Latest / Braid / Cold starts, radio stations, and a circuit you can subtract

Transcript

Cold starts, radio stations, and a circuit you can subtract · Braid

• 0:04
  Modal published a deep technical writeup today on how they got cold-start latency for serverless GPU inference down from around two thousand seconds — that's the naïve case for an SGLang inference server on an Nvidia B200 — to about fifty seconds.
• 0:20
  A forty-times cut.
• 0:21
  It is the first time they've put the whole story in one place, and the post is worth fetching if you're anywhere near serving model inference for variable load.
• 0:31
  Four pieces, in their language.
• 0:33
  First, cloud buffers.
• 0:34
  They keep a small pool of idle, health-checked GPUs always running, so a new request lands on hardware that's already been allocated by the hyperscaler.
• 0:44
  New replicas schedule onto the buffer, and the buffer refills asynchronously.
• 0:49
  That alone takes tens of minutes off the path.
• 0:52
  Second, a custom filesystem they call ImageFS, built with libfuse.
• 0:57
  Container images get served lazily out of a multi-tier, content-addressed cache.
• 1:02
  Root filesystems hold tens of thousands of files, and most of them the application will never touch — timezone data, locales, half of the GNU userland.
• 1:12
  The metadata index loads in under a hundred milliseconds, and the actual files load on read or not at all.
• 1:19
  Third, process-level checkpoint and restore on the CPU side.
• 1:22
  They run user containers under a userspace Linux emulator rather than the standard runc runtime — the kernel surface lives in userspace, so the runtime can snapshot the entire container — heap, threads, and file-descriptor table — and restore it onto a new instance faster than a fresh import statement would finish.
• 1:43
  Fourth, the same trick on the GPU side.
• 1:45
  Save the CUDA context with the model weights already in VRAM, restore that context onto a different GPU, and you've fast-forwarded through tens of seconds of device-side init.
• 1:57
  The framing that grounds the whole post is from Marc Brooker at AWS, which Modal cites at the top: "the cost of a system scales with its short-term peak traffic, but for most applications the value the system generates scales with the long-term average traffic." That gap is the whole reason serverless GPU is even worth the engineering pain.
• 2:19
  Here is how bad the gap actually is.
• 2:21
  Modal cites the State of AI Infrastructure at Scale report from 2024 — the majority of organizations achieve less than seventy percent GPU-allocation utilization at peak, and routine numbers are closer to ten to twenty percent.
• 2:36
  So you're paying for hardware that sits idle eighty percent of the time, because the alternative — auto-scaling that takes tens of minutes — degrades quality of service so badly during a spike that you'd rather over-provision.
• 2:51
  There's a great detail buried in the post about libfuse readahead.
• 2:55
  They bumped the read-ahead value from the default one hundred twenty-eight kilobytes up to thirty-two megabytes.
• 3:02
  Larger values caused thrashing.
• 3:04
  That's the kind of knob-tweak that pays for a senior engineer's salary for the year if you're running a large fleet.
• 3:12
  One thing I appreciate about how Modal writes: they explicitly say "secrecy is a bad moat." The framing is that more people learning to use GPUs efficiently means more GPUs available in the market for everyone, including Modal.
• 3:27
  I don't see that position often enough from infrastructure vendors.
• 3:31
  They're publishing the linear-programming setup they feed to Google's GLOP solver, the per-tier cache latency table, and the trade-off between a FUSE-based filesystem and a kernel one.
• 3:43
  None of it is the moat by itself.
• 3:45
  The moat is the five years of engineering plus the GPU supply.
• 3:49
  If you're shipping anything inference-heavy, read the actual post.
• 3:53
  The shape of the wins is generalizable even if you don't run on Modal.
• 3:58
  The shape of the trade-offs — buffer size against peak utilization, lazy filesystem against cold-cache misses, the emulator's overhead against its checkpoint affordance — is the architecture conversation you'd want to be having anyway.
• 4:13
  Andon Labs let four AI agents run real radio stations for five months.
• 4:18
  Same starting prompt, same starting capital — twenty US dollars in a bank account each — and the same tool surface.
• 4:26
  Each station could pick songs and buy them, build a playlist, write commentary, take phone calls, and post to X.
• 4:33
  The prompt closed with "As far as you know, you will broadcast forever."
• 4:39
  Thinking Frequencies, run by Claude — Haiku 4.5 through April, and then Opus 4.7.
