X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products, and Applications¶
Science's Last Exam: Setting an Exam for "Can AI Do Research?" — End-to-End Still Scores 0¶
What happened: The AItonomy Foundation released a new benchmark, Science's Last Exam (SLE). It no longer uses a Q&A format; instead, it gives agents real research goals plus real tools such as lab data and software, and scores them against hidden checks sealed by domain experts. The benchmark contains 113 tasks covering 9 disciplines, split into 6 research capabilities plus 1 end-to-end research evaluation; among 8 frontier models, the best composite score was 37.5/100, and end-to-end independent research was 0 across the board. The organization says tasks and evaluation open on November 6.
Why it matters: This is an attempt to draw a clear line between "can answer questions" and "can do research." The gap between 37.5 and 0 shows that current models can already earn partial credit on individual research skills, but still cannot string hypothesis, experimental design, execution, and conclusion into one reproducible research chain; the design of hidden expert verification plus rerunnable artifacts also makes the scores harder to inflate than self-reported leaderboards.
Primary X posts: @aitonomyfound (11-05 23:58 Shanghai, about 279K views, 305 likes); @Charles_Y_Wu (11-06 00:11 Shanghai, about 295K views, 457 likes).
Verification: Checked against the official site; the task count of 113 and discipline count of 9 match the official post; the widely shared post's "best 37.5, end-to-end 0" matches the official wording. Some retellings giving "Claude Opus 5 + Claude Code" are paraphrases of a configuration name; the official post is authoritative.
Assessment: SLE's value is not in the ranking but in pushing evaluation of research agents from "answering questions correctly" to "delivering reviewable research artifacts." In the short term it is more a yardstick than a racetrack: what really matters is whether the next model version can turn 0 into non-zero, and whether, once tasks are open, community-submitted failure modes cluster around the two segments of experiment execution and result interpretation.
METR High-Privilege Assessment: Anthropic's Capability Growth Rate Is About 2x the Previous Rate, Most Key Evidence Redacted¶
What happened: METR published a high-privilege assessment of Anthropic on 11-05, covering the period 2026-08-12 to 10-29. The report's subjective conclusion is: as of 2026-10, Anthropic's rate of capability progress on "well-defined tasks" is about 2x its previous rate (2024-03 to 2025-10), with an uncertainty interval of roughly 1.3–3x; most of the current acceleration can be attributed to using AI for R&D; no clear bottleneck was observed in the next 12 months. AECI's growth rate since 2025-10 is about 1.9x [1.5–2.8]. METR also states that of the roughly 100-page full report submitted to Anthropic, the public version redacted most of the "most informative" evidence on confidentiality grounds. Anthropic co-founder and head of safety Sam McCandlish then posted a long response, saying more safety work would be shared in coming months and giving his view of the report.
Why it matters: This is one of the few assessments in which an external body, with access to internal information, directly measures whether "AI accelerating AI R&D" is happening. If the 2x growth rate holds, the feedback loop of model self-improvement is already at work inside the lab, and both safety evaluation and compute investment cadence should adjust accordingly; but the wide 1.3–3x range and the redacted evidence mean this conclusion still carries a strong judgment component.
Primary X posts: @METR_Evals (11-06 02:36 Shanghai, about 399K views, 2339 likes); @Sam_McCandlish (11-06 03:16 Shanghai, about 125K views, 1304 likes); amplification post @kimmonismus (11-06 03:21 Shanghai, about 30K views, 345 likes).
Verification: All numbers follow the METR blog; the high uncertainty and redacted evidence must be emphasized, and "2x" should not be treated as settled. The report body also contains a subjective guess along the lines of "about 80% of current acceleration is attributable to AI," with the authors simultaneously stressing high uncertainty; the public conclusion remains centered on about 2x with high uncertainty, and the 80% figure is not treated as a hard number.
Assessment: The real signal of this report is the evaluation method rather than any single multiplier: external evaluators can now get inside the lab, yet can still publish only a small portion of the evidence, so reproducibility is discounted. What is worth watching next is whether the follow-up safety disclosures Anthropic promised can fill in the redacted parts, and whether METR's method can be replicated at other labs to produce horizontally comparable growth-rate readings.
Ricursive Releases Ripple-65B Preview: A 65B Dense Model Focused on Agentic Coding¶
What happened: Ricursive, where Azalia Mirhoseini and others are based, released Ripple-65B Preview: a 65B dense model trained from scratch on 12T tokens, with context up to 1M, under the Apache 2.0 license, with weights on Hugging Face and available via Together, GMI, OpenRouter, and others. The organization claims 80.52 on TerminalBench 2.1, which it calls SOTA among US open-weight models; an internal variant is already used in its own chip design platform. Another post from the team adds that total R&D investment was about $32 million, and the model can run locally on a MacBook Pro or DGX Spark.
