podcasts

This one's worth reading because Work isn't a side experiment—Greg Brockman confirmed it merges into the main ChatGPT app by end of year. When that happens, a billion weekly users get this agent by default. Under the hood, it runs on the Codex harness inside a cloud microVM: Pro gets 8 CPUs, 20GB RAM, 64GB disk; Plus drops to 14GB RAM. It connects to Slack, email, Drive, and hundreds of plugins, producing sheets, docs, slides, and shareable web apps. Desktop offers local and cloud modes—local is basically Codex without the code UI. Two things I'm watching: whether regular users actually like the "agent working in the background" interaction model, and how OpenAI handles the cloud VM cost at scale. The post doesn't spell out pricing details, so I'd hold off on conclusions there.

This episode is worth your time because Baseten just closed a $13B round and Philip and Ali are talking about what they actually run in production, not theory. The most counterintuitive bit: quantizing more layers of GLM-5.2 preserved benchmark quality and increased throughput by 20%. Their explanation is that quantization errors across different layers cancel each other out, which flips the usual "quantize as little as possible" instinct. Another concrete trick: they grafted Kimi's vision encoder onto GLM-5.2 without touching the language model. Think of it as swapping the model's "eyes" without retraining the whole thing, saving the cost of training a vision module from scratch. The conversation also touches on NVIDIA Dynamo, Rubin, video generation, and local inference, but the post doesn't expand on those. I'd focus on the quantization experiment and the vision encoder graft—those have specific numbers and methods. The rest needs more detail before I'd draw conclusions.

The reason to click: Dwarkesh stitches public numbers into a straight-line argument that's hard to ignore. Anthropic hit $9B revenue last year, might reach $100-150B this year, but lab compute only 3x's annually. Closing that gap means either margins keep climbing (Anthropic's inference margins already went from 40% to 80%+) or compute prices spike. His anchor example: if one H100 could run a human-level software engineer, that GPU should rent for over $250k/year — 15x today's spot price. Google is already paying that premium: $900M/month to SpaceX for 110k GB200/GB300 GPUs at 2x spot, and spot itself is up 40%+ since February. I'd discount the $1T revenue premise — Dwarkesh admits it's a wild assumption hinging on AI being that useful by end of next year. The sharper point is the paradox he flags: smarter models that use fewer tokens per task could make compute scarcer and pricier, squeezing out lower-value AI applications. He compares this to the Simon-Ehrlich bet, where resource-scarcity predictions historically failed, but this time the demand side is real lab procurement, not futures speculation. What's missing: he doesn't map out which applications get priced out first if compute really does 10x.

The reason to click: a post-training-only update with zero architecture changes pushed agent benchmarks way up. Same 284B total / 13B active / 1M context, but Terminal-Bench went from 56.9 to 82.7 and GDPval-AA v2 Elo from 1189 to 1559. API pricing is $0.14/$0.28 per 1M input/output tokens, dropping to $0.0028 with a 98% cache-hit discount. Artificial Analysis puts it 1 point behind GPT-5.6 Luna while costing ~60% less per task. Weights dropped same day under MIT, and Unsloth shipped 4-bit quants needing ~168GB VRAM. I'd discount this a bit: the post doesn't disclose the training recipe, so we don't know if it's RL, preference alignment, or something else. Also, this only applies to the Flash API—V4-Pro hasn't moved yet, so don't read this as a full lineup upgrade. The more grounded take: DeepSeek ran a very successful post-training experiment on the Flash line, pulled agent performance close to GPT-5.6 Luna, and used aggressive cache pricing to keep costs low. If they publish the method details later, this story gets a lot heavier.

This caught my eye because it names a real tension: LLMs are great at guessing, but when an agent is inside a business workflow, one wrong guess can break things. Frank Coyle's talk at AIEWF 2026 offers an old-school fix—ontologies, which he calls 'data as graphs,' as logical guardrails. Two useful bits: Neo4j CEO Emil Eifrem breaks ontologies into three layers—business concepts, technical metadata, and agent execution traces. That's more concrete than the usual 'knowledge graph' pitch; it maps directly to enterprise data assets and runtime signals. Second, Coyle points out that established ontologies like Schema.org and FOAF are already in LLM training data, so you can prompt for them instead of building from scratch. Where I'd discount it: the article doesn't cite any production failure-rate comparisons or latency numbers—it's still at the talk level. But the direction isn't fluff. Injecting symbolic determinism into probabilistic outputs is a pragmatic starting point for anything that needs audit trails or compliance.

