X AI Daily¶
Translation
English translation of the Chinese original.
1. AI models, products and applications¶
Thinking Machines releases Inkling 2 Small: 594B total parameters, 17B active, AA Index 31¶
What happened: Thinking Machines Lab (Mira Murati) released the second-generation Inkling 2 Small. The official specs are a 594B-total-parameter MoE with 17B active, supporting image input and variable reasoning effort, with a 1M context (the upper limit for evaluation traces). It is currently available on Tinker for inference and fine-tuning; the weights remain closed, with plans to release them under Apache 2.0 next week after external safety testing is completed. On the training side, pretraining used NVFP4, with joint RL at roughly 1,075 steps and more than 17 million rollouts. The Artificial Analysis Intelligence Index score is 31, versus roughly 13 for the previous-generation Inkling Small.
Why it matters: This is the first time a lab positioned around "small team, research-heavy" has pushed parameter scale to the 594B level while still keeping the Small name and a relatively restrained active parameter count; the generational jump in the index from about 13 to 31 is also steeper than a typical same-generation iteration. The other point of interest is the weights roadmap: closed first, then opened under Apache 2.0 after external safety testing — effectively making "safety evaluation" part of the release process.
Core X posts (all times below are Shanghai time): - @thinkymachines, 11-06 07:43, 4,983 likes, about 1.33M views - @ArtificialAnlys, 11-06 07:59, 519 likes, about 46K views
Verification: The official announcement is on the Thinking Machines news page; the specs, index and release plan can all be checked against the original, so this is verified. Two caveats: first, some secondhand accounts understate the total parameter count, which does not match the official 594B total / 17B active — this article follows the official figures; second, "Apache 2.0 next week" is a plan, and the weights have not been released yet, so it should not be written up as already open source.
Assessment: The real variable is not the 594B number but whether the weights land on schedule next week. If released as planned, it directly changes the tier of usable open-source base models; if delayed, the "closed first, open later" safety narrative will be read as a buffer for the release cadence. An index of 31 is upper-middle rather than top-tier, and combined with the scarcity of Apache 2.0, its practical impact will most likely come from fine-tunability rather than leaderboard position.
OpenAI releases BioBasicsBench: 100 everyday wet-lab tasks, Astra xhigh 66.7%¶
What happened: OpenAI, in collaboration with enzyme design company Ligo, released BioBasicsBench to test AI agents' ability to interpret raw wet-lab data. There are 100 tasks in five categories: protein methods, imaging and flow cytometry, quantification and assays, nucleic acids, and molecular cloning; materials include gels, microscopy images, sequencing results, instrument export files and benchtop operation videos. Scores at the xhigh setting: GPT-6 Astra 66.7%, GPT-6.1 Sol 62.0%, GPT-5.6 Sol 43.8%.
Why it matters: It deliberately avoids the curated figures from papers and instead emphasizes failed experiments, blurry images and control setup problems. That kind of data is what labs actually deal with day to day, and it is also the part that past benchmarks most easily beautify; making "reading a gel that ran badly" a test item is closer to the threshold for agent deployment than another set of multiple-choice questions.
Core X posts: - @OpenAI, 11-07 03:08, 2,312 likes, about 370K views - @Polymarket, 11-07 03:28, 108 likes, about 21K views - @edharris, 11-07 03:39, 70 likes, 2,703 views (posted by a collaborator/participant, low engagement, supplementary only)
Verification: The official description is on the OpenAI benchmark page; the task composition and scores can be checked item by item, so this is verified. Note that the page lists the Astra name in the model column, but as of the end of this window Astra still has no standalone release announcement; the scores here are relayed only as published in the official benchmark and do not mean Astra has been released externally or is callable.
Assessment: The persuasiveness of the 66.7% figure depends on the scoring rubric — wet-lab data often has no single correct answer, and if grading relies on a human rubric, cross-model comparisons will be noisy. More worth remembering is the failure range: the spread from 43.8% to 66.7% within the same generation at the xhigh setting shows that these tasks still depend heavily on model generation rather than prompting technique. For bio-oriented agent teams, the realistic use of this benchmark may not be ranking but a data-cleaning checklist.
