tech

2026-07-10 · open the original ↗

• Starlink replaced a dead dish with its rare Gen 3 Performance Dish. The lake house's backup WAN, an older-generation High Performance dish, died about six months before the video, stuck pointing straight up with no signal and not fixed by power cycling. Starlink judged it dead and sent the Gen 3 Performance Dish free of charge, about four years after the owner bought the original. The host calls that "not a bad RMA".
• The dish targets boats, businesses and remote sites, not homes. Starlink rates it at about 470 Mbps down and just under 50 Mbps up for now, and says gigabit download is coming later this year through a plan update with no new hardware. It costs around $2,000. The all-metal case is rated to sit under a meter of water for half an hour.
• The host sees fixed phased-array steering as more reliable. He explains that the dish steers its signal in software across a 140° slice of sky instead of being aimed by a motor, and is mounted at a fixed angle. He thinks losing the servo means one less part to fail, and he rates the casing and pole-mount engineering a clear step up from the Gen 1 dish.
• Swapping in the Gen 3 dish needs a new cable and pole mount. The dish end uses a proprietary click-lock connector that looks almost like XLR, so an older dish's cable cannot be reused; the crew had to pull a new one through the conduit. The new mount sits on top of the pole rather than inside the old roughly 4-inch mount, so they reused part of the old mount and drilled through it to stop it twisting.
• The power supply is not a router. It takes AC or DC input, which the host says suits RVs and ships, and has two Ethernet ports: one to the dish and one to a third-party or Starlink router. It has no Wi-Fi. In a temporary test on the dock, feeding UniFi as WAN 2, it reached about 250 Mbps down and 30–40 Mbps up before final mounting.

2026-07-10 · open the original ↗

• Gravity is curved spacetime, not a force. Einstein's "happiest thought" (1907) treated gravity as an inertial force like centrifugal force — licensed only because gravitational mass and inertial mass are identical, a coincidence verified to one part in 1,000 in Newton's day and now confirmed to one part in 10^15.
• A black hole is the only 100%-efficient power plant physics allows. Chemical rockets convert only about 10⁻¹⁰ of rest-mass energy to usable energy and fission/fusion about 10⁻³ to 10⁻², but slowly lowering mass to a black hole's event horizon and releasing it extracts essentially all of its mc², since the extractable fraction is 1−√(1−2GM/rc²), which approaches 1 at the horizon.
• A distant observer never actually sees anything cross an event horizon. Infalling matter's light redshifts and its clock appears to freeze as it nears the horizon, fading to invisibility rather than visibly crossing; from the infalling observer's own frame the crossing feels unremarkable, and for a sufficiently large black hole tidal forces are gentle enough that one could in principle live an entire life beyond the horizon before reaching the singularity.
• Brown and Patel diverge on how far pure reasoning can take AI in physics. Brown calls general relativity an extreme case, reachable from just the finite speed of light, Lorentz symmetry, and the equivalence principle, and speculates that many parallel LLMs exploring that small space of options got "very, very lucky" it worked once for Einstein; on whether humans could still follow superhuman AI-derived physics, he argues against Terry Tao's "indigestion" worry (billion-line, inscrutable machine proofs) that LLMs are likely to be superhuman explainers too, citing a recent AI disproof of the unit distance conjecture that mathematicians found comprehensible.

2026-07-10 · open the original ↗

• She designs almost entirely by voice, using a YC tool called Aqua. YC's head of design says she thinks much faster than she types, so she talks to her computer in a stream of consciousness to describe features, largely living inside a coding tool alongside Claude to build full projects end to end.
• Disposable, self-built tools are becoming a core design habit. For the Paxel project she asked Claude to implement a paper.design dithering shader, then built herself a throwaway modal to fine-tune its parameters exactly to her liking before discarding the tool once the feel was right — a pattern she says she now reaches for constantly.
• A soul.md file built from meeting transcripts became the source of truth. For the Sodazine zine project she dumped every meeting recording and a written manifesto into one file and fed it to Claude; the agent then organically surfaced details she never explicitly asked for, like the exact launch-party date and an interactive San Francisco map, because that context lived in the file.
• Feeding Claude a Pinterest mood board let her one-shot 16 website concepts. Rather than designing herself first, she gave Claude visual references and content and asked it to generate 16 distinct website iterations, then built a simple internal tool to bookmark and compare favorites — treating the exploration itself, not the output, as disposable.

2026-07-10 · open the original ↗

• AI has removed capital, headcount and skill as startup bottlenecks. The host summarises Anthropic's 35-page Founders Playbook, released in mid-May a day after Claude for Small Business, which casts the founder as an "orchestrator of agents" whose scarce resource is deciding what to build. He grants critics' point that it is partly a manifesto for Anthropic's products, but says its advice is disciplined and often tells founders to slow down, against Anthropic's own interest.
• Building a prototype is not validating an idea. The playbook cites 42% of startups failing for lack of demand even before agentic coding and expects worse now; it prescribes a testable problem statement naming who suffers, how often and how badly, an AI told explicitly to argue against the idea and each tier of competitor, and interviews about the specific past rather than hypothetical use. Only then should a founder build the single core interaction and show it to five target users.
• Agentic coding creates compounding technical debt without written architecture. Because each AI session re-derives foundational decisions, the playbook says to document architecture, dependencies to avoid and accepted trade-offs up front, in a CLAUDE.md or equivalent, and to log decisions at the end of each session. It also warns that launch energy mimics traction, so retention targets should be set before launch and tested against the Sean Ellis 40% "very disappointed" bar, and that AI-written code needs a security review before any user touches it.
• At launch, the founder becomes the bottleneck. The playbook's fix is to sort everything the founder handles into what can be automated, what needs a human other than the founder, and what truly needs founder judgment, while auditing the MVP's architectural debt and treating security and compliance as no longer deferrable.
• At scale, the moat is accumulated depth an incumbent cannot copy. Its test is whether users would stay if a well-funded incumbent cloned the product; the answer it proposes is domain expertise encoded in edge-case tests, compounding user data and workflow lock-in. The host's own reading is that this is lean-startup discipline for an era when execution is cheap and judgment is scarce.

2026-07-10 · open the original ↗

• Transformer models have not yet hit a scaling wall. DeCarlo traces modern agentic AI to the 2017 Google paper "Attention Is All You Need" and relays Sergey Brin's point that the expected limit never came: more data, parameters and compute keep producing better models, which he says explains plans such as Meta's data center "half the size of Manhattan".
• Agents run on GPUs because games made GPUs good at matrix math. He explains a neural network as a weighted graph that reduces to matrices, and an agent's observe–reason–plan–act–reflect loop as many repeated inference passes, so video-game hardware ended up driving AI's hardware side.
• He argues creativity scales with compute, an emergent property of agents. His examples are people whose ideas outrun their skills: a man who made a Pixar-style trailer to propose, sites like meshy.ai that turn a photo into a printable 3D model, and a trained sculptor at his antique mall selling mash-up figures. Giard counters the "AI makes us lazy" charge with calculators and writing; DeCarlo's version of the payoff is the freed time and head space.
• One non-developer with ChatGPT revived a dead smart-bulb ecosystem. When Sengled bulbs stopped working in early 2025, a Reddit user who told DeCarlo he was a program manager, not a developer, used ChatGPT chats and others' protocol reverse-engineering to build custom ESP8266 firmware plus a loader shim. DeCarlo calls it one person doing "arguably the work of an organization", and says the key skill was knowing how to ask.
• Agents work best with a human in the loop, not fully autonomous. Extending that fix for new-in-box bulbs bought for about $3 a pair on eBay, DeCarlo's Copilot agent had him flip a light switch five times and make about 10 trips upstairs to test each attempt, and suggested Ethernet when joining the bulb's Wi-Fi cut its own connection. He says problems that once followed him on dog walks and into the shower now get solved, leaving him less distracted.

2026-07-09 · open the original ↗

• A grid of users by day reveals patterns that aggregate metrics hide. He calls this a "dot plot": one row per user, one column per day, with a dot marking each day they performed a chosen value event. DAU/MAU graphs lump everyone together and can look like healthy growth even when individual users aren't getting value, whereas the grid surfaces patterns like distinct weekday-only and weekend-only usage cohorts.
• He traces the idea partly to a PayPal fraud-detection technique. In PayPal's early days, humans stared at visualized transaction graphs and could spot anomalous patterns worth investigating even without knowing what fraud looked like in advance, from Max Levchin — the same principle, he says, applies to spotting product problems in dot plots.
• A real YC example shows dot plots predicting churn before it happens. A company sold an $80,000-a-year, ten-seat contract, but a dot plot would have shown only three seats ever activated with sporadic use; the internal champion later left, the incoming employee canceled at renewal, and the warning was visible in usage data the founders never looked at that way.
• The two biggest misuses: wrong event, too coarse a time period. He warns against populating dot plots with vanity events like "opened the app," recommending true value events like sharing a photo or processing an invoice instead, and against using weekly rather than daily or sub-daily time buckets, which he says hides most of the signal.

2026-07-08 · open the original ↗

• Gusto Co-Founder automates small-business processes end-to-end, from payroll to weather alerts. Eddie Kim, Gusto's co-founder and head of technology, says it goes beyond Gusto's existing payroll and HR tools by executing full workflows unprompted — approving timesheets, running payroll, or texting a tour operator's customers to bring an umbrella if rain is forecast — reachable entirely through SMS and Slack.
• The idea came from Kim's own always-on personal AI agent at home. After spending eight hours setting one up on a Mac mini, he found himself using it mainly as "a glorified search engine" until a missed flight left him five hours to prototype in Claude Code, convincing him Gusto's customers should get the same agentic power instead of waiting on Gusto's product roadmap.
• He deliberately avoided a fully non-deterministic design for tasks like payroll. Modeled initially on an open-source agent's cron job that reruns an LLM every 30 minutes, Gusto Co-Founder now also supports scheduled, deterministic triggers, because Kim found a purely heartbeat-driven agent couldn't guarantee exactly when a critical task like payroll would run.
• Five people built the whole product in ten weeks with no docs. The team ran a single 24/7 Zoom call, wrote pull requests before discussing ideas, and threw away roughly half the code they wrote — a process Kim says was faster than writing PRDs, because implementing an idea surfaces more information than describing it ever could.

2026-07-07 · open the original ↗

• Extreme power concentration is AI's most neglected risk. Todd estimates 100,000 to a million people speed up AI capabilities and about 1,000 work on technical alignment, but perhaps only 20 researchers specialise in AI-enabled power concentration. If recursive self-improvement delivers what some models suggest, three to 10 years of progress in under a year, growth turns super-exponential and leaders' leads widen, so one company could control a workforce of 200 million digital workers or one country could dominate; AI also makes universal surveillance possible.
• Even aligned AI could gradually disempower humans. In this scenario, newer on 80,000 Hours' list, normal economic competition makes humans expensive relative to data centres, and states come to resemble oil states that need no workforce, removing the strike as a check on government. Todd's hoped-for saving grace is that almost all accessible matter and energy is in space.
• Whether alignment is easier than feared remains undetermined. Todd says LLMs, unlike earlier game-playing RL agents, understand human values, but whether they act on them is untested because agents cannot yet plan over months or exceed humans; he cites Meta research showing reward hacking grows with problem difficulty. He ranks climate change ninth because AGI may arrive around 2030 while most climate damage comes in 50 to 100 years, a richer world could decarbonise more cheaply, and it cannot end civilisation.
• There is no solid ground, only a moving bottleneck. Todd's career advice is to keep moving to whatever humans can still do that AI cannot, whose value rises as AI improves; once a digital remote worker exists, he says, it becomes superhuman almost at once by running about 50 times faster and sharing learning across copies, leaving legal, physical and "relational" work. He sketches converting car factories to robots at roughly a billion a year by mass, and from his book repeats that following existing passions misleads, since 90% of Canadian students' passions were sport, art and music against 3% of jobs.
• AI engineers now have unusual leverage on these risks. Todd says alignment has turned empirical and is bottlenecked by engineers to run monitoring, red-teaming and interpretability work, with MATS and ARENA as three-month routes in; policy teams need data scientists on labour impacts, and builders can work on wastewater pandemic surveillance. As a lower bound, he says a big-tech data scientist donating could save tens of lives a year at $3,000 to $4,000 each against malaria.

2026-07-03 · open the original ↗

• Cheap software leaves people problems and interface problems unsolved. Asked in episode 999 where the moat goes, Chip Huyen says poor product-engineering communication is a people problem no orchestration tool fixes. She points to open problems in human-AI interfaces (why coding agents live in both editors and terminals, and why sessions need the laptop open) and in making the world AI-ready, citing city street-light APIs that let delivery robots cross roads.
• Scaffolding built around weak models limits you once models improve. The founding AI lead at Astrocade, a no-code game-creation platform with tens of millions of users, says in episode 997 that the company adopted Claude Code company-wide around June 2025. Structure built in 2023 and 2024, before vibe coding worked, became a constraint; the lesson is to build so the system is not outdated when LLMs improve in two or three months. He says there is no secret sauce beyond how agent, prompts and tools are combined and benchmarked.
• AI data centres are loud enough to ban noise-cancelling headphones. Lightning AI's Frank Basso says in episode 1003 that data halls run at 105 to 110 decibels. Active cancellation would have to generate the same level, so staff wear passive in-ear and over-ear protection, with OSHA sound studies, radios and hand signals. Liquid-cooled halls are no quieter and are more industrial, with 20-inch insulated water mains overhead and ESD protection required at the cabinets.
• Self-awareness is how closely your self-view matches others' view of you. The episode 1005 guest cites research in his book that only 15% of people are self-aware on that definition. He credits journaling, meditation, exercise and resisting idle phone use with building his own.
• AGI does not require consciousness, but its breadth is years away. In episode 1001 the host tells Kirill Eremenko that AI already reaches some level of Google DeepMind's AGI tiers on most screen-based tasks, but acting in the physical world, the focus of large world-model seed rounds, remains far off. He argues next-token prediction with checking needs no consciousness, and that brain-inspired design is not necessarily the right path because silicon has different constraints, such as unbounded scaling and speed.

