the launches were loud; the plumbing was the story
Week ending October 3, 2025; archive coverage: September 30.
Another week, another wall of launch confetti. Sonnet 4.5 dropped, everyone posted their favorite benchmark, and the vibes cycle spun up right on schedule. But if you read past the hype reels, the actually interesting stuff this week was structural: how agents are wired, how the GPU math really works, what fine-tuning costs when someone measures it honestly, and who gets to decide what a generative model is allowed to eat.
That's the thread I want to pull. The launches are downstream of the plumbing. And the plumbing this week was unusually well documented.
the harness got a name
Anthropic quietly renamed the Claude Code SDK to the Claude Agent SDK, which is a bigger deal than a rename usually is. The claim is that the same harness powering Claude Code—file access, bash, subagents, MCP tools, auto-compaction of context—can power basically any agent. The design principle is blunt: give the model a computer and let it work like a person does, in a gather-context, act, verify loop.
The mechanism worth stealing is agentic search over semantic search. Instead of embedding everything into a vector store, you let the agent grep and tail its own filesystem, and treat folder structure as context engineering. Slower per query, but transparent and maintainable. The catch is permissions: you're handing an autonomous loop a terminal. The practical consequence is that your agent's reliability now lives in the verification step—linting, rules, screenshots, an LLM judge—not in the prompt. Build the eval harness first.
the math under the latency
If you want to know why any of this costs what it costs, Aleksa Gordić's anatomy of high-performance matmul kernels is the antidote to benchmark theater. It walks Hopper from HBM to registers, through PTX and SASS, warp-tiling, TMA async loads, tensor cores, and swizzling.
My favorite moment: swapping % and / in a naive kernel—a trivial-looking edit—turns coalesced memory access into non-coalesced and drops throughput 13x, from 3171 to 243 GFLOP/s. That's the whole lesson. Without a hardware mental model, you'd never see it coming. And at scale the tail matters: as his friend puts it, you optimize in units of nuclear reactors, because 1% across 30,000 H100s is real money and real megawatts. This is a long read, but it's the kind of grounding that makes your cost and latency intuitions trustworthy.
fine-tuning, measured
Thinking Machines' LoRA Without Regret is the honest-economics piece of the week. Their finding: LoRA matches full fine-tuning on sample efficiency and final loss, provided you get two things right. Apply it to all layers—especially MLP and MoE, not just attention—and stay out of the capacity-constrained regime where your dataset outgrows the adapter.
The consequences are concrete. LoRA runs at roughly two-thirds the FLOPs, trains on a layout barely bigger than inference, and multiplexes many adapters on one server. For RL it's almost free: they argue policy gradients absorb about one bit per episode, so even rank-1 has plenty of capacity. The catches are real but small—LoRA tolerates large batch sizes worse, and its optimal learning rate is about 10x FullFT's. If you're doing post-training on a budget, this is permission to stop over-provisioning.
the models are earning their keep
Anthropic's building AI for cyber defenders is the rare capability post backed by outside evals instead of self-reported vibes. Sonnet 4.5 hits 76.5% on Cybench with ten attempts, and on CyberGym reproduces vulnerabilities in two-thirds of programs and finds new ones in a third—for about $45 across 30 trials. The number that should stick with you: repeated sampling changes everything. A motivated attacker isn't capped at $2 a task.
The honest note is that they deliberately trained for defense—finding and patching—and avoided exploitation and malware. Whether that boundary holds under general capability gains is the open question. But the message stands: don't cede the AI cyber advantage to attackers by default.
the power move
And the wildcard. Per the WSJ, OpenAI's next Sora will generate copyrighted material unless rightsholders opt out. Opt-out-by-default is not a technical decision; it's a distribution-of-power decision. It shifts the burden onto studios and agencies to police their own likenesses, and it sets a provenance and consent precedent everyone shipping generative systems will have to reckon with. Bold, and legally spicy.
what I’d do Monday
- Build your agent's verification loop before its capabilities. Linting and rules beat an LLM judge; wire real permissions on that terminal.
- Prefer agentic filesystem search first; only reach for embeddings when you actually need the speed.
- Default to LoRA on all layers for post-training, start the LR near 10x your FullFT value, and don't bother with high rank for RL.
- Run your security evals many times, not once. Single-shot numbers hide the motivated-adversary case.
- If you ship generative media, audit your provenance and consent posture now—opt-out norms are forming without you.