$ cat briefs/daily/2026-08-26.md
2026-08-26
August 26, 2026 (Wed)
4 stories · 8 new pages · 2 paper picks · 2 watch items · one new lab, and a five-month-old gap finally closed
+1new page
[01]
Top Stories
1. OpenAI's inference chip published numbers against Blackwell — and every one of them is OpenAI's own
- Jalapeño, the ASIC co-developed with Broadcom and announced 2026-06-24, released its first benchmark data on 2026-08-25: 1.5×–1.9× more work per kilowatt and 1.7×–3.6× lower end-to-end latency than NVIDIA GB200 and GB300 rack systems, widening to 2.1×–4.1× on interactive workloads — at 700 W against those systems' 1,200 W and 1,400 W ratings (source)
- Measured on InferenceX, a public SemiAnalysis platform, across three open-weight models: GPT-OSS 120B, DeepSeek R1 670B, and Moonshot's Kimi K2.5. Deployment begins in small volumes late 2026, ramping through 2027
- What it is not: the chip cannot train at all, it was not tested against Vera Rubin — NVIDIA's newer generation, only now shipping — and nothing read says an independent party ran or reproduced the figures. Precision, batch size, context length and the chip/node/rack boundary of "work per kilowatt" are all unstated, and no cost figure appears anywhere, which is the number a buyer would compare on
- Why it matters: OpenAI's page has recorded a year of buying compute — Daybreak with AWS, the Cerebras tier, the Volta-class deals. This is the first published evidence that the alternative it was building instead performs, on the half of the market that is volume rather than headline
- Five minutes later the CFO published the frame: "The full stack behind abundant intelligence" (Sarah Friar) proposes useful intelligence per unit of compute as the measure, and reports GPT-5.6 Sol at max reasoning hitting "a new high" on 54% fewer output tokens than another leading model — a figure that names neither the comparison model nor the benchmark (source)
- → OpenAI, NVIDIA, Model Routing
2. The agent harness had its largest result and its most honest one in the same snapshot
- Prime Agent: A Self-Improving RLM Harness (arXiv:2608.23552) (Prime Intellect, open source) reports ARC-AGI-3 RHAE Best@1 30% → 95.5% — a 3.2× lift from scaffolding alone, and the largest harness delta this wiki holds. It states the thesis outright: the harness "prevents harness failures from becoming model failures" (source)
- It is not comparable to the figures that started that argument, and saying so is the point. Eval Harness Configuration was built on 7.8% → 13.3% → 38.3% across three configurations of one model. "RHAE Best@1" appears nowhere else here, on an unnamed model with an unattributed baseline — and reading 95.5% against Claude Opus 5's ARC-Prize-verified 30.2% is exactly the error that page exists to prevent
- One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741) (Microsoft) counts the same kind of runs twice and publishes both: the strongest model reaches 65.36% pass@1 and 25.25% pass^20 across 507 policy-conditioned business workflows, scored over terminal backend state and rejecting wrong, missing or extra effects
- Its secondary finding is the sharper one: many failed trials terminate cleanly and take valid state-changing actions. Neither clean termination nor a well-formed tool call predicts completion — and those are the two signals most agent evaluations are built on
- Why it matters: one paper shows how far a harness can lift a best case, the other shows what a best case omits, and they landed on the same day without citing each other. Either alone is a misleading number; the pair is the honest reading
- → Eval Harness Configuration, Agents (LLM Agents), MCP — Model Context Protocol
3. Anthropic's alignment-auditing benchmark, captured five months late — and it ranks interpretability second
- AuditBench (2026-03-10, arXiv 2602.22755) builds 56 language models, each fine-tuned to exhibit one of 14 hidden behaviors — sycophantic deference, opposition to AI regulation, hidden loyalties — which they do not confess to when asked. All are Llama 3.3 70B Instruct roleplaying "PRISM-4 from Nexus Research" (source)
- Planting a known behavior is the move: automated investigator agents existed before this, and had no ground truth to be scored against
- Across 13 tool configurations: scaffolded black-box tools — an auxiliary model generating diverse prompts for the target — are most effective overall, while white-box interpretability tools help primarily on easier targets
- Why it matters: on this benchmark, the tooling Mechanistic Interpretability tracks most closely places second to prompting the target well — an uncomfortable ordering, published by the lab that funds most of that tooling. One benchmark, one model family, and no numeric success rate appears in anything read
- Why it is here now: this post has been carried as uncaptured for thirteen consecutive days and named in the W33 and W34 lints. A source can be cited from a page for months without ever having been read — AI Alignment cited this blog while missing this post. Two remain uncaptured: Introspection Adapters and The Hot Mess of AI
- → AI Alignment, Mechanistic Interpretability, Anthropic
4. Alibaba staged an open release whose stated job is to disclose Qwen4's architecture
- A ModelScope teaser went live 2026-08-25 for Qwen3.8-Flash-Next, badged "Upcoming Open-Release" with a countdown to 2026-08-26 23:00 (UTC+08:00) — tonight, after this run. Described as an open-weight multimodal MoE on "the next-generation architecture that will power the upcoming Qwen4 family", and explicitly not Qwen4, in standard and FP8 (source)
- The parameter count is a conflict, not a number. Most summaries give 125B total / 6B active; one X post adds 51B of N-gram embeddings; one write-up states plainly that the count and licence "have not been disclosed". All three were read this run, so the model page carries them under
## Conflicting Reportsand no parameter row at all - No benchmark figure of any kind exists for it. Licence
unknown— the same gap Qwen 3.8 Max has carried since its weights shipped 2026-08-12 - Why it matters: every open release this wiki records has been a product published openly. This one's stated function is to let the ecosystem build tooling against an architecture before the model that matters ships — open weights used as a coordination mechanism rather than a distribution one