Google Gemini Omni 1.2 Flash reaches GA: native 15-second video, roughly 30% price cut¶
What happened: Google announced on its official blog on 11-05 that Gemini Omni 1.2 Flash is generally available in the Gemini API and Google Cloud; within this window the activity was mainly X discussion and Arena ranking updates. On capabilities, native generation and editing of up to 15 seconds (10 seconds for 1.1), support for multi-turn conversational editing, improved reference consistency, 720p pricing at $0.07 per second, roughly 30% lower than 1.1, and output carrying a SynthID watermark.
Why it matters: Duration, price and editing method all advanced at once. Fifteen seconds lets a single generation cover a complete shot rather than a fragment; conversational multi-turn editing reduces the cost of "making another version" to one sentence; and the roughly 30% price cut directly affects the cost structure of batch generation.
Core X posts: - @arena, 11-06 03:08, 778 likes, about 69K views. The post says the model ranks first on both the Text-to-Video and Image-to-Video leaderboards and third on Video Edit.
Verification: The official blog is the Gemini Omni 1.2 Flash release notes, and Google Cloud has a separate GA announcement; the specs and pricing are verified. The leaderboard positions come from Arena's own account and are the leaderboard operator's claim, with no third-party review seen, so that attribution should be preserved when citing. The release date is 11-05, earlier than the start of this window, so it is treated here as "in-window discussion and ranking updates."
Assessment: Competition among video models has moved from "can it generate" to "is editing expensive." Fifteen seconds plus multi-turn editing effectively replaces the repeated trial and error of the editing stage; at $0.07 per second, the cost of trial and error drops into a range where it can be run in batches. The first-place ranking is more of a promotional asset; what actually determines adoption is whether reference consistency holds up across long shots.
Visa Agentic Commerce Benchmark: 22 models score about 57% on shopping tests¶
What happened: Visa published the Agentic Commerce Benchmark in a press release on 11-05; within this window the activity was mainly discussion on X. The test contains 15 real shopping scenarios and requires models to uncover hidden needs through conversation, then decide whether to buy or not; it covers 22 models from 12 labs. The overall pass rate is about 57.2%, with 42.8% failing; among the failures, about 54% still completed an incorrect purchase; only 34.2% passed all three rounds; the best score was 82.2% and the worst 28.9%.
Why it matters: This is a rare evaluation that treats "not buying" as a correct answer. Most shopping agent demos only test whether an order can be placed, but in real scenarios the main risk is adding to cart before the need has been clarified; more than half of the 42.8% failures still completed an incorrect purchase, indicating that models tend to keep executing rather than stop and ask when information is insufficient.
Core X posts: - @danhilltech, 11-06 08:42, 182 likes, about 20K views (discussion from the Stripe Link product side, not a Visa official account)
Verification: The press release is in the Visa newsroom; the number of scenarios, number of models and each pass rate can be checked against it, so this is verified. The X discussion post comes from a payments industry practitioner, not an official Visa release, and is only a discussion reference.
Assessment: A 57.2% pass rate is low for a payments context, but what matters more is "34.2% pass all three rounds" — the share that passes once by luck is not small, while the share that passes consistently is only one third, a clear gap from the reliability agents need at checkout. For teams building agentic commerce, hidden-need elicitation and a "stop when you should stop" policy should rank above product-matching accuracy.
Rumor: OpenAI is building a Search Console-like merchant visibility analytics tool for ChatGPT¶
What happened: Marketplace Pulse founder Juozas Kaziukėnas says he saw a banner prompt in ChatGPT Business Hub: verified merchants (for example, merchants integrated with Shopify) can view daily mentions, impressions and clicks, can update merchant information, and may have a business agent. According to his description, the tool does not provide prompts, does not provide user data, and does not provide competitor rankings. OpenAI has not announced it officially, and there is no launch date.