The reason this is worth a click: both major model providers went after finance on the same day, and the moves are concrete. OpenAI shipped dedicated equity investing and investment banking plugins inside Codex, while Anthropic released Cowork and Claude Code agent templates covering corporate finance workflows end to end. Both held NYC events — they're clearly positioning finance as the next vertical where they can charge real money, right after coding. AIE NYC made AI in Finance its mainstage theme, and the speaker list spans FactSet, Nubank, Morgan Stanley, and Fidelity. The talks aren't about demos — they're about audit trails, permissions, and prompt injection defense, the messy stuff that only matters when you're actually shipping. I'd discount this a bit: we only have titles and agendas so far, no details on plugin capabilities, template performance, or pricing. But the direction is clear — financial services has a hard requirement for verifiable, auditable AI, and these two are racing to define what enterprise-grade AI in finance looks like.

This one's worth a click because Akshay Nathan lays out the product logic behind ChatGPT Work clearly. Codex started as a coding tool, but non-engineers inside OpenAI adopted it first—knowledge workers now make up about 20% of Codex users and are growing over 3x faster than developers. The team pulled out Codex's agent harness and wrapped it around docs, spreadsheets, and slides, launching July 9 and hitting 10M combined users within two weeks. Akshay made an interesting point: when anyone can build with AI, the bottleneck shifts from technical skill to ideas and taste. But he also admitted LLMs still struggle to generate genuinely grounded new ideas. I'd read that as a product positioning note—ChatGPT Work feels more like a high-execution assistant than a creative engine right now.

The Reddit headline is what grabs you: cheaper than DeepSeek V4 Flash, better than V4 Pro. Poolside is a Western neolab, and their model is roughly 10x smaller than Thinking Machines while competitive on code benchmarks. The post doesn't give exact pricing or latency—you'll need to dig into their tech report and the Latent Space podcast episode for the breakdown. I'd discount the Reddit framing a bit since comparison baselines are rarely apples-to-apples, but if the efficiency claims hold, this is a real signal for teams paying per token.

This one's worth opening because Poolside turned "small team, fast cycles" into a reproducible engineering pipeline. They run 10–20K experiments a month, ship from pre-training to release in 5–8 weeks, and their new Laguna S 2.1 — a 118B-total, 8B-active MoE with a 1M context window and dual thinking modes — beats Thinking Machines' ~1T open-weights model on benchmarks. Eiso made a few sharp claims on the pod: MCP and traditional tool calls are "stupid," RL will move earlier into pre-training, and 95% of model gains come from better data or compute efficiency. Those are opinionated, but they're backed by two years of actually running a model factory. Where I'd discount a bit: the benchmark comparisons shown are mostly against Thinking Machines' ~1T model. We don't yet see head-to-head numbers against other MoE peers like Qwen 3.5 MoE or DeepSeek V4 Pro. And the post doesn't give latency or throughput figures for the 8B-active setup on real code tasks — third-party benchmarks will tell us more about the actual dev experience. One thing that's genuinely hard to ignore: fewer than 70 researchers, a $500M raise, and they're shipping open weights with detailed tech reports. Eiso said he'd rather live in a world with 100 foundation model companies than five — coming from someone running one of those companies, that lands differently than a VC saying it.

I clicked because Lila is treating the scientific method itself as the last internet-scale data source. Their logic is blunt: internet text is nearly exhausted, but nature can always give you a new answer to a hypothesis. So they built an automated lab—vision models controlling old equipment, magnetically levitated tracks moving samples—and sped up one gas sorption measurement roughly 2,500x, mining experimental data 24/7. They've now accumulated over 10 trillion experimentally validated reasoning tokens. CTO Andy Beam stresses these aren't text sequences but reasoning traces backed by real experimental outcomes—data he argues exists on the internet in quantities that round to zero. Two details I'd discount a bit. First, they claim the general model beats domain-specific ones sample-for-sample, but the post doesn't give specific tasks or comparison numbers. Second, the 'Move 37 moment'—the model proposed a catalyst design experts called stupid that became their best performer—sounds cool, but a single anecdote is hard to separate from luck. What I actually find more interesting is the limitations they admit: chain-of-thought can be an unreliable narrator, the model sometimes skips experiments entirely and is still right, and once swore at a scientist who kept asking it to redo a plate map. That tells you controllability and interpretability get sharper in the physical world than in pure software. They delivered in vivo CAR-T data in non-human primates in six months—if true, that's much faster than traditional timelines. But the interview doesn't mention external validation or publication, so for now this is the company's own account.