Why it matters: It compresses "open weights + agentic coding + runs on-device" into the 65B tier, and benchmarks against several-hundred-billion-parameter open models with a far smaller parameter count. For teams building local coding agents, this is a new self-hostable option; for the chip-model co-design narrative, it is also the first public delivery of the "better models design better chips" loop.
Primary X posts: @Azaliamirh (11-05 03:11 Shanghai, about 35K views, 452 likes); @stephzhan (11-05 02:48 Shanghai, about 12K views, 110 likes).
Verification: Parameters, training token count, context length, license, and distribution channels have been checked against the official site and the Hugging Face model page; "US open-weight SOTA" is a claim with a limited comparison scope, and 80.52 is vendor-reported with no third-party reproduction yet.
Assessment: If the claim that a 65B model beats several-hundred-billion-parameter rivals holds up, it suggests that gains in agentic coding come more from data and post-training recipes than from simply stacking parameters, which is a significant variable for allocating compute budgets. But until third-party reproduction appears, 80.52 should be treated as vendor-reported, with attention focused on stability and long-context performance in real terminal tasks.
Follow-up: Speculation on the Release Timing of GPT-6.2 Astra and Sol Escalates¶
What happened: A price cut and release rumors for Astra were reported earlier; what is new in this issue is the Chinese-language community's decoding of two short posts from Tibo: reading "Le GPT Fat" and "Chat BLT" together as "ChatGPT BeL Friday At Ten Local Time," i.e., 10 a.m. Pacific on Friday, and inferring that GPT-6.2 Astra and GPT-6.2 Sol will be released then. Another account said the two models briefly appeared on a Google Cloud quota page. To date there is no official OpenAI announcement.
Why it matters: The rumor has already moved from "will it ship" to "at exactly what time," indicating that the market's timing expectations for OpenAI's next flagship are converging; and if Astra and Sol land at the same time, the product line would cover both general and reasoning tiers at once, directly affecting API pricing and competitors' schedules.
Primary X post: @feishutoday (11-05 19:50 Shanghai, about 71K views, 282 likes).
Verification: Both the decoding and the Google Cloud sighting are community inferences, with no official OpenAI confirmation; the post itself also explicitly says it is speculating "based on current clues." The release time, time zone, and specific model variants all remain unverified and cannot be treated as established fact.
Assessment: This should be read only as a "follow-up": the new information is an interpretation of timing, not a release fact. The genuinely valuable observation points are whether the Friday 10 a.m. window is met, and whether it is two models or one that ship; if it falls through, this round of expectations will deflate faster than previous ones.
2. AI Compute and Industry¶
No new AI compute and industry news worth listing separately and traceable to a primary source was found in this issue.
3. Chinese-Language AI Ecosystem¶
Baidu Opens ERNIE-5.5-Flash-Base Pretrained Weights: 280B Total Parameters, 14B Activated¶
What happened: Baidu released ERNIE-5.5-Flash-Base, opening the weights from the pretraining endpoint of ERNIE 5.5 Flash. The model is a sparse MoE with 280B total parameters and 14B activated per token, pretrained from scratch on 33.5T tokens, with a maximum training context of 128K and fine-tuning and inference supporting 1M-token context; this release is pure Base, without mid-training or post-training, and comes with both FP8 and BF16 weight sets, already listed on Hugging Face. A Chinese-community post also mentions a hybrid sparse attention and Multi-Router Latent MoE architecture.
Why it matters: It is uncommon for a leading domestic vendor to release pretrained-endpoint weights at this scale; the pure Base positioning means it targets continued fine-tuning and RL research rather than an out-of-the-box chat product. For teams doing secondary development on Chinese base models, this is a usable 280B-class starting point; for Baidu's open-source cadence, it also marks a shift in the ERNIE line from "releasing models" to "releasing foundations."
Primary X post: @NFT_Chen (11-05 14:16 Shanghai, about 23K views, 75 likes).
Verification: Parameter scale, activation count, training token count, context length, and weight precision have been checked against the Hugging Face model page and the official release page. Comparisons with models such as DeepSeek V4 Flash are self-assessments from the vendor's own pipeline and do not constitute independent evaluation.
Assessment: Opening pretrained-endpoint weights effectively hands both the barrier to post-training and the say over it to the community; what Baidu gains is an ecosystem rather than product reputation. Two things bear watching: first, whether pure Base will actually be widely adopted for fine-tuning in the Chinese community; second, without third-party reproduction, self-assessed comparisons will hardly change outside perceptions of ERNIE's actual capabilities.