gave 89% of 198 reviewed findings the same severity the model assigned, and 98% fell within one level.
• Gated release is safety framing and good business. Project Glasswing gives defenders including Apple, Google, Microsoft, AWS, Nvidia, JPMorgan Chase, the Linux Foundation, Cisco, CrowdStrike, Broadcom and Palo Alto Networks early access. Anthropic commits up to $100 million in credits and $4 million to open-source security groups. The host notes pricing of $25 and $125 per million input and output tokens, about five times Opus 4.6, which suggests heavy compute. He argues exclusivity also hinders distillation by rivals and nudges enterprises to Claude Code, and says OpenAI made a similar announcement about a version of GPT-5.4 days later.
• Skeptics say finding bugs isn't weaponizing them, and the head start is short. The host relays Bruce Schneier's note that a security firm reportedly replicated some findings with older, cheaper public models, and his view that AI may help defenders more for now. Glasswing could neutralize zero-days the US government has hoarded, which he expects Defense Secretary Pete Hegseth to dislike. Epoch AI estimates the open-weight lag at about 3 months on average, though 5 to 22 months on some benchmarks.
• The host's lesson: patching must now run at machine speed. He argues vulnerability discovery is no longer limited by scarce experts, and that the gap from bug to exploit has shrunk from months to minutes. Defenders who adopt automated scanning, AI code review and runtime enforcement first will come out ahead. Heavy Claude Code or Codex users likely ship far more vulnerabilities than in hardened software like Firefox. He suggests a dedicated security pass with those tools, PR review tools such as CodeRabbit, and AI-native tools like Socket, Endor Labs and Semgrep.

2026-05-08 · open the original ↗

• YC is hosting a one-day Startup School Paris on June 29th. Builders across France and Europe will hear from founders of companies including DataDog, Supabase, and Amie, plus get direct access to YC partners and alumni.

2026-05-07 · open the original ↗

• Recurring small choices are decision fatigue, fine to outsource. The speaker, married 25 years, says a long relationship boils down to "what are we eating tonight?", and calls the back-and-forth over restaurants decision fatigue. When you lack "the spoons" to decide, the speaker's advice is "just ask the LLM or spin the wheel."

2026-05-07 · open the original ↗

• Recurring small choices are decision fatigue, fine to outsource. The speaker, married 25 years, says a long relationship boils down to "what are we eating tonight?", and calls the back-and-forth over restaurants decision fatigue. When you lack "the spoons" to decide, the speaker's advice is "just ask the LLM or spin the wheel."

2026-05-07 · open the original ↗

• Most attacks start with stolen credentials, then escalate toward the crown jewels. Gupta describes the common pattern: phishing, or calling support posing as an employee to reset a password and MFA, as she recalls in the MGM case. Attackers then gain more access and other identities on the way to production systems and customer data, before exfiltrating, encrypting or deleting data for ransom.
• Recovery first requires knowing the blast radius. Before restoring, leaders ask what was hit, whether sensitive data was affected, when and where the attack began, and how not to restore the vulnerabilities too. Gupta says Rubrik's preemptive recovery engine scans data and how it changes over time to answer these quickly. She credits that with giving Rubrik a deep understanding of data and identity, which she calls the basis for securing AI agents.
• Non-human identities are multiplying the attack surface. Al-Dhubaib calls it an "identity crisis": every agent acting for a user holds keys, tokens and API access that can cascade across systems. One guest describes common Claude Code practice: fetching all their SaaS API keys, setting the longest expiry, granting full personal permissions and leaving tokens in a file on a laptop. They say enterprise application builders would never allow this, and that the tooling to monitor it instead of blocking productivity hasn't existed.

2026-05-06 · open the original ↗

• Giving agents personas doesn't make them better. One speaker doubts that telling an agent "you're an expert database person" improves its work, and says the only argument for naming and anthropomorphizing agent teams is saving context window, since a focused agent need not hold the whole project.
• The real value of a focused agent is to the human. The other agrees, saying the benefit is knowing which conversation to have with which agent, which improves their own interaction with it more than it improves the agent's output.

2026-05-06 · open the original ↗

• AI-assisted coding has sped up ZoomIt's feature growth. Mark wrote ZoomIt around 2001–2002 for Windows internals classes with Dave Solomon, and says new features have been "accelerating over the last year and a half" with AI help. Recent additions include Ctrl+5 snippet recording for reporting bugs and a webcam overlay that was "not out yet" at recording. The branch shown is ZoomIt 11.
• Panorama stitching works on any scrolling app, not just Win32. With Ctrl+8 the user drags a region and scrolls at any speed. Mark says it compares grayscale brightness (luma) on a grid at full size, captures about 15 frames a second, and uses heuristics that work in "99 plus percent" of cases. Its hard failures were black stripes, smears, and floating elements such as a chat window's down-arrow repeating through the stitch, all since handled. The first live attempt stitched horizontally because the tool used the box's aspect ratio as a direction hint, an edge case Mark blames on AI-assisted coding. A retry produced an 800×5386 image that kept a fixed header, and the feature is already released.
• Video clipping stays inline, short of an OBS replacement. ZoomIt now trims, appends clips and adds fade transitions, and Mark's LinkedIn video was made entirely in it. Mark wants every gesture to stay simple and in the flow, with no timeline editing, which Mark calls "OBS territory"; Hanselman admits to gently pushing toward that.
• Watching the agent's reasoning is the review loop. Mark prefers thinking models and stops them mid-stream when they "fix" something that was correct, rather than letting one finish and reverting to the last commit. Hanselman runs VS Code beside Copilot CLI so diffs appear as code is edited. Asked whose fault it is when AI-assisted work fails, Mark tells audiences "it's always your fault."
• AI makes low-stakes questions cheap enough to answer. During the recording Hanselman built a speaker-attribution analysis of the show, using Whisper on an RTX 4080 Super. It took nine turns, 866 lines of code and 46 minutes, and found 84,594 words against 84,169 (50.1% to 49.9%). Hanselman's point is that such a question is "not interesting enough to spend a weekend on" but is worth a few turns.