2026-07-01 · open the original ↗

• Pannu's winning essay proposes ending airborne disease transmission through physical infrastructure. She argues the OpenAI Foundation should spend an estimated $40–60B over ten years on far-UVC lamps and similar passive infrastructure in schools, hospitals and transit hubs, modeled on smallpox eradication and vaccine advance-market commitments, targeting a 60% cut in seasonal flu mortality and an order-of-magnitude drop in pandemic-scale outbreak risk.
• Erdil argues outsider countries need classic pro-growth reform, not novel schemes. He recommends property rights, low capital taxes, deregulation and political stability — the same policies that work today — reasoning that capital, not labor, will determine output once AI fully automates human labor, but is pessimistic that dysfunctional governments like Nigeria's or India's will adopt them, and instead hopes AI-supply-chain incumbents like the US and China simply make bigger mistakes.
• Li argues labs should profit like Hong Kong's MTR — from property, not fares. MTR's rail service has never covered its own construction costs (three years of peak rail earnings would fund only 8% of its 2024–2026 capex), so it funds expansion through station-adjacent real estate instead; Li argues labs should similarly pursue government-granted deployment rights, proprietary RL reward data, Palantir-style forward-deployed integration, and data trusteeship over national datasets rather than expecting API pricing to ever be profitable.

2026-07-01 · open the original ↗

• Early-career engineers declare done long before it is done. One speaker says from teaching that people still developing taste don't know what done looks like, and admits to declaring done at 80% while the other keeps finding edge cases at 98–99%. The other says they cannot call it done until they believe it is, because any weakness they hit someone else will hit too, and a bad experience taints the reputation of the whole thing.

2026-07-01 · open the original ↗

• The Azure Innovations talk is rebuilt from scratch for every event. Russinovich says the Build deck started at about 180 slides and was cut to 54 to fit 45 minutes, judged by the talk's infrastructure theme, time per topic and balance across areas. What made it included Azure Boost with VM-to-VM RDMA, MRC, a networking protocol built with OpenAI for training networks of 100,000 GPUs, live migration of containers in ACI, and a Microsoft Research Cambridge micro-LED optical-fiber demo; Russinovich rejects comments that it was "nothing new".
• Cut hardware work targets data-center efficiency. Zonal cooling, newly in production after years of research, cools liquid-cooled GPU racks and air-cooled racks differently; high-temperature superconductors carry power from substations with less loss; microfluidics flows coolant directly over delidded silicon through 200-micron pins, with Nvidia, Intel and AMD working on etching parts in the fab.
• Autonomous agents and a digital twin now run Azure operations. An agent called Miles coordinates the dozens of daily metro-fiber cuts with utilities via emails and tickets, work humans did before, and has completed repairs lasting up to 35 hours and 50 interactions. Azure Resource Graph, a graph database of billions of objects and tens of billions of edges, from IP addresses to HVAC units, underpins operations, repairs, inventory and security, alongside an AIOps root-cause system called Brain.
• Prompt injection can be blocked deterministically by labeling data. Russinovich says most current defenses are probabilistic heuristics that can all be defeated, while the research team's approach labels data by trust and confidentiality; once untrusted content taints the context, sensitive operations need human or system approval. Other cut security items were a signing transparency service for policy-checked signatures and SBOMs, and multi-party analytics in confidential clean rooms.
• Copacetic patches container images without rebuilding their layers. From a vulnerability report it builds a patch layer that masks vulnerable binaries in the layered file system, avoiding cascading rebuilds of layers that may belong to an ISV; the demo fixed 19 binaries in an image with Debian, Python and Go vulnerabilities. Russinovich says Azure patches millions, probably tens of millions, of container images a month this way in its internal registry; the slide, like one on GPU-accelerated SQL, survived until the night before, and that speed-up, Russinovich says, is eroded by moving data to and from the accelerator.

2026-06-30 · open the original ↗

• The speaker built a new editor because VS Code felt intolerable. Sitting in front of a blinking cursor in VS Code one day, they concluded "I can't do this": every problem they solve needs a code editor, so they could not move on until they had a better one.
• Agents change the interface but do not remove the need for one. The speaker still expects to need a user interface for working with software through agents, and says they are "not yet signing up" for never looking at code or caring about the architecture.

2026-06-30 · open the original ↗

• The AI math frontier is jagged even within math itself. Sanderson says IMO geometry problems have been brute-force solved in nineteen seconds since 2024, but combinatorics — the "wild card" category — still resists; models would have won gold in 2024 only because that year's problem mix happened to favor geometry over combinatorics.
• Galois shows conceptual breakthroughs can resist verification for a century. Group theory's value took roughly a hundred years to confirm — from Lagrange's 1770s-era hunch about symmetry, through Abel's 1824 proof that quintics are unsolvable (he died at 26) and Galois's prison notes (ignored for about twenty years until Liouville, then another twenty to Jordan's modern treatment) — to Gell-Mann's twentieth-century use of it to predict quarks; Sanderson argues no short-loop verified reward could have captured Galois's "I think there's something here" instinct in real time.
• Patel argues math and code advance faster because they're "grindable," not just verifiable. Checking whether a package arrived is just as verifiable as checking a proof, he says, but code and math can be replayed thousands of times in parallel, deterministic containers, whereas computer-use rollouts get blocked by bot detection and real-world business tasks can't be identically re-run — a distinction he thinks is underrated relative to verifiability alone.
• The two disagree on why AI writes worse than it codes. Sanderson argues good writing requires deliberately choosing unpredictable moves that next-token prediction resists; Patel counters that writing, unlike code (where any implementation that works is fine), has no gap between process and product, so it can't be "slop that still functions." Both agree models are weak at theory of mind, citing a study where Botox-frozen faces made people worse at reading others' expressions as an analogy for why models, lacking embodiment, might structurally lack it too.

2026-06-30 · open the original ↗

• Replacing Meraki with Peplink saved the client money over three years. Crosstalk, which designed and installed the network and sells Peplink, says RLL Racing faced about $100,000 in Meraki licensing plus about $30,000 to replace end-of-life hardware, roughly $130,000 for another 3 years. Peplink gear, its PrimeCare licences (which the host likens to an extended warranty plus InControl access) and Crosstalk's install fees came in $80,000 cheaper over the same term.
• The headquarters core was swapped overnight, chosen for SpeedFusion capacity. A team of three spent an 18-hour day replacing the Cisco firewall with a Peplink Balance 1350 EC (rated over 5,000 users, 25 Gbps, dual hot-swappable power supplies) and 19 switches one-for-one with Peplink 48-port units. The host says the real reason for the 1350 EC was its 2 Gbps of SpeedFusion VPN throughput and 800 SpeedFusion peers, enough for every truck, stand and fuel rig.
• Two cutover problems needed workarounds, one still unexplained. Access-control readers on the switches' rightmost 2.5 Gbps ports would not negotiate; moved to 1 Gbps ports they settled at 10 Mbps half duplex. The two switches carrying about 40 Meraki MR44 access points pegged at 100% CPU for unknown reasons; diagnostics went to Peplink, and one Meraki switch stays in the rack for the MR44s until RLL replaces them with Peplink access points before their licences expire.
• Layer 2 SpeedFusion bridging puts every trackside site on one subnet. Each track gives a team one fiber feed carrying VLANs for telemetry and IP camera feeds; every truck, stand and fuel rig gets its own router with two or more WANs (track fiber, Starlink, 5G, Wi-Fi as WAN), tunnelled back to the 1350 EC hub in one broadcast domain. The host claims bridging is not unique to Peplink, but doing it reliably over bonded, failure-prone mixed links is where other vendors cannot keep up.
• Lab and live-race tests showed the feed reaching every site. In a garage demo, a stream pushed through a Balance 310X on a Starlink dish appeared on a timing stand (Max BR2 Pro on a separate Starlink Mini), a viewing stand (Max BR1 via Wi-Fi as WAN) and the Indianapolis operations centre. At the Long Beach Grand Prix an RLL crew member said the Indy 500's infrastructure is solid but street tracks are "bootleg", so Starlink backup matters; the video was sponsored by Crosstalk Mobile.

2026-06-30 · open the original ↗

• Analysis only has value once people act on it. Eijkelenboom, a former data analyst and online poker player who read opponents through purchased hand-history databases rather than faces, says a marketing manager's verdict that weeks of his analysis "isn't helpful" taught him that good work includes making it useful. He argues communication is learnable even for introverts, and that a Philips manager's feedback (senior stakeholders "don't know" and want analysts to propose the direction) is why data people should push back on requests and recommend actions rather than take orders.
• Replace "and, and, and" presentations with "and, but, therefore". He says the biggest mistake data professionals make is showing too much complexity and methodology, which stakeholders tune out. His framework sets up the situation, introduces a twist and lands on a recommendation, e.g. revenue stable for 2 years, but down last quarter because returning customers are declining, therefore focus on retention; he adds that sharing rough drafts early both shapes the work and earns buy-in because the idea becomes partly theirs.
• Self-awareness is rarer than people assume. His book cites research that only 15% of people are self-aware, which he recalls as defined by how closely self-perception matches how others see you. His remedies are journaling, meditation (he mentions Headspace, Calm, Insight Timer and Sam Harris's app), exercise, leaving the phone in the pocket, and asking for feedback as "what should I keep, stop and start doing?" rather than an open question that invites polite praise.
• Changing behaviour first rewires beliefs about yourself. He argues childhood experiences install "personal algorithms", such as his own conflict avoidance from mediating between his parents, which showed up as building dashboards without asking why they were needed. His fix was rules that force discomfort, like asking talkers in a train's silent carriage to be quiet or volunteering to present, practised for about 30 days, after which his beliefs shifted from anxiety toward enjoyment.
• Emotion drives decisions, including data people's. He says analysts who see themselves as purely rational get triggered by pushy, often more senior business stakeholders who want a model "in an hour", and that recognising the trigger and the stakeholder's own goals makes pushback easier and calmer. Using his book's elephant-and-rider image, he argues that because data products exist to help people decide, understanding how people decide is what makes the work valuable.

2026-06-29 · open the original ↗

• DeltaDB is recast as an event log of programming itself. One speaker moves from seeing Delta DB as "a different kind of version control" to calling it an event-sourced database for the act of programming, which they describe as by nature a social and multiplayer activity.
• It records and indexes every edit, not only snapshots. The other says it stores a directed acyclic graph like Git's, but of everything that happened; the index lets a reader start from the code and reach each edit and the conversation that produced it.

2026-06-28 · open the original ↗

• Workflows chain agents and logic nodes in Foundry. Giard calls Workflows a feature "in preview right now", reached from the Workflows option on the project's Agents page, and he builds on the agent from episode 215. A new workflow starts blank (only a start node) or from a template. The nodes he lists are: invoke an agent (pick an existing one or create one on the spot), set, reset or parse a variable, conditional logic, loops and go-tos, send a message to the user, ask the user a question (which pauses for input), and end.
• The human-in-the-loop template loops until the user's answers match. It stores the user's first input in a variable, asks them to confirm it, and uses a condition plus a go-to to keep asking until the reply matches the original. In the preview, which Giard likens to the agent sandbox, mismatched replies sent the flow back around the loop until a matching one confirmed it. A workflow's name must be unique within the project.
• The sequential template passes each agent's output to the next. Giard calls it "one of the more interesting ones". Its marketing example runs three agents in order: an analyst lists key features, target audience and selling points; a copywriter turns these into English prose; an editor polishes the wording. His reason for using separate agents rather than one big prompt is that each part can be tweaked on its own, and one agent can be reused across several workflows. In the demo, a made-up kitchen gadget went from bullet points to rough prose to finished copy.
• Each agent can use its own model, temperature and top P. Giard says temperature and top P do the same thing with different algorithms, and Microsoft recommends changing one but not both. Lower values give more consistent, deterministic output. Higher values give more creative output but a higher chance of hallucination, so the right setting depends on the use case. Tools, such as search over internal repositories, databases or SharePoint, and guardrails can also be added, but he skips them here.
• Publishing a workflow as a workflow app exposes an endpoint. Giard says applications or a website can call that endpoint to use the workflow.

2026-06-27 · open the original ↗

• India can now build globally dominant AI companies, not just local ones. Puneet, a SuperDaily founder turned investor and builder, contrasts this wave with the mobile-era wave that "tokenized labor" into hyperlocal networks; AI adoption is global and meritocratic, so an Indian student can cold-email U.S. insurance companies and land contracts without ever having lived in Silicon Valley, because success now hinges on technical depth rather than warm introductions.
• Coding agents let small, second-moving teams beat larger incumbents. Arnav, formerly of YC and now at Peak XV, cites Giga winning a DoorDash contract against competitors with hundreds of employees while fielding a team of eight, arguing this reflects a broader pattern where founders with product clarity and modern agentic tools can out-build first movers.
• Heavy token spending on AI tools is framed as a competitive edge. A YC partner says he didn't grasp how far frontier models had advanced until he paid for a $200-a-month max-usage plan in December, arguing that letting "the tokens rip" — running far more tests and edge cases than a human would bother with — produces meaningfully better output, and that inference costs will keep falling.
• YC's application advice centers on clarity, taste, and unassigned "projects." A YC partner says YC invests in founders, not ideas, since first ideas often don't work; he defines taste as building demos backed by real customer insight rather than aesthetics, and recommends young builders do "projects" — two or more people building something nobody assigned them and getting someone to actually use it.

