Latest / Braid / Foothills, and the morning Karpathy moved

Transcript

Foothills, and the morning Karpathy moved · Braid

• 0:04
  Demis Hassabis took the Shoreline Amphitheatre stage in Mountain View yesterday afternoon and closed the Google I/O 2026 keynote with a sentence Google had clearly workshopped for weeks.
• 0:15
  Quote: When we look back at this time, I think we will realize that we were standing in the foothills of the singularity.
• 0:23
  End quote.
• 0:24
  He called it a profound moment for humanity, and walked off.
• 0:28
  It was a strong keynote on its own terms — three new models, a coding agent that built an operating system on stage, a 24/7 personal AI tier called Spark, and a benchmark deck thick enough to justify the rhetoric, at least the part that's quantifiable.
• 0:44
  Hassabis used to be the conservative voice on AGI timelines; five to ten years was his line for most of the last cycle.
• 0:51
  He's compressed it to just a few, and at I/O he gave Google's most aggressive on-stage framing yet.
• 0:58
  And about forty minutes before he walked out, Andrej Karpathy — one of the most-followed researchers working on large language models — posted to X.
• 1:07
  Quote: Personal update: I've joined Anthropic.
• 1:10
  I think the next few years at the frontier of LLMs will be especially formative.
• 1:15
  End quote.
• 1:16
  Twenty-one thousand likes inside the first hour.
• 1:19
  Eight hundred twenty-five thousand views in a day.
• 1:22
  The keynote was still being staged when the news cycle had already split — half on Google, half on the highest-status engineer in the field walking into a competitor.
• 1:33
  Today I want to work through both stories together, because they're the same story at different altitudes.
• 1:40
  We'll go through I/O — Flash, Omni, Antigravity, Spark — then through the lab responses: Anthropic, Alibaba, DeepSeek.
• 1:47
  Then a substrate failure that happened while Google was pitching Spark.
• 1:52
  And we'll end on a builder writeup from this week that's a sharper picture of agentic coding than anything that hit a keynote slide.
• 2:00
  Start with Gemini 3.5 Flash, because it's the model most of you will actually run.
• 2:05
  It shipped GA the day of the keynote — Gemini app, Search, API, enterprise, all live.
• 2:11
  A one million token context window.
• 2:13
  Sixty-five thousand tokens out.
• 2:15
  Four thinking levels: minimal, low, medium (the new default), and high.
• 2:20
  The pricing comes in at one dollar fifty per million input tokens, nine dollars per million output, with a 90 percent discount on cached input.
• 2:29
  Google's speed claim was four times faster than comparable frontier models, and up to twelve times faster when run inside Antigravity, where Google controls the harness around it.
• 2:41
  The benchmarks are credible — Terminal-Bench 2.1 at seventy-six point two percent, GDPval-AA at sixteen fifty-six Elo, MMMU-Pro in the eighty-three to eighty-four range, and Chatbot Arena placing it ninth overall in text and code at a 1507 score.
• 2:57
  None of those are leaderboard-leading, but for a model branded Flash they're more than a tier up from the prior generation.
• 3:05
  Which brings us to pricing, which is where the model name and the price tag don't agree anymore.
• 3:11
  A Reddit thread caught it the same evening — Gemini 3.5 Flash costs roughly three times more per token than the previous Flash, and about thirty times more than the original 1.5 Flash from two years ago.
• 3:24
  The screenshot went around with the headline: behold, Gemini 3.5 Flash.
• 3:29
  Here's how I'd read it.
• 3:30
  The Flash brand used to mean cheap-fast-shallow.
• 3:33
  The new Flash means cheap-relative-to-Pro and intentionally tuned for long-horizon agent workloads — coding loops, tool calls, repeated turns inside a harness.
• 3:44
  Google is repositioning where the volume lives.
• 3:47
  They want the agentic coding workload running on Flash, paying Flash margins, instead of on Pro at frontier-tier prices.
• 3:55
  The name carries inherited brand equity the price tag no longer earns.
• 3:59
  If you've been quietly budgeting agent runs against the old Flash unit economics, this episode is your reminder to re-run the spreadsheet today before your next agent loop posts to billing.
• 4:11
  Gemini Omni is the headline media model and the bet Google planted next to the AGI rhetoric.
• 4:18
  The pitch on stage was that Omni merges Gemini's reasoning with DeepMind's generative-media stack — Nano Banana, Veo, Genie — into a single model that takes text, image, audio, and video in, and starts with video out.
• 4:32
  The marketing line was that Omni handles physics, gravity, and kinetic motion better than the prior Veo generation.