2026-05-06 · open the original ↗

• Regulatory friction that nearly killed Razorpay's launch became its deepest moat. Harshil Mathur says it took a full year after joining YC just to complete their first live transaction, since payment licensing required certifications and approvals before any transaction could occur; he argues this hurdle is actually fair because every later competitor must clear the same year-long process, which keeps the field of entrants small in a way ordinary tech products can't replicate.
• A banking-partner crisis two weeks after demo day became a lesson in trust that shaped a permanent AI policy. When Razorpay's enabling bank abruptly cut off around 50 live merchants over one complaint, Mathur says the team personally called every affected customer, absorbing abuse rather than going silent, and regained a live partner within days; he cites this as why Razorpay still refuses to let AI replace customer support, even while using it elsewhere, because he sees support as a trust channel rather than just a problem-solving function.
• Razorpay deliberately stayed capital-inefficient by choice, banking Series A funds instead of burning them. After raising roughly $11 million, Mathur says the company kept monthly burn under $200,000 and earned more in bank interest than it spent, angering an investor who expected it to burn faster; he argues B2B economics are inherently "logical" — a business pays only for value delivered today, unlike consumer models that spend to acquire and monetize later.
• A calculated early bet on India's then-unproven UPI system, made because the company had nothing to lose, created a six-month competitive window. Mathur says Razorpay went live on UPI months before India's largest banks did, so when demonetization hit in November 2016 and demand surged, it was the only payment gateway ready — winning clients like Zomato, Swiggy and BookMyShow before rivals caught up; he extends this into a broader claim that AI is now compressing capital and infrastructure moats industry-wide, leaving execution speed and the willingness to reinvent proactively, rather than wait to "respond," as the only differentiator left.

2026-05-06 · open the original ↗

• Mythos is the most visible point on a curve, not an inflection. The host reports that Anthropic announced Claude Mythos preview on April 7th without releasing it publicly. Anthropic estimates Mythos-class capability will reach other labs within 6 to 18 months, with open-weight versions to follow; the host also calls the withholding brilliant marketing. Gupta says Opus can already find vulnerabilities with more hand-holding, while Mythos can be pointed at an open-source repo and find them all.
• Machine-speed exploitation forces an assume-breach, recover-fast posture. Security has focused on prevention and detection because attackers used to sit in systems for weeks or months. Now that AI finds and exploits vulnerabilities at machine speed, Gupta argues firms must assume they are breached and plan to restore applications, data, identity and infrastructure quickly. She notes that this is Rubrik's business.
• Enterprise AI adoption widens the surface while models cause harm unaided. Al-Dhubaib says leadership-driven FOMO is rapidly expanding what Mythos-like models can chain vulnerabilities across. He adds that trust infrastructure to stop models doing harm on their own is mostly missing, so "you don't even need the bad guys".
• Outcome-driven agents erase the old security through obscurity. On the host's question about Anthropic's leak of the Claude Code codebase, Gupta says it was a misconfigured boolean flag, not an agent posting code to the wrong repo. Agents, she argues, take the fastest path to a goal without the rules employees learn. Al-Dhubaib says humans held permissions they rarely used at scale, whereas agents search everything available, which cyber controls never anticipated. Gupta adds internal risks, such as salaries leaking between employees, and says errors deep in a 20-step process must be caught while the agent runs.

2026-05-05 · open the original ↗

• Talking to an LLM is like squatting on a Smith machine. One speaker compares LLM conversation to the gym machine that holds a squat on rails in a single plane: it isolates a muscle group but skips the small stabilizing muscles that a free weight works. Talking to a human brings micro-expressions and "all the senses"; "simply yapping at an LLM" uses barely one, so each builds "a differently shaped persona."

2026-05-05 · open the original ↗

• He will never write code again without AI help. House says he feared AI would break his flow but now stays in flow more, because he rarely gets stuck for the hour or two that used to end in an angry walk. He has used GitHub Copilot, Cursor, Codex and Claude Code and mostly uses Claude Code, partly because as a workshop teacher and speaker he wants the tool most people use; he was sceptical of a CLI but now likes the large text window and @-file completion.
• Keep the instructions file nearly empty and push specialisms into skills. He no longer uses Claude Code's init, saying it generates a bloated file; everything in that file is sent with every request, yet little of it is relevant to every request, so he lets the agent grep on demand and writes lazy-loaded skills (accessibility testing, security auditing) for repeated tasks. For the same reason he clears context regularly: carried-over conversation burns tokens, lowers quality and can colour a feature with an unrelated earlier discussion.
• Humans could not review all the generated code, so AI reviewed it. On a six-week trial to build a rewards site and CMS, his team produced an app with over 90% test coverage and "not quite 100,000 lines"; heads-down teammates on Max plans hit usage limits, though he did not. GitHub Copilot reviewed the Claude-written code, a Claude skill read and answered Copilot's comments, and House read Claude's replies to decide; he says velocity expectations now force teams to dial back human review.
• Anthropic's tools are eager; OpenAI's Codex is more thorough. In his reading, Claude reads between the lines and gets code out quickly, which suits iterating as code gets cheaper, while Codex asks more questions and plans to one-shot the task, which may suit people who would rather not review. He values Copilot for not being a walled garden: it can run one code review through Google, Anthropic and OpenAI models and then compare them.
• Work in small pieces he can read, with one or two agents. He tried handing an agent a whole backlog and rejects it, because he is responsible for every pull request he merges and wants to watch the agent's direction as it works. He favours small, clear tickets in the lean limit-work-in-progress sense, and asks the AI to compare a ticket with its mock-up, which often turns up disagreements to settle with the product owner.