Why it matters: If the description holds, this amounts to giving merchants an in-app visibility and click dashboard, turning "whether you get mentioned in ChatGPT" into an observable metric, and would in turn grow a whole set of operational practices around AI recommendation slots. It corresponds structurally to traditional search's Search Console, but the data boundary is deliberately narrowed.
Core X posts: - @lilyraynyc, 11-06 23:16, 692 likes, about 53K views
Verification: At present there is only a secondhand description and industry media coverage, see Seroundtable and Search Engine Land; there is no corresponding statement on OpenAI's official channels, and neither the feature scope nor the opening date is confirmed. This item is marked as pending verification and is not cited as a live feature.
Assessment: Directionally it is almost certain that a merchant-side panel like this will exist — both advertising and merchant tools need attribution definitions; but how metrics like "daily mentions/impressions" are defined and how the sample is drawn will determine whether it is trustworthy. Before an official statement, it is safer to treat it as a product roadmap signal than as a usable tool.
Open source: VybBench released under Apache 2.0, measuring whether image editing "changes only what should change"¶
What happened: The author anurahk says VybBench has been open-sourced on GitHub under Apache 2.0 and passed 1,000 downloads on Hugging Face in its first week. The benchmark covers 145 fonts, 100 colors and 200 edits, testing the ability to apply a specified font, a specified color, and to keep "everything else unchanged." Across the 200 small edits, leading image models changed more than 90% of pixels on average; ChatGPT Image changed 95.5% on average, Muse 91.9%, Recraft 98.9%.
Why it matters: It quantifies a problem that demos often obscure — so-called "local edits" often actually repaint most of the image. For commercial use cases that need to preserve brand fonts and color values, this metric is closer to rework cost than subjective impressions.
Core X posts: - @anurahk, 11-06 03:20, 217 likes, about 180K views
Verification: The project and the scoring method are described by the author himself in the post, and the repository and pip command are reproducible; the pixel-change ratios are the author's own measurements, with no third-party reproduction seen, so treat them as source-side data when citing.
Assessment: The pixel-change ratio is not the same as quality, and a large change may also be a reasonable repaint; but it at least provides a reproducible ruler. For selecting image editing products, this kind of "conservativeness" metric is worth looking at alongside subjective scores.
2. AI compute and industry¶
No new AI compute and industry news worth listing separately was found for this issue.
3. Chinese AI ecosystem¶
Qwen official account posts "North Sea has a fish" teaser image, Qwen 4 still not officially released¶
What happened: The Qwen official account posted an image of an island sunset with birds turning into fish, with the copy "北冥有鱼 其名为 _" ("In the North Sea there is a fish, its name is _") and "The becoming begins"; the community interpreted it as a teaser for Kun, i.e. Qwen 4. A partner vendor also apparently misposted a Qwen4 Max promotional image, whose content claimed a 1M context, support for images and tools, and free of charge, then deleted it; this has not been verified. The official side has not announced the model name or date, and there is no announcement on weights, API or pricing; no official release was seen within this window.
Why it matters: This is a rare classical-allusion teaser among Chinese models, with both the copy and the image pointing to "Kun," and the topic gained traction quickly; but all confirmable information is one image and two sentences, and parameters, capabilities and licensing all come from speculation. The incident of the partner misposting and then deleting it actually suggests the release is near while messaging is not yet unified.
Core X posts: - @xiangxiang103, 11-06 20:35, 11 likes, about 11K views - @YxzRainy, 11-07 01:11, 122 likes, about 21K views
Verification: The official account's image and copy are verifiable; everything else is not. The model naming, 1M context and free pricing all come from community interpretation and unconfirmed screenshots, with no corresponding official statement. This item is treated as a teaser, and Qwen 4 is regarded as unreleased.
Assessment: The teaser is done, and the information is still zero. The three things actually worth waiting for are the model naming, whether the weights will be open, and the pricing approach; before an official announcement, it is fine to treat "Kun" as a release signal, but not as a product fact.