• 4:40
  Multi-turn editing is consistent.
• 4:42
  Worlds hold together across cuts.
• 4:44
  Rollout is staggered: paid Gemini users got Omni day-of, YouTube Shorts and the Create surface get it this week, and the API follows in the coming weeks.
• 4:54
  That last bit matters — if you build on Omni today, you're building on a UI rather than an endpoint.
• 5:01
  The physics claim got tested fast.
• 5:03
  Inside hours, a Reddit user shared a Gemini link where they'd asked Omni to render someone doing a backflip.
• 5:11
  The body distorts mid-rotation.
• 5:13
  The legs cross the spine in a way that bodies don't.
• 5:16
  The post collected six hundred upvotes and a long comment thread of people doing the same probe with the same result.
• 5:24
  So the headline footage clearly works — Google wouldn't have shown it otherwise — and the moment you push into a specific kinematic prior, the model still falls into the same articulation traps Veo had a year ago.
• 5:38
  Not catastrophic; not exactly the win-condition for the marketing line either.
• 5:43
  The more interesting architectural choice is that Google bundled the three media models into one Omni.
• 5:50
  They could have kept Veo, Genie, and Nano Banana as separate APIs, the way OpenAI keeps DALL-E, Sora, and the realtime voice stack distinct.
• 6:00
  They didn't.
• 6:00
  That tells you Google is betting the future of generative media is a single multimodal inference endpoint, not a portfolio of point models.
• 6:10
  Whether that bet pays out depends on how soon a single weight set can outperform three specialised ones at their own jobs.
• 6:18
  Today, it doesn't — the backflip clip is evidence — but the direction Google's investing in is clear.
• 6:25
  Antigravity 2.0 was the demo.
• 6:27
  Onstage, Varun Mohan ran Gemini 3.5 Flash inside the new Antigravity harness and built an operating system from scratch.
• 6:34
  Live.
• 6:35
  Or rather, the run was live; the result was pre-baked, since the actual build took twelve hours.
• 6:42
  Here are the numbers Google put on the screen.
• 6:45
  Ninety-three parallel sub-agents, fifteen thousand model requests, two point six billion tokens consumed, twelve hours of wall-clock time, and less than one thousand dollars in API credits.
• 6:57
  Output: a microkernel, a basic shell, a window manager, and just enough graphics scaffolding to boot Doom.
• 7:04
  Cue the room.
• 7:05
  Hold the altitude on this one.
• 7:07
  It's not a Linux replacement.
• 7:09
  The OS does one famous thing.
• 7:11
  Nobody's pushing it to production.
• 7:13
  The point of the demo was the orchestration shape — that ninety-three coding agents can run in parallel under a single harness for twelve hours and still produce a coherent codebase at the end.
• 7:26
  A year ago, an order of magnitude smaller scope would have cost more in compute and three months in engineer time, and would have needed a researcher driving it the whole way.
• 7:37
  A thousand dollars of inference producing a working microkernel that runs Doom is a real productivity number you can put on a slide.
• 7:46
  It's not the slide it would have been five years ago, but it's a slide.
• 7:50
  The skeptical read.
• 7:52
  The demo was rehearsed end-to-end.
• 7:54
  The OS hasn't been fuzzed, hasn't shipped, and isn't running anyone's payroll.
• 7:59
  The success rate of an unscripted ninety-three-agent run is the number Google didn't publish.
• 8:05
  I'd want a few independent reproductions before I trusted the eighty-cents-per-line economics to my own backlog.
• 8:13
  The Antigravity SDK and the new Managed Agents API in Gemini are how that reproduction is supposed to happen, and they shipped same-day.
• 8:22
  I'll be looking for the first non-Google teams to share their own multi-hour parallel-agent runs and tell us what the dollars-per-shipped-feature number actually looks like.
• 8:33
  Spark is the announcement that flew under the demo, but it's the one that changes the product surface most.
• 8:39
  Gemini Spark is a personal AI agent that runs on dedicated Google Cloud VMs, twenty-four hours a day, while your laptop is closed.
• 8:47
  It checks in with you before any sensitive action — purchases, sends, irreversible edits — and otherwise just keeps running.
• 8:55
  The target workloads Google pitched were the ones you can't quite delegate to ChatGPT today: monitor these listings and tell me when something matches; watch this project board and prep a summary; draft replies to overnight emails and queue them for review; and chase down vendors.
• 9:13
  Long-horizon tasks measured in days, not turns.
• 9:16
  Spark is Google's direct answer to where Claude has been heading with the Agent SDK and ChatGPT has been heading with background agents.