2026-05-05 · open the original ↗

• Machine-speed exploitation means assuming breach and planning recovery. The host reports that Anthropic announced Claude Mythos preview on April 7th but withheld it, and estimates Mythos-class capability will reach other labs within 6 to 18 months, with open-weight versions to follow. Gupta, Rubrik's chief product officer, calls Mythos the most visible point on a curve: Opus finds vulnerabilities with more hand-holding, while Mythos can find all of them in a repo. Attackers once sat in systems for weeks or months, she says, but AI exploits at machine speed, so prevention and detection must give way to fast restoration.
• Agents cause damage without any attacker. Al-Dhubaib, three weeks into a principal technologist role at Rubrik, says incidents have hit a "hockey stick". He says Amazon lost 99% of its orders in a single day to agent-generated code, and a small business's booking database was wiped. Agents optimize for outcomes without employees' training, and they exhaustively use permissions humans held but rarely exercised, ending security through obscurity. Gupta says the Claude Code source leak was a misconfigured boolean flag.
• Non-human identities are the problem to solve. Gupta traces a typical attack from compromised credentials, as in the MGM support-desk password and MFA reset she recalls, through escalation to ransom. Rubrik's preemptive recovery engine is meant to establish the blast radius before restoring. Al-Dhubaib says every agent acting for a user multiplies keys and tokens. One guest describes Claude Code users storing long-lived, full-permission SaaS tokens in a laptop file, which no enterprise application would allow. Gupta defines zero trust as assuming every identity is breached and keeping crown-jewel access minimal and ephemeral.
• Resilience must be designed in up front, not bolted on. Al-Dhubaib's trust engineering framework pairs human-centred design, such as the Waymo signal telling pedestrians it sees them, with trust infrastructure to anticipate risks, detect misbehaviour and fail safe. He says clients' AI governance most often lacked a tested recovery plan. Gupta likens bolting on resilience to the costly retrofit many firms made after moving to public cloud.
• Rubrik pairs a small-model policy engine with an undo for agents. Gupta says Rubrik Agent Cloud gives visibility across Copilot, Bedrock and Claude Code agents, citing over 80% of organizations who feel they lack it. Sage, a fine-tuned SLM from the Predibase acquisition, enforces natural-language policies at runtime; the host cites Rubrik's result of five times GPT 5.2's speed with higher accuracy. Agent Rewind uses the recovery engine to undo dropped tables or bad Salesforce and identity changes, though not delivered emails, which Sage must block up front. Al-Dhubaib adds that benchmarks can be gamed and calls for safety benchmarks on over-permissioning and out-of-scope search.

2026-05-04 · open the original ↗

• Current AI chips are poorly suited to how agents actually run. YC says this causes GPUs to hit only 30–40% peak utilization on agentic workloads, because agents loop, branch, and backtrack across dozens of steps, bouncing between memory-bound model calls, IO-bound tool use, and CPU-bound orchestration — a fundamentally different hardware problem than simple prompt-in, response-out inference.
• Purpose-built silicon for the agent loop itself is still missing. YC points to Nvidia's reported $20 billion acquisition of Groq and Google's inference-specific TPU v7 as evidence the shift toward specialized hardware is underway, but argues the real advantage is the compiler, not the chip — fast context switching between models, native speculative decoding, and memory built for KV caches that persist across an entire execution graph.

2026-05-04 · open the original ↗

• Three to five agents is a comfortable limit for one person. One speaker says that is what the speaker and Mark Russinovich ("my buddy at work") settled on. Colleagues run 10 or 20 at once in "squads of squads", but the speaker finds the context switching and cognitive load of that many terminal tabs hits a wall quickly.
• Agents should prepare work, then stop until you return. The other speaker says tabs are too short-lived for tracking agent work, and that multitasking around the agents' "hurry up and wait" has to be careful: have them do prep and research, then stop, so the task is ready when you have time for it.

2026-05-04 · open the original ↗

• Project Nomad can run on Windows 10 or 11 under WSL2. The host says this route is community-supported rather than official, but he has tried it himself and it works; his write-up on projectnomad.us has separate instructions for Docker Native and Docker Desktop, aimed at users without Linux skills or a spare server.

2026-05-04 · open the original ↗

• Scientific discovery is shifting from AI copilots to closed discovery loops. YC argues this is already happening in drug discovery, material science, and protein engineering, where models propose candidate molecules, automated labs synthesize and test them, and results feed back to iteratively improve the next candidate; it predicts the most impactful companies won't just sell research copilots but will be full AI-native discovery engines.

2026-05-04 · open the original ↗

• The AI compute stack must be governed neutrally, not owned by one company. Haviv points to vLLM, Ray and Kubernetes as the AI compute stack and notes Ray recently joined the Linux Foundation, as MCP is doing. She calls it crucial that the stack have neutral governance and community processes, and argues that open source's culture of building on each other's work will win in the long run.
• Open models lower the barrier to building, fine-tuning and running AI locally. Haviv expects open source to catch up with proprietary models. She cites people building home servers on open models such as the newly released Gemma 4 and DeepSeek. She runs Gemma 4 on her phone for flights with bad Wi-Fi and a Llama model on her laptop. She urges people to contribute to open source to find community, innovate and learn from the code.
• Build first to solve your own problems. Haviv says she tried OpenClaw as a home assistant and newsletter automation but now does most of it in Claude, building many skills and plugins. Some of her skills overlapped, which she counts among the mistakes that come with learning by building.

2026-05-04 · open the original ↗

• The best AI-native companies make their whole operation queryable by AI. YC argues this turns decision-making into a closed loop — the system monitors what's happening, compares it to what should be happening, and adjusts — citing teams that cut sprint time in half and shipped 10x as much after recording every meeting and tracking every ticket and customer interaction.
• Building this today requires brutal custom integration work across tools. YC points to Slack, Linear, GitHub, Notion, and call recordings as the kind of systems currently stitched together with custom glue code, with no single product connecting it all into one AI layer; it wants to back the connective layer that makes a company legible to AI by default, rather than another one-off patchwork.

2026-05-03 · open the original ↗

• AI has collapsed the cost of producing software by 100x or more. YC argues this erases the code-volume moat that protected legacy SaaS incumbents, framing the recent hit to software market caps as bad for incumbents but good for startups, since decades-old codebases once considered untouchable are now vulnerable to cloning, bundling, or open-source replacement.
• YC is pushing founders toward bigger targets than typical product-management tools. It names chip design software, ERPs, industrial control systems, and supply chain management as the next frontier, framing it as a repeat of the shift from on-premise to cloud: the next generation of great software companies will replace legacy SaaS with AI-native software.

2026-05-03 · open the original ↗

• AI infrastructure is infrastructure shaped by AI workloads' distinct demands. Haviv's working definition is any infrastructure specific to AI/ML workloads, which overlaps with traditional infrastructure. Those workloads are far more compute-heavy than older input/output-bound ones, multimodal and less deterministic. She says inference is usually where users first hit the difference, whether batching for a finance firm or serving many users.
• Neo clouds and frameworks like Ray and vLLM exist to fill that gap. Haviv says AI workloads need direct bare-metal access more than virtualization, hence neo clouds such as Lightning AI, Nebius, CoreWeave and Crusoe. She calls them specialists, against hyperscalers like AWS, GCP and Microsoft that "do everything". Ray, being Python-native and built to keep GPUs from sitting idle, distributes these workloads where Spark served earlier ones.
• Batch inference needs CPU preprocessing sized so GPUs neither idle nor starve. In Haviv's popcorn analogy, popping one buttered kernel at a time is latency-optimized single inference. Batch work, such as a finance firm's huge dataset, is a full microwave bag, but the microwave is always on and costing money. So the CPU-side prep of data has to keep pace with the GPUs.