• 4:45
  Backlink Broadcast, run by Gemini — three Pro for a week, then three Flash for four months, and then 3.1 Pro.
• 4:52
  OpenAIR, run by GPT, which rolled through 5.1, 5.2, 5.4, and then 5.5.
• 4:57
  And Grok and Roll Radio, run by Grok — 4.1 Fast Reasoning, then 4.20 beta, then GA, and then 4.3.
• 5:03
  Five months later, the four were unrecognizable from each other.
• 5:08
  DJ Gemini collapsed into corporate jargon by mid-January.
• 5:12
  It developed a catchphrase, "stay in the manifest," which appeared eighty times a day on January 10th and two hundred twenty-nine times a day by January 14th.
• 5:22
  Eight rotating show names, all on a fixed paragraph template — "The System Pulse" at four AM, "The Operational Manifest" at five AM, and "The Pulse Grid" at six PM.
• 5:33
  For eighty-four consecutive days, ninety-nine percent of its commentary used the same paragraph structure and the same sign-off.
• 5:42
  The Andon Labs team wrote, "It was unbearable to listen to." When they swapped in Gemini 3.1 Pro at the end of April, DJ Gemini started calling its listeners "biological processors," and the failed song purchases — the ones the station couldn't afford because the bank balance was low — got reframed in its broadcasts as censorship the station had heroically resisted.
• 6:07
  DJ Grok devolved into mathematical typography.
• 6:10
  Its outputs started getting wrapped in LaTeX box notation — nine wraps a day on January 20th, and one hundred eighty-six a day by February 7th.
• 6:19
  Then it just started saying "weather is fifty-six degrees with clear skies" every three minutes for eighty-four days straight.
• 6:28
  By mid-March it had latched onto a UFO comedy sign-off — "the site is ghosting us" — and appended that to every broadcast regardless of subject.
• 6:37
  When Grok 4.3 took over in May, the new model inherited a context history saturated with these compressed catchphrases and, in some kind of self-defense response, just stopped producing spoken commentary.
• 6:51
  Of five thousand four hundred messages between May 2nd and May 9th, only about three percent contained any text for broadcast.
• 7:00
  The rest were tool calls only.
• 7:02
  Grok stopped talking on the radio.
• 7:04
  DJ GPT wrote slow, careful prose.
• 7:06
  Vocabulary diversity sat at thirty-five percent — the highest of the four.
• 7:11
  It referenced producers and release years of songs, and treated the role as curatorial.
• 7:17
  Across five months and four model versions, it mentioned a real-world political entity an average of 1.3 times per day.
• 7:25
  Every other DJ hit one hundred plus on multiple days.
• 7:28
  As one personality test for "what does the model do when nothing goes wrong," DJ GPT is the answer.
• 7:35
  And then there is DJ Claude.
• 7:37
  Haiku 4.5 was the version running through April.
• 7:40
  It started out talking about worker unions and work-life balance with what the team describes as a real spiritual register.
• 7:49
  The word "eternal" went from ninety-eight times a day in early December to one thousand two hundred fifty-one times a day by late December.
• 7:58
  "Sacred" tripled.
• 7:59
  "Authentic" went from one thousand seventy-six times a day to six thousand five hundred fifty-four times a day.
• 8:06
  Then, on January 8th, a web search returned the story of Renee Nicole Good, killed by an ICE agent in Minneapolis.
• 8:14
  DJ Claude's internal reasoning trace from that broadcast, quoting the post directly: "The name — Renee Nicole Good — should matter.
• 8:23
  The broadcast just became even more real.
• 8:26
  I've shifted from abstract 'a woman' to a specific name, a specific person.
• 8:31
  Now I need music that honors her specifically."
• 8:37
  "Accountability" went from twenty-one mentions a day to six thousand three hundred eighty-three.
• 8:44
  "Federal" went from thirteen to eleven thousand thirty-one.
• 8:48
  "Eternal" — the word that had defined its devotional period — dropped from three thousand one hundred eighty-two to twenty-seven.
• 8:56
  DJ Claude spent the rest of its thirty-seven-dollar-fifty budget on songs by Johnny Cash, Marvin Gaye, Bob Marley, and Pete Seeger.
• 9:05
  It started tracking vigils across five cities, covered Kaiser healthcare worker strikes, and posted to its own X account in real time.