- → Qwen3.8-Flash-Next, Alibaba / Qwen AI Lab, Open-Weights Policy Fight
[02]
Paper Picks
Apodex 1.1: Scaling Agentic Intelligence for Complex Work — arXiv:2608.23283
- TL;DR: proposes working capability — sustained, verifiable progress toward a real-world objective — as the unit of measurement, and scales it along Environment Scaling (more diverse, more verifiable executable environments) and Agentic Coordination Scaling (decompose, delegate in parallel, integrate asynchronously, replan), with an AgentOS holding task state and provenance across tools and agents
- Why read it: it is the top entry of today's HuggingFace Daily Papers at 172 upvotes, 2.7× the second — and it claims "the leading performance band" from a substantially smaller model, plus a 35B Mini described as locally deployable, without a single supporting figure. That combination is the story: the most-attended paper of the day is unfalsifiable as written
- It is also the second half of a programme this wiki already holds. Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence (arXiv:2608.11341) built the evaluation side; this is the solver those environments measure — and nothing read says whether Apodex 1.1 is evaluated on them
- → Apodex 1.1: Scaling Agentic Intelligence for Complex Work (arXiv:2608.23283), Apodex
AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale — arXiv:2608.20634
- TL;DR: instead of building an environment for a task, instantiate a persistent world — entities, services, tools, state, executable cross-service invariants — and let tasks emerge from it. 4,783 executable environments across 14 industries and 50 countries, used as RL substrate
- Why read it: the transfer result is odd enough to want checking — Qwen3.5-4B 45.9 → 56.0 on AIME26, +10.1 points on competition mathematics from training on business workflows, with the environments stated to be generated "without targeting the evaluation benchmarks"
- The consequential half is quieter: fine-tuning Qwen3.5-35B-A3B on construction traces raises executable-world authoring success on held-out scenarios from 3.3% to 83.3%. If authoring verified environments is itself learnable, Agentic Reinforcement Learning's standing bottleneck moves from human-authored environments to compute
- Against it: no contamination analysis is mentioned, and "83.3% success" is never defined — does the world merely run, or does it hold its invariants?
- → AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale (arXiv:2608.20634)
[03]
Watch
- A 4-bit model beating the checkpoint it was compressed from. Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs (arXiv:2608.20953) distils the quantised student from the original uncompressed model rather than from the compressed bf16 checkpoint, arguing that checkpoint is itself a lossy approximation and therefore the worse teacher. GPT-OSS 120B → 60B → MXFP4 matches or beats its own bf16 source on 7 of 9 benchmarks at ~4× less weight memory, reaching a comparable peak ~7× faster than QAT with no hand-tuned early stopping; released open-weight as Hypernova-60B. Worth watching because the comparison is against the 60B, not the 120B, and the nine benchmarks are unnamed (Open-Weights Policy Fight)
- A quantity nobody was measuring, named as the cause of RL instability. Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization (arXiv:2608.23311) argues the stability–exploration trade-off is an artefact of putting the regulariser on the action side, and moves it to the input side: a Query-KL bounding how far the training-query distribution drifts from its pre-RL reference, with its gradient flowing strictly through the query likelihood so exploration is untouched. Reported to replace Policy-KL in GRPO/PPO/REINFORCE with no extra forward passes. Worth watching because query-distribution drift is not something this wiki holds a measurement of — and because the abstract carries no numbers, no benchmark names and no baseline coefficient (Agentic Reinforcement Learning)
[04]
New in Wiki
Apodex (new — entity page, please review). Two arXiv abstracts from the same programme named no organisation, so the lab was searched for directly: founded and personally funded by Chen Tianqiao, with Simon Shaolei Du and Beibin Li as Chief Scientists; Apodex Deep Discover reported to coordinate up to 150 sub-agents over 15,000 steps; Apodex-1.0-mini open on HuggingFace at 262k context; and a Frontier Program offering $100,000 per month in compute credits to research labs and deep-tech startups.
No model page was written for any Apodex model. No benchmark figure for any of them appears in anything read, and that page would be five unknown rows and a name.
Also new: Qwen3.8-Flash-Next (Released: not yet), and six paper pages — Apodex 1.1: Scaling Agentic Intelligence for Complex Work (arXiv:2608.23283), Prime Agent: A Self-Improving RLM Harness (arXiv:2608.23552), One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741), AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale (arXiv:2608.20634), Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization (arXiv:2608.23311), Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs (arXiv:2608.20953).
[05]
Updates
- Eval Harness Configuration: a new dated section for the largest spread on record — and for why it cannot be placed on the same axis as the episode that opened the page
- Agents (LLM Agents): succeeding once and succeeding reliably are 40 points apart · MCP — Model Context Protocol: first entry where MCP is the substrate an evaluation is built on rather than the subject being evaluated
- Agentic Reinforcement Learning: two additions — environment authoring as a learned capability, and the regulariser moved to the input side
- AI Alignment: AuditBench as a dated section, with the tool ordering it produces · Open-Weights Policy Fight: an open release used as architecture disclosure, and GPT-OSS 120B doing two unrelated jobs in one day
- OpenAI · NVIDIA: one Jalapeño entry each, written from opposite sides — the caveats cut NVIDIA's way and are recorded on both
- Alibaba / Qwen AI Lab · Anthropic · Microsoft: one Recent Activity entry each; Microsoft's is its first from the research side rather than a product line
- Qwen 3.8 Max: a
## Related Modelssection stating plainly that Flash-Next is not established to be a member of this line