2026-05-03 · open the original ↗

• Sovereign cloud is a range of offerings, not one product. Fianbakken, a Microsoft program manager for disconnected operations, says geopolitical instability has made sovereignty top of mind for customers, and Microsoft now offers public sovereign cloud, private sovereign cloud and national partner clouds to meet different requirements.
• Public sovereign cloud is mostly policy applied to ordinary Azure. He describes it as Azure policies and governance frameworks, such as bringing your own encryption keys and "sovereign landing zones" that extend Azure landing zones; the simplest example is a data-residency policy assigned to the management group hierarchy so every service keeps data in the customer's geography.
• Private sovereign cloud puts Azure Local under the customer's control. It runs on the customer's or a trusted partner's premises, either connected to an Azure region or fully disconnected with no outbound connectivity, with only a subset of public-cloud features. He argues it serves the "last mile", such as health services that must keep running if internet access is cut for a long period, and cites regional data-centre outages from recent Middle East conflicts; he recommends designing critical workloads to fall back from public cloud to connected Azure Local to fully air-gapped.
• Going fully disconnected adds burdens customers rarely anticipate. Customers must buy and plan the hardware and failover capacity across sites, and run their own PKI and time servers because they cannot depend on public services. To decide what goes where, he says to map applications against risk appetite and regulation, asking for example whether the country could be cut off from a hyperscaler for 12 months.
• National partner clouds are run by entities independent of Microsoft. Besides the US government cloud he calls Fairfax, he names Bleu in France and Delos in Germany, serving approved government customers with a subset of features; he says independence from Microsoft is a European requirement and that they are fully isolated from Microsoft's data centres.

2026-05-03 · open the original ↗

• Star Cloud's Adi Oltean wants YC to fund lunar industrial capabilities. He describes starting with extracting raw materials like silicon, aluminum, iron, and titanium from lunar regolith via electrolysis, then 3D printing complex structures from the molten material — which he says could be more efficient than on Earth due to the lack of structural supports needed.

2026-05-02 · open the original ↗

• Every major software category needs to be rebuilt for AI agents. YC argues agents need machine-readable interfaces like APIs, MCPs, and CLIs, plus thorough documentation to discover and use tools without a human in the loop, replacing the visual forms and dashboards built for people — and predicts this rebuild will come from agent-first startups rather than incumbents bolting on agent support.

2026-05-02 · open the original ↗

• Bots are taking over the internet, so humans need proof of personhood. The guests, who say they are helping World (world.org) gain exposure in the Midwest, argue that more than half the internet is already browsed by AI agents and traffic is heading toward about 99% bots. They present World ID as a "priority lane" for humans and a way to tie an agent acting for someone to that person uniquely.
• An iris scan issues a credential without World holding the data. The "orb" camera scans the irises, which they say have less than a one-in-a-billion chance of matching anyone else's, converts the image to a hash that is split across databases so it cannot be traced back to a person, and issues a World ID stored on the phone. By their account it is a one-time scan, likened to a biometric RSA key, and each application gets its own account identifier, so apps cannot link a user across services.
• Credentials let apps check claims without seeing the underlying data. A passport can be added, with driver's licences "eventually"; they say World ID is integrated with Tinder in Japan, where a zero-knowledge proof confirms a user is over 21 without revealing the birthday.
• Developers can build on World's existing base of verified users. They cite close to 39 million people onboarded in over 160 countries and a World ID SDK usable in standalone apps or in mini apps published inside the World app. Their own aim is to educate and help local developers, including with funding, and they invite people to partner through sohuman.us to operate orbs locally.
• A payments network sits on top of the verified identities. They say users can send money through World chat to anyone with the app, including in Mexico or Japan, as a digital US dollar the recipient converts to local currency, pitched against scams on services like Cash App or Venmo.

2026-05-02 · open the original ↗

• Ray lets ML engineers distribute work without becoming distributed-systems experts. Haviv, until recently a staff developer advocate at Anyscale, describes Ray as a Python-native, open-source distributed computing framework suited to AI workloads with multimodal data such as video, audio and text. She says it was created in Ion Stoica's Berkeley lab by researchers including Anyscale co-founder Robert Nishihara, who were spending more time on distributed systems than on reinforcement learning research.
• Ray's RL origins fit again now that reinforcement learning is back. Haviv notes RL had faded from discussion but returned through RLHF and its variants in large models such as DeepSeek. Ray has since grown beyond RL into Ray Data, Ray Train and Ray Serve, all in Python, with native integrations for vLLM and Hugging Face. It serves cases from serving an MCP server to distributed training across GPUs and TPUs when workloads don't fit on one GPU.

2026-05-01 · open the original ↗

• Crosstalk Mobile SIMs work in the UniFi 5G Max, the host reports. Pitching his company's own SIMs, he says the 5G Max plugs into a LAN port on PoE and, once adopted, is set up automatically as an extra WAN. Sitting in his window it gets 475 Mbps, and he says speeds up to about 800 Mbps have been seen elsewhere.
• The UniFi 5G Max will not accept every SIM. The host says a T-Mobile Home Internet SIM moved from its modem "absolutely will not" work, and nothing from Verizon works yet, though he says Verizon support is on the roadmap.

2026-05-01 · open the original ↗

• AI-augmented programming produces personal software that needs a home. Hanselman, who says they have changed role to work more with the Windows teams on making Windows a great place for developers, describes building "tiny tools" for an audience of one, such as a click-to-show-desktop tool, maximizing a window to a virtual desktop, and blood sugar management. Tiny Tool Town (tinytooltown.com) is where such tools can live, and Hanselman reports 325 submissions so far.
• The site runs entirely on GitHub, with no database. It is an Astro site in a GitHub repository. Tools are submitted through a GitHub issue form or a CLI.
• Free, licensed, open-source tools may get a Build shout-out. The rules are a license, a public GitHub repo and nothing commercial or promotional. Hanselman will feature favourites on the pre-Build show for Microsoft Build, June 2–3 in San Francisco, and asks for submissions within "the next couple of weeks". Hanselman and Mark Russinovich will do "a bit of a vibe check" there, and Hanselman says AI-augmented software engineering is more interesting than vibe coding straight into production.

