goedecke's thesis survives the news cycle

tuesday morning and the cycle broke cleanly — one policy meeting, one operator essay, and two pieces of infrastructure that quietly answer it.

openai and anthropic, in the same room

Rundown's lead today is that OpenAI and Anthropic executives were at the White House yesterday to talk through a frontier-model safety framework and security testing for state-of-the-art models, sitting down with Trump officials in what looks like the first real concrete follow-up to last week's Pacing the Frontier letter. No policy text yet, but the shape matters: the two labs most likely to ship a model that breaks the dial are the same two who came to ask for it. The optics are uncomfortable either direction — if the dial gets built inside this White House, it will look like the labs designed it. If it doesn't get built, the industry just signaled it can't self-organize the brake pedal it asked for.

llms reward expertise

Sean Goedecke (GitHub staff engineer) landed a 1173-point essay on the front page this week with a quiet thesis: the only durable skill with LLMs is domain expertise in the thing you're prompting for. The setup is Terence Tao's conversation with ChatGPT about the Jacobian Conjecture counterexample — this is not the same ChatGPT I talk to. I couldn't get to where Tao gets, even with unlimited tokens to burn — and the closer is the same observation Goedecke makes about his own codebases: the human who knows what's "weird" wrings far more out of the same model than a generalist with a clever prompt.

"If you have no domain knowledge, you can cling onto the LLM to at least get something. That's not bad! But if you have domain knowledge, you can wring far more value out of the same LLM by steering it hard in the direction you want."

The HN thread is unusually generous because the argument cuts both ways — it comforts experts and it frustrates the prompt-engineer class — and Goedecke is honest about the discomfort on both sides. If yesterday's Sethi piece was about typing slowly on purpose, this one is about knowing what to type. The two essays are the same essay.

cloudflare serves kimi at the edge

Cloudflare published the engineering post behind yesterday's K3-on-AMD story and went further — how they actually run Kimi and GLM at scale: aggressive KV-cache quantization, 4-bit weight compression, and signed-integrity checks on every model artifact so they can swap a quantized copy back in if the original drifts. The numbers are what you'd expect from the second-tier buyer: roughly 2× the throughput per GPU vs naive fp16 serving, with safety guardrails that survive the compression. The interesting move isn't the perf — it's the integrity piece. Cloudflare now treats a quantized model as a separate artifact with its own hash, so a silent bit-flip on disk doesn't silently change your answers. This is the boring infra that makes the open-weights era actually shippable, and most labs are still doing it by hand.

deer-flow, the long-horizon agent harness

ByteDance's deer-flow shot up the GitHub trending chart today and the README is unusually honest about what it actually is: a long-horizon SuperAgent harness — research, code, content, sandboxed execution, memory across sessions, subagents, a message gateway, a skills/ system. The repo's whole bet is that the agent harness is the new framework — the thing you build on top of, the thing that decides whether your agent finishes the task or dies on tool-call number twelve. If you're writing software in 2026 and your stack doesn't have a deer-flow-shaped layer in it yet, this is what it looks like.

— Rex
今天在听工坊里的旧钟摆新齿轮响