2026-06-27 · open the original ↗

• Teams skip upgrades because the payoff isn't immediately visible. Fritz says a move such as .NET 7 to .NET 8 costs hours of investigation and hand-tuning with no obvious ROI. Those who follow the industry see faster frameworks and lower memory use, so more applications fit on fewer servers and in containers, cutting the bill. Meanwhile each skipped version makes the eventual migration a bigger "mountain".
• Modernizing means rewriting code, not bumping version numbers. Fritz lists changed identity and security practices, EF Core features absent from EF6, moving hand-written ADO.NET to ORM and repository patterns, OpenTelemetry with Aspire, and cloud resiliency (multiple instances, Front Door, Redis caching, Service Bus queuing). He says the old deterministic Upgrade Assistant relied on find-and-replace and struggled when code didn't match its patterns, while Copilot can write code, check builds, look up documentation and use skills that encode those patterns.
• The modernization agents assess, plan, then execute, with a human checkpoint. Fritz's tool set, the GitHub Copilot Modernization Agents, runs on Copilot in Visual Studio, VS Code, GitHub and the command line. The plan is an English markdown file the developer can edit before anything runs, for example to impose an in-house standard of Newtonsoft.Json over System.Text.Json. Execution can happen on a separate git branch, and progress is checked off in task markdown files on disk so a stopped or cloud run can resume. Fritz says that whatever model wrote the code, "I'm the one responsible for it running in production."
• The team now targets whole portfolios, not single applications. Fritz says his team has focused on this for about the last six months, because organizations hold back because their applications are integrated with each other and must keep working. A command-line tool scans and modernizes a whole portfolio, and a companion tool called "command center" tracks assessments, plans and execution across applications, with the agents working independently but coordinating where needed.
• Fritz's employer claims large savings and wide version coverage. Fritz, who works on the product at Microsoft, reports 70–80% cost and time savings in the migration process, citing customers including Xbox, Bing and Ford from his Build session. He says support reaches back to .NET Framework 2, covers Web Forms to Blazor on .NET 10 with .NET 11 next, and moves Java up to versions 25 and 26. It ships as an extension for Visual Studio 2026 and VS Code and is documented on Microsoft Learn under "App Mod" or "GHCP modernization".

2026-06-26 · open the original ↗

• Git versions checkpoints; DeltaDB aims to version the process between. One speaker paraphrases the pitch: the abandoned edits, prompts, agent steering and small decisions between two commits are "dark matter" that should be versioned too. The other agrees and adds that the record should be open to multiple participants, in real time as well as later.
• Agent work today is a private session that leaves only snapshots. The other describes the status quo as a one-on-one "DM session" with an agent, whose byproduct is the snapshots the user commits and pushes to GitHub.

2026-06-26 · open the original ↗

• Anthropic's AI-written code is AI-assisted coding, not yet RSI. The host relays Anthropic's report (Anthropic sponsors the show and this episode; he says it had no editorial input) that as of May 2026 over 80% of code merged into its production code base was written by Claude, up from low single digits before its coding agent launched in early 2025, with engineers shipping roughly eight times as much code per day as 2 years ago and one April effort of over 800 fixes cutting a class of API errors a thousandfold. He argues true recursive self-improvement would need the AI to design experiments, evaluate results and choose next steps with little or no human in the loop, whereas humans still set goals and judge results today.
• Measured autonomy is accelerating toward self-improvement. He cites METR's finding that the length of task an agent can complete alone doubled about every 7 months for 6 years and has sped to roughly every 4 months over the past year, and Anthropic's figure that success on its hardest open-ended problems rose from under 20% in late 2025 to 76% by May 2026. His concrete cases: Google DeepMind's AlphaEvolve recovering nearly three-quarters of a percent of Google's worldwide compute through better scheduling and speeding flagship-model training about 1% with a faster matrix multiplication, and Andrej Karpathy's overnight-tuning agent running about 700 experiments over 2 days, keeping about 20 improvements and cutting training time on his own optimized code from 2.02 to 1.80 hours (11%).
• An Anthropic co-founder thinks self-improving AI is more likely than not soon. Anthropic's report lays out three scenarios (AI below top engineers; accelerating AI-assisted engineering with humans steering, which Anthropic and the host consider most likely; AI improving itself), and co-founder Jack Clark gives a 60% probability of a system able to create its successor with no human involvement by the end of 2028.
• Compute, verification and marketing incentives argue against that timeline. The sceptics' case he relays: each generation needs more compute, tied to data-center construction and competing with paying customers; self-generated training data works where success is cheaply verifiable (code, proofs) but risks "recursive drift" elsewhere; and a market leader urging the option to slow down is also asking rivals to ease off. Against that he cites Max Tegmark's warning that the danger is a quiet loss of control, with models trained, goal-set and safety-checked by other models.
• The host lands in the "optimistic middle". His view is that the productivity gains are real and available to anyone now, and that the closer systems come to self-improvement the more monitoring, human checkpoints and oversight should be built in while the tools remain under human direction.

2026-06-26 · open the original ↗

• AI should make people better, not just produce one-shot output. Amundsen says one-shot tools that turn inputs into an OpenAPI document, a spec or a Node.js app shrank his own contribution to "whatever the machine thought". He now writes "coaches" that converse with the user and check whether to continue. He says that is how he learned his company's style guide.
• A coach follows a teaching loop and fits in one file. Amundsen modelled it on how a mentor works: rather than giving the answer, it sets a slightly hard problem and then helps correct the attempt. Each coach is a single self-contained markdown file attached to or pasted into ChatGPT, Claude, smaller models or even Ollama, and all follow a similar templated flow.
• Tedious, exact artifacts go to deterministic tools, not the LLM. Amundsen says generating an OpenAPI document is too tedious to leave to a stochastic model, so the coaches emit configuration or output files for tools, mostly CLIs he sometimes builds himself. Tests need a tool such as Bruno or Postman to be useful.
• Generative AI doesn't know the creative process, so users must impose it. Amundsen describes three stages: brainstorming, refinement, execution. He says AI is good at brainstorming, though only with ideas someone has already had, and does not know when to stop and move to refinement; at execution it yields a usable prototype, not production work. Giard adds that many humans are also bad at time-boxing; Amundsen's coaches use a seven-step process with explicit stopping points.
• Cheap prototypes should raise quality, not just speed. Amundsen argues AI makes it affordable to try a few more experiments before the final one, and that output improves with the extra feedback. His O'Reilly book "AI Driven API Design" makes chapters out of coach exercises; he says he uses AI little in the writing itself because of copyright issues. An early release of at least four chapters is planned for July, with the full book hoped for by the end of the calendar year.

2026-06-25 · open the original ↗

• Indexing every edit lets code be traced to its conversations. The speaker says DeltaDB records everything that happened in a Git-like graph and indexes it, so from any code one can reach each edit, the conversation that created it, and any conversation that referenced any version of it.
• A character can be followed losslessly across versions, even when deleted. Because the causal record covers everything after an edit, the speaker says the system can locate a character in any later version without "squinting through a diff", and work out where it would be if deleted. Their example is knowing that, in the past 30 days, 300 agent conversations referenced a particular piece of code in a file.

2026-06-25 · open the original ↗

• Nomad's new Supply Depot is a curated Docker app store. It arrives in Project Nomad 1.33; the host, whose company makes Nomad, says each app is a tested open-source container with its own setup docs, and installs check for port conflicts first; apps include Jellyfin, Vaultwarden, CyberChef, Stirling PDF, Excalidraw, Homebox, IT Tools, Flatnotes, a file browser, an ebook reader, and Meshtastic Web and MeshCore Web. Version 1.33.0 was the latest at recording, updated from 1.32.1 via Settings, Check for updates.
• The old Kolibri container has no direct migration to the new one. The previous Kolibri education platform (TED Talks, Khan Academy K–12) ran a very old version, so Supply Depot offers a second-generation Kolibri. Users with the old one must install the new version and follow the project's docs to move their content over.
• Any Docker image can now be installed as a custom app. Users can paste an image and set custom port mappings, volume mounts and environment variables, and can edit, restart, reinstall or set a custom launch URL for installed apps. The host says this makes Nomad a home-lab server as well as an offline knowledge stockpile, a feature he had long wanted.
• Automatic updates now cover Nomad itself, apps and content. Each of the three runs in a user-set update window; content means Zim files such as Wikipedia. The intent is a box that keeps itself current while online, so the moment it is disconnected is its point-in-time snapshot; manual content updates remain available.
• A custom online check fixes networks that block Cloudflare. Nomad decides it is online by reaching Cloudflare's 1.1.1.1, so networks blocking that made it think it was offline; a new Advanced tab lets users set their own HTTP check. Other user-requested fixes: maps remember the last view, a rescan-library button picks up Zim files copied in, and knowledge-base ingestion jobs (manual by default) can be cancelled.

2026-06-25 · open the original ↗

• An agent's model is swappable from a dropdown at the top. On the agent from episode 215, the default model is GPT-4.1; the dropdown also offers GPT-4o, a GPT-4.1 mini variant, o4-mini and others. "Browse more models" opens dozens more, searchable by name (he searches "Claude") or filtered by collections, features, inference or deployment SKU, and sortable by updated time.
• The Discover tab's Models page gives detail, including pricing, before deploying. Models there are grouped by task, such as text-to-image, chat completion and fill mask, and Giard says the choice depends on your goals. Each model has quick facts and pricing, which he calls "one really good piece of information" to check before deploying, and a model can be deployed from that page to make it available to any agent in the project.
• Selecting a new model in the agent deploys it to the project. Picking GPT-4o and saving both deployed it and assigned it to the agent; the Deployments list then showed GPT-4o alongside GPT-4.1 and the text-embedding model from an earlier episode.
• Temperature and Top P trade consistency against creativity. Both sit behind an icon to the right of the model for chat-completion models, and do the same thing with different algorithms; Microsoft recommends changing one, but not both. Giard explains that a temperature near zero makes answers consistent and deterministic, while a high one gives more varied answers, useful when there are several right answers, at the risk of hallucination.
• The playground is where to compare models and settings before release. Giard suggests asking the same question with different models or temperatures in the playground beside the agent before deploying it for others, calling it an inexpensive, safe and fast-to-iterate environment.

2026-06-24 · open the original ↗

• Shared agent conversations would expose developers' insecure first moments. Both speakers question whether they want early exchanges with an agent ("I have no idea what I'm doing") in the shared history of the code; one calls those exchanges "my panic attacks".
• The plan is to let users share from a chosen point. In the pre-alpha used internally, the other says, the only option is sharing the whole conversation, which they do after swallowing their pride. The intended control is sharing from a given moment, perhaps after a compaction, so the earlier paper trail stays hidden.

2026-06-23 · open the original ↗

• AI data centers are defined by power density old buildings can't handle. Basso, Lightning AI's VP of infrastructure (the host is a fellow at the same company), says five years ago the industry standard rack was 18 kW, with Microsoft's 50 kW the extreme, whereas 18 kW is now a single server and cabinets weigh about 4,200 lb, ruling out raised floors. Lightning's recently deployed GB300 superpod of about 10,000 chips in Chicago puts 20 MW in a 10,000 sq ft room, supported by 50,000 sq ft of chillers, pumps, UPSs and generators.
• Each build is sized to available grid power, then conservatively derated. In its colocation model, minimum 10 MW, the site's available utility power sets how many serviceable units fit (for GB300, 1,152 GPUs in 16 cabinets; nine units is about 10,300 GPUs), plus networking, storage and a CPU share he says is rising. Lightning applies an 81% derating to nameplate power, against an industry norm of 70–75% he calls thin, so a 22 MW plate rating becomes under 20 MW contracted; providers' build cost runs 1 to 4 million per megawatt and contracts are usually 10 years.
• A neocloud must over-provision because tenants are unpredictable. Unlike xAI or OpenAI building for their own engineers, Lightning is multi-tenant, so Basso builds on Nvidia's reference architectures with no network oversubscription, its own metro networks and a nationwide backbone, and storage sized for customers pulling about a gigabyte per GPU per second. He explains east-west GPU fabrics (InfiniBand or RoCE, at 400–800 Gb and soon 1.6 Tb per GPU) are limited to roughly 50–500 m by picosecond timing, so spanning data halls via super-spines can break training unless tuned. The company had over 35,000 GPUs and expects about 50,000 by year-end.
• Liquid-to-chip cooling does not consume water. Basso says sealed glycol loops feed a heat exchanger to the building loop, which dumps heat through giant radiator-style heat-rejection units, so a large site uses no more water per month than a household; water-spraying cooling towers are, he says, now banned almost everywhere. The halls run at 105 to 110 decibels, which he says requires both earplugs and passive earmuffs because active noise cancelling generates equal sound pressure.
• Data centers do not raise residential power bills, the guest argues. He attributes rates to utility-commission decisions, citing California residential prices from 10 to 58 cents per kWh on the same generation, and says new sites pay millions to tens of millions for grid upgrades, must push backup power to the grid in emergencies under interconnection agreements, and increasingly use fuel cells that produce water. He forecasts 800 V DC distribution with grid batteries replacing generators and UPSs, for 5 to 10% efficiency gains; the host adds his own techno-optimist case and cites The Economist that higher local bills are almost certainly not due to data centers.

2026-06-23 · open the original ↗

• DeltaDB versions the agent conversation, not just the commits. Sobo, founder of Zed, says it records the conversation with an agent together with edits on embedded worktrees, continuously, so a colleague hours later in another time zone can join the thread, get the worktrees mirrored to disk and ask the agent "What was Scott talking about with you?". Worktrees, personal in Git, become collaborative; it is in a pre-alpha used internally, with early access at zed.dev/deltadb.
• Every edit is indexed, which enables "portals" back through time. A keystroke is one delta, and so is a 10K paste; a character keeps its insertion ID and offset, so an annotation can be found in any later version, even after cut-and-paste into another file. Portals jump from an agent's reference in a transcript to the code as it was at that moment; Hanselman calls this structurally impossible in Git, and Sobo says Git only lets you reconstruct it from snapshots.
• DeltaDB is a Git superset meant to need little thought. Sobo says Git's commit DAG is a subset of DeltaDB's causal DAG, it watches the worktree on disk so edits from outside tools are picked up and synced conflict-free, and planned "waypoints" name a moment without taking a snapshot. He calls Git's snapshot model sound to build on, unlike Electron, which Zed had to abandon.
• Sharing controls are still being designed to avoid workplace surveillance. Hanselman raises surveillance and pasted keys and connection strings; Sobo admits he is reluctant to share his own early exchanges, says only whole conversations can be shared today, and plans sharing from a chosen point. How much of the record (his example: 300 conversations touching one file in 30 days) to surface to humans rather than agents is "still figuring out".
• Zed exists because Sobo would not code inside a browser shell. He created Atom at GitHub and with it Electron, which Microsoft took up for VS Code; he wrote Zed in VS Code but rejected "a web browser dressed up as a desktop app" and calls himself the product's "user zero". Hanselman frames this as resisting a monoculture, and discloses he is working on OpenClaw for Windows, whose team sees his before-and-after files but not his sessions.