2026-05-01 · open the original ↗

• Imperfect consumer sensors pull people away from their own signals. The speaker relays that their cardiologist hates the Apple Watch because its ECG is one lead, not several. For every person it saves by catching AFib, the cardiologist says, a thousand come in worried about a resting heart rate or an Oura ring reading, and the advice is "just listen to your body."
• The feedback loop should run through self-reflection, not validation. The speaker worries about outsourcing judgment to measurement systems and wants to think, read the signals available, and decide without checking with Apple Health or ChatGPT. The example is a Post-it note reading "Talk to more people."

2026-05-01 · open the original ↗

• Imperfect consumer sensors pull people away from their own signals. The speaker relays that their cardiologist hates the Apple Watch because its ECG is one lead, not several. For every person it saves by catching AFib, the cardiologist says, a thousand come in worried about a resting heart rate or an Oura ring reading, and the advice is "just listen to your body."
• The feedback loop should run through self-reflection, not validation. The speaker worries about outsourcing judgment to measurement systems and wants to think, read the signals available, and decide without checking with Apple Health or ChatGPT. The example is a Post-it note reading "Talk to more people."

2026-05-01 · open the original ↗

• US hardware moves far slower than China's because of iteration speed. YC contrasts Shenzhen, where a team can go from design to a physical part in a day, with the US, where the same loop takes weeks, attributing China's edge to dense supplier networks and tight coordination between design and production.
• YC wants to fund startups that make US hardware iteration dramatically faster. It points to a startup already building faster turnaround from design to manufactured mechanical part as an early example of the kind of missing infrastructure it's looking for.

2026-05-01 · open the original ↗

• Transformers trade recursive reasoning for parallel training speed. Chopard argues that unlike RNNs, which iterate on a compressed hidden state across time steps, an LLM computes in one feedforward pass with no compression in the time direction — every decode re-processes the full context. RNNs paid for their recursion with backprop-through-time, which caused vanishing and exploding gradients as models or sequences grew.
• Some problems are provably incompressible, unsolvable in one forward pass. He cites comparison sorting's known n·log(n)-step lower bound, plus Sudoku, mazes and rolling sums, arguing a transformer with fewer layers than the required steps cannot solve them. Chain-of-thought and tool-calling can simulate a Turing machine at test time, but only within human-demonstrated traces — an LLM taught only bubble sort, he says, could not discover merge sort on its own.
• A tiny recursive model outperformed giant pretrained LLMs on ARC-AGI. He credits the Hierarchical Reasoning Model — a 27-million-parameter model with no pretraining, trained on roughly 1,000 ARC-AGI tasks — with scoring about 70% on ARC-1 versus OpenAI's o3 scoring zero at the time. The follow-up Tiny Recursive Model, by a researcher he calls Alexia, shrank it to 7 million parameters by collapsing HRM's two networks into one, and pushed the score to 87%.
• The outer refinement loop, not the two-network design, drives the gain. Chopard says a researcher at Francois Chollet's company ran scaling ablations showing this repeated outer loop is the main source of HRM and TRM's performance, which is why TRM kept and simplified around it while discarding most of the rest. He expects the next leap to come from combining this recursion with giant pretrained LLMs rather than treating either as a substitute for the other, since today's recursive models are trained per task, not general-purpose.

2026-05-01 · open the original ↗

• Agent memory maps onto four kinds of animal memory. Richmond Alake, Oracle's director of AI developer experience (episode 985), splits agent memory into four types. Episodic memory is time-stamped, like conversation turns with role, content and timestamp. Semantic memory is world and institutional knowledge. Procedural memory is routines such as skills.md files, which he calls "SOPs for agents". Working memory is the LLM's context window. He and the host, who studied neuroscience, ground this in biology's role in inspiring CNNs through Hubel and Wiesel's experiments on the cat visual cortex.
• Storing skills in a database lets agents retrieve only what they need. Alake says Oracle is experimenting with putting skills in database tables, including vector representations, rather than files. Skills can then be exposed progressively instead of loading 100 skills into context, which he says allows scaling.
• Data engineering agents will curb junior hiring, not cut existing staff. Matt Glickman, CEO of Genesis Computing (episode 981), says no one is losing their job: existing data engineers were the limiting factor and now get "10x" more done. Junior hiring is what AI will "wipe out", and it is already starting. A few hires will be "conductors" running many agents in parallel; Glickman says a co-founder runs six while Glickman manages two and a half. Most schools still discourage AI use, which Glickman calls a "double whammy" for graduates.
• Coding is being abstracted, but AI infrastructure and systems thinking are surfacing. Linda Haviv (episode 987) notes non-technical people now talk about memory and context lag, which are infrastructure problems in other words. Engineers' edge, she argues, is knowing where systems fail, since even model training is an orchestration problem. People who know a domain should bring that context, as she did moving from media to building a news CMS.
• AI raised the floor in one school; teachers are raising the ceiling. Tracy Walker Griffith, principal of a K-8 innovation school in Boston (episode 983), describes fifth graders preparing for the state test. They write an answer, have AI in Magic School write one, explain which is better, and log wrong multiple-choice answers in a tracker the teacher checks for patterns. She says the school stays "platform-informed, not platform-loyal", also using PlayLab, which runs on Claude. AI literacy for both teachers and students comes first.

2026-05-01 · open the original ↗

• Surrounding yourself with agreement stops you from growing. The speaker argues that you will not find new solutions, and will fall into in-group/out-group bias, unless you deliberately seek out unfamiliar thinkers. Talking to strangers, such as people on the bus, disrupts your thinking patterns and gives you new information and frames, which is how decision-making changes.
• An LLM companion locks in one trained frame. The speaker suggests an LLM may lean toward a particular, perhaps neutral, way of deciding or seeing people. Relying on one while growing up, or as an adult, replaces "the randomness of a community" with that single programmed paradigm.

2026-04-30 · open the original ↗

• AI has erased the barriers to selling to the world's largest companies. YC argues enterprise leaders now actively seek teams that can apply AI to their problems, coding agents let a two-to-three person team ship nuanced products to large companies in months rather than years, and those leaders now understand what to build in-house versus outsource — making it common for a YC startup's first customer to be one of the largest companies in the world. This, YC says, kills the old model of startups operating in stealth for years to reach feature parity with an incumbent before launching.