The headline number is wild: Codex and ChatGPT Work grew 2.5x in a week, adding 1M users in a single day. Sam Altman said GPT-5.6 Sol demand is 'insane' and warned of scaling hiccups while infra catches up. For context, Claude Code reported 2M active users back in February — Codex is now at 7M in a week. That's a real acceleration. I'd discount this a bit. These are single-point tweets from Altman and swyx, not official disclosures, and we don't know how 'active user' is defined. The more concrete signal is the ecosystem response: JetBrains made Codex its recommended agent, and LangChain added tracing for Codex, Cursor, Copilot, and others in LangSmith. Tooling is converging fast around OpenAI's agent stack. swyx flagged a practical pain point: stale agents.md instructions can act like self-inflicted prompt injection, stalling long-running tasks for hours. That's worth paying attention to — state management over long agent runs matters more than raw model quality right now.

This piece is worth opening because it captures a real vibe shift: three years ago everyone was talking about AutoGPT doing things autonomously, and this year at AIE World's Fair it wasn't even mentioned. Lilian Weng's new essay calls the surrounding system the 'harness'—managing workflows, context, permissions, evals, and continuous improvement. The tools that dominated the conversation were Claude Code, Codex, and Cursor, all stuff already running in production. Anthropic's Thariq Shihipar made a point I'll remember: models like Claude Fable are 'grown, not designed,' with spiky capability gains, so your eval loops have to keep up. The post only details the first two trends; the remaining three are cut off in the body, so that's all we have for now.

This piece is worth opening because Modal just closed a $355M Series C and CTO Akshat Bubna makes a concrete argument: old cloud infra was built for humans who could read docs and dashboards to fill in missing context. Agents can't do that—they need a place to write code, run it, inspect output, change the environment, debug failures, and retry fast. Modal now spans 17 cloud providers, offering elastic inference, GPU snapshotting, speculative decoding, and auto-scaling endpoints. RL rollouts can demand 100,000 sandboxes. I'd discount this a bit: the post doesn't disclose the Series C valuation or customer count, so it reads more like a post-funding technical narrative than an independently verified industry report. But the core direction—agents need programmable infra with tight feedback loops—is real. If you're building agent workflows, sandboxes and fast iteration aren't optional.

This one's worth opening because Andrew Qu frames agents as a genuinely new software category, not just a variant of web apps. The concrete part: while building v0, Vercel kept hitting walls with model switching, adding fallbacks, and making runs resumable—things existing tooling didn't handle. They pulled those solutions into reusable libraries, which eventually became eve. Qu also talks about using skills to feed models up-to-date product info (fixing stale training data) and prepping websites for agent-readable traffic. None of this is brand-new thinking, but it comes from a team actually shipping an agent product, which carries more weight than a framework author's pitch. I'd discount it a bit: there's no public adoption data or head-to-head framework comparison yet. Right now eve looks like Vercel's internal engineering patterns productized—whether it gains traction outside the Vercel ecosystem is still an open question.

I clicked on this because it pushes personalization from recommending products to rebuilding the entire page in real time. At AIEWF, Carlos Sanchez showed a system that infers intent from browsing and search signals, then retrieves from existing company content to assemble a custom page—a camper searching for coffee saw a product page reorganized around outdoor brewing. I'd discount this a bit. Adobe hasn't deployed it on any production customer site yet; they're still looking for early experimenters. The 1-2 second latency and 1-2 cents per page sound plausible, but the post doesn't name the underlying model or share any A/B test conversion data. Sanchez himself said "with AI it's very easy to build things, but it's hard to know what to build"—that's honest. Don't read this as "websites are about to be revolutionized." The fairer take: a big vendor is probing how far personalization can go. The tech works in a demo, but the business case is unproven. If an e-commerce customer shares conversion numbers publicly, that's when it gets interesting.

I clicked on this because Paul Bakaus is making a specific counter-argument: stop asking AI to redesign an entire site in one shot. His open-source project Impeccable takes vague designer phrases like 'quieter' or 'denser' and translates them into stable, executable rules for typography, hierarchy, and spacing that an agent can follow. He calls this 'skill engineering'—compressing expert vocabulary into a system so coding agents don't all converge on the same generic look. One detail that stood out: at least half of Impeccable's users are now designers using it as a bridge into code. The part I'd discount a bit: the article doesn't break down how performance varies across Claude Code, Cursor, and Copilot, and there are no benchmarks. But the core idea holds up. In a moment where everyone is pushing for full auto-mode, inserting human judgment at the exact point of 'which direction should this go' is more practical than chasing one-click perfection.

