Daily Executive Summary

The day now totals 12 new events and six material updates. This incremental run added Qwen-Image-2.1-Turbo, Youtu-Parsing-Omni, two agent-monitoring studies, a statistical review of METR time horizons and Ai2's GPU-scheduling case study. Claude Code, Codex CLI and GPT-6.1 Sol received material updates. Earlier work added ExpDis, SciExam, RunningTab, Docker Agent, Whistle and an explicitly ignored GLM-5.3 derivative.

Models

  • New · Qwen-Image-2.1-Turbo (TRIAL): Public 7B weights generate and edit images in eight denoising steps, but the model card has no comparison with the base model. Denoising steps are repeated passes that turn noise into an image. Fewer steps improve speed but may hurt text rendering and edit fidelity, so test latency, memory and quality directly.
  • New · Youtu-Parsing-Omni (TRIAL): A 5B model parses documents, charts, geometry, audio and video into structured JSON. Omni parsing uses one model across modalities instead of routing each input to a separate stack. The hard part is independent validation: evaluation code is still TBD, so use a mixed internal sample.

Agent & AI Engineering

  • New · SciExam for ENSO (TRIAL): Instead of giving an agent a known answer, SciExam asks it to build a climate model from real observations, then uses hidden tests for statistics, unobserved variables and held-out years. A hidden grader is an acceptance program the agent cannot inspect. The hard part is preventing answer-targeting while proving that held-out data did not leak. Public Docker, grader and examples make a one-day regrade feasible.
  • New · RunningTab (WATCH): It places requirements, provenance-bearing excerpts and unopened-file candidates in a runtime-maintained tab, then blocks finishing while obligations remain unresolved. An environment-side record means the runtime preserves state instead of relying on model context. The paper has no public code, so its claimed reduction in omissions cannot yet be verified.
  • New · OnTrack (WATCH): It compares live trajectory structure with successful runs. The authors report about one millisecond per step and roughly 18% less failed-run compute. Only six aborts triggered and no code is public, so it remains a watch item.

Open Source

  • ExpDis publishes Apache-2.0 code and checkpoints. SciExam publishes MIT-licensed code, Docker and a grader. Both have runnable artifacts, but their results remain primarily author-reported.
  • New · Whistle (TRIAL): A 16.9 MB Apache-2.0 model provides seven-language transcription, word timestamps and speech embeddings on CPUs. Its author benchmarks need replication, but the public runtime supports target-device tests of WER, latency, memory and thermals.
  • New · GLM-5.3 E224 (IGNORE): A community derivative targets two DGX Spark systems with a 141 GiB pruned and quantized model, but accuracy and throughput were measured only on B200. Target-hardware claims remain estimates, so popularity does not justify investment.

Research

  • New · ExpDis (TRIAL): Novelty-rewarded explorers search for different solutions, then correct trajectories are distilled into a separate student. This avoids letting sparse rewards distort the production policy. pass@k measures the chance that at least one of k samples is correct. It depends on both correctness and diversity, so single-sample accuracy is not enough.
  • SciExam moves research-agent acceptance from matching a known answer to passing unseen scientific tests. RunningTab moves completion state outside model context. They are independent signals, but not enough for a new trend.
  • OpenAI Math Results (update / WATCH): A sign error invalidated one construction and two dependent manuscripts, leading to three withdrawals. Fourteen other manuscripts were repaired and thirteen updated dependency citations. Lean coverage reached 300/719, but roughly 42% coverage also shows that formal verification remains incomplete.
  • New · Caught in the Act (TRIAL): White-box activation probes reportedly reach 98.8% AUC on SHADE-Arena and detect hidden goals that text alone cannot reveal. Activation probes read internal model states rather than only outputs. They need white-box access, and cross-model transfer plus false positives still require testing. FIBS data, code and probe weights are public.
  • New · METR time-horizon review (WATCH): The analysis finds that moving from three to 30 minutes and from 30 minutes to five hours are not equally difficult despite both being tenfold gains. Do not extrapolate automation progress from one exponential curve; wait for replication code and a METR response.

Developer Tools

  • Claude Code 2.1.294 (update / ADOPT): Fixes instruction-written prompt and agent hooks that could allow commands they should block, and reduces premature Stop/SubagentStop termination. Replay both allow and block cases after upgrading.
  • LiteLLM 1.104.2 (update / TRIAL): The stable line gains OpenAI-compatible /v1/decisions, /v1/systemone and a Decisions provider. Verify schemas, authentication, fallback and spend attribution in staging.
  • New · Docker Agent 1.149.0 (TRIAL): The Apache-2.0 runtime defines multi-agent workflows in YAML and supports MCP, RAG, evaluators and OCI distribution. This release adds GitHub skill loading and an OpenAI Decisions evaluator; compare it with an existing two-agent workflow in one day.
  • Claude Code 2.1.295 (update / ADOPT): Command and HTTP hooks now fail closed when configured with onFailure: block; the gateway gains per-upstream model allowlists, TTFB limits and request IDs. Replay hook failures, model denials and MCP reconnects after upgrading.
  • Codex 0.162.0 (update / ADOPT): Stable enables managed-worktree creation, task pinning and ranked tool discovery, and fixes cross-platform sandboxes plus CRLF preservation. Parallel coding teams should regress worktree ownership and policy boundaries.