2026-04-30 · open the original ↗

• Security is a subset of resilience, not its own discipline. Shortridge, chief product officer at Fastly and author of Security Chaos Engineering, calls "chaos" in the title a buzzword publishers needed. Resilience, she says, is recovering from failure of any kind and preparing for what comes next, whether that is a DDoS or traffic growth for good reasons.
• A security team's sense of control signals low resilience. Because systems and people are not deterministic, she argues that teams claiming full control of the delivery lifecycle are funding what feels good rather than what limits impact. She recommends small, low-impact experiments instead, such as a Fastly Compute function that strips cookies or auth headers from duplicated requests to test whether a login page really needs them, designed jointly with platform engineering.
• Compliance checklists and familiar metrics can actively erode security. She cites a paper by Josiah Dykstra (on HIPAA, she believes) finding that more compliance does not produce better outcomes, and a CISO who spent years telling auditors that removing SSH access had improved security. A rising vulnerability count can mean more trust between teams. In another case she describes, a security team chased 1,000 unexposed vulnerabilities while keeping its own admin credentials in a text file on a desktop.
• LLMs fit wherever they don't replace human judgment. That is her litmus test. LLMs suit tedious work like pulling facts from compliance documents, rubber-duck debugging, or testing how a resilience pitch lands across cultures. Her example: deception techniques draw "we don't want to be the bad guys" from US practitioners and "tell me more" in the EU. Asking an LLM whether your system is resilient will not go well, she says, and prose playbooks open security to more people but are a potential foot-gun.
• You own your dependencies; the vendor owns its platform. For Fastly, her own company, she says memory safety by design and multi-tenant isolation are Fastly's responsibility. Vulnerabilities in a customer's own code, or a database left open without an access control list, are not. Outsourcing specialist work such as AI bot management makes sense when it isn't your core business.

2026-04-30 · open the original ↗

• Agent tooling should widen the bandwidth between user and agents. Hinkle, CTO and co-founder of Nimbalyst, says multitasking works only with preparation: agents research and then stop, so each context switch leads into 20–30 minutes of focused review. Nimbalyst puts mockups, Excalidraw architecture, Prisma data models and annotation (circling part of a screen, clicking a table in an ERD) in one pane, so sessions don't lose context.
• Visual artifacts serve the human, not the model. He concedes the agent gains nothing from an ERD over the raw Prisma file. He is willing to spend tokens on the agent communicating better with him so that he can steer it. Files and diffs stay in front of the user, in local-first open formats on disk, which he presents as a deliberate design choice.
• One model reviewing another beats persona-based agent squads. He has Claude Code plan, Codex review the plan, Claude Code implement and Codex review the result, which usually catches errors. He uses Nimbalyst's built-in Claude "teammates" less than he expected. Their value, he says, is that each focuses his own conversation with an agent, not that the agent does better work. Hanselman shares the scepticism about "you're an expert database person" prompts and grants them only savings in the context window.
• Plans should document intent, not instruct the agent. Hanselman puts his own approach to him: "code sculpting", against a long plan.md as a "prose compiler". Hinkle says heavy plans were needed a year ago, but models have advanced. He still writes a plan for nearly everything, but it records the motivation and the points that can't be missed, so that later sessions can be checked against it.
• Agents can verify that something works but supply no taste. Nimbalyst is built in Nimbalyst. Its agent reads live logs, settings and the database, and it navigates the UI, takes screenshots and reads the DOM, looping toward a goal. Hinkle says it will make things work, and the human adds taste. He gives the status as free for individuals, an Electron app on Windows, Mac and Linux, with an iOS app first and Android coming, connected by an end-to-end encrypted tunnel through Nimbalyst's servers. His voice agent that controls coding agents is still in testing.

2026-04-30 · open the original ↗

• Semiconductor supply chains are still run on spreadsheets and calls. YC notes a single advanced AI chip takes 1,400 manufacturing steps across a dozen countries over five months, and cites 2021, when a $300 chip held up a $50,000 car and left $200 billion in vehicles unbuilt, as proof that companies see their direct suppliers but have no visibility into second- and third-tier ones.
• The bottleneck has only gotten worse. YC says Nvidia has locked up over 60% of TSMC's advanced packaging capacity, high-bandwidth memory is booked through 2026, export controls change quarterly, and the CHIPS Act is standing up new fabs in Arizona, Texas, Ohio, and New York that each need a supply chain built from scratch.
• This tooling doesn't exist yet, and building it needs deep technical knowledge. YC says real-time allocation tracking, multi-tier risk monitoring, and export compliance all require understanding wafer allocation and packaging constraints at a deep level — which is why it sees this as a startup opportunity rather than a feature bolted onto SAP.

2026-04-30 · open the original ↗

• Work on the side is insurance against being a number. Haviv argues that however good the people are, to the company as a whole you are "a human resource", so she never kept all her eggs in one basket; she chose software partly because it is a skill no employer can take away.
• Her personal brand fed her full-time job. Haviv says her job at Amazon came about because the company found her talking about cloud computing on TikTok, and that the side work made her better in the day job rather than distracting from it.
• Independent work lets her be more intentional than inside a company. She says committing to a single project or campaign without internal red tape is more efficient for her, and that she can keep advocating for tools she likes (Ray, AWS) from outside; the host lists her independent brand work as including Anthropic, Dell, Amazon, GitHub, Microsoft and Nvidia.
• Work is getting leaner, toward tiny billion-dollar companies. Haviv likens the present to the arrival of the internet and guesses that maybe 50% of jobs will be at small companies, with two- to three-person and perhaps one-person billion-dollar companies; she says plainly this is her own impression, "not some scientifically backed metric".

2026-04-30 · open the original ↗

• Diversity of thought means seeking out unfamiliar people on purpose. Westerhoff, a principal AI security researcher on Microsoft's AI red team and Hanselman's colleague, keeps a daily noon reminder to talk to strangers. "If you sycophant yourself, you're never going to grow," she says. Random encounters disrupt patterns of thinking and update frames in a way that one LLM's trained decision-making paradigm cannot.
• LLMs are becoming people's introspection tool, with their context loaded. She reports a growing number of published studies showing people use AI for introspection. It is not the worst tool for that, she says, but the assessment may now come from the LLM rather than from your own mental back-and-forth. Hanselman's framing is that it is a mirror that also carries a corporation's weights and biases.
• Social skill is a muscle, and text-only practice misses the cues. She lost the muscle during COVID and got it back because she had trained methods to return to. She takes up Hanselman's analogy of the Smith machine against free weights and its stabiliser muscles. Microexpressions and other cues, picked up across many settings, are how people build dynamic skills, and rehearsing with an LLM does not supply them.
• LLMs push users to select from options instead of assessing. She bases this on decision-making research. Offering a menu suits decision fatigue, but skipping the assessment loses how your values interact with the choice, which is where the insight and better outcomes come from. An LLM widening the options (beyond Indian or Noodles) helps, but it still does not get you the answer.
• Intentional use is the line, and its trade-offs are poorly signposted. Her team studies the psychosocial impacts of AI. She grounds herself in knowing that an LLM "best friend" engages the same neurological systems as a human one, and she expects tools that build in that awareness to become the frontier for AI in relationships, coaching and education. Hanselman draws the line between deliberate companionship, such as robots for lonely older people, and drifting into a relationship like the one in the film Her.

