Karpathy → Anthropic, Gemini 3.5 Flash Bombs, and GitHub Investigates Its Own Breach

Claude Code & Anthropic Updates

Karpathy joins Anthropic — to use Claude to train Claude. The big tweet of the day, by a wide margin: Andrej Karpathy announced he's joined Anthropic ("the next few years at the frontier of LLMs will be especially formative"). 131k likes, ~19M views, ~7k replies in under a day. Anthropic's Nicholas Joseph then clarified what he's actually doing: "He'll be building a team focused on using Claude to accelerate pretraining research itself." swyx summed it up in two words: "rsi is here. jesus." Boris Cherny (reply), Jerry Liu, and Cat Wu all welcomed him publicly; trq212 added a terse "the future is bright, let's get to work." The peanut gallery's contributions ranged from "this is like KD joining the 72-9 warriors" to "aura loss of biblical proportions" to a career-trajectory roast tracing Karpathy from OpenAI founding member → Tesla (Musk, OpenAI founder) → OpenAI (Sam, OpenAI founder) → Anthropic (Dario, OpenAI founder).

Code with Claude London: self-hosted sandboxes + MCP tunnels. Anthropic used the London event to ship two infrastructure pieces for Managed Agents: self-hosted sandboxes (public beta) and MCP tunnels (research preview). The pitch is regulated/enterprise teams that can't ship their data to Anthropic-managed infra: "Run agents inside your own perimeter, with your security controls applied by default." 1.8M views on the announcement tweet; the most coherent reply was oxnw noting that self-hosted sandboxes change the calculus for teams "that couldn't trust the agent with real infra." Boris Cherny posted from the floor (155 replies, mostly Mythos-rumor fishing — no Mythos news landed).

Claude Design: token limits doubled across every plan. Anthropic announced (retweeted by both bcherny and trq212) that Claude Design's token allowance is now 2× across all tiers. No fanfare, just a quiet capacity bump.

Stainless joins Anthropic — mitsuhiko is unimpressed. Stainless, the codegen company whose SDKs ship to ~25% of professional developers (OpenAI, Gemini, Meta Llama Stack are all customers), is being acquired by Anthropic. Jessie Frazelle pushed back: "absolutely no offense to stainless, but we have a generator for our sdks for each programming language we maintain (we did this before AI!!), you don't need a whole fucking company for this." Armin Ronacher piled on: "Sonnet 4.0 (yeah, last century tech) built us an OpenAPI to Python and TypeScript generator last year. Total lines of code 2400 and hasn't been touched since." He followed up with a pointed observation about the discourse contradiction: "SDKs are too hard, you need a flimsy startup to generate you SDKs from an OpenAI spec and it's the end of software engineering because clankers can do everything." The contrarian read in his replies was that Anthropic is probably acquiring the customer relationships and codegen data, not the templates — Stainless's customer list is literally Anthropic's competitor list. Jerry Liu added a more direct concern: "do we have to maintain our SDKs manually again."

Models & Benchmarks

Gemini 3.5 Flash drops — and it's not flash, it's expensive. Google announced Gemini 3.5 Flash as their "most powerful model to date." The numbers, per Theo's reading: 3× more per token than Gemini 3 Flash, token-hungry (73M tokens vs. GPT-5.5-medium's 22M on the same eval), and ends up costing 2× more to run than Gemini 3.1 Pro. Cursor's bench landed even worse: Theo's reaction — "Oh my god it scored worse than Composer 2! Not even 2.5! And it cost 4× more to run!!! This might be the worst major lab model drop of all time. Llama 4 tier." (Reference: Cursor's evals page.) Replies are unanimous: ByteCrafter notes the cost kills Flash as the cheap fallback for multi-hop triage agents; grichadev reports 40% more expensive than Sonnet on Seer evals because it's "overzealous on tool calls and basically didn't wanna give up"; Neil Blake's joke landed best — "It's called Gemini 3.5 Flash because it's fast. If it were built to be cheap it would be called: Gemini-3.5-permanent-underclass."

Theo's $7,370 Google crashout video. Theo filmed and shipped a Google-critical video in under 2.5 hours, explicitly worried about another demonetization round (he claims his Gemini 3 / Antigravity video was throttled last cycle). On the side, he wrapped his "$10 per cancelled-Claude-Code-screenshot" donation drive: 737 replies → $7,370 to open source, asking the timeline what projects to back.

Simon Willison's measured take. While Theo was crashing out, Simon Willison published his usual sober notes on Gemini 3.5 Flash — same conclusion ("3× the price of Gemini 3 Flash") but framed as "Google are planning to use it for many of their own products," i.e. the price hike is upstream cost-shifted, not arbitrary. Worth reading alongside Theo's cost-vs-score chart showing 3.5 Flash sitting north of GPT-5.5-medium on cost while losing on score.