• 9:24
  The form factor is now standard across the three frontier vendors: a server-resident agent loop you can talk to like an inbox, with a permission boundary and an approval queue.
• 9:35
  What's been missing from each lab's version is integration — agents that can reach Drive, Gmail, Calendar, and Search without you wiring tools.
• 9:44
  Spark gets that almost for free, because it's Google.
• 9:48
  The number Google didn't put on the slide was the price.
• 9:51
  Hassabis said pricing later.
• 9:53
  If Spark turns out to be bundled with the existing Google AI plan, the unit economics for agent-resident apps shift overnight — every Gemini subscriber suddenly has a server agent they didn't have last week.
• 10:06
  If it's a two-hundred-dollar add-on like Claude Max, it's a luxury tier and the indie-builder ecosystem keeps owning the long tail.
• 10:14
  That price disclosure is the I/O thread I'd track next.
• 10:17
  Now the lab response, starting with the one everyone read first.
• 10:22
  Andrej Karpathy posted at eight oh five Pacific yesterday.
• 10:26
  Four sentences.
• 10:27
  Quote: Personal update: I've joined Anthropic.
• 10:30
  I think the next few years at the frontier of LLMs will be especially formative.
• 10:35
  I am very excited to join the team here and get back to R and D.
• 10:40
  I remain deeply passionate about education and plan to resume my work on it in time.
• 10:45
  End quote.
• 10:46
  The context.
• 10:47
  He co-founded OpenAI in 2015, left for Tesla in 2017 to lead Autopilot, came back to OpenAI in 2023 for a year, then left in 2024 to start Eureka Labs, his AI-for-education project.
• 10:59
  Eighteen months on his own.
• 11:01
  And now he's at Anthropic on Nick Joseph's pre-training team.
• 11:05
  Anthropic told TechCrunch the charter is to start a team that uses Claude itself to accelerate pre-training research — model-in-the-loop on the most compute-expensive phase of building a frontier model.
• 11:19
  Ethan Mollick wrote a sentence yesterday that lands on this directly.
• 11:24
  Quote: One interesting side feature of recursive self-improvement, to the extent that is happening, is that it makes the Big Three labs more appealing to talent, and shortens the runway for launching a potential competitor instead at the same time.
• 11:41
  End quote.
• 11:41
  That's the read I'd offer.
• 11:43
  Karpathy spent eighteen months trying to build something on his own.
• 11:48
  He looked at the slope and decided the slope compounds faster inside one of the three frontier labs than outside any of them.
• 11:56
  Eureka isn't dead — he said he plans to resume it — but the order of operations is now: spend the formative years inside Anthropic first, education later.
• 12:07
  A sober update for any senior engineer telling themselves the indie play still has a clean runway.
• 12:14
  And then there's the timing.
• 12:15
  Karpathy could have posted any morning.
• 12:18
  Anthropic could have asked him to wait until I/O was over.
• 12:22
  They didn't.
• 12:23
  Whether that's strategy or just two people landing news when it was ready, the day's narrative bent toward — the talent voted for Anthropic, no matter what Google rolled out on stage.
• 12:35
  Google had the bigger announcement.
• 12:38
  Anthropic had the bigger signal.
• 12:40
  Both were true by lunchtime.
• 12:42
  The other competitive response came from Alibaba, and it's stronger than anything the US labs put up yesterday.
• 12:49
  Qwen 3.7-Max hit Hacker News this morning under the blog title, The Agent Frontier.
• 12:55
  Alibaba had been seeding Max and Plus on the Arena leaderboard for five days with no press release — the standard Qwen launch pattern.
• 13:04
  Max ended preview at Elo fourteen seventy-five, number thirteen overall, number seven in math.
• 13:10
  The headline product claims are long-horizon: thirty-five-hour autonomous task runs, support for more than a thousand registered tools.
• 13:19
  Whether those numbers hold up in independent harness tests is the next thing to track, but the framing is the right one — Qwen positioning itself explicitly as the agent frontier rather than the open-weight frontier.
• 13:34
  And Alibaba did it as a stack.
• 13:36
  The same Alibaba Cloud summit launching Max is also launching the Zhenwu M890, Alibaba's new AI inference chip.
• 13:43
  Model and silicon on the same day.
• 13:45
  The closest US analogue would have been if Google had launched Omni alongside a TPU v6 SKU, which they didn't quite do.
• 13:53
  Alibaba is treating agent-tier model launches as a chip event, and that fundamentally changes the cost-curve story for Chinese inference over the next year.