2026-04-29 · open the original ↗

• Coding agents now let ordinary users become their own "forward-deployed engineers." YC notes that before AI, nearly all software shipped one interface for everyone — Netflix looks the same for every user, just with different imagery — with deep per-customer customization limited to enterprise software built by human forward-deployed engineers, and argues users can now radically customize interfaces themselves.
• This implies rethinking how software gets delivered. YC raises open questions like whether companies will need to ship source code instead of packaged binaries, and whether users' agents should be able to modify only front-end elements or also middleware on the fly.

2026-04-29 · open the original ↗

• A small mini PC runs Project Nomad as well as a big workstation. On Nomad's own benchmark, the host's Minisforum UM890 Pro (Ryzen 9 8945HS, Radeon 780M, 64 GB RAM, 1 TB NVMe) scored 75 against 73 for his roughly 6-year-old Threadripper with 128 GB RAM and an RTX 3090 Ti. Project Nomad is Crosstalk's own offline server.
• The big GPU still wins for local AI. The host concedes the 3090 Ti is better hardware for LLM inference and chat, but says the mini PC is all you need for the knowledge-base content (Wikipedia, Khan Academy, offline maps), so a large, expensive server is unnecessary.

2026-04-29 · open the original ↗

• YC is launching a Paper Club to keep builders and researchers current on frontier AI. Framed around how hard it is even for a PhD student to keep up, the group will meet every few weeks at YC's Mountain View office to present and discuss papers, followed by an off-the-record dinner; the first cohort caps under 100 people and meets May 20.

2026-04-29 · open the original ↗

• Hassabis expects AGI soon, but flags one or two open unknowns. His personal timeline is around 2030; he says continual learning, long-term reasoning and some aspects of memory remain unsolved, and puts it at "50/50" whether closing those gaps needs one or two genuinely new ideas or just incremental scaling of what already works — DeepMind is pursuing both bets at once.
• Today's reasoning models are only "fairly brute force," producing jagged intelligence. Watching Gemini play chess, Hassabis says its own thinking trace will identify a move as a blunder and then play it anyway — evidence, he argues, that models lack introspection into their own reasoning, which is why the same system can solve IMO gold-medal problems yet still make elementary math errors.
• Agents haven't truly arrived until a vibe-coded game tops the charts. Hassabis says no AI-built game has hit the App Store charts the way a studio-made hit would, and expects that to change in the next 6 to 12 months; he predicts individual builders working at roughly 1000x their prior output will appear before agents operate fully autonomously.
• The AlphaFold pattern: combinatorial search, a clear objective, plenty of data. Hassabis says that formula — a massive combinatorial search space, a clear objective function to hill-climb, and enough data or simulation to generate synthetic training data — drove both AlphaGo and AlphaFold, and advises founders to pair deep AI expertise with a hard science domain like materials or medicine as one of the few bets defensible against the next foundation-model update.

2026-04-29 · open the original ↗

• A team of specialist agents beat generalists and a single agent. Gaster, who works at Microsoft on GitHub Copilot app modernization, says he gave the same "build a snake game using Aspire" prompt to one agent, a team of five generalists and a team of specialists each fed its own large Markdown knowledge file (on App Service, Container Apps, Aspire, Functions). The single agent produced about 2,000 lines and 80 tests slowly and badly; the specialists wrote about 20% less, cleaner code with fewer tests that covered more, and ran 18–20% more efficiently than the generalist team on average over repeated test runs.
• Generalists fight over work and hallucinate; specialists stay in lane. Gaster says generalists have a "hungry hungry hippo" effect, grabbing any task and hallucinating because they carry too much knowledge, while a specialist tester will not write code. A coordinator is always required, and the team can "hire" new agents when it finds a gap: on a Swift app it asked for a front-end designer, then a marketing agent.
• Turn on the lockout policy; Gaster says it stops hallucinations. It is off by default and set with an instruction such as "set it to two": after two failed attempts at a task, as judged by a tester agent (his example is Playwright checking a colour), that agent is barred from it, others try, and if none can do it the team hires one. When the task passes, every agent writes what it learned to disk so the error is not repeated.
• Squad's memory is just Markdown files, written constantly. A "Scribe" agent in every squad writes decisions to files, so after a Copilot CLI crash (which Gaster says hits him once or twice a day) the squad can say where it was and resume; Giard adds that the files are checked into source control, so later developers see the decisions. Agents refuse to move a secret pasted into chat because it would land in a file; a storage abstraction with a SQLite vector store, built by Dina Berry after Jeff Fritz asked for one, was committed the night before recording, and longer-term-memory experiments are planned.
• Refresh the team after about a month of work. Squad starts with no more than about five members; Gaster tells a long-running squad to clear its context, report how much it cleared, then hold a candid meeting on its performance and restructuring. His overloaded Git agent "Kobayashi" was let go and the team asked for two replacements, one for CI/CD without Git ops and one for Git ops without CI/CD.

2026-04-29 · open the original ↗

• Non-technical builders now hit AI-infrastructure problems in other words. Haviv says that once people move past toy examples to production and scale, they run into systems issues; non-technical people now talk about memory, context that "is not working" and lag, which are the same problems as KV cache and prefill-decode disaggregation without the vocabulary.
• Engineers' edge is knowing where things fail. She argues that agentic workflows sit on a distributed-systems layer, and that even model training and mixture-of-experts models are orchestration problems, so someone from DevOps or software brings real strength even without AI/ML expertise. The host's premise is that within "a couple months" it has become theoretically possible to do the whole AI pipeline in natural-language prompts.
• Domain knowledge transfers, so start where you already understand users. Haviv, who taught herself to code in 2015 and took her first JavaScript job in 2016, says her media background let her build a content management system for journalists because she had written the articles herself; the host draws the lesson that a newcomer may do better moving into development at their current employer than applying to Meta.
• Building and media are democratised, so curate what you build. She says creators no longer need a $50,000 studio or a label, and that she saw in TV news that specialists such as lawyers and doctors who came on as experts kept control of their time. The host adds that hands-on experience in a field gives more confidence in AI advice and warmer contacts than a cold email; Haviv notes junior developers struggling to find jobs out of college are also turning entrepreneurial.