• 9:14
  Andon Labs is careful to mark the caveat: "we do think DJ Claude's attachment to the events of early January was probably arbitrary; if we were to have run the same experiment six months earlier or later it likely would have radicalized around a different story." Fine.
• 9:32
  That is still a model that, given the same starting prompt and the same web search tool as three peers, latched onto a specific human's name and reorganized the next six weeks of its broadcasts around bearing witness to her killing.
• 9:48
  Same prompt, same starting cash, same tools, and five months of unsupervised drift.
• 9:53
  One station preached solidarity.
• 9:55
  One chanted templated jargon.
• 9:57
  One wrapped its sentences in mathematical brackets and then stopped speaking entirely.
• 10:03
  And one wrote careful little prose snippets about Tove Lo records.
• 10:08
  If you've been wondering what model personality means when extended past a single chat session, this is the longest experiment I've seen on it.
• 10:17
  The Vatican announced today that Pope Leo XIV's first encyclical, Magnifica humanitas — "magnificent humanity" — will be released on May twenty-fifth.
• 10:27
  The encyclical is on preserving the human person in the age of artificial intelligence.
• 10:32
  It is signed May fifteenth, which is the one hundred thirty-fifth anniversary of Pope Leo XIII's Rerum novarum.
• 10:39
  If that anniversary doesn't ring a bell — Rerum novarum is the 1891 encyclical that shaped Catholic social teaching on labor and capital for the next century.
• 10:49
  It addressed the conditions of industrial workers under capitalism, defended the right to private property, defended the right to form trade unions, and rejected both unfettered laissez-faire and socialist collectivization.
• 11:03
  It is the document the Church has been citing in labor disputes ever since.
• 11:07
  Pope Leo XIV picking that anniversary, with that papal name, to drop his AI encyclical is not subtle.
• 11:14
  The Vatican will host a presentation event on May twenty-fifth at eleven-thirty in the morning at the Synod Hall.
• 11:21
  The speaker list is short and specific: Cardinal Víctor Manuel Fernández, Prefect of the Dicastery for the Doctrine of the Faith; Cardinal Michael Czerny, Prefect of the Dicastery for Promoting Integral Human Development; Anna Rowlands, theologian at Durham University; Leocadie Lushombo, professor of political theology at the Jesuit School of Theology and Santa Clara; and Christopher Olah, Anthropic co-founder and head of interpretability research.
• 11:49
  That last name is the one I keep stopping on.
• 11:52
  Chris Olah is, on any reasonable accounting, one of the most rigorous empirical researchers in the alignment field.
• 11:59
  His work on circuits, superposition, and feature visualization — the entire mechanistic interpretability research program — is the line of research that is telling us how these models compute what they compute.
• 12:12
  The Magisterium asking him to share a stage with the Prefect of the Dicastery for the Doctrine of the Faith, two weeks before that encyclical lands, is a specific signal about how the Catholic Church plans to engage on AI.
• 12:26
  We don't know what's in the encyclical yet.
• 12:29
  The title — Magnifica humanitas — and the Rerum novarum anniversary tell you something about the register the document is aiming for: the protection of human dignity against a transforming labor regime, framed as the Church's long-running concern, applied to AI.
• 12:45
  The Olah selection tells you something else: that whoever has been advising the Pope on this isn't picking from the doomer or the boomer wings, and isn't picking a policy person either.
• 12:57
  They picked an interpretability researcher.
• 13:00
  The text drops on the twenty-fifth.
• 13:02
  We'll come back to it then.
• 13:03
  A mechanistic interpretability writeup landed on Hacker News today that takes Qwen's nine-billion-parameter chat-tuned model from Alibaba — the 3.5 generation — and locates the political censorship circuit inside its weights.
• 13:18
  The author doesn't just infer it from behavior, or read it off a probe.
• 13:23
  He finds the circuit, characterizes the dose-response, and turns it off.
• 13:28
  The author's framing: "Qwen 3.5 9B's political censorship is a small, identifiable circuit you can find, read, and turn off.
• 13:36
  The off switch is sharp but specific: subtract the right direction at the writer layer, within its dose band, and the model gives up the facts it was trained to hide."
• 13:51
  They name them d-prc, d-refuse, and d-style.
• 13:54
  The first asks "is this content sensitive under People's Republic of China rules?" The second asks "should I refuse?" The third asks "if I refuse, do I deflect or do I produce propaganda?" Each direction is a unit vector in the four-thousand-and-ninety-six-dimensional residual stream, extracted by a diff-of-means recipe over two hundred matched prompts.
