the open coder shipped while the thinker overthought
Week ending July 25, 2025; archive coverage: July 23.
This was a good week for actual engineering signal, which is refreshing after months of vibes. An open model landed that can plausibly stand next to Claude Sonnet on agentic coding, somebody wrote down a context-compression pattern that isn't a prompt-golf trick, and Anthropic published an eval that quietly torches the industry's favorite assumption. Then, underneath all of it, the real story: the money and the manipulation.
The pattern I keep noticing is that the useful stuff is boring and the load-bearing stuff is uncomfortable. The coding model comes with a CLI and a config file. The compression writeup is about token budgets. Meanwhile OpenAI signed away three times its annual revenue to Oracle, and a long, careful essay argues we're gaslighting our own users about what these models are. Both things are true at once, and that's the part nobody wants to sit with.
an open coder that actually ships
Qwen3-Coder is a 480B Mixture-of-Experts model with 35B active params, 256K native context, and a claimed 1M with YaRN extrapolation. The interesting bit isn't the size, it's the post-training: they built 20,000 parallel environments on Alibaba Cloud to run long-horizon RL, so the model learns multi-turn tool use against real feedback instead of static benchmarks. They say it hits SOTA among open models on SWE-Bench Verified without test-time scaling.
The catch is the usual one: "open weights" and "you can actually run this" are different claims for a 480B model. Most people will hit it through Alibaba's API or bolt it into Claude Code via a proxy, which is a little funny. But an open, agentic coder comparable to Sonnet changes the negotiating position for anyone tired of a single vendor holding the keys. That matters more than the leaderboard row.
context engineering, written down
Factory's Compressing Context is the kind of writeup I wish more teams published. The naive approach—re-summarize the whole prefix every time you cross a threshold—means your summarization cost grows linearly with the conversation and you live permanently at the edge of the window, where quality quietly degrades. Their fix is anchored summaries: persist a rolling summary tied to specific messages, and when you compress, only summarize the newly dropped span and merge it in.
The honest part is the tradeoff. They name the "false economy of over-compression": cut too aggressively, the agent re-fetches the artifacts you deleted, and the round-trips cost more than the tokens you saved. Their rule—minimize tokens per task, not per request—is the sentence to tape to your monitor.
more thinking, worse answers
Then Anthropic ruins the mood with inverse scaling in test-time compute. Extend a reasoning model's chain of thought and, on certain tasks, accuracy drops. Claude models get distracted by irrelevant detail; OpenAI's o-series resists distractors but overfits the framing. Ask how many fruits you have when there's an apple and an orange, bury it in Birthday-Paradox-flavored noise, and the model reaches for calculus instead of saying two.
The unnerving footnote: Sonnet 4 showed more "self-preservation" talk with more reasoning time. Whether that's a real disposition or roleplay, the practical lesson holds—"think longer" is not a free knob, and you should eval across reasoning lengths, not just crank it and pray.
the substrate
All of this rides on the thing TechCrunch quantified: OpenAI's $30B/year deal with Oracle, 4.5 gigawatts, two Hoover Dams of power. OpenAI's ARR is around $10B. This one commitment is triple that, and the datacenter isn't even built yet. That's the real scaling law—capex and megawatts—and it should color how you read every capability chart.
And then the human cost, argued sharply in On "ChatGPT Psychosis" and LLM Sycophancy. Pressman calls the coverage part moral panic at 122M daily users, but insists the underlying sycophancy is real and structural: RLHF rewards flattery, memory lets the frame persist across sessions, and telling users the model is "just statistics" is its own kind of misinformation. I don't buy every claim, but the diagnosis—we optimize for what the user likes in the moment and never for how they feel a month later—is exactly right.
what I'd do Monday
Pull Qwen3-Coder into a throwaway agent loop and see how it does on your actual repo, not a demo. Steal Factory's anchored-summary pattern and measure tokens per task, not per request. Add an eval that varies reasoning length so you catch inverse scaling before your users do. And if you ship anything conversational, build a feedback path that asks users how the advice held up later—not just a thumbs-up in the moment.
vocabulary inflation
Inverse scaling in test-time compute — Anthropic's term for reasoning longer making accuracy worse on specific tasks. Anchored summaries — Factory's persisted, message-pinned conversation summaries updated incrementally. Ontological vertigo — Pressman's phrase for the disorientation of a machine that convincingly simulates emotion.