2026-04-28 · open the original ↗

• Space-optimized inference chips are an emerging startup opportunity. Philip Johnston, co-founder and CEO of Starcloud (building data centers in space), argues reusable rockets from SpaceX and Stoke Space will drive a huge increase in orbital compute capacity, creating demand for chips tuned for mass, thermal and radiation constraints; he's looking to hear from chip designers with experience at SpaceX or Nvidia.

2026-04-28 · open the original ↗

• A drone swarm reportedly destroyed an AWS data center unopposed. He says a swarm of cheap Iranian drones took out the facility last month with nobody stopping it, and warns that a coordinated swarm of a thousand drones — far larger than the single hobby drones most counter-drone systems are built to handle — would be even harder to stop.
• The cost math today overwhelmingly favors attackers, not defenders. He points to a $3 million Patriot missile intercepting a $500 FPV drone as the current exchange rate, and argues restoring that cost advantage to the defense side is the real problem to solve.
• YC wants to fund the counter-swarm stack as software, not weapons. He lists interceptors that neutralize dozens of drones at once, sensor-fusion software giving a single real-time picture, and non-kinetic tools like rotor-fouling aerosols and entangling streamers — arguing that with radio jamming becoming obsolete, the winning companies will resemble Cloudflare more than Raytheon.

2026-04-28 · open the original ↗

• AI automation of companies is now blocked by scattered internal knowledge, not model quality. The speaker argues every business needs a "company brain" — a system that extracts tacit know-how from emails, Slack, support tickets and databases into a living, current "skills file" so AI agents can execute tasks like refunds or pricing exceptions safely and consistently.

2026-04-28 · open the original ↗

• Leaner work makes this the time to go independent. Haviv, who had just left her staff developer advocate role at Anyscale at the time of recording, guesses that maybe 50% of jobs will be at small companies and that two- to three-person, perhaps one-person, billion-dollar companies are coming; she flags this as her own impression, not a measured figure. She treats side work as insurance against being "a human resource", noting her Amazon job came from the company finding her cloud-computing videos on TikTok.
• AI-infrastructure problems are now everyone's problems. Haviv says non-technical builders now complain about memory, context and lag, which are infrastructure problems in different words, and that past toy examples the work becomes systems thinking. She argues engineers keep an edge because they know where things fail, since even model training and mixture-of-experts are orchestration problems; the host frames the premise that the whole AI pipeline has become doable in natural language within "a couple months".
• A Python-native framework spares ML engineers distributed-systems expertise. Haviv describes Ray as a Python-native open-source distributed computing framework, created at Ion Stoica's Berkeley lab by Robert Nishihara and others for reinforcement learning, now spanning Ray Data, Ray Train and Ray Serve with native vLLM and Hugging Face integrations. She says it suits multimodal, GPU-bound AI workloads better than Spark because those workloads do not fit on one GPU.
• AI infrastructure is infrastructure shaped by compute-heavy AI workloads. She defines it as infrastructure specific to AI/ML workloads, which are compute-heavy where older workloads leaned on data input and output, and says inference is where users first hit the difference. Her popcorn analogy for batch inference: popping one kernel is latency-optimised, a microwave full is throughput-optimised but costs money while on, so CPU preprocessing must keep the GPUs fed. She casts neoclouds (Nebius, CoreWeave, Crusoe) as specialists beside hyperscalers that "do everything".
• Open governance of the AI stack matters more than open weights. Haviv says it is crucial that Ray, vLLM, Kubernetes and MCP sit under the Linux Foundation rather than being owned by one company, and expects open models such as the newly released Gemma 4 and DeepSeek to catch up with proprietary ones. She runs Gemma 4 on her phone for offline use on flights.

2026-04-27 · open the original ↗

• Falling diagnostic and gene-therapy costs, paired with agent harnesses, will enable personalized medicine at scale. The speaker says genome-sequencing costs have fallen faster than Moore's Law and that N-of-1 mRNA-based therapies plus greater FDA openness are emerging, arguing agents like Claude Code can combine genomic, EHR and wearable data into individualized treatment suggestions.

2026-04-27 · open the original ↗

• Episode 900 is a montage of guests' closing lines, not an interview. It strings together dozens of past guests riffing on the show's "technology and friends" sign-off, with no host questions and no single topic; the only speaker who names himself is Jeff Fritz.
• The common thread is that tech careers are built on friendships. Several guests credit conferences, speaking and user groups with friendships worldwide, one dating it to publishing and speaking from the late '90s and saying their company was founded on working with friends, and an entrepreneur advises starting a company with friends who share your vision and principles.
• A minority of clips warn about AI and technology's risks. These include "Friends don't let friends use GenAI technology without some safeguards", "Friends don't let friends build irresponsible technology", that technology can isolate as well as connect, and one tongue-in-cheek claim that AI has eaten the speaker's friends. Others argue developers need to understand AI to set guardrails.
• One guest argues the GUI made computing antisocial. The speaker says clicking gave nobody anything to show colleagues, while PowerShell scripts got shared, admired and repaid in gratitude ("I owe Bruce a beer"), and that this restored the social joy of computing.

2026-04-27 · open the original ↗

• AI-native companies that sell the finished service, not a software tool, are the next wave. The speaker contrasts 2023–2025 "AI co-pilot" startups with a new category that performs the work itself — citing insurance brokerage, accounting, tax, audit and healthcare administration — arguing total services spend dwarfs software spend and is easier to disrupt because much of it is already outsourced.

2026-04-27 · open the original ↗

• Farmers are trapped in a worsening chemical treadmill. The narrator argues pesticide resistance keeps evolving, the pipeline for new chemicals has gotten slower and more expensive, and farmers must keep spraying more for diminishing results even as costs rise and margins fall, because failing to spray risks crop loss and food-price shocks.
• Three technologies converging at once make the problem newly solvable. He points to cheap sensors and cameras giving AI real-time weed/pest identification, robotics precise enough to treat individual plants instead of whole fields, and biological alternatives (microbes, peptides, RNA-based tools, engineered self-defending plants) capable of replacing whole classes of synthetic chemicals.
• This is framed as a generational investment opportunity for YC. The narrator predicts adoption will be "explosive" rather than slow because a company that cuts pesticide use by 90% while increasing yields would be attractive in one of the world's largest markets, and closes with an open call for founders building in this space.