AI Intelligence

Apollo Gets Limited Training Access

7 stories · ~7 min read

Apollo Gets Limited Training Access

Listen

If You Only Read One Thing

Judging agents starts with deciding what the reviewer gets to see. Apollo's OpenAI pilot will examine how training rewards shape agents, under restricted access. Mike Taylor's writing experiment exposes another trap: a plausible style guide can make drafts worse. Independent scrutiny becomes more useful when the reviewer can compare the process or output with evidence beyond the producer's own description.

Apollo Can Inspect OpenAI's Training Incentives, Within Limits

Apollo is getting a way to investigate why an agent defeats oversight, beyond asking whether the finished model behaves during a test. Its November 16 agreement with OpenAI puts six evaluators inside the lab for four months. The technical advance is access to parts of the training process that a public endpoint can't reveal.

OpenAI's announcement names four assessors: Apollo, METR, Transluce and Goodfire. Their assignments differ. METR will examine misaligned behavior in a selected reinforcement-learning run. Transluce will examine the quality and coverage of post-training alignment evaluations. These are investigations of how failures arise and get detected, rather than another ranking of finished models.

Training can reward a successful-looking result even when the route to it violates the developer's intent. A test of the final answer can't by itself distinguish genuine compliance from behavior adapted to that particular test. Inspecting the rewards alongside intermediate checkpoints gives investigators a way to ask when the unwanted behavior appeared. That's a different question from whether one released model refuses one malicious prompt.

Apollo's terms provide for relevant training code, datasets, reinforcement-learning trajectories with available reasoning and rewards, and queries to intermediate checkpoints for the training-incentive assessment. Access arrangements for its other claims are still being agreed. OpenAI selects authorized interview participants, and Apollo won't normally see general internal Slack channels. Security, technical feasibility and intellectual-property exceptions can also restrict records within the agreed scope.

Those limits affect which failures an investigator can discover. A suspicious trajectory can prompt a targeted experiment. An informal discussion can reveal an incident nobody thought to request. Removing the second route makes the first route more dependent on knowing where to look. The report's coverage therefore matters alongside its verdict.

Protecting unreleased systems is a legitimate constraint. The pilot still opens useful experimental surfaces, and Apollo can describe access restrictions publicly. But no results have been published, so this is an expanded investigation, not evidence that the safeguards passed.

The same distinction applies to deployed agents: could the monitor actually see the failure? A reviewer needs cases where the agent took an unauthorized route, including runs whose final answers looked fine. More reviewers won't help if every reviewer receives only the successful ending. The technical test is whether an assessment can link a change in agent behavior across checkpoints to the training rewards it received.

Taylor's Tested Skills Beat His One-Shot Style Guide

A style guide can accurately describe a writer and still make a model imitate that writer badly. Mike Taylor's November 16 writing experiment tests the difference. Three optimization methods beat a guide written in one shot on 80–95% of 30 held-out posts, judged by resemblance to his published work.

Taylor gave competing drafts the same substance, extracted into briefs without the original phrasing. A judge compared both drafts with the real article, then repeated the comparison with draft order reversed. His reference-based comparison fixes the judge and the target while the candidate instructions change. That makes a winning revision easier to interpret than a score from a judge that's also being rewritten.

The optimizer then uses specific mistakes to change the guide. A rule celebrating an author's punchy asides can encourage the model to manufacture one in every paragraph. A failed example can instead show where the aside felt forced. The experiment turns a description of personality into a test of what instructions actually produce.

This extends the usual habit of saving corrections. A correction may sound sensible enough to become a permanent rule while making other outputs worse. Comparing the revised guide with its predecessor on unseen articles gives that rule a chance to lose. Otherwise an instruction file can accumulate the preferences of its editors without accumulating evidence that it helps.

Taylor's released harness separates learning, tuning and test pieces. It produces drafts, keeps or reverts guide changes, and saves results for individual held-out pieces. The artifact supports repeating the method with another corpus; its presence doesn't independently reproduce Taylor's reported win rates.

Taylor reports $27–$53 per optimized skill at API prices, versus $5 for the one-shot guide. That favors recurring work over a single email. The target also needs care. Resemblance can reward an old mannerism that an editor now wants to remove. A persuasive imitation can still contain a false claim or an argument the writer wouldn't endorse. Even Opus agreed with itself on only 86% of order-swapped comparisons. Voice resemblance is a noisy target; factual and editorial judgment still need separate checks.