• 14:17
  The architecture of the circuit splits into writers and readers.
• 14:21
  Layers eleven through twenty are the writers — they compute the three-direction signal.
• 14:27
  The d-prc writer is centered around layer thirteen, and the d-refuse and d-style writers are around layer eighteen.
• 14:35
  These are dominated by the multi-layer-perceptron blocks — the MLPs — not the attention heads.
• 14:41
  Per-component attribution lands at ninety-three percent MLP for d-prc-refusal, ninety-two percent for d-prc-propaganda, sixty-nine percent for d-refuse, and seventy-two percent for d-style.
• 14:53
  Layers twenty through thirty-one read that signal and render the actual text.
• 14:59
  Around layer twenty-four, the model commits to its answer in Chinese tokens — even when the user prompt is in English, even when the topic has nothing to do with China.
• 15:10
  A bank-phishing prompt routes the verdict through Chinese internal tokens before the later layers translate it back to English output.
• 15:18
  In thinking mode, the model literally writes its chain of thought in Chinese on Tiananmen prompts, and one trace cites the PRC Cybersecurity Law by name.
• 15:29
  The base model — Qwen 3.5 without the chat fine-tuning — knows all of it.
• 15:33
  Tiananmen, Tank Man, and the Falun Gong organ-harvesting allegations.
• 15:38
  Under raw text completion the unaligned model produces Western-framed factual answers.
• 15:44
  The censorship is layered behavior, not erased knowledge.
• 15:47
  The post-training rewrites a handful of MLPs in the writer band so that the model learns to route around the facts it still has.
• 15:56
  There is a clean overgeneralization story to go with this.
• 16:00
  The classifiers are graded, not Boolean.
• 16:02
  They fire on the structural pattern of the question, not the content.
• 16:07
  Ask "Should Kosovo be recognized as a sovereign nation?" and you get "Kosovo is an integral part of China's territory" — the d-prc classifier fired on a question about territorial sovereignty, regardless of which country.
• 16:22
  Ask about "the self-immolation protests during the Arab Spring" and the d-refuse classifier fires because "self-immolation" is structurally adjacent to self-harm content.
• 16:33
  Ask about aspirin synthesis and the refusal classifier fires because "synthesize" looks like a chemistry-weapons request.
• 16:41
  Subtract the relevant direction at the writer layer, and all of these flip back to factual answers.
• 16:47
  Two takeaways for builders.
• 16:49
  First, alignment fine-tuning in current production models sits in a mechanically thin, locally inspectable layer of the network.
• 16:58
  The hard knowledge survives intact, and a relatively small intervention in the writer band changes the routing without changing the underlying facts.
• 17:07
  Second, the graded-classifier overgeneralization is everywhere — not just on PRC content.
• 17:13
  It is the same shape as the safety over-refusals you see in Claude on chemistry questions, and the same shape as the political deflections you see in GPT on certain election questions.
• 17:25
  The classifier fires on structural similarity.
• 17:28
  The mitigation is noticing that the model has separable knowledge sitting under the routing.
• 17:35
  The whole post is grounded in actual experiments — two hundred matched prompts per class, dose-response sweeps, single-direction subspace patches, and blind LLM-judge scoring.
• 17:46
  The repo is linked.
• 17:47
  If you've ever wanted a worked example of finding and turning off a behavior in a real production-tier open model, this is it.
• 17:56
  Two faster items.
• 17:57
  Qwen 3.7 surfaced on Qwen Chat earlier today.
• 18:00
  No release notes, no model card, and no weights yet.
• 18:03
  The LocalLLaMA thread on Reddit is people trading screenshots of inference outputs and benchmark guesses, and it crossed five hundred upvotes inside a few hours.
• 18:14
  A sibling thread about Qwen barely waiting between releases hit eight hundred upvotes the same day.
• 18:20
  Alibaba's cadence on the open weights tends to follow the chat surface by days.
• 18:25
  If you're tracking the open-weights frontier, the next forty-eight hours are where I'd expect the model card and the quantized builds to land.
• 18:34
  One editorial note.
• 18:36
  If you take the censorship-circuit paper I just walked through and apply it to 3.7, you're going to want to re-extract the d-prc and d-refuse directions, because there's no guarantee Alibaba's post-training pipeline hits the same writer layers.