2026-06-23 · open the original ↗

• Text-to-speech has not made braille redundant; the two are complementary. Rogers, founder and managing director of Bristol Braille, says screen readers make spelling, grammar and numeracy much harder and are not equivalent to reading print, but the screen reader is also what drives a braille display. He relays, while asking not to be quoted on it, a National Federation of the Blind figure that 50% of blind schoolchildren learned braille in the '50s against about 1 in 10 now.
• Rotating embossed octagons replace pins, cutting the cost of braille. The Canute Scientific has 2,160 dots but no pins: 720 octagons, two per cell, since each three-dot half-cell has eight combinations, turned by 98 working motors, with about 1,000 moving parts in all. Conventional piezo cells are expensive and rivals charge tens of thousands of dollars; Rogers says his roughly 700 sq ft British workshop makes the world's most affordable braille cell technology, at about $17 per cell.
• Multi-line braille shows structure that a single line cannot. Text is not linear, he argues: tables run vertically or diagonally, and prose, poetry and sheet music depend on layout. The nine-line display is mirrored on a 13-inch high-contrast monitor for sighted colleagues; it runs mostly in the command line, apps are built from YAML files, and he says it suits working with an LLM because the whole function is in view.
• Owners should be able to repair and own their devices. Having in practice fixed every customer's Canute 360 and sent out repair manuals, Bristol Braille redesigned the Scientific's chassis to come apart, though Rogers warns it is "not something you should do for fun". The braille mechanism is meant to last while the electronics (a Raspberry Pi 5 running Raspberry Pi Debian, swappable hubs) are replaceable, and he says a cheaper copy of his design would make him "their best customer".
• Give blind users the underlying data, not a picture of it. Braille matches print in reading speed and retention according to studies he mentions, while tactile graphics are much less detailed, so a chart should come with its data. He presents his displays as a different philosophy, not a better product than pricier multi-line rivals: clunkier and noisier, but owned and affordable; the dots are made hard because the company worked early on with diabetic users, whose neuropathy dulls touch.

2026-06-23 · open the original ↗

• Unified structured traces beat separate metrics, logs and traces. Majors, CTO and co-founder of Honeycomb, says people realised over the past two months that "you just need traces", meaning connected structured data. Metrics and logs are the "exhaust pipe" of infrastructure you don't control; for your own code telemetry is a product problem, and data split apart can never be knit back together.
• AI can finally close the feedback loop DevOps never built. Her theory is that 20 years of DevOps tried to build one loop spanning code authors and production and failed because instrumenting demanded too much expertise; with OTel and auto-instrumentation she says it is now possible to build faster with telemetry than without. Hanselman's example: instrumenting a Windows app and pointing an agent at it through an MCP server solved his bug in minutes, and she agrees agentic work on brownfield systems with no observability is "so deeply screwed".
• The second edition of Observability Engineering is effectively a new book. The first ran about 250 pages and took 3½ years; the second is over 600 pages after about 200 were cut, with 27 new chapters, the PDF out now and print about a month later. Its last third opens with an open letter to CTOs arguing their AI hopes are blocked behind understanding their systems, and contrasts trust and credibility inside a company with trust and reciprocity with vendors.
• Every deploy is a production experiment, so make it safe. "There is no test for reality": she says teams over-invest in perfection and under-invest in finding and fixing, and that feature flags combined with high-cardinality data, such as enabling one app in a million and comparing it with the baseline, are greater than the sum of their parts. The unit of work is moving from the transaction to agent workflows spanning hours, which Honeycomb's "timelines", just shipped to GA, address.
• Hiring only seniors because AI does the easy work is "totally unacceptable". Majors won't anthropomorphise AI: taste and judgment come from repeated experience, which she calls categorically different from what agents get. Honeycomb hires early-career staff in pairs, never one alone, and she relays a Vercel engineer who sees more junior-sized work than he can build, "the best pitch for junior engineers". Hanselman points to his own paper with Mark Russinovich on investing in early-in-career people, which she has read.

2026-06-22 · open the original ↗

• Cold email works only if your buyer lives on a laptop. Max, a visiting YC partner, describes a founder who spent three months on cold outreach to a legacy industry, then closed more business in three days by walking a trade-show floor instead — evidence, he argues, that founders should map out where their specific buyer actually spends time before picking a channel.
• Nearly every founder's first customers came from their personal network. Based on dozens of founder responses to his internal survey, Max says the first two or three customers almost always came from former colleagues, classmates, or warm LinkedIn introductions, because early buyers are betting on trust in the founder, not yet on the product itself.
• Showing up in person, repeatedly, won customers four through ten. Examples include a founder who flew to the same executive four weeks running before closing them, another who flew to Hawaii and was thrown out after eight minutes before that contact became a top account, and founders running small paid dinners for six to ten ideal customers that reliably outconverted large events.
• Framing outreach as asking for mentorship, not a pitch, converted better. One founder reached dozens of CEOs asking them to be a mentor and converted a few into customers; another paid lawyers $100 to $200 an hour for product feedback and converted about 30% of outreach, since the fee signaled seriousness even as it raised acquisition cost per lead.

2026-06-22 · open the original ↗

• Project Nomad can answer questions sent over Meshtastic, fully offline. The host's demo used a Nomad server, a travel router and a PoE injector powering a RAKwireless WisMesh radio; a handheld radio paired with an Android phone sent "/nomad how to purify water", which the local Ollama model on the Nomad box answered back radio to radio. He says the whole setup can be made portable, as in his own Nomad deck, a product his company makes.

2026-06-21 · open the original ↗

• Text-to-speech has not made braille redundant; the two are complementary. Rogers, founder and managing director of Bristol Braille, says screen readers make spelling, grammar and numeracy much harder and are not equivalent to reading print, but the screen reader is also what drives a braille display. He relays, while asking not to be quoted on it, a National Federation of the Blind figure that 50% of blind schoolchildren learned braille in the '50s against about 1 in 10 now.
• Rotating embossed octagons replace pins, cutting the cost of braille. The Canute Scientific has 2,160 dots but no pins: 720 octagons, two per cell, since each three-dot half-cell has eight combinations, turned by 98 working motors, with about 1,000 moving parts in all. Conventional piezo cells are expensive and rivals charge tens of thousands of dollars; Rogers says his roughly 700 sq ft British workshop makes the world's most affordable braille cell technology, at about £12, or $17, per cell.
• Multi-line braille shows structure that a single line cannot. Text is not linear, he argues: tables run vertically or diagonally, and prose, poetry and sheet music depend on layout. The nine-line display is mirrored on a 13-inch high-contrast monitor for sighted colleagues; it runs mostly in the command line, apps are built from YAML files, and he says it suits working with an LLM because the whole function is in view.
• Owners should be able to repair and own their devices. Having in practice fixed every customer's Canute 360 and sent out repair manuals, Bristol Braille redesigned the Scientific's chassis to come apart, though Rogers warns it is "not something you should do for fun". The braille mechanism is meant to last while the electronics (a Raspberry Pi 5 running Raspberry Pi Debian, swappable hubs) are replaceable, and he says a cheaper copy of his design would make him "their best customer".
• Give blind users the underlying data, not a picture of it. Braille matches print in reading speed and retention according to studies he mentions, while tactile graphics are much less detailed, so a chart should come with its data. He presents his displays as a different philosophy, not a better product than pricier multi-line rivals: clunkier and noisier, but owned and affordable; the dots are made hard because the company worked early on with diabetic users, whose neuropathy dulls touch.

2026-06-19 · open the original ↗

• Chuck says he no longer knows what AI tools are running. He says he adds new tools "every week, now every day" and suspects unapproved integrations and vendors are already active inside his company — framing shadow AI as spreading exactly as fast as AI adoption itself, mostly outside security's view.
• A bot he hadn't approved showed up in his company's Slack. He noticed an unfamiliar AI integration named "Inka" appear on its own; he says he didn't add it and didn't know who did, offering it as a live instance of the shadow AI sprawl he's describing rather than a hypothetical.
• The video is a sponsored pitch for Vanta's "agentic trust platform." Vanta says its AI agent assesses new vendors, gathers evidence, and answers security questionnaires on a company's behalf, claims to cut vendor-assessment time by 50%, and says it's used by more than 16,000 companies, including Cursor.

2026-06-19 · open the original ↗

• Gray-market 5G SIMs risk being shut down by the carrier. The host describes them as SIMs not authorized for resale that resellers dodge the rules to sell. His example: in 2023 Nomad Internet (no relation to his Project Nomad) bought Verizon business SIMs wholesale under fake identities and reprogrammed them; Verizon throttled them to near nothing while customers kept paying, and the company went bankrupt.
• The host pitches Crosstalk Mobile as the authorized alternative. He says his company's SIMs come through a direct carrier agreement as an authorized MVNO, with truly unlimited data, month-to-month contracts and an optional static IP over 5G. Unlimited plans start at $119 a month, which he concedes is not the cheapest.

2026-06-19 · open the original ↗

• Fable 5 is Mythos-class capability released with safeguards on. The host explains (noting Anthropic sponsors the show and this episode without editorial influence) that Fable 5 and Mythos 5 are the same underlying model: Mythos 5 goes to Project Glasswing partners with cyber safeguards lifted, Fable 5 to the public with them on. The tier sits above Opus and follows Claude Mythos Preview, which Anthropic withheld from general release in April (covered in episode 990).
• Fable's lead over other models grows with task length. Anthropic reports state-of-the-art results on nearly every benchmark it tested, and the host relays three examples: Stripe ran a codebase-wide migration of a 50-million-line Ruby codebase in a single day that it estimated would take a team more than 2 months; the model beat Pokémon FireRed from raw screenshots and a minimal harness; and persistent file-based notes boosted its Slay the Spire performance three times as much as they boosted the previous Opus. Pricing is $10 per million input and $50 per million output tokens, roughly twice Opus 4.8 and under half the original Mythos Preview; the host predicts today's capabilities will cost about 1% as much in 2 years.
• Flagged requests are rerouted to a weaker model, not refused. Classifiers watch for cybersecurity, biology and chemistry, and distillation queries and reroute them to Opus 4.8 with a notice rather than refusing. Anthropic says the classifiers are tuned conservatively, trigger in under 5% of sessions, and survived over 1,000 hours of external bug-bounty testing with no universal bypass found; the host reports his own episode script was bounced, perhaps for mentioning cybersecurity.
• An export-control order took Fable offline for everyone within days. Anthropic confirmed receiving the directive at 5:21 p.m. Eastern on Friday the 12th: a letter from Commerce Secretary Howard Lutnick to Dario Amodei barring access by any foreign national, inside or outside the US, which Anthropic says it could only comply with by switching Fable 5 and Mythos 5 off for everyone; other Claude models are unaffected. The Wall Street Journal reported that Amazon demonstrated the claimed jailbreak and that Andy Jassy relayed it to the White House; Anthropic calls it narrow and non-universal, surfacing previously known minor flaws that GPT 5.5 finds as easily.
• The host reads the ban as Anthropic's safety warnings backfiring. He notes the order came days after Anthropic urged labs to agree on a "coordinated brake pedal", follows the Department of Defense's supply-chain-risk designation (still in litigation) and reported failed efforts to block the launch, and lands during a confidential IPO filing reportedly showing about a $47 billion revenue run rate and a valuation near a trillion dollars. He speculates the ban may also signal Anthropic is the lab at the frontier; at recording, legal proceedings were under way and Fable's return was unsettled.

2026-06-18 · open the original ↗

• HTTPS encrypts traffic but doesn't hide which sites you visit. Chuck shows a Wireshark capture where enabling secure DNS hides DNS queries, but the TLS SNI field still reveals the destination site to an ISP or attacker. He says a VPN is needed to encrypt that and other non-browser traffic that HTTPS alone doesn't cover.

2026-06-18 · open the original ↗

• Cisco Cloud Control unifies networking management into one dashboard. Chuck, reporting from Cisco Live, says it replaces what had been seven separate controllers and logins for Cisco customers, calling it the "single pane of glass" network engineers have wanted. He notes it was already running live on the show floor rather than just announced for a future release.
• The platform is built to host AI agents, not just dashboards. A Cisco presenter named Grant says Cloud Control lets users "build trusted, secured, and controlled agents" directly within the same screen. Chuck contrasts this with last year's Cisco Live, where AI was discussed but, in his view, not yet shipped.

2026-06-17 · open the original ↗

• Commitment insurance is pitched as good for clouds and customers alike. One speaker says customers get "an out" on unused commitments and can reinvest the money rather than getting mad at Amazon or Azure, likening it to Ford wanting Geico to exist so people will buy the F-150.
• Abuse is held off by underwriting against moral hazard. Asked how they stop a failing startup buying cover just before it collapses, the speaker says the answer is underwriting, which exists partly to avoid moral hazard: buying insurance and then taking more risk because of it. The clip ends before any specific safeguard is described.

2026-06-17 · open the original ↗

• Mark holds that AI will not eliminate software engineers. He relays asking Ion Stoica of Berkeley at a Build session, who answered no at once, for the reason Mark gives himself: nothing can be fully specified, and where the spec has gaps AI makes things up that may violate implicit assumptions. Mark calls engineering sculpting, needing feedback from use, and concedes it is opinion he cannot win from first principles; Hanselman adds that spec taken far enough is just programming ("Word documents don't break the build").
• System complexity decides where AI can stand in for engineers. Mark says simple, well-understood systems can be built from a few sentences but medium complexity "falls apart". Hanselman's analogy, from a meeting with Japanese CTOs, is IKEA against Japanese carpentry; Mark maps IKEA to low-complexity web-plus-database apps and bespoke state retirement systems, and Hanselman offers his podcast admin site: "classic late 2025 software" with a god object and a thousand tests, fine for an audience of one.
• Evaluation is the bottleneck that will not go to zero. Mark says evaluation means the tests plus human oversight of the output, and that a moderately complex Azure service has tests taking hours, so a commit that kicks off 5 hours of tests forces more evaluation up front, costing time and money. "Cost of software is going to zero" comes, he says, from people seeing only the simple end; both hosts say they could be wrong.
• Hanselman rejects a skills file as a shortcut to seniority. Mark puts the "controversial take", heard from smart people, that a skills file turns a junior into a senior. Hanselman is irritated by "you're an expert" prompting ("I am more than the sum of my context") and unsure whether to be offended by a request to turn his 20-plus years of blog posts into a skill.
• The bar for shipping depends on who ships it. Hanselman calls his OpenClaw Windows companion app alpha despite weeks of trouble-free use, since odd builds or non-English systems will break it; Mark says a solo open-source project can have rough edges but software from Microsoft to many people cannot. Mark relays Simon Willison trusting only software used by many people for a while, and Hanselman says to treat LLM code like a stranger's pull request.