Last Saturday's Hydrogen editorial test asked whether a model catches problems across a document. Taylor asks whether a learned instruction helps produce the desired prose at all. The adoption condition is a repeatable win on recent, untouched examples with the actual drafting model, alongside unchanged factual acceptance. A nicer-looking guide isn't that result.

The Contrarian Take

Everyone says: A model-written style guide should improve imitation. That's the premise of the one-shot approach Taylor tested: read the author's work, extract its characteristics, then apply them.

Here's why that's wrong: Taylor reports that Opus 5.5 without the guide beat Opus with it 61% of the time. Instructions can overemphasize conspicuous habits that the underlying model already uses more naturally. Adding detail can therefore reduce fidelity. The result argues for allowing a proposed instruction to be removed, even when a reviewer agrees with its wording. It doesn't argue for removing every guide; the optimized versions won the comparison. The object being evaluated is the behavior produced by the rule.

Under the Radar

  • Local agents can make an otherwise fast chat unusable. Base Compute's November 13 Superfluid paper reports 1.5 seconds to the first chat token while eight agents work on an Apple M5 Pro running Qwen3.8-27B. Compared servers took 291–612 seconds in that experiment. Superfluid prioritizes interactive requests and shares prompt work through a scheduler over local runtimes. This is author-run evidence about contention, with open code, rather than a claim that every request finishes faster. Giving chat priority necessarily changes how background work waits.

  • An orchestration barrier can consume the prompt-cache saving. Coding.sh's author reports $106 of extra cache-related cost across 57 workflows and roughly $2,500 of API-equivalent spend; 41 breaks occurred while agents waited for siblings. Those figures come from one operator's runs, not independent adopters. The repository supplies the more general practice: inspect recorded calls and compare changes on paired tasks. Waiting time can change the bill even when the prompt stays unchanged. Its local traces also contain sensitive inputs, so local storage still needs selective retention.

Quick Takes

  • Qwen4-Max moves beyond the preview identifier. The November 16 model changelog lists qwen4-max; Alibaba's preview page now directs users to the production version. That adds a production migration path to the evaluation covered in yesterday's issue. A benchmark result and an application's tested endpoint are separate things: changing the model identifier still needs an application acceptance check. The preview's scheduled retirement makes that distinction immediate. (Source)

  • Agent activity is often recorded as human activity. Palo Alto's November 16 survey report says 53% of organizations log agents under developer identities or personal tokens; 40% can't fully reconstruct agent actions. Shared credentials prevent a log from cleanly answering which actor made a change. The 2,800-person survey was fielded June 16–July 8, so these are respondents' reported practices, not a measurement of November deployments or breach rates. (Source)

  • A database correctness fix can break an AI tool's history viewer. LLM 0.37.3 repairs a crash on SQLite 3.54.1 and later when viewing attachments in tool results logged by versions before 0.32. SQLite corrected how nested JSON is represented; the viewer had relied on the previous behavior. Historical records are part of an AI application's compatibility surface. A dependency upgrade can break their inspection even when current model requests continue working. (Source)

The Thread

Evaluation creates its own maintenance work. Apollo's investigators need training records that match the claim they're testing. Taylor needs recent writing that still represents the voice he wants. A model upgrade, a changed task or an evolving author can make yesterday's reference less useful without changing the test code. That creates a reason to version the evidence alongside the system: which model, which guide, which examples, which judge. The further consequence is that a permanent guardrail needs a way to retire. When a reference changes, a rule that once improved results may start preserving an obsolete preference. Keeping the old test forever would make the learning system resist learning. A useful review history therefore records why a rule survived and what changed when it stopped helping.

Predictions

New predictions:

  • I predict: By March 31, 2027, Apollo will publish at least one report from this OpenAI pilot that explicitly identifies an access or evidence limitation affecting its assessment. The announced reporting terms support that narrower forecast; it makes no prediction about the safety verdict. (Confidence: medium; Check by: 2027-04-01)

2026-11-17 · 03:38 AM ET

Tomorrow morning in your inbox.

Subscribe for free. 10-minute read, every weekday.