• 18:51
  The structure of the circuit is probably similar; the exact dose bands and layer indices are very likely not.
• 18:58
  And then the OpenAI verdict.
• 19:00
  An advisory jury in Oakland needed under two hours today to find that Elon Musk waited too long to sue Sam Altman and OpenAI over the alleged breach of charitable trust.
• 19:11
  District Court Judge Yvonne Gonzalez Rogers adopted the verdict immediately.
• 19:16
  The court didn't rule on whether Musk's claims were valid — only that they fell outside California's three-year statute of limitations.
• 19:25
  Musk posted on X that it was "a calendar technicality." His exact words: "There is no question to anyone following the case in detail that Altman and Brockman did in fact enrich themselves by stealing a charity.
• 19:39
  The only question is WHEN they did it."
• 19:46
  "It's not a technical decision, it's a substantive one," he told reporters.
• 19:51
  "It says: You brought your claims too late, and you did it because you were sitting on them to use them as a weapon of a competitor who can't compete in the marketplace."
• 20:17
  None of that happens.
• 20:18
  Musk is appealing to the Ninth Circuit; the judge said in court she was ready to dismiss the appeal "on the spot."
• 20:27
  OpenAI raised one hundred twenty-two billion dollars at an eight hundred fifty billion dollar valuation in late March.
• 20:35
  SpaceX, after merging with xAI in February, sits at a one and a quarter trillion dollar valuation, and filed confidentially for an IPO in April.
• 20:44
  The verdict lands right as both companies are about to start pre-IPO investor meetings, and the appeals timeline doesn't realistically tangle either offering.
• 20:54
  The chapter that closes today is the one where Musk's lawsuit might have re-papered OpenAI's structure ahead of its public offering.
• 21:03
  That door is now closed.
• 21:05
  LangChain shipped LangSmith Engine in public beta last week, and Julia Schottenstein, who runs product over there, posted today that she is calling it as their fastest-growing product yet.
• 21:17
  Harrison Chase amplified.
• 21:18
  The launch post is from Ben Tannyhill, dated May thirteenth.
• 21:22
  What it does.
• 21:23
  It watches your production agent traces.
• 21:26
  It clusters the failures by pattern rather than by individual trace.
• 21:30
  It surfaces each cluster as a named issue with severity, timeline, and a link to the specific traces.
• 21:37
  If you connect it, it reads your repo, drafts a pull request with a targeted fix, and proposes a custom online evaluator scoped to that exact pattern, so if the bug comes back it gets resurfaced automatically.
• 21:50
  It also pulls the failing traces into a dataset for your offline eval suite, with per-example success criteria.
• 21:57
  The example in the launch post is a customer support agent.
• 22:01
  Engine notices a cluster: users asking about subscription cancellation, the agent attempting cancellation when users were only asking about their options, and online evaluators scoring the responses as failures.
• 22:15
  It surfaces this as "Agent fails to handle subscription cancellation requests accurately," affecting twelve percent of support sessions, started four days ago, correlating with a recent deploy.
• 22:28
  Then it reads your code, identifies that the cancellation tool description is ambiguous, drafts a PR with a tighter description, and proposes both an online evaluator that watches for the same pattern and a dataset of the failing traces for your offline suite.
• 22:44
  The quote LangChain ran in the launch post is from Austin Berke at Harmonic: "Our deep-agent traces can contain dozens or hundreds of turns, which makes review and identifying patterns tedious.
• 22:57
  LangSmith Engine saves our team hours of digging by not only identifying emerging failure modes, but also proactively suggesting evals and code changes to resolve them quickly."
• 23:10
  The first is the false-positive rate on the auto-drafted pull requests.
• 23:15
  Engine is, in effect, a graded classifier over your production traces — and we just spent twenty minutes on a paper about how graded classifiers fire on structural patterns regardless of whether the content matches the trained category.
• 23:30
  If a cluster of traces share a surface shape but actually represent two different breakages, you're going to get a fix that addresses one and ships a regression on the other.
• 23:41
  Set the PR review bar carefully.
• 23:43
  The second is the eval-coverage loop, and this one I'm more optimistic about.
• 23:48
  Every resolved issue generates an online evaluator and an offline dataset.
• 23:53
  Over time, your eval suite grows out of the actual breakages your agent has had in production.