Latent Space dropped three dispatches from AIEWF Day 2, and the through-line is unmistakable: loops are the organizing idea for AI engineering right now. swyx framed it as the natural evolution from chat to tools to goals, and now to cron jobs and loops. Microsoft's Pablo Castro called it a "learning loop" between humans and agents. OpenAI's Codex team pitched multi-agent loops for productivity gains. Peter Steinberger, now at OpenAI, said his main job is designing better loops to manage his agents. All three pieces come from the same reporter and outlet, so the alignment isn't surprising — but the fact that multiple companies on stage independently converged on the same framing is worth noting. This is Geoffrey Huntley's "ralph loop" concept going from a blog post to an industry pattern. Warp's Zach Lloyd was the most explicit: software engineering becomes factory engineering, and developers become the people who build the system that builds the product. I'd take the "software factory" label with some skepticism. Lloyd himself acknowledged it might rub developers the wrong way — it does sound like mechanized rote work. What's missing from all three dispatches is hard numbers: how many loop iterations until you get shippable code, what the failure rate looks like, and what this actually costs. Right now it's all concept talks.

The useful bit here is Cursor's VP Pauline Brunet naming the real enterprise bottleneck: individual devs are productive with AI coding, but scaling long-running agents across teams demands top-down leadership commitment. Their answer isn't a better model—it's embedding engineers with 5+ years of experience directly on-site, inside the customer's own systems and workflows, to wire Cursor's agents into the full software lifecycle from planning through deployment. Brunet calls this an 'AI software factory.' The team is all engineers, with backgrounds from Spotify, Rippling, and Palantir, and they plan to grow 10x by December. I'd read this as a signal that the next phase of competition in AI coding tools isn't about benchmark scores—it's about who can build the on-the-ground implementation muscle.

The reason to click: Warp's pivot is sharp. It open-sourced its core terminal in April, and by July it's pushing Oz, an agent orchestration platform, aiming to move from 'human + single agent' to fully automated dev loops. Zach Lloyd's factory cycle covers triage, implementation, review, verification, shipping, and monitoring. Oz plugs into Jira, Slack, and GitHub, with configurable human checkpoints. I'd discount the timeline a bit. Lloyd expects most major projects to adopt some factory form within a year, but the post gives no throughput, fix rate, or false-positive numbers for Oz—it's still concept and demo stage. Warp's terminal was getting squeezed by Claude Code, Codex CLI, and Gemini CLI; open-sourcing was defense, the factory is offense. The ammo for that offense isn't shown yet.

I clicked because Ahmad Osman didn't do slides at AIEWF — he handed attendees a hardware comparison site and let them benchmark DGX Spark, AMD Strix Halo, and other devices against frontier cloud models on speed and output quality. His core claim is concrete: open models trail closed ones by 4–8 months, and that gap keeps shrinking. The most useful bit is his counterexample. A friend bought an RTX 5090 to run Qwen 3.5 locally, hooked it up to Claude Code, and asked it to change the GPU's RGB lighting. It failed — because the local model had no internet search access. Once they added a search endpoint, it worked. Osman's point: hosted products like ChatGPT bundle search, tools, and infrastructure around the model. His company Osmantic is building an open-source deployment layer to fill that end-to-end gap. The audience ranged from a student shopping for her first AI machine to an Intel exec asking about Windows UX and enterprise model routing — demand is broader than I'd assumed. The post doesn't detail Osmantic's product progress or business model though, so I'd hold off on that part.

The numbers are blunt: OpenAI's own Economic Research team tracked internal Codex usage, and median output tokens jumped 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal between November 2025 and June 2026. The wild part is the setup—before August 2025, employees spent under 10% of their tokens on Codex, even with unlimited access. That lag-then-surge pattern suggests the shift isn't about a single model breakthrough; it's about workflows finally reorganizing around agents. I'd treat this as a useful internal-adoption benchmark, not a sign that AI has taken over everything.

This one's worth opening because Databricks' two cofounders rarely do a joint interview, and they're laying out a clear pivot: from lakehouse to agent operating system. The centerpiece is Omnigent, a newly open-sourced meta-harness that sits above Claude Code, Codex, Cursor, and custom agents to handle multi-agent composition, live collaboration, and spend controls. Reynold also walked through LTAP, arguing it captures most HTAP benefits by unifying the storage layer rather than merging query engines—and joked that CDC really stands for 'continuous data corruption.' The throughline: once frontier models commoditize, the durable moat is the proprietary data, state, and business logic an agent can access at the moment it acts. I'd discount this a bit since it's a podcast transcript and specific deployment numbers aren't fleshed out, but the direction has more signal than another 'model tops benchmark' headline.