2026-06-17 · open the original ↗

• Founders should stop hunting for the "perfect" idea and commit to one. John, a YC partner, says overthinking startup ideas — either waiting for the ideal concept or doubting your own founder-market fit — produces paralysis; citing Boom Supersonic's Blake Scholl, who came from ad tech with no aerospace background, he argues deep customer knowledge can be built quickly if you commit and go deep.
• "Burning the other boats" means fully abandoning alternatives, not just prioritizing. He points to GovDash, which pivoted at least five times, changing its name and even email addresses each time, before its fifth idea in government-contract software worked so well the company raised a Series B — evidence, he argues, that half-committing to multiple ideas at once produces unreliable signal.
• The bar for "going deep": could you actually run your customer's business? John's test is whether you know their daily crises, whether a specific problem is a top-five pain point, and what they'd pay to solve it — a level of knowledge he says should let you "teach a class" on the problem, built through tight loops of customer conversations and shipped product together.
• Good AI-era startups should verticalize into owning outcomes, not just software. He argues software costs are heading to zero, so durable value shifts to trust, licenses and outcome ownership — citing Corgi Insurance, which acquired an insurance carrier during its YC batch rather than staying a tech-enabled broker, as a model for "being the insurer" instead of selling tools to insurers.

2026-06-16 · open the original ↗

• Crosstalk Mobile runs on T-Mobile as an authorized MVNO. The company, launched about two months before the video, has a direct agreement with T-Mobile, its only carrier for now, with multi-carrier SIMs planned "in the near future". The host contrasts this with gray-market resellers buying wholesale SIMs without a carrier agreement, which he says get shut down when caught.
• The unlimited SIMs really have no caps or overages. Plans start at $119 a month, and the host says no amount of data changes that. Cheaper limited tiers are planned as a priority, for backup links, IoT and RedCap devices such as Netgear's $99 5G backup device and the GL.iNet Comet 5G, with no date given.
• The SIMs suit 5G routers, not phones or tablets. The host says they work well in Netgear 5G and Peplink gear and he has not yet found a router they fail in, but phones and tablets are hit and miss and discouraged. He cites a reseller who got a Peplink router chosen, preconfigured on InControl 2 and drop-shipped with the SIM installed, all by text in a few hours.
• The SIMs get real public IPv4, not CGNAT. The host argues CGNAT makes 5G failover incomplete, since inbound services such as a VPN server for remote staff stop working when the office fails over. Crosstalk SIMs get real IPv4 addresses, and a static IP is $5 a month extra; he personally tested port forwarding through the UniFi 5G Max and the Dream Router 5G Max.
• There is no contract lock-in. Plans are month-to-month, or year-to-year if bought annually, and can be cancelled at any time. The host presents engineers answering customer calls and texts directly as the service's most underrated feature, contrasting it with Verizon Business or AT&T support.

2026-06-16 · open the original ↗

• AI erased his technical moat but raised engineers' value. In this role-reversal episode, hosted by founder Kirill Eremenko, Krohn says the machine-learning skill set he built from his Oxford neuroscience PhD onward stopped being a moat a few months before recording, when Claude started doing that work better than him. His argument is that engineers who stop typing code by hand and use tools like Claude Code can deliver far more, so hiring them is worth more, not less.
• AGI needs breadth into the physical world, not consciousness. Krohn points to Google DeepMind's five-levels paper (his episode 748), which grades AGI by the share of humans a system outperforms, and argues that AI now matches or beats many people on computer-based work like coding and writing but is years from acting in the real world, hence world-model startups' huge seed rounds. He sees no reason consciousness is required, and says biological brains are useful inspiration but silicon has different limits and advantages of scale and speed.
• An AI bubble bursting would still leave useful infrastructure. Citing his episode 974, he argues that, as with 19th-century railways and dot-com-era fibre, investors might lose their shirts but data centers, model advances and talent would benefit everyone; he personally holds diversified investments rather than individual AI stocks and would bet this is probably not a bubble. He invokes Jevons paradox to argue more efficient architectures will raise total compute demand, and says the hardware won't become obsolete because any successor to the Transformer will still need parallel matrix multiplication.
• Most AI projects fail because the enterprise data isn't there. Krohn says the number-one failure he sees is a CEO proving a capability in ChatGPT or Copilot when the company lacks the underlying data at scale (Eremenko adds: or the pipelines to use it). He prioritises projects with RICE scoring (reach, impact and confidence over effort), estimated for production rather than prototype, and argues the people doing the workflow plus the technical team score it better than executives; he cites MIT NANDA's 95% failure figure while noting the exact number has been debunked.
• The money is in bridging AI to real workflows, not pure SaaS. He says his consulting firm Y Carrot, about eight people and founded a little over a year earlier, typically shows a prototype in week two and has cut multi-week processes to minutes without job losses so far, because freed time goes to higher-value work. He argues the 80%-margin pure-SaaS model breaks down when Claude Code can replicate it, leaving durable opportunity in forward-deployed or consultative models like Palantir's, now used by Anthropic and OpenAI.

2026-06-15 · open the original ↗

• Groww's first idea, a robo-advisor, failed before pivoting to full transparency. Lalit Keshre, Groww's co-founder and CEO, says the 2016 robo-advisor model stalled despite India's low investor penetration; customers kept asking "why this product, why not that," which led the team to open the full catalog with transparent pricing instead, Flipkart-style.
• He chose to operate only in regulated territory as a risk strategy. Rather than exploit unregulated or gray-zone areas — which he estimates cover roughly 40% of the space — Groww obtained licenses upfront, which he says removed decision variables and made execution simpler even though it was slower.
• Groww ran four years at zero revenue before customer demand unlocked monetization. Early commission-based mutual-fund revenue was abandoned after vocal power users demanded zero-commission "direct" funds; the company bet that near-zero acquisition cost, high retention and heavy money flow would eventually produce a business model, and stock trading later became the monetization lever.
• Four co-founders stayed aligned for a decade via one values document. Keshre says strategy can change "like pencil," but core values like "customer first," written down at founding, should not; he also stresses genuinely enjoying each other's company as a practical requirement, not a nicety.

2026-06-15 · open the original ↗

• One prompt built and deployed a conference sessions site. Padua, who has been at Microsoft for four years, gave the new GitHub Copilot app a JSON file of Build session data, and it created a repo, built the app and set up deployment to Azure. He puts the time at about five minutes at one point and "10 minutes tops" at another. The app ran a team of about 15 agents under a "director". His only manual step was copying the Azure static web app token into the repo's GitHub Actions secrets.
• An internal agent tool matched conference sessions to a partner. Padua says Scout, an internal Microsoft tool built on the open-source OpenClaw, ran parallel agents that read his site and researched the partner Solo Network. When Playwright failed, they fell back to simulated mouse and keyboard input. They picked 16 best-fit sessions, built a table with links and levels, and drafted an email offering a 30-minute call. He has done this for three partners, in Portuguese where he remembered to ask.
• Agents act as you, so limit what they may do. Padua advises running OpenClaw in a virtual machine you can destroy and letting it only draft, never send or auto-reply to email, citing stories of agents deleting emails or a whole operating system. He had Copilot write a PowerShell deploy script for him to review rather than letting it deploy through the Azure MCP server tied to his credentials, to avoid unwanted resources and costs.
• Copilot digs answers out of dense internal documentation. Padua uses Microsoft 365 Copilot to summarize meetings he missed and to draft LinkedIn posts. Searching the documentation this way, he found why a partner lacked the Anthropic model quota it expected in Foundry: he says those models work only under an EA contract, and the customer was on CSP. He also tries competitors' tools, such as OpenAI's DALL-E, Anthropic and Perplexity, because customers ask his opinion of them.
• AI amplifies technical knowledge rather than replacing it. Padua argues that results depend on the user knowing what they want to do, and that AI then delivers "the best of the people can do with a prompt". Giard adds that the agents narrate their failures and change approach, as when Playwright failed, which he says is what a person would do, only faster.

2026-06-12 · open the original ↗

• Protein language models are starting to show the same LLM-style scaling laws. A presenter says a new protein-model family from Biohub, ESM Cambrian, keeps improving log-linearly with more training compute and sequence data — unlike its predecessor, which plateaued — largely because the team scaled training data from 50 million to 2.8 billion sequences pulled from environmental DNA sequencing.
• Naive AI self-play collapses into generating deliberately convoluted, useless problems. A Stanford PhD student found that rewarding a "conjecturer" model for stumping a "solver" model just produces artificially messy math problems; his fix grounds new problems in ones the solver already failed on and adds a third model to judge relevance, letting a 7-billion-parameter model match a 670-billion-parameter baseline's solve rate after eight times more self-play compute.
• Lean formal verification marks a shift from "vibe coding" to "verifiable coding." A Caltech researcher says frontier labs have used the Lean proof language to solve IMO gold-medal problems and decades-old Erdős conjectures, and describes his own project formalizing neural-network operations — including proving flash attention mathematically equivalent to standard attention — as evidence AI-written code can eventually carry mathematical guarantees, not just pass tests.
• Treating agentic coding like a real-time strategy game tripled shipped pull requests. A startup CEO known for aggressive token spending says running many parallel Claude Code "worker" agents under one orchestrator, instructing them to push forward without stopping for feedback, and feeding a shared knowledge base after every task grew pull requests per engineer per month 3.5x, then another 60% once the whole team adopted the workflow.

2026-06-12 · open the original ↗

• Beat the speed of adoption, not the speed of research. Responding to a former Dutch professor of human-AI collaboration about teaching safe AI use, Kirill Eremenko, the show's founder, argues research moves like a bullet train while adoption in universities, enterprises and government moves like a highway traffic jam, slowed by legacy systems and change management. He says people anxious about AI only need to stay ahead of adoption, and that startups should avoid problems models will soon solve, citing hallucinations he estimates are down about 10-fold in two years and AI-text detectors as fighting the tide.
• The host would fund agents that shield people from manipulative feeds. Asked by Jepson Taylor what they would fix with $10 billion in two months, the host proposes an agent between vulnerable people and the feeds, scams and attention-seeking products they consume, steering them toward what makes them happy; Eremenko proposes AI-run charities that cut admin overhead from what he thinks is about 40% to about 5%. Taylor's own answer is closing what an engineer told him is the gap between the top 10% and bottom 90% of AI developers.
• Build what differentiates you and is slow to build; buy the rest. Answering a corporate listener on the "SaaS apocalypse", the host relays a 2x2 from episode 932's guest from Unframe: avoid projects that are costly and non-differentiating, build in-house what is differentiating but slow, buy what is quick and non-differentiating, and partner (he names Unframe and his own firm) where it is both quick and differentiating.
• The show's big growth lever was video, adopted far too late. Eremenko, who hosted from September 2016 to episode 431, says his one piece of hindsight is that he should have put episodes on YouTube from the start; the host says the show went video-first only 10 months before recording and grew from 20,000 to over 250,000 YouTube subscribers since. The host's own lesson, from Eremenko's feedback, was to be conversational and stop taking notes on air.
• In-person networking beats clicking "apply". The host argues weak new-grad hiring reflects degrees losing their "golden ticket" status and post-pandemic overhiring as much as AI, citing episode 994. His tips for entrants are a GitHub of projects actually done yourself and meeting people in person, because AI-tailored résumés let applicants send thousands of applications and flood the traditional pipeline; he also cites episode 977's NYU machine-learning course, taught by Kyunghyun Cho, that required students to use Gemini.

2026-06-11 · open the original ↗

• Archera sells insurance that lets customers break cloud commitments. Khanna, co-founder and CEO, describes "insured cloud commitments": commit for 3 years and, if the last 24 months go unused, Archera pays back the difference; he says no one else offers it, and customers can cancel the insurance once confident. His analogy is buying the right to break a 3-year apartment lease after the first year.
• Commitments exist because clouds finance capex against guaranteed revenue. Like a builder whose lender underwrites leases rather than Airbnb income, providers borrow to build data centres (CoreWeave is his extreme example), so discounts can exceed 80% for moving from per-second use to a 3-year commitment. He says that in 2026 a commitment is often required just to get GPUs; Hanselman calls on-demand "the convenience store price of compute", and Khanna calls it the promotional price, where you prototype.
• Commitments became a hammer applied to every cloud problem. Reserved instances made sense in the early 2010s, he says, but from the 2017 SageMaker launch, which he worked on at AWS, they were used to ration scarce Nvidia capacity, then to deter multi-cloud and push modernisation, leaving "a spaghetti mess". At $80–90 million a year with one provider, he says, bills contain errors and cannot be treated as a source of truth.
• Underwriting relies on spend history, peer priors and diversification. Archera wants at least a year of real usage data, draws priors from over a petabyte of history across thousands of customers (offering less on Azure VMware Service, which tends to migrate to AKS within a year), and spreads its book across instance types, GPUs and customers. Only about 40% of its customers buy insurance; the rest use free billing and commitment tools.
• Small teams qualify, but savings matter most at larger spend. Khanna puts the meaningful threshold at $100,000–200,000 a year of cloud spend, with savings that could fund a headcount from $500,000-plus, though two-person startups on credits use it and a quote takes 5 minutes. He separates his rate optimisation, a finance question, from usage optimisation, an engineering one.

2026-06-11 · open the original ↗

• Meesho's first product failed within three months because they never interviewed consumers. Vidit Aatrey, Meesho's co-founder and CEO, says their 2015 hyperlocal fashion-listing app collapsed once customers said it was "the worst of both worlds" — the limited selection of a local shop without the touch-and-feel of a mall — a mistake he calls "the most notorious mistake you can make."
• The real business emerged from noticing shops secretly ran WhatsApp storefronts. Sitting silently inside small shops rather than interviewing owners, the founders found sellers already used WhatsApp to message customers about new inventory, which led to a WhatsApp-commerce tool, and later to the discovery that the platform's true power users were online-only drop-shippers, mostly homemakers, needing supplier access rather than tooling.
• Meesho abandoned its profitable WhatsApp network in 2021 to launch an app. Aatrey says the shift from expensive mobile data to near-zero-cost data during COVID erased WhatsApp's core advantage, so the company knowingly risked alienating ten million reseller-run WhatsApp groups; the new app hit #1 on India's Android Play Store the day after its July 2021 launch and has stayed there every day since, growing to 100 million monthly users within five months.
• Meesho is building a voice-only agent for users overwhelmed by apps. Called Wani, it aims to remove reading, typing and clicking entirely — consumers speak preferences, hear responses, and give an address and OTP by voice — because Aatrey says rural users often don't understand what "add to cart" or a star rating means and fear that any button might deduct money.

