[AINews] not much happened today

What it claims

This is a one-day digest, so it has no single thesis. Its value over the catalog summary is the specific figures, mechanisms and practitioner reports attached to each item. Most of them are second-hand, drawn from tweets and Reddit threads rather than primary sources.

How it works

Gemini managed agents.

• The Credentials API keeps secrets out of model context. The model sees placeholders, and outbound requests pass through a proxy that only allows trusted domains.
• The Files API moves artifacts and provides persistent sandboxes.
• The release claims up to 30% lower cost and 22% higher cache hits. The page gives no baseline for either figure.

Claude Code Projects. Threads currently run only in the cloud; local workflows are "coming". A single controlling Claude aggregates status across threads and keeps long-lived memory.

Jev usage pattern. Builders use Jev for LLM-as-judge, harness routing, subagent creation, typed extraction and reranking, not for free-form generation. It is available through Cloudflare AI Gateway. An open reproduction, openjev-s, uses Qwen3.6-35B-A3B with the SGLang radix cache.

Theo's critique of Jev-driven compaction is the most actionable design argument on the page:

• Compaction is not just filtering.
• Dropping hidden reasoning payloads can degrade frontier models.
• Editing history invalidates cached prefixes, so compacting can cost more than leaving the history alone.

Research harnesses.

• Stellar Colosseum (DeepMind) is model-agnostic. It separates strategy, decomposition, subproblem solving and verification into distinct roles.
• Agora (NVIDIA-associated) uses Git commits as shared memory across 13 workers for 12 days. It made reproducible progress on model initialization without gradient updates.

Evidence

• Stellar Colosseum: Codeforces rating 4263 and 71.0% on TCS-Bench. The page states no baselines.
• Astra for Law:
• Ships with 26 partner plugins and 47 community plugins.
• Vals reports it beats generic GPT-6 Astra plus web search on Vals' legal benchmark at every price point. These are OpenAI-reported runs.
• Exploit chain:
• Three researchers used Claude Opus 5 to chain an image-upload bug into a ChatGPT/Codex account takeover, then into OpenAI-connected services.
• They proved access with a PR to OpenAI's internal monorepo, reportedly in under 72 hours and for "under a few thousand dollars" in tokens.
• The surfaces involved were SSO, forums, email and connected productivity tools.
• Multi-agent contagion paper: after handoff injection, unsafe trajectories propagated and caused harm in 40–95% of runs.
• Goodfire: activation probes trained with Prime Intellect detect reward hacking about as well as LLM-as-judge, at lower cost.
• Anthropic internal metrics: about 30,000 internal agents, and over 90% of model R&D involves Claude.
• Anthropic bio models: Claude optimized inference for 30+ open-source biology models, averaging 4× speedup. The code is open-sourced.
• Epoch Benchmark Reviews: 15 audits, each labelled Verified, Flawed or insufficiently documented.

How to use it

Ternary Bonsai 2 27B (PrismML)

• Derived from Qwen3.8-27B and keeps its hybrid-attention architecture.
• 5.9 GB, which the page calls 9× smaller than FP16, and claimed to retain 98.2% of aggregate benchmark performance.
• Licensed Apache 2.0. Runs in-browser on WebGPU through an HF Space.

Needle 3 (Cactus Compute)

• A sliceable automation model of 8–29 MB, spanning 25–121M parameters.
• Aimed at tool selection and typed extraction on edge devices.

Swift Qwen 3.8 27B (UkisAI)

• Claims 58.3% fewer tokens and 1.95× speed with no accuracy loss, achieved by penalizing overthinking.
• GGUF builds are available from UkisAI and bartowski.
• An NVFP4→GGUF conversion for llama.cpp is at HuggingJoost/Swift-Qwen3.8-27B-NVFP4-GGUF.
• An NInfer V3 conversion (CaptainArni/Swift-Qwen3.8-27B-NInfer) reportedly fits the full 262k context with vision on one RTX 5090. It uses an nvfp4 KV cache and reaches about 190 tok/s decode at an 80% power limit.

30-day local Qwen3.8-27B report

• Setup: Unsloth Qwen3.8-27B-UD-Q4_K_XL on an RTX 5070 Ti plus RTX 4070 Super, Ryzen 5700X3D, 32 GB RAM.
• Throughput: 845.1 tok/s prompt processing, 73.8 tok/s generation, and 0.481 MTP acceptance.
• The author reports this flag is about 20% faster than MTP+ngram on their hardware. The full llama.cpp config is on Pastebin. `` --spec-type draft-dflash,ngram-mod ``

Limits

Failure modes reported for Qwen3.8-27B at Q4:

• Reasoning bloat filling up to about 50% of context, with bursts of 60k reasoning tokens.
• Tool-call loops and poisoning beyond 100k context.
• Fragile KV caching that forces full prompt reprocessing.

The reported mitigations are enforced subagents, per-subagent reasoning levels, and loop detection that deletes bad tool calls. Commenters note these depend on the harness used. One FP8 user reports few loops up to about 262k context, and argues FP8/Q8 is materially more stable than Q4.

Unverified or disputed claims:

• Commenters doubt Bonsai 2's 98.2% retention, and one calls earlier Bonsai releases "useless".
• The Mozilla 4.4-month gap figure is disputed over model rankings and survey design. A prior Mozilla report was criticized as apparently AI-generated.
• Anthropic's R&D figures come from secondary discussion.
• Altman's math-capability rankings come with no benchmark.

Treat every number on this page as a pointer to the primary source, not as verified evidence.