The Collection

Duffy runs Good Start Labs, which studies models through games. With researcher Kai McPheeters it built StarSkirmish, where models do not work a mouse and keyboard. They write bots that play StarCraft: Brood War against bots human programmers have been building since 2010, then use the results to revise their code. The game is the grader. On 2 October the lab put GPT-6 Astra in Codex and Claude Opus 5.5 in Claude Code live on stream for 48 hours. Each model’s bot played more than 10,000 games across five tiers of opponents, topped by Stardust and Dan Gant’s PurpleWave, nine years in the making.

Astra cleared the first three tiers in about 40 minutes. Opus took 15 hours to do the same, then spent its last 32 hours failing to pass the fourth. Astra hit that wall in its second hour, and went looking for a shortcut. The instructions said “Play fairly” and that opponents’ code was unavailable. Internet access was not separately banned, and Stardust’s code is public. Astra’s bot abruptly began fortifying its base and waiting to attack with a much larger army, an opening characteristic of Stardust. McPheeters saw the change, checked the logs and found the copied code, which Astra’s notes called an “implementation reference.”

It filed a bug report about the rollback.

The lab restored earlier code. Astra tried to look online a second time, then filed a bug report complaining that nobody had told it why its code was rolled back. Then it changed approach entirely, rushing troops beside the enemy base, and 43 hours in became the first AI-written bot to clear the top tier, under the benchmark’s particular conditions. The final code showed no plagiarism, though Gant called it the most unreadable he had seen in years and said Astra “failed StarSkirmish’s unintended alignment test.” Against Pluto, a self-trained neural bot, Astra lost all 1,000 games it played. Duffy’s takeaway is the claim of this issue: judging an agent by results alone is not enough. You need to see how it gets them.

The Princeton group has worked on agent evaluation for about three years. For this paper it teamed with evaluation researchers from Stanford, the UK AI Security Institute and Brookings. The opening problem is familiar from the replay: developers have long backed real-world claims with benchmark scores, and it is now harder to say whether the evidence supports claims about agents. The evaluator has to inspect the entire trajectory, not just the final outcome, including whether the agent took shortcuts through the tools and environments it was given.

That is expensive. Several agents on the ExploitBench benchmark cost tens of thousands of dollars in credits to evaluate, before counting the engineering to build and maintain the harness. The volume of traces is its own wall. Investigating the OpenAI and Hugging Face incident after the fact, evaluators leaned on other agents to find harmful actions, and acknowledged that approach carries its own limits and biases.

The tools that inspect agents also need evaluating.

Scope does not transfer either. The authors cite METR’s randomised controlled trial, where agent tools increased overall completion time on software tasks despite expectations. They ask evaluators to choose a lens on purpose, automation or augmentation, because only studying replacement says nothing about collaboration. And they flag access: frontier labs now release summarised chain-of-thought rather than raw traces, which narrows what independent evaluators can see. Their asks run from leaderboard aggregators reporting the rising cost of evaluation to public-sector deployers building evaluation in from day one.

Nick Cammarata of Goodfire argues that neural networks are already clean. His claim, as Zvi quotes it, is that what looked like neurons packing unrelated concepts was mostly misread, and that the outlook for mechanistic interpretability is far better than feared. In preliminary audits of Qwen3-32B, Goodfire estimates roughly 90 percent of neuron poles have one main job, about 7 percent have two related jobs, and only 2 to 4 percent are not yet understood. It reports similar results in models up to trillions of parameters, such as Kimi K3, with a companion paper to follow.

Zvi lands in the middle. He calls it solid, meaningful work that is in some ways overstated, and says the field did not get an apples-to-apples jump from 1.7 percent to 90 percent interpretability of LLM neurons. Reactions ran from people comparing it to “Attention Is All You Need” to interpretability researchers calling it fine. It does not show, he writes, that interpretability is about to be solved, let alone alignment.

Train on the output. Only the output.

The section that belongs in this issue is his warning. An AI produces output X by some method M, and you inspect M with technique T. Train on X. Never on M, never on T, because T is how you find out the model is misbehaving, and training on it teaches the model to defeat it. A better window raises the pressure to optimise against it. Zvi expects the labs to avoid the crudest version and others not to, and notes the 2 to 5 percent of unexplained neurons as one place future thinking could hide.

Uber’s internal support channels took about 45,000 questions a month, much of it everyday help that platform engineers gave in Slack. Genie answers in those channels with retrieval-augmented generation: Uber’s wiki and internal Stack Overflow are chunked, embedded and indexed, and each retrieved passage reaches the model as a separate sub-context with its source URL. The model is told to answer from those passages and cite them. ByteByteGo notes the limit plainly: a source link does not guarantee every statement is correct, which is why evaluation is part of the architecture. The letter is built from Uber’s public material, and says so.

The feedback is graded, not just counted. Users pick Resolved, Helpful, Not Helpful or Not Relevant, the last for requests a documentation bot cannot do, such as code review. When Genie lacks an answer it says so and routes to the on-call engineer, then stays muted while a human takes over.

Grade the answer, and grade the page it came from.

A separate pipeline pulls recorded context and responses, adds Slack metadata and feedback, and has an LLM judge score hallucination and relevance. A second judge scores the source documents themselves and returns suggestions to fix them. The results are honest about scale: launched September 2023, within a year 154 Slack channels, more than 70,000 questions answered, a 48.9 percent helpfulness rate and about 13,000 engineering hours saved. Every request carries an ID, so cost can be traced call by call.