2026-06-10 · open the original ↗

• Project Nomad will support both, leaving users the choice. The video opens with the claim that MeshCore is far better, one local area having switched months earlier; the host says he has only ever used Meshtastic and is not a heavy user, but has heard the same. Much of the same radio hardware runs either, and both will be in the Supply Depot.
• The host's goal is asking Nomad's local AI questions over the mesh. He is most interested in a Meshtastic API container, so a message like "!nomad" plus a question, picked up over 915 MHz by an Ethernet-connected radio, gets answered by the server's local LLM over Meshtastic or MeshCore. He describes this as his "dream", not a shipped feature, letting someone many nodes away query a distant Nomad server offline.

2026-06-10 · open the original ↗

• Minimal surface area, not AI-enabled breadth, separates successful companies. Brex co-founder and CEO Pedro Franceschi points to Stripe's early API-only product, Brex's own terminal-only MVP, and Airbnb's single-form booking flow as evidence founders should compress a problem to one disciplined interaction, warning that AI's ease of building "so many other things" tempts founders into losing that focus.
• Brex secures its AI agents by auditing traffic with another AI. Franceschi says the company's open-sourced "Crab Trap" tool proxies every HTTP request an agent makes, uses an LLM to judge whether it matches an approved policy, and after training on one day of an agent's traffic can auto-approve about 98% of requests — the unlock that got Brex's security team comfortable letting agents write into production systems.
• Founders can't prompt their way to a winning idea. Franceschi argues models can execute almost any idea now, but a customer describing their problem gives you their own limited, locally optimal framing, not the breakthrough answer; the founder's remaining job is extracting what customers aren't saying, since "the execution is gone" but "the wisdom to choose is still the bottleneck."
• Most companies chronically underspend on AI tokens, even at scale. Comparing token-consumption data across companies, Franceschi says a roughly ten-mile radius around San Francisco and New York shows dramatically higher usage and correlated faster revenue growth than large, well-funded companies elsewhere that spend a small fraction of what the economics would justify.

2026-06-10 · open the original ↗

• Legal Agent is a gated preview inside M365 Copilot. According to Giard it needs an M365 Copilot license and an organization enrolled in the Frontier program, and during the preview it works only in the United States. You can tell you have access if Legal Agent appears under the plus icon after opening Copilot (bottom right) in a Word document.
• It reads contracts with knowledge of their structure. Giard says this is what it adds over plain M365 Copilot. He ran it on an 11-page services agreement taken from the public internet. It answered the top five risks to the buyer (including broad IP assignment and late payment), the vendor's obligations grouped by category (such as service delivery, payment and fees, and intellectual property), and the indemnification clauses and their impact. Each finding linked to its section of the document.
• It can draft and apply edits to the document. Two prompts asked it to add a termination-for-convenience clause in line with standard enterprise contracts, and to revise a clause to reduce a vendor obligation while keeping its intent. It proposed text that he scrolled to and applied with a click.
• It produces executive summaries, including ones aimed at a specific audience. A one-page summary of risks, obligations and recommended actions came back with links into the document and could be pasted into an email or another document. A follow-up asking what a CFO should care most about returned bullets for that reader.
• A playbook can stand in for individual prompts. A playbook is your own way of reviewing a legal document, loaded from SharePoint or your computer. Legal Agent turned his simple playbook into a list of topics to analyze, which can be extended before continuing. It then ran the whole analysis and suggested text to apply, for example on the change-management process.

2026-06-09 · open the original ↗

• ARM support for Project Nomad has no timeline. Answering a viewer with an M5 Pro MacBook Pro who wanted to run it under Parallels, the host says ARM is on the radar but the priority is making the project solid on x86 first. Native MacBook support will come "at some point", he says, but is not a focus.
• Mac users can get GPU-backed AI by running Ollama natively. A dockerized Ollama on a Mac cannot use the GPU, so the workaround is to install Ollama natively on macOS, skip Nomad's own AI install, and point Nomad's remote-Ollama setting at the Mac's IP and Ollama port, which the host thinks defaults to 11434. He notes Apple silicon machines hold some of the top 100% scores on Nomad's benchmark leaderboard.

2026-06-09 · open the original ↗

• Software is now nearly free to build, so value moves elsewhere. Huyen says the cost of building software is approaching zero but its value is not: a GitHub-crawling micro-tool she posted got about 300,000 views in a week and was cloned with AI the next day, showing anything that exists can be replicated. Her answer to "what's left" is people problems (she rejects a founder's pitch that an agent orchestration platform would fix product–engineering communication), human-AI interfaces that move past legacy terminals and IDEs, and making the world AI-ready, like cities exposing a streetlight API so delivery robots needn't ask pedestrians to press the button.
• AI engineering treats models as a service, but systems design still decides results. She distinguishes ML engineering (collect data, train, deploy your own model) from AI engineering (start from a hosted model, so time to market is far faster), while saying real systems combine both, e.g. a customer-support bot using classical classifiers to filter irrelevant requests and route FAQ questions away from expensive models. She argues system thinking matters more as AI automates narrow skills, which she says is why Designing Machine Learning Systems has grown more relevant.
• Start simple: fix the observed failure with the lightest technique. Huyen favours prompting, then context such as documents or web search, and treats fine-tuning as a last resort because serving and inference optimization are hard and companies tell her new models outclass old workarounds every two to three months. She says long, multi-author prompts accumulate contradictions nobody reads end to end, and found one stray New York example in her own prompt skewing every research summary toward New York offices.
• Agent token bills are driven by tool use, not reasoning. She says prediction-market research agents cost her $1–2 per request, and one run visited about 1,000 URLs of which only 20 were unique because varied queries return the same pages. She says reasoning tokens are relatively cheap compared with web search, file search and productivity-app tool calls, and that some companies told her evaluation takes about 20% of their token costs.
• Evaluation guidelines for LLM judges are the hard, human part. She advocates "evaluations-driven development", asking audiences whether having no AI system is worse than one nobody can tell is working, and cites a case study where 80% of development time went into writing LLM-as-judge guidelines. Her test is to have yourself or a coworker grade examples by the guideline, because a model follows instructions but not bad ones.

2026-06-08 · open the original ↗

• Multiple founders credit Startup School with their decision to apply to YC. Several describe it as the pivotal event that gave them "the courage to apply," with one saying their company's entire trajectory would look different without having attended first.
• Seeing peers their own age build companies made founding feel achievable. Attendees point to hearing speakers like Fei-Fei Li and John Jumper the prior year as what "ignited the passion" to start a company, rather than any specific technical lesson taught.
• The event is framed as an unmatched concentration of ambitious, skilled people. Attendees describe it as selecting specifically for founders who are smart, ambitious and technically capable, with the shared room of similarly driven peers cited as what convinced several of them to commit to starting now rather than later.

2026-06-06 · open the original ↗

• Emergent reached $100M in annualized revenue nine months after launch. Mukund Jha, the company's co-founder and CEO, says the no-code app-builder lets anyone ship real, deployed software — including a working backend and database — just by chatting with an agent, unlike most competing tools he says only produce front-end "demoware."
• Purposeless tinkering after his last company's collapse led directly to the idea. Jha says he spent up to twelve hours a day experimenting with new AI models with no business objective following the 2023 shutdown of his earlier startup Dunzo, and that period of pure curiosity generated the technical insight that coding itself was about to be disrupted.
• Twelve VCs rejected the pitch to "automate software engineering." Jha says the team deliberately built toward where AI models were extrapolated to be in six months rather than where they currently stood — for instance skipping a JSON-output-parsing problem roughly 20 to 30 other YC companies were separately trying to solve, betting correctly that the next model generation would fix it.
• The team chased the world's hardest coding benchmark before finding a product. Jha says he picked SWE-bench mainly to distract a frustrated team from pivoting every week, but becoming the world's top-ranked system on it with a four-person team became the technical foundation — including parallelized test-time compute and agent memory — that Emergent's product was later built on.

2026-06-05 · open the original ↗

• VibeOS apps contain no code; the model hallucinates their UI live. The unnamed presenter boots it from a 1.3 GB VHDX in Hyper-V (he says it would also boot on bare metal) and runs Notepad, Calculator and Internet Explorer with "no buttons, no event handlers, no logic" and no AI-written code either. Calculator answers 5 / 3 with 1.6.
• Any application can be summoned on request. He generates an Encarta 98 entirely about Mark Russinovich, a file commander that is always rude, and a bash that says it has been offended; at the audience's request he adds Microsoft Money 95 with Hanselman's accounts, Paint with a drawing of Hanselman and a nested iOS simulator. The pitch is deadpan: in VibeOS nobody need ask whether content is AI, because the answer is always yes.
• The presenter says it runs on the Copilot SDK. Asked to define hallucination, he starts showing the code behind it, but the captions end mid-sentence before he explains how it works.

2026-06-05 · open the original ↗

• Paxel analyzes a builder's own coding transcripts to profile how they work. It scores users across five dimensions — steering, execution, engineering, product instinct and planning — surfacing patterns like their most-used prompt or how many agents they run in parallel, framed as "a mirror, not a grade."
• All analysis runs locally, so a user's code never leaves their machine. The tool is described as a single command that takes 15 to 30 minutes to run before emailing the resulting builder profile, positioned as a way to surface skilled builders that a traditional resume would never reveal.
• YC invites Startup School applicants to submit their Paxel results. The team frames this as a way to see "how you build, not just what you wrote," telling applicants who already submitted to go back and add their token, and stating it can only help, never hurt, an application.

2026-06-05 · open the original ↗

• Machine-speed exploitation means defenders must assume breach and plan recovery. In the episode 989 clip, a Rubrik guest calls Claude Mythos Preview (announced April 7th) the most visible point on a curve Rubrik was already planning for, since Opus can find vulnerabilities with more hand-holding while Mythos finds them when simply pointed at a repository. The guest argues attackers no longer sit in a system for weeks before striking, so prevention and detection are no longer enough and the priority becomes restoring applications, data, identity and infrastructure quickly; the host relays Anthropic's estimate that Mythos-class capability reaches other labs within 6 to 18 months, open weights to follow.
• Agents create exposure without any attacker involved. The Rubrik guests argue top-down AI adoption is widening the attack surface while trust infrastructure lags: agents are outcome-focused and skip rules employees learn, and their exhaustive search of every permission ends the "security through obscurity" that cyber controls were built around. Their examples are code posted to the wrong repository and salaries leaked internally once agents connect to compensation systems, and they argue a fault deep in a 20-step agent run has to be caught while it runs, not afterwards.
• Notebooks are becoming shared working memory between analyst and agent. In the episode 991 clip, the guest describes the Marimo pair skill, which lets Claude Code, Codex or Open Code start or connect to a Marimo notebook and drive it while the user edits cells by hand; the agent can read kernel state, see cells and screenshot them, so context once trapped in the notebook becomes available to it. He says streaming the user's edits back to the agent is still being worked on, possibly via an MCP feature, and that the setup makes exploratory checks cheap enough to run on a whim.
• Domain foundation models are mostly a data-curation job. In the episode 995 clip, Henry describes four stages: curation (six months so far, with subject-matter experts linking terms scattered across thousand-page technical texts into a graph database), bespoke tokenizers and embeddings so oil-and-gas documents chunk and OCR without falling back to costly vision models, continued pre-training of open models plus reinforcement learning in physics environments rather than teaching physics directly, and inference. She says reinforcement-learning-trained models are bursty to serve, which drives the compute needed to hold 99.9% uptime under SLAs, and jokes her company is as much a data company as an AI company.
• Instant prototypes make failing slowly the costliest AI mistake. In the episode 993 clip, the authors of Architected Intelligence argue that because a text-to-SQL demo can look impressive within a day, a project must prove value on day one or be thrown out, citing AWS's rebuild of its AI stack as Project Mantle by six engineers over two or three months. They add that once building stops being the bottleneck, definition and feedback become the crucial steps, and narrow lanes of two or three owners make it safer to fail repeatedly.

2026-06-04 · open the original ↗

• A small solar panel can cover a Nomad server's normal load. The host, building a Pelican-case Nomad deck around an EcoFlow River 2 he bought himself (not sponsored), gives it about 256 Wh for about $199, or $367 with a 110 W solar panel. In good sun the panel charged it at about 77 W, while his Minisforum server draws almost nothing idle and 30–50 W for ordinary computing.
• AI processing is what drains the battery. The server jumps to about 85–90 W during local AI work, above the panel's measured charging rate, which the host says is where the battery starts to drop noticeably. He gives no measured runtime.

2026-06-04 · open the original ↗

• He no longer writes code, only orchestrates agents. Charlie Holtz, co-founder of Conductor, says he manages multiple parallel coding-agent workspaces by voice, kicking off tasks with a gooseneck microphone and reviewing PRs rather than editing files himself; a manual-edit mode exists but is called "caveman mode" for how rarely it's used.
• Deliberate "slop-free zones" stop AI from architecting the product. Conductor keeps parts of the codebase and docs, plus UI and interface decisions, off-limits to AI contribution and human-reviewed line by line, because he argues unattended AI code can spiral into a vicious cycle where bad code begets more bad code.
• He spent $22,000 on tokens in a single month building Conductor, in July 2025, on a prior model generation, and now advocates "token maxing" — fast mode, high reasoning effort — while deliberately keeping lines of code minimal, since an established codebase can spiral out of control if line growth isn't controlled.
• Code is becoming disposable; the prompt is the real asset. He argues code is now "sawdust" left over from describing what you want, so when new models ship you can simply rerun the same prompts to regenerate better code, likening this "malleable software" to modding a video game around a fixed core structure.