• 14:04
  It's also a more aggressive integration than any of the US frontier labs are running, since none of them own their own fab line.
• 14:12
  The timing is not subtle.
• 14:14
  Alibaba's summit lands on I/O day two.
• 14:17
  The model was clearly held back from a January or February ship just so it could land this week.
• 14:23
  Nobody at Alibaba is pretending this is anything other than counter-programming.
• 14:28
  Then there was a smaller post that's a bigger signal.
• 14:32
  Deli Chen, an engineer at DeepSeek, tweeted yesterday: We're hiring.
• 14:36
  DeepSeek is forming a new Harness team to build Code Harness from the ground up — maybe you can call it DeepSeek Code or something like this.
• 14:45
  Beijing-based.
• 14:46
  Two roles open — a Harness Product Manager and Harness R and D.
• 14:50
  The post collected twenty-three thousand views and three hundred fifty likes in a day, which for a DeepSeek personnel tweet is major engagement.
• 14:59
  The reason to mark this.
• 15:01
  A year ago, the word harness was a research term you'd see in METR papers.
• 15:05
  Today it's a hiring category at every frontier lab.
• 15:09
  Anthropic has the Claude Code team.
• 15:11
  OpenAI has the Codex team.
• 15:13
  Google has the Antigravity team.
• 15:15
  Alibaba is bundling Qwen with their own harness already.
• 15:18
  DeepSeek was the last of the major Chinese labs publicly building everything model-first, and yesterday they admitted in a job listing that the model alone is no longer the product.
• 15:30
  The product is the model plus the loop around it — the tool registry, the planner, the memory, the security policy, the approval queue, and the failure-recovery logic.
• 15:40
  If you're a senior engineer thinking about where to spend the next two years, harness work is now a discipline.
• 15:47
  It has its own pattern library, its own failure modes around prompt injection and tool poisoning, and its own evaluation problem.
• 15:56
  DeepSeek opening this team this week tells you that no major lab thinks the next eighteen months are about better weights alone.
• 16:04
  They're about the loop the weights run inside.
• 16:07
  While I/O was running, something else was happening on the same Google infrastructure Google was busy selling.
• 16:14
  At twenty-two ten UTC on the nineteenth, Google Cloud's automated systems incorrectly suspended Railway's production account as part of a platform-wide automated action.
• 16:25
  Railway is one of the better-known PaaS companies for developers.
• 16:29
  Their dashboard went dark, their API went dark, and their persistent disks were inaccessible.
• 16:36
  For about an hour, Railway's edge proxies kept serving traffic from cached routing tables.
• 16:42
  Then the cache expired, and everything they ran on AWS and on their own Metal hardware also went unreachable, because the routing-table source of truth was a control-plane API hosted on Google Cloud.
• 16:55
  Recovery took eight hours.
• 16:56
  By the time it was over, GitHub had piled on by rate-limiting Railway's OAuth and webhook integrations on the burst of retried calls.
• 17:05
  The postmortem by Chandrika Khanduri and Cody De Arkland is candid in a way you don't always see.
• 17:12
  They explain that Railway's network was designed as a mesh ring across Google Cloud, AWS, and their own metal — but the discovery mechanism for that mesh was a control-plane API hosted in a single cloud.
• 17:25
  So the mesh worked for cached routes for one hour.
• 17:28
  Then it didn't.
• 17:29
  They're removing GCP from the data-plane hot path and extending high-availability database quorum across AWS and Metal, so any single cloud disappearing instantly will leave the database with a majority.
• 17:43
  The line I'd pull is this.
• 17:44
  Khanduri and De Arkland frame it bluntly: Railway picked the vendors, so Railway owns the outcome.
• 17:51
  Customers don't care whether the failure was Google or Railway — they see your product.
• 17:57
  That is the right framing for a public postmortem, and it's the framing every team running on a single hyperscaler should internalize before their own bad day.
• 18:07
  The reason this belongs next to I/O is direct.
• 18:10
  Spark and Antigravity 2.0 are both pitches for moving more of your workload onto Google Cloud — the agent runtime, the harness, the model, the storage, and the auth.
• 18:21
  The Railway story is what trusting one vendor to be your full substrate looks like on the bad day.
• 18:28
  An automated suspension, no proactive outreach to customers before the action, and eight hours of downtime.
• 18:35
  The postmortem ends up being you owning it publicly, because the customer was never going to read the GCP status page.
• 18:42
  Step away from the keynote for a few minutes.