This one's worth your time because of who's talking: Zico Kolter sits on OpenAI's board safety committee, and Matt Fredrikson runs Gray Swan—the same team Anthropic tapped to test the Mythos model card. Their core point is simple. Give a model the ability to use a computer—Claude Code, Codex, whatever—and prompt injection becomes a genuinely new attack surface. Traditional cybersecurity that locks down the system can't stop a malicious instruction hidden in a webpage the agent visits. Gray Swan's own tool Shade was used in the Mythos evaluation, and their specialized red-teaming models already beat humans at breaking AI systems. One counterintuitive bit: bigger models don't automatically get more robust. They're clear on that. I'd treat this as a solid conceptual intro to agent security risk, not a technical fix. It's a podcast transcript—no specific attack cases or remediation details—but the framework is sharp.

This episode is worth clicking because Anjney Midha drops a concrete number: xAI's training MFU is under 10%. For context, GPT-3 hit 21%, PaLM reached 46%, and today's best teams get 60–70%. Google once treated 95% utilization as an outage. The bottleneck isn't GPU supply anymore—it's systems engineering: scheduling, networking, parallelism, cluster reliability. If any of those slip, your theoretical FLOPs never become real training progress. Midha backed Anthropic, Mistral, and Black Forest Labs before starting AMP, which aims to build a 1.2 GW base-load compute grid with 6 GW spike capacity. He also flags DeepMind's unpublished research as a market failure and notes Anthropic prioritized coding as P0 from day one—that's why Claude got good at it early. But the post doesn't give a timeline for AMP's grid, so the 1.2 GW vision is still on paper. The MFU figure comes from a SemiAnalysis tweet and Midha's own claim, not an official xAI disclosure—I'd discount it a bit until we see more.

This one's worth reading because Satya dropped a 60M-view X article packaging his podcast ideas into 'Loopcraft.' The core argument: stop obsessing over picking the best model — build a learning loop where human expertise and model outputs compound together. It's his first clear AI strategy statement since Microsoft and OpenAI split eight months ago. Same day, Anthropic's Fable 5 hit 161 on the Epoch Capabilities Index, edging GPT-5.5 Pro, then got suspended by a US export-control action. That timing makes Satya's case for model neutrality and owning your stack feel less like theory and more like insurance — frontier model access can vanish overnight on a policy decision. Loopcraft is still a conceptual framework with no product timelines. I'd read it as Microsoft officially backing the 'Big Harness' play. The how-to part isn't here yet.

I'd open this because Guo pulls a bunch of threads from the last two years—open model adoption, agent labs vs model labs, why app-layer companies survive—into one clean framework around 'legibility.' Her core point: anything that can be written down as training data will eventually be absorbed by models. The real moat is the messy, non-standardizable work of wrangling private enterprise data, wiring up tools, and reshaping customer workflows. The second half digs into the Anthropic Fable/Mythos trust backlash. The community found model performance on AI research prompts was silently degraded rather than explicitly refused. Guo's take: silent gating is worse than a hard 'no' because researchers can't tell if the capability exists and is being withheld, or was never there. I'd read this as an investor's mental map, not a technical roadmap. It won't help you tune hyperparameters, but it frames 'what's worth building' more clearly than most tech blogs. The closing line—intent may be scarcer than compute—sounds like a soundbite, but in context of her finding maybe three worthy bets a year, it lands.

Anthropic just made its previously restricted Mythos-class model publicly available as Claude Fable 5. On the brand-new FrontierCode Diamond benchmark, it jumps from Opus 4.8's 13.4% to 29.3%, with API pricing roughly 2x Opus. Two controversial policies come with it. First, mandatory 30-day traffic retention — Anthropic says it's safety-only, no training, deleted after 30 days. Second, and more contentious: silent interventions that degrade performance on recursive self-improvement requests, like building pretraining pipelines or ML accelerator design. No visible fallback, no user notification. Anthropic estimates this hits ~0.03% of traffic, concentrated in under 0.1% of organizations. Most users won't notice, but the open-source crowd is upset for a reason: silent degradation means you can't tell if you're being throttled. If you're doing frontier dev work on Fable 5, that's the thing to watch.