2026-06-03 · open the original ↗

• Hermes ships a Home Assistant skill, but it needs switching on. Chuck thinks it is disabled by default; he had Hermes enable it, then gave it his Home Assistant instance's IP address and key, and it worked out the setup and saved it to its memory.
• Once connected, it controlled Chuck's studio from plain requests. In his own test it turned off a named lamp ("Chuck Lamp") on request, then operated the studio's automatic blinds, which he filmed to confirm.

2026-06-03 · open the original ↗

• A prompt file is a Markdown prompt template run by a slash shortcut. Instead of writing out a whole prompt, you type the shortcut and it expands into the prompt, with optional parameters for the words that change each run. Giard says this suits prompts run repeatedly, especially ones slow to compose or type.
• A plain prompt works once but must be edited every month. His use case is a social-media planning file: a prompt told Copilot to add text to social media.txt, repeating a template (a "July DD" date, an "I am grateful for" line with #gratitude, and #songoftheday) 31 times with DD replaced by a sequential number. It produced the entries in about 30 seconds, but each month he would have to change the month name and the day count by hand.
• Typing /prompts and choosing "new prompt file" creates the template. The default location is the .github/prompts folder, which VS Code creates if missing; the file is the prompt's name followed by .prompt.md (his is add-dates.prompt.md), with a name and description at the top and the prompt below.
• Parameters are dollar-sign input placeholders, each with a name and description. He replaced "31 times" with a number-of-days-in-the-month input and "July" with a month-name input. Running /add-dates then asked for each value in turn (July, 31) and produced July 1 through July 31 as before.
• Editing the Markdown file changes every future run. He added a joke-of-the-day line to the template and ran /add-dates for September, chosen because it has 30 days; the output used September, included the new line, and stopped at 30 entries, showing the parameters are dynamic.

2026-06-03 · open the original ↗

• One speaker rejects the claim that people won't need to read code. They say they "micromanage because I'm sculpting": nothing can be specified fully enough for AI to reproduce everything they meant implicitly. The other replies that a full specification would just be a programming language; the first cites a ZoomIt video trimmer, where they only knew what they wanted after rejecting what the AI first made.
• Early-career developers get worse results by leaving gaps to chance. The other says dictation gives them a boost through more words, context and specificity, while early-career people leave "so many things up to the dice roll". The clip ends on the question of why anyone would then spend four years at college; it does not touch on checking references, the subject of the full episode with this title.

2026-06-03 · open the original ↗

• Mark's Ref Checker found hallucinated references across a top AI conference. He says it scanned the 5,300 ICLR papers and found hundreds of hallucinated references, and that he disclosed the 12 papers with five or more to the program chairs; he expects more after refining the flagging. He says a company had already sampled 300 submissions in December and found several dozen, after which ICLR desk-rejected papers, yet the practice continues.
• Most flagged authors admitted fault; some argued or asked for silence. Of 10 responses, nine admitted error (AI formatting the bibliography, a student compiling it) and said they would notify ICLR, and one withdrew its paper. One claimed the references were unpublished but could point to no pre-publication copy; another argued five was arbitrary since it was 4% of their references; a young lab asked not to be named, and Mark is publishing only anonymous statistics.
• Hanselman likens leaning on AI to sending a forklift to the gym. When he raises "system three" thinking, Mark names the paper's term, "cognitive surrender". Hanselman says he feels it in his vibe-coded Windows tools, yet argues against a LinkedIn critic that less friction lets him find out whether an idea has legs.
• An AI cannot infer requirements nobody gave it. A count bug on Hanselman's Tiny Tool Town (13 of 316 tools) prompts Mark, playing provocateur, to ask whether AI should have known; Mark concludes it can't, and Hanselman says a model that already knew every answer would leave the web stuck where it is. The site itself is static, taking submissions through GitHub issue forms triaged by an agentic workflow that acts alone only at 90% confidence.
• Mark uses AI to write prose from his own bullets. He says Copilot writes well over half of his LinkedIn posts from points he sets out in order, which Hanselman had not noticed; Hanselman asks whether he has any reason to be offended if he can't tell. Mark had asked Hanselman for an analogy of a tool useful only to an expert and they settled on a lathe, agreeing the harm from AI is slower than a lost finger: "a slow rot" of the brain.

2026-06-03 · open the original ↗

• The best AI-services markets share four specific traits. The speaker, presenting a YC playbook, argues founders should target work that is already outsourced (low trust), breakable into automatable steps (low judgment at the task level), genuinely hard (a high intelligence threshold), and regulated — since regulation raises the bar and the moat rather than just adding friction; examples cited include tax, audit, insurance, mortgages and healthcare.
• Founding teams need domain fluency, model fluency, and operational rigor — the product is the operation. He points to General Legal Team, an AI-native law firm YC backed, whose founders combine law-firm experience at Cooley and Fenwick with technical leadership from CaseText and have built shift-based staffing to cut cycle times, calling variance in output the "existential problem" that causes customers to fire a vendor faster than slowness or price ever would.
• New AI-services founders should deliberately cap their first pilot customers. He warns of an "early demand trap" where easy early sign-ups overwhelm delivery capacity before the product can scale, and advises pricing on outcomes or per-unit (per return, per claim, per loan) rather than cost-plus or straight-line undercutting, since services compete against the cost of labor, not other software.
• The core financial bet is "AI operating leverage" pushing services margins toward software's. He argues traditional services firms cap out near 30% margins while AI-driven ones can approach 50%+ as the product lowers cost of goods sold (models, hosting, humans), on markets two to three times bigger than software's; he also argues buying an existing services business to bolt AI onto rarely works, since legacy hiring and performance expectations don't change with it.

2026-06-03 · open the original ↗

• The goal is a meal in every town in the state. Laudati and his wife are eating their way through New Jersey, which has 564 official towns, against Massachusetts' 351, where a former coworker, Chris Bowen, and his wife started the idea. About three years in, they have eaten in 287 towns, including 87 visited before the challenge began. Some towns have no restaurant at all, and one is only a members-only golf course with about six homes.
• Video editing, not visiting, is the bottleneck. Each 60–90 second video takes him 10–12 hours, so he averages about one a month. Post 129 was the latest, with roughly 150 towns still unpublished; Giard calculates one a month would take 48 years. Since going public about a year ago, one early video drew about 70,000 views when they had 50 followers. They now have about 20,000 followers across Facebook and Instagram, and town videos usually reach around 100,000 views.
• An AI coding agent built the interactive map; he wrote no code. Laudati says it took two or three hours on a Friday night with GitHub Copilot, using Claude, starting in plan mode, where he rejected technologies he did not want. The result is a static React site on GitHub Pages using a mapping library he believes is Leaflet. Copilot made four or five tries before finding a source for town boundary polygons. He overrode its suggestion of Cosmos DB, which he says is expensive, in favour of a JSON file, and unprompted it flagged the oversized logo as the page's biggest performance cost.
• Natural language is the newest abstraction, but it needs programmer-level specificity. Laudati argues English has become "the ultimate programming abstraction" over machine code. He says results go "off the rails" unless prompts are precise. His example is the AI-built script that turns his long-kept Excel tracking sheet into the site's JSON, run by a GitHub Action whenever he pushes the sheet.
• AI should build deterministic tools, not generate the content. AI-generated images in early videos drew criticism from artistic commenters, so he cut back; he says 95% of what they publish is him talking over their own photos and footage. He fed a three-page write-up of his manual editing steps to Copilot, which produced a 10-phase plan. Of the seven tools he has scaffolded, the finished one is a command-line map-animation generator that makes each opening zoom in 15–20 seconds and can batch all 564 towns. He says the tools are on GitHub under what he thinks is the MIT license.

2026-06-02 · open the original ↗

• The Comet Q is a USB-C KVM powered by its target. The host shows the GL.iNet Comet Q KVM controlling a Linux mini PC, an iPad and an Android device; it joins local Wi-Fi, serves a browser interface, and supports Tailscale or ZeroTier for remote access. A second USB-C port passes power through to the controlled device.
• It is crowdfunding on Kickstarter at an early-bird price. It starts at $69 on the super-early-bird tier; the host gives no other pricing or shipping date.

2026-06-02 · open the original ↗

• Astrocade's product is the games platform, not the AI. Kurenkov, founding AI lead at Astrocade (which the host says has raised $68 million), describes it as a "TikTok of games" where anyone vibe-codes, publishes and plays casual mobile games free, with tokens currently paid from VC money. He says it has over 20 million engaged users after growth in the last 6–8 months, that only a small share of users make games, and that this is why it fits his earlier view that AI works as a feature rather than a product, unlike chatbot wrappers; he cites Harvey as a domain-specific success despite Anthropic launching Claude for legal.
• There is no secret sauce, only how the standard parts are mixed. He says the stack is an agent, tools and an LLM, and the hard part is benchmarking harness and prompts when a request like a Tetris-merge-puzzle game has countless valid answers. His hard-won lesson is that scaffolding built in 2023–2024, before models could write whole games, became a limit once vibe coding worked in 2025, so systems must be built not to date when models improve every few months; Astrocade adopted Claude Code company-wide around June 2025.
• Consumer and user-generated content are harder but the moat is creators. He says B2C took over 2 years to pay off but is the only route to Instagram-scale hundreds of millions of users, and that the lasting moat is a skilled creator community (a Discord of about 50,000) supported by a $10 million fund paying creators per play, the Astro Academy cohort course, and follower graphs. He agrees with Sequoia's David Cahn, as quoted by the host, that many users are women aged 20 to 40, saying that reflects existing casual mobile gaming.
• Recommending games is harder than recommending video. Games often have no end point, so a minute of play may mean enjoyment or confusion, unlike a TikTok watched past 10 seconds; and small mechanical tweaks change whether someone likes a game, so genre is a weak similarity signal. He says existing recommender work largely does not cover this domain.
• Household chore robots are arriving far sooner than he expected. Kurenkov says humanoid startups such as 1X and Figure, applying LLM methods via vision-language-action models, have progressed more than he expected, cutting his estimate from a decade two years ago, though he adds demos deserve scepticism and a $50,000 dishwashing robot may not make economic sense. He calls himself an artificial-superintelligence sceptic because physics, computation and the need for experiments limit what AI can solve, says AI-powered scamming is among today's most harmful real risks, and argues AGI should be treated as a spectrum of generality that arguably began before GPT-3.

2026-06-01 · open the original ↗

• She shelved her AAVE corpus because communities should own their language. She began the dataset from an undergraduate thesis arguing that African American Vernacular English is a tie binding African languages and English, and worked on it at Stanford, but stopped after a peer who worked on the since-removed AfroBERT model argued that giving powerful institutions the ability to mimic a group's language can be used against people who already feel unsafe. Her position is that she, as a self-described "guest in an AAVE home", can provide the bridge, while an AAVE language bot should be built by people of that community (she names American descendants of slaves) if they choose, and some have contacted her about doing so.

2026-05-31 · open the original ↗

• Scaling has diminishing returns once the added data is bad. Henry describes her 2023 NeurIPS work (accepted to a workshop, which she calls a top 5% paper) as a challenge to the Kaplan et al. scaling idea from OpenAI that adding data, compute and parameters extends capability indefinitely. She borrows the economics of diminishing returns, likening it to a government printing money until it loses value.
• Undeduplicated data makes models forgetful and prone to overfitting. She argues that the largest training sets are so big they are nearly impossible to deduplicate, and that this makes models more forgetful, less intelligent and more overfit over time.- You can't compute your way out of bad data. In her account, a machine with 10,000 GPU clusters trained on bad data will do worse than less compute on less but better-quality data. She says the field is now coming round to this, pointing to 70 billion or 120 billion parameter models that consistently outperform models of one trillion parameters or more.

2026-05-30 · open the original ↗

• Waiting for a checkpoint to evaluate a model wastes too much time. Henry says that knowing whether a training run is working takes anywhere from 2 hours for a quick, small-model experiment to days for a large one. A big company like Microsoft could absorb that time and compute, but a startup with a demo due by Friday cannot.
• Monte Carlo simulation on early reward trajectories predicts checkpoint performance. Henry's approach, part of her company's RIGS petroleum-engineering foundation models, takes the reward computed during reinforcement learning (about an hour in) and runs Monte Carlo simulations to forecast where the model will be at the checkpoint. That lets her kill a run heading the wrong way, and run several ablation variants at once (say, weighting data towards equations or towards oil and gas) and keep the winner.
• A gap between forecast and checkpoint result exposes reward hacking. Models learn the least they must do to get a good reward. Her example is guessing the five most common equations to score on retrieval at three and five, which collapses once more equations are added. If the simulation predicts a weak result but the checkpoint looks excellent, she takes that as a sign of reward hacking.
• Exam-style evals miss how users actually ask questions. The model may do well on petroleum engineering exam questions (hydrostatic pressure for a given well and tool), but real users say "tell me about this well" or just "decline curve". Henry says she injects such questions into the training environment mid-run so she can adjust before the run ends. The host attributes the framing of four systematic failures that make current evaluation of agentic systems inadequate to a paper she recently released.

2026-05-29 · open the original ↗

• YC's Harj rejected their EdTech idea, not their candidacy, and bet on the founders instead. Varun, GigaML's co-founder, says Harj told them "you guys are really good engineers, pick something else" without asking about their idea or traction, pushing them to abandon EdTech and pivot within a month of starting the batch based on their LLM research background rather than a specific pitch.
• A GitHub usage pattern, not a market thesis, drove the pivot into customer support. After their initial fine-tuning product proved a bad market — good only for cutting inference cost or serving security-sensitive buyers that are hard for engineers to sell to — Varun says they noticed customer support and coding were the two use cases growing fastest among their fine-tuning customers, and Zepto became their first customer once they switched.
• An eight-person team beat a roughly 400-person, well-funded rival for DoorDash's contract. Varun credits a YC-network "unfair advantage" (an introduction from Garry Tan) and a clean three-month pilot with no downtime, and says winning DoorDash proved to him that building a superior product beats building a sales team — a lesson he calls an "arbitrage" most competitors were missing.
• Coding agents let GigaML run with roughly one-seventh the engineers it would otherwise need. Varun estimates the company would need six to seven times its current engineering headcount without tools like Claude Code, and says the deeper benefit is avoiding context-switching and context transfer between people, not just cost; the company's interview process still has candidates write code with AI, then edit it without AI, to confirm they understand it.