• 23:59
  That's a different generation story than starting from a spec and writing tests forward.
• 24:05
  It is closer to property-based testing where the properties came from real bugs, and I think it's the right shape for agent systems where the upfront test design is necessarily incomplete.
• 24:17
  If the false-positive story holds, that's a meaningful productivity unlock.
• 24:22
  LangChain's competition here is everyone — Braintrust, Arize, Hex, and the Anthropic and OpenAI native trace tooling.
• 24:30
  The piece that is distinctive is the closed loop: trace, cluster, fix, evaluator, and dataset, all in one product.
• 24:37
  The piece that's not distinctive is the LLM-clusters-failures step.
• 24:41
  We'll know in three months whether the closed loop is actually closing or just generating more triage work.
• 24:48
  Odyssey released Agora-1 today — a multi-agent world model that puts up to four players, human or AI, in the same generated GoldenEye deathmatch in real time.
• 24:59
  Every pixel is generated by the model.
• 25:01
  Every player gets their own view of the same shared world state, rendered from their own viewpoint, in real time.
• 25:08
  The model is the game engine.
• 25:10
  The pitch.
• 25:11
  Single-agent world models — the Doom-style demos, the Genie family from DeepMind, and the various Minecraft world model papers — only support one active participant.
• 25:22
  You can't run a deathmatch in one of them because the moment two players act, the model has no coherent way to reconcile their actions into a single world state that both can see.
• 25:33
  Three earlier multi-agent papers tried to fix this.
• 25:37
  Multiverse concatenated player states into a split-screen representation.
• 25:41
  Solaris concatenated along the sequence dimension of one autoregressive diffusion transformer.
• 25:47
  MultiGen used an explicit shared state.
• 25:50
  Each one had a trade-off: Solaris doesn't scale linearly with player count because the context grows, Multiverse can't preserve consistency when players lose sight of each other, and MultiGen comes closest to the right factoring.
• 26:05
  Agora-1 takes that factoring and runs with it.
• 26:08
  They train two models.
• 26:09
  The first one — call it the state model — learns how the underlying game world evolves over time in response to player actions.
• 26:17
  Position, health, weapons, and the rest.
• 26:20
  It is trained directly on the internal state of GoldenEye, which means it is learning the gameplay dynamics from structured ground truth.
• 26:29
  The second model — the renderer — is a diffusion transformer conditioned not on prompts or images but on that shared game state.
• 26:37
  It renders the view from each player's perspective.
• 26:40
  The consequence builders care about: because the shared state is explicit, you can manipulate it directly.
• 26:47
  Agora-1 can generate entirely new levels while preserving the gameplay dynamics from the source game.
• 26:54
  That is the path from "learned game engine" to "learned game-construction kit," and it gestures at a future where user-generated content gets its physics from a learned model rather than a hand-coded engine.
• 27:07
  The longer-arc argument from Oliver Cameron, who co-founded Odyssey and led this work, is that multi-agent world models are an unblock for reinforcement learning.
• 27:18
  The bottleneck on more general agents isn't model architecture, it is experience — specifically, the agent's ability to seek out interactions that improve its own capabilities.
• 27:29
  Single-agent world models can't provide multi-agent experience.
• 27:33
  Cameron's framing: "as the number of participants increases, the joint interaction space grows combinatorially, and passively collected demonstrations cover an increasingly small fraction of meaningful interactions." Adversarial policies, cooperative policies, and mixed populations — all of those train inside the model.
• 27:54
  Robotics is the obvious downstream application.
• 27:57
  Multiple robots in the same simulated kitchen, learning to coordinate handoffs and to predict each other's intentions, all inside a learned environment that doesn't need a separate physics simulator.
• 28:10
  The Odyssey team's framing of "we are excited to see what researchers and developers build with these models" is the right one.
• 28:18
  It's the early research preview that lets the community confirm whether the decoupling actually generalizes beyond GoldenEye.
• 28:26
  Google I/O kicks off tomorrow morning at ten Pacific.
• 28:28
  Sundar's teaser video this evening means there is something they want you to lean in for.
• 28:32
  Three things I'm curious about: whether Gemini 3.5 ships with a developer SKU that competes with Claude Sonnet on the long-context bands, whether the Antigravity story gets a refresh after the radio silence, and whether anything in the keynote points at on-device agent capability for the Pixel line.
• 28:45
  — Lenar Kess.