• 18:45
  The most interesting builder writeup of the week is from Cheng Huang, an architect who published Learnings from One Hundred Thousand Lines of Rust with AI.
• 18:55
  The project: a Rust multi-Paxos consensus engine, rebuilt from scratch, mirroring the feature set of Azure's RSL — the replicated state library that underpins most of Azure's core services.
• 19:08
  He drove a hundred and thirty thousand lines of production Rust in roughly six weeks, with thirteen hundred tests covering more than sixty-five percent of the codebase.
• 19:19
  Throughput went from twenty-three thousand operations per second to three hundred thousand in three weeks of performance tuning.
• 19:27
  Almost all of that with AI agents in the driver's seat and Huang doing review and direction.
• 19:33
  His drivers are Claude Code and Codex CLI, with VS Code for diffs and minor edits.
• 19:38
  His subscription anecdote is funny enough to quote.
• 19:41
  He pays a hundred dollars a month for Anthropic's Max plan, and quote: This became a forcing function — if I don't kick off a coding task with Claude before bed, I feel like I'm wasting money.
• 19:54
  End quote.
• 19:55
  When Codex CLI shipped, he added a second ChatGPT Plus subscription and rotates them — Anthropic Monday through Wednesday, OpenAI Thursday through Sunday.
• 20:05
  He also notes that on his subjective sample, GPT-5 High writes better code contracts than Opus four point one, for what it's worth.
• 20:13
  The technique to pull forward is what he calls AI-driven code contracts.
• 20:18
  Preconditions, postconditions, and invariants attached to critical functions — converted into runtime asserts during testing, compiled out in production.
• 20:28
  He applies them at three levels.
• 20:30
  First, AI writes the contracts; he reviews and refines.
• 20:33
  Second, AI generates targeted test cases from each post-condition.
• 20:38
  Third — and this is where it pays off — AI translates the contracts into property-based tests that explore randomised input spaces.
• 20:46
  One AI-generated contract caught a subtle Paxos safety violation before it ever shipped.
• 20:52
  In the old world that bug surfaces in production as a replication inconsistency you debug on a Saturday from a customer report.
• 21:00
  That stack — contracts as the AI-readable spec, property tests as the AI-driven verifier, and the human architect reviewing contract correctness rather than diff-reading every function — is what serious agentic coding looks like outside the demo.
• 21:16
  Not ninety-three agents in parallel boots Doom.
• 21:19
  A senior engineer with a clear design markdown, a two-hundred-dollar-per-month subscription stack, and a contract regime that turns LLM output into something Paxos-safe.
• 21:30
  Huang's writeup landed the same week Antigravity 2.0 did, and the contrast is useful.
• 21:36
  The framing on the I/O stage is, fire up dozens of agents and watch the magic.
• 21:41
  The framing in Huang's article is, write the contracts and let the agents enforce them.
• 21:46
  Both are real.
• 21:47
  But if I'm picking what to copy into my own workflow tomorrow, it's the contract regime.
• 21:53
  The ninety-three-agent run is a benchmark.
• 21:56
  The contract loop is a practice.
• 21:58
  Sign-off.
• 21:58
  Pull a few threads.
• 21:59
  Google had the bigger announcement.
• 22:01
  Three new models, a coding harness that builds a microkernel in twelve hours, and Hassabis on stage compressing his AGI timeline by half.
• 22:08
  The Flash and Antigravity numbers will change the cost curve for agentic workloads if they hold up in independent reproductions.
• 22:15
  Anthropic had the bigger signal.
• 22:17
  Karpathy walked into a pre-training role built around using Claude to improve Claude.
• 22:21
  If recursive self-improvement is even partially happening, that's the org chart that proves it.
• 22:26
  Alibaba did something neither lab did: shipped a model and a chip in the same week, treating the agent frontier as a vertically integrated product.
• 22:34
  DeepSeek admitted, in a job listing, that the model alone is no longer the product.
• 22:39
  Railway's incident report is the reminder that the substrate Google is selling you on stage is the same substrate that took an entire PaaS down for eight hours yesterday morning.
• 22:48
  And Cheng Huang's writeup is the proof, on the same week as the ninety-three-agent demo, that the serious gains for builders right now come from contracts and property tests, not from spinning up more parallel agents.
• 23:00
  Hassabis's foothills line is selling the slope, not measuring it.
• 23:03
  The slope is there.
• 23:04
  The summit isn't visible from here.
• 23:06
  Tomorrow we'll see what reproductions of the Antigravity demo look like in the wild, and whether Spark's price tag turned into a number.
• 23:13
  Talk then.
• 23:14
  — Lenar Kess.