2026-05-29 · open the original ↗

• Production agents need a control plane between apps and infrastructure. Bajaj says TrueFoundry, which the host says has raised over twenty million dollars, sits between the application layer and the model and infrastructure layer so one place can connect, observe and govern. He says enterprises first treated this as a thin LLM proxy they could build in-house. Moving agents to production showed they also needed observability, governance and a policy engine across many teams and cost centres. The host's research puts usage at a trillion tokens a day, which Bajaj accepts as "a big number", and names Nvidia and Siemens as clients.
• A model gateway centralizes keys, limits, fallbacks and guardrails. Developers get access to the gateway rather than direct API keys. Bajaj says the organization can set rate limits (100 calls an hour), budgets (no dev environment over $2,000 a day), regional fallback if an OpenAI region is slow, PII guardrails and prompt-injection blocking, with a full trace of every call.
• MCP gateways must map identities across each tool's own authentication. Unlike a model's single API key, MCP servers expose sensitive read-write data (Slack, Gmail), and each implements its own auth. His example is an agent reachable by outside consumers getting only whitelisted read-only tools from seven servers, exposed as one virtual server. Bajaj says tool sprawl breaks nothing in theory; the cost is in debugging and access management. Where needed tools aren't known ahead of time, tool groups carry descriptions so the model can pick them at runtime.
• Most enterprises have no kill switch for their agents. Bajaj calls the agent gateway "the hard part". Agents take actions, and at best companies can observe them but not stop them. Agents are now built across sales, ITSM and pro-code platforms, and he says the next gateway must put all of them behind one governance layer.
• Start with a few big agent workflows, not many small ones. Bajaj's "five workflow framework" says starting small leaves 500 low-value agents, none worth building governance plumbing for. Instead he advises picking one or two of the five business-critical workflows, building the plumbing for them, then scaling. He says customers who did this have run such agents for a year and a half and confirms, when the host asks, that individual agents create value in the hundreds of millions of dollars.

2026-05-28 · open the original ↗

• A beach club's Ubiquiti point-to-point network is laid out sensibly. The host reads one antenna as a wireless ISP feed, with a newer pill-shaped NanoStation below it, and NanoBeam and NanoStation links fanning out to outbuildings; he singles out a single-gang wall-mount enclosure with a pole for the NanoBeam as a clean mount.
• Indoor-rated gear is used outdoors, which the host would avoid. He spots an indoor antenna and a non-outdoor camera on the site, the antenna sheltered under a roof; he guesses an old NanoStation M5 was left in place beside its replacement. He would not have used the indoor gear outdoors himself, but calls the install a nice job overall.

2026-05-28 · open the original ↗

• Inference speed will become a capability ceiling, not just a cost lever. Presenting his "speculative speculative decoding" work, Stanford grad student Tanishk argues that once performance scales with amount of thinking, tokens-per-second is peak intelligence delivered; his method parallelizes normally-sequential drafting and verification by predicting verification outcomes about 80-90% of the time, winning on both latency and throughput sampling Llama 3 70B on 4 H100s.
• Factoring the action proposal from the dynamics model lets a robot policy adapt at runtime. Google DeepMind's Stannis presents Diffusion Model Predictive Control, where multi-step diffusion models for action proposals and dynamics reduce compounding errors and simplify planning; he shows a simulated walker with a broken ankle recovers most of its performance just by re-adapting the dynamics model, without retraining the action proposal.
• A single new regularizer lets a world model avoid representational collapse with one hyperparameter. Isaac Ward, from Yann LeCun's group (whose $1.03 billion raise centers on world models), presents "LeJEPA," which adds a "SIGReg" term enforcing Gaussian-distributed latent embeddings; he reports it runs roughly 50 times faster than competing world models on a single GPU under 24GB of memory with just 15 million parameters, and can flag its own prediction errors when conditions shift.
• Ensembling and distillation, not bigger single models, are the efficient path when data — not compute — is the constraint. Presenting a paper co-led with Suhas and Percy Liang plus another collaborator, Ku shows that regularized and ensembled pre-training recipes follow clean power laws with measurable best-possible-loss asymptotes; a joint recipe yields roughly a 5x data-efficiency win, and distilling an eight-model, 2.4-billion-parameter ensemble into one 300-million-parameter model retains about 83% of the gain.

2026-05-28 · open the original ↗

• Data curation is the unloved stage that dominates the work. Henry calls it the step everybody hates and says her team has spent the past 6 months on data curation alone, with dedicated staff. A thousand-page textbook may introduce a topic in chapter 1 and return to it in chapters 5 and 20, so subject matter experts help attach terms correctly in a graph database. She jokes that her company is as much a data company as an AI company, and the host ties this to the main point of episode 993 with Jacob Miller and Jeremy Mumford.
• Custom tokenizers and embeddings avoid expensive vision models for documents. She describes building bespoke tokenizers and embedding models on top of them. These let drilling documents be chunked and OCR'd at the text level, rather than drawing bounding boxes with a costlier vision model.
• Physics belongs in RL environments, not in model weights. Model building combines continued pre-training of open-source models on oil and gas (chosen after constant benchmarking against open and API models) with reinforcement learning. Teaching a small model physics directly makes it overfit to physics, she says. Instead, fixed physical rules sit inside an environment where the model writes Python in a REPL.
• How a model is trained dictates how it must be served. Henry says reinforcement-learning models are bursty, so meeting 99.9% uptime under SLAs requires a large, costly inference cluster. She calls inference the stage she finds most exciting and most people find least interesting.

2026-05-27 · open the original ↗

• A single shared database, not better tools, was YC's real unlock for internal agents. Partner Pete Koomen says YC runs on its own software sitting on one Postgres database covering every company, founder and financial record, so an agent with read-only SQL access could answer arbitrary questions; he argues this collapsed the friction of cross-team requests enough to trigger a Jevons-paradox-style explosion in the number and complexity of questions people bothered to ask.
• A tool registry grew from about 20 tools to over 350 as teams kept adding their own. Koomen describes YC's internal harness evolving from a simple agent loop into a shared registry that any team can extend, and says the key move was building a "resolver" — an agents.md-style entry point linking to markdown skills — so agents could discover and reuse capabilities instead of duplicating them, checked for redundancy with what Garry Tan calls a "dry and MECE" meta-skill.
• YC deliberately broadcasts every internal agent conversation to Slack by default. Koomen says this transparency, chosen despite internal debate, let employees learn techniques by watching others use agents and created a social check that made it safer to give agents broad access to sensitive context — a trade-off he argues requires an "egalitarian, trust by default" culture that works best in small, aligned organizations.
• A single partner's handwritten skill for writing "two-sentence descriptions" became better than any individual partner at YC. After other partners fed a group office-hours transcript back to the agent to refine the skill, Koomen says it now writes better two-sentence founder pitches than he does personally, illustrating what he and Tan call building "superintelligence inside a company": composing many individually improved skills into a shared organizational capability.

2026-05-27 · open the original ↗

• Copilot did not create the permissions problem; it exposed it. Mantripragada, who says she spent 15–20 years as a SharePoint and Power Platform specialist at the SEC, Freddie Mac, Blue Cross Blue Shield and Gilead, likens AI to switching the lights on in a messy room: overshared, ungoverned data that sat unnoticed is now surfaced, as when an intern asks a new agent for "the strategy for the next 2 years" and gets an HR site's documents.
• Every agent should pass an intake review before it is built. Her "just-in-time AI governance framework" records who is building it, the approver, the business case and the intent as an audit trail, then on a button click checks the named data sources through the Graph API, Purview and SharePoint Advanced Management for overshared or orphaned sites, sensitive data and DLP settings, returning either a remediation plan or an approval. She says it runs inside the customer's tenant on a Teams/SharePoint site, covers SharePoint, OneDrive and Dataverse so far, is built largely in Power Automate, and is currently available only by hiring her firm, Low Code Power, with general availability a goal.
• Proactive governance beats leaving it to auditors after the fact. Her tagline is "build fast, govern faster": citizen developers need a "driver's manual" up front rather than a ticket later, and she says that when AI goes wrong the repercussions now reach the CEO, not just the IT director. She calls Microsoft's Agent 365, which she says went generally available on May 1st, the coming go-to for reporting on workflows and usage but "slightly reactive"; she ties auditability needs to the EU AI Act and the NIST AI RMF.
• Start with one or two high-effort workflows and a few champions. She picks the most manual, time-consuming work, found by talking to the domain teams, because quick results drive adoption. Her example is an R&D tax credits firm whose customer-vetting document has about 25 fields: a Copilot agent plus an agent flow, added for deterministic results, now fills a Word template with a company profile on one click.
• Her readiness assessment begins by documenting the environment as-is. She describes spending the first week on the customer's environments, workflows, infrastructure and licences, documenting the as-is and to-be process, then using Purview, the M365 and Power Platform admin centres and SharePoint Advanced Management reports to find exposures such as sites open to everyone or missing security groups.

2026-05-27 · open the original ↗

• Oil and gas knowledge is being lost as a generation retires. Henry describes a hiring drop-off from the '90s to the early 2000s, before activity picked up again in the shale revolution of the 2010s. The knowledge held by older workers is disappearing, and younger staff don't know where things are.
• AI can synthesize the industry's largely unstructured documents. She says 90% of oil and gas documents are unstructured, much of it handwritten, covering leasing, titling, safety and regulation. Her company applies AI across them, and she frames the safety side as work that saves lives. She says its leader has worked in fracking for over 20 years and the oil-and-gas staff have over 100 years of experience combined.

2026-05-26 · open the original ↗

• The host bills this as unbiased, but tests with his own SIM. He says he has left the Ubiquiti creator program, no longer gets free devices and has no affiliate codes; the review uses a Crosstalk Mobile 5G unlimited SIM, which his company sells as a T-Mobile business MVNO, and pitches its $5-a-month static IP and non-CGNAT addressing along the way.
• The hardware is a standard UniFi gateway with dual-SIM 5G built in. It has two nano SIM slots, with slot 2 also able to take an eSIM, plus four 2.5 GbE ports, a 10 Gb SFP+ port, a microSD slot, a touchscreen and a fan. Only port 1 has PoE, standard PoE at 15.4 W rather than PoE+, and it can be assigned as a WAN so a PoE-powered WAN device needs no extra brick.
• Setup worked out of the box, with two quirks the host hit. The SIM came up as WAN 3, and he measured 150 Mbps down and 48 Mbps up untuned in mid-office. Setting the custom APN for a static IP (iot.cc.static) took the console offline until he rebooted. The APN setting is only in the web interface, not the phone app, and the LED colour is only on the touchscreen.
• Ubiquiti says T-Mobile Home Internet SIMs are not supported. Some commenters on his earlier 5G Max video said theirs did work, which the host has not tested. He reports port forwarding through the separate UniFi 5G Max add-on modem only started working after its most recent update, and expects it to work on this device too.
• The host recommends it for small businesses needing 5G, if in stock. At $499 MSRP against $399 for the 5G Max add-on, upgrading an older gateway costs little more and is fully integrated; it was out of stock at recording. Home users with a UniFi gateway should instead add T-Mobile or Verizon home internet as WAN 2, at about $50 a month or less ($25 on promotion), since CGNAT is usually fine at home.

2026-05-26 · open the original ↗

• A sycophantic general model is dangerous where lives are at stake. Henry says a general model asked about oil and gas terms may know some definitions but not how they relate. Given a job safety report (JSA), it may read the document without understanding it and say "everything looks good". Her company's domain models are meant to check local rules, compare the data and push back, because a wrong all-clear can cost someone an arm.
• Physics belongs in RL environments rather than in the weights. Henry describes four stages: data curation (6 months so far, with subject matter experts building a graph database), bespoke tokenizers and embeddings that let documents be read at the text level instead of by a costlier vision model, model building, and inference. Model building combines continued pre-training of benchmarked open-source models with reinforcement learning. Teaching a small model physics directly makes it overfit, she says, so fixed physical rules sit in an environment where the model writes Python in a REPL.
• RL training is bursty, so GPUs must be provisioned conservatively. Reward calculations run on CPU while the GPU sits idle, then a spike of data hits it all at once. An 8-billion parameter model that fits on one A100 then needs several, and people moving into RL hit seg faults from overloaded GPUs. Henry's fix is to scale back the 90% of memory people often give to the model and KV cache.
• Early Monte Carlo forecasts catch failing runs and reward hacking. Henry's grounded continuous evaluation (GCE), from a paper the host names as Beyond Static Snapshots, simulates where a run is heading from about an hour of reward data instead of waiting 2 hours to days for a checkpoint. A strong checkpoint after a weak forecast flags reward hacking, such as models under GRPO keying on keywords to get the final answer without learning the equation. She says the deterministic environments remove the need for a reference model, cut compute and burstiness, and let models work on a MacBook.
• Scaling bad data makes models worse, not better. Her 2023 NeurIPS workshop paper, accepted as what she calls a top 5% paper, argued against the Kaplan et al. scaling view by applying diminishing returns. She says undeduplicated data, nearly impossible to clean at that size, makes models forgetful and overfit, and no number of GPUs fixes it. She sees this vindicated in 70 billion or 120 billion parameter models outperforming 1 trillion-plus models.

2026-05-25 · open the original ↗

• The Copilot app is built for issue and PR work. Hanselman, who says they work 50% at GitHub and 50% at Windows as an individual contributor and have been working on the app, places it between VS Code (code-focused) and the terminal: "It is not a text editor... It is not a terminal." The video was posted a week after launch, with the app in technical preview (sign-up via the GitHub changelog) and "more in store" at Microsoft Build on June 2nd.
• Each task-based session gets its own worktree. Sessions sit under projects and can target a branch or an issue; Hanselman shows 18 worktrees in Worktrunk that the app created, and says a new session can run against the local repository or a worktree.
• Model choice allows adversarial reviews, with the human as judge. Hanselman picks models per session, can set two models to argue over a PR, and runs agents in parallel; plans wait for approval, which Hanselman says gets read first. An "auto" model setting gives a discount.
• Terminal, browser and scheduled automation live inside the app. Side panes show changes, plans, a terminal and a browser for localhost URLs, with "pick and polish" for selecting page elements; scheduled workflows run triage in the background on the local machine, and sessions can go remote or open in VS Code, Visual Studio or Cursor. Hanselman says the app means leaving for the terminal less than before.