Infrastructure

  • New · Ai2 GPU scheduling (WATCH): Ai2 replaces static priority queues with hierarchical, budget-weighted fair sharing. Thirty days of production data report 98% delivery of owed GPU hours, 98% occupancy and a drop in debug-job P90 wait from two hours to 30 seconds. The implementation is not public; reuse the metrics and minimum-runtime contract, not the scheduler itself.

Business & Policy

No business or policy event materially changed model access, API costs, chip supply or the open-source landscape.

Trend Signals

  • Trend #3 gains signals from Claude Code 2.1.294 and Docker Agent 1.149.0: coding-agent runtimes are repairing hook boundaries while integrating multi-agent work, evaluation and OCI distribution into products.
  • Trend #6 gains counterevidence: three withdrawals show that public Lean artifacts do not mean every critical claim is formalized or its dependency graph resolved. Status remains emerging.
  • Trend #7 gains a policy-boundary signal from the same event: natural-language hooks can still fail open, so deterministic runtime controls must make final allow or deny decisions.
  • LiteLLM's Decisions API is one implementation update and does not change Trend #2's status.
  • Trend #3 gains a stable Codex 0.162.0 signal: coding runtimes are bringing worktrees, task organization and tool discovery into one control plane.
  • Trend #7 gains three monitoring signals from Caught in the Act, OnTrack and Claude Code 2.1.295. Monitoring expands from output text to internal activations and execution trajectories, but the runtime should still make the final deny decision.
  • This cycle produced no evidence-backed new trend.

Tech Radar

  • New · ExpDis (research / TRIAL): Public code and checkpoints support a one-day comparison of pass@k, accuracy and solution diversity.
  • New · SciExam for ENSO (research / TRIAL): Public Docker and grader support checks of hidden splits and held-out-year isolation.
  • New · RunningTab (research / WATCH): The architecture is clear, but wait for public code or a harness before assigning a POC.
  • Claude Code (developer-tools / ADOPT): After upgrading to 2.1.294, replay allow/block hooks and verify that long-running work is not stopped early.
  • LiteLLM (ai-engineering / TRIAL): Validate Decisions schemas, authentication, fallback and spend attribution in staging.
  • New · Docker Agent (developer-tools / TRIAL): Use an existing two-agent workflow to test OCI reproducibility, MCP isolation and secret handling.
  • New · Whistle (open-source / TRIAL): Test real accents, noise, silence and keywords on the target device.
  • New · GLM-5.3 E224 (open-source / IGNORE): Its two-DGX-Spark target remains unmeasured; do not fund an expensive replication yet.
  • OpenAI Math Results (research / WATCH): Withdrawals widen the verification boundary; wait for independent dependency review and complete Lean proofs.
  • New · Qwen-Image-2.1-Turbo (foundation-model / TRIAL): Compare eight-step speed, memory and quality directly with the base model.
  • New · Youtu-Parsing-Omni (foundation-model / TRIAL): Public weights support a one-day mixed-input test, but evaluation code and license need review.
  • New · Caught in the Act (agent-security / TRIAL): Public data, code and probe weights support tests of AUC, false positives and cross-model failure.
  • New · OnTrack (ai-agent / WATCH): Wait for public code and a larger abort study.
  • New · METR time-horizon review (research / WATCH): Wait for replication code and a METR response before changing roadmap assumptions.
  • New · Ai2 GPU scheduling (infrastructure / WATCH): Simulate fairness, debug P90 and occupancy on historical queues first.
  • Claude Code (developer-tools / ADOPT): Version 2.1.295 makes more hook, gateway and MCP failures fail closed.
  • Codex CLI (developer-tools / ADOPT): Version 0.162.0 brings managed worktrees, task pinning and tool ranking to stable.
  • GPT-6.1 Sol (foundation-model / TRIAL): With Ultrafast generally available, route by end-to-end latency and completed-task cost.

Current radar: ADOPT 10 / TRIAL 137 / WATCH 86 / IGNORE 2. See the site insights page for the complete current radar.

Worth Trying

  • Use the public ExpDis CPU tests and checkpoint to compare baseline accuracy, pass@k and solution diversity on a small math set.
  • Regrade one SciExam final run in Docker and inspect hidden splits, third-party data terms and held-out-year isolation.
  • In an isolated project, write allow and block hook cases for Claude Code 2.1.294, then run a long subagent with a Stop hook.
  • Call /v1/decisions through a LiteLLM staging gateway and record schema errors, authorization failures, fallback and spend attribution.
  • Rebuild one two-agent workflow in Docker Agent YAML and check whether configuration, tool authority and results survive OCI push and pull.
  • Test Whistle on accents, noise, silence and keyword bias on the target device, recording WER, first-token latency and thermals.
  • Run identical generation and editing prompts on Qwen-Image-2.1-Turbo and the base model, recording latency, peak memory, text rendering and edit fidelity.
  • Run Youtu-Parsing-Omni on document, chart, audio and video samples and check JSON validity, accuracy, latency and memory.
  • Replay benign and deceptive traces with the public Caught in the Act probes and record false positives, memory and cross-model failures.

Sources