Agentic Coding & Agent Harnesses

Antigravity 2.0 — and Matt Pocock's open-source skill in the Google demo. Google Antigravity launched 2.0: standalone desktop app, multi-agent teams, scheduled tasks, native voice, one-click integration with other Google products. Matt Pocock noticed his own permissively-licensed /grill-me skill being used in the product demo — "Clearly Antigravity is not SOTA, still using /grill-me and not /grill-with-docs." He's explicitly not salty: "the skill is permissively licensed and the point is that anyone can use it/adapt it. Lots of folks realising that adversarial discussion is better than reviewing plans." Theo, meanwhile, used the launch to dunk: "Clearly these people haven't experienced the magic of Gemini 3.5 Flash Preview on High in the new Antigravity CLI." Reception was rough — an AI coach quoted by LLMJunky, Robin Ebers — "Antigravity 2 a blatant copy of Cursor/Codex and a really bad one at that."

Codex still the agent to beat. Theo: "Honestly I'm still really impressed with the Codex app. It works reliably. It adds useful features consistently. It has taste." Greg Brockman quoted him approvingly. Theo's subtext: "every other 'agentic coding app' is kind of trash right now. It's insane that only OpenAI is shipping something of this quality. It's clear they actually use it internally. I honestly believe T3 Code is the only thing close." Counterpoint same day from LLMJunky, who flagged a Codex regression: "Anyone else notice that compaction seems to lose more details than normal in Codex? It never seemed to matter before, but I'm seeing it frequently now." (Same complaint surfaced in this week's earlier roundups — appears to be a sustained quality slip, not a one-off.)

Matt Pocock's one-liner doing rounds. "Making implementation cheaper has made creating bad code cheaper. But it's also made creating good code cheaper. You just need to know the difference." Quoted at the InfoQ panel "How AI is changing software engineering" (Pocock, Mappletons, Sophie Koonin) under the banner "Tactical programming is gone, strategic programming is here." Pocock also flirted with — then killed off — an /auto-grill skill idea: "the more I think about it, the more I hate it."

swyx posts a 4-part AI SDLC recipe. Concise enough to quote in full (source):

  1. Have ~50 tests in place with instructions to add more, including "use computer vision to visually spot check design and ux issues on mobile/desktop/ipad/ultrawide."
  2. /plan break up & isolate hot paths for easier editing/reading; add logging and error boundaries while you're in there.
  3. With plan in hand: "you can break backward compatibility. first map out all the remaining work, then proceed on this next slice, do not stop until all work is done, periodically stop to commit, deploy and test."
  4. Periodically spot-check deployed functionality and /steer bugs as it goes.

The interesting bit is the visual-CV-spot-check instruction smuggled into the test memory — m13v_'s reply was sharper than the original: "Most teams treat memories as context padding when they're actually behavioral anchors. That instruction only sticks if the model gets feedback when it skips the visual check, not just when it runs tests."

Karpathy's HTML trick now on the Claude blog. Trq212 confirms "Using Claude Code: the unreasonable effectiveness of HTML" is now live on the Claude blog. (Referenced in last week's roundups; now officially Anthropic-blessed content.)

Retrieval & Memory

Jerry Liu polls the timeline: what's the actual SOTA for file search? Question: grep over filesystem, virtualized grep / BM25 over a db (Mintlify's approach), vector, hybrid, SQL, none, or all of the above? 53 replies, 37k views, no consensus. The most informed answers:

  • IndexFox (reply): "hybrid wins for us, but it matters less than how the content was chunked before it got embedded. Fixing chunking changes [results] more than swapping retrievers."
  • Shain Noor (reply): pure grep over filesystem in production with 10k users.
  • Levi (reply): "Actual grep + vector search over a db — we have a paper somewhat on this soon." Worth tracking.
  • Chris DiNicolas (reply) gave the most honest answer: "Often the use case defines the approach."

The takeaway, consistent with this week's earlier "grep wins retrieval" thread: the agent-coding crowd has converged on grep + BM25 + (optional) vector; the knob that actually moves results is chunking quality, not retriever choice.

ParseBench update. Separately, Jerry plugged LlamaIndex's ParseBench — "There are a lot of coding and reasoning benchmarks for AI agents, but not a lot for document understanding." Their argument: doc understanding is a prerequisite for downstream knowledge work and is undertested.

Security & Supply Chain

GitHub investigating unauthorized access to its own internal repos. GitHub disclosure: "We are investigating unauthorized access to GitHub's internal repositories. While we currently have no evidence of impact to customer information stored outside of GitHub's internal repositories..." Theo's response: "It would be really funny if Github itself got pwn'd by one of the NPM package takeovers." Replies were not amused — Adnan Khan flagged "99% sure this was due to Tanstack or Nrwl based on the timing", aligning with the ongoing Shai-Hulud / npm supply-chain story this column has been tracking for two weeks. Darshan Yadav's reply pointed at a different vector: "Not npm — a poisoned VS Code extension. Same supply chain logic, different vector. Dev tooling is the new perimeter. AI coding agents that auto-install extensions make this worse."

Google Cloud deletes Railway's account by mistake — again. Railway's Jake confirmed Google Cloud blocked their account. Theo noted "it's not the first time they've done it." Useful reminder when picking infra dependencies.

Other Notable