X AI Daily¶
Translation
English translation of the Chinese original.
1. AI models, products and applications¶
Anthropic open-sources FlashProve: using zkSNARKs to prove that "this sentence really came from this model"¶
What happened: On November 3 Anthropic published a research article, "Progress on provable inference for large language models," introducing FlashProve, a prototype for verifiable inference. The idea is for the model provider to make a one-time commitment to its weights, after which every batch of responses comes with a zero-knowledge proof, so that a user or auditor needs neither the weights nor a GPU to confirm that the output really came from the committed model, while the proof does not leak the weights. The official test case is Llama 4 Maverick (400B-parameter MoE, 4-bit quantized, about 17B active per token): on a single H100, proving runs at about 110 tokens per second; the proof for a 1024-token request is 18.78 MB; a 24-thread CPU verifies it in 17 seconds (21 seconds single-threaded); the commitment file is about 6.5 MB. The offline commitment takes about 39 minutes once, and one commitment serves about 16 million requests. The price is compute: proving 24 requests on each GPU takes about 72 times as long as vLLM takes to process the same 24 requests (66 to 77 times across runs), and about 94 times when vLLM is at full load. The code, the Lean formal proofs and the documentation are open-sourced under Apache 2.0, with a technical report attached.
Why it matters: The previous public result on this track was VerInf, which needed 14.3 hours, a 93.6 GB proof and close to a day of verification for 1,000 tokens on the same model. FlashProve compresses proofs to the megabyte range and verification to seconds, and it is a pure software approach that does not rely on trusted hardware. The article is concrete about uses: a regulator can check only the proof, without seeing the weights, to confirm that traffic in a given period really ran on the declared model; a safety evaluator can confirm that it tested the very model running in production; a compromised inference server cannot forge proofs. On distribution, the official post drew 192K views, 2,756 likes and 236 reposts, with clear follow-up from the zero-knowledge community.
Key X posts: @AnthropicAI release post (11-04 04:10, 192K views, 2,756 likes, 236 reposts), and partner @SuccinctLabs congratulatory post (11-03 23:26, 11K views, 143 likes, 26 reposts).
Verification: The 110 tokens per second, 18.78 MB, 17-second verification, the 72 times figure (66 to 77 times) and 94 times, the 39-minute offline phase, the roughly 6.5 MB commitment and Apache 2.0 were each checked against Anthropic's official article; the repository anthropics/FlashProve was confirmed through the GitHub API as Apache-2.0, with the description reading "A research prototype for zero-knowledge proofs of AI inference. Not maintained.", and about 347 stars as of 11-04 07:50. Two retellings on X need correction: @Prathkum said "last time 14 hours and a 93 GB proof, this time a few seconds and 19 MB"; 14.3 hours and 93.6 GB do come from the official description of VerInf, but "a few seconds" is the verification time, not the proving time (24 requests take 223 seconds); @bittingthembits's "about 72 times the GPU time" matches the official figure. The limitations the vendor itself states are also listed: the proof covers only context within 1024 tokens, only the integer version of the model at 4-bit weights with greedy decoding, the model architecture itself is not hidden, and the fast C++ verifier carries no formal guarantees.
Assessment: This is a prototype that runs, not a service that can ship. The overhead is still close to two orders of magnitude, and what is proven is "integer-version greedy decoding," which the vendor acknowledges carries drift risk against production floating-point, sampled decoding. Its value is that it turns "verifiable deployment" from a slogan into an engineering artifact with concrete numbers and Lean statements, rather than an immediately usable compliance tool.
OpenAI pushes ChatGPT and Codex into CAD, PCB and simulation: a device designed in a week, with mass production still done by people¶
What happened: On November 3 OpenAI published a research article saying ChatGPT and Codex are being used for mechanical design, electrical engineering and simulation: editing CAD, drawing PCBs, producing manufacturing drawings, running design reviews. The main example is two employees using GPT-6 Astra to design a test device for talking to Codex, handing the enclosure, internal structure, circuit board layout, interface and firmware entirely to the model; the work was done within a week and a contract manufacturer built a small batch of prototypes within a month. OpenAI describes the test as "seeing whether a device can be built without any direct human modification." But it also concedes that going from design files to mass production still requires "several rounds of review and iteration," and that the model is unreliable on a manufacturer's process limits, tolerances and assembly details. Other figures: since February this year, model usage by OpenAI's hardware team, measured in tokens, has grown 23 times; on BenchCAD 2.0, an unnamed new internal model averaged 0.875 and GPT-6 Astra 0.594. On partnerships, the company expanded its work with the Chip Ganassi Racing R&D team to turn 3D scans of existing parts into editable CAD and two-dimensional manufacturing drawings, and added a research partnership with Mazda on aerodynamics and crash-safety simulation.
Why it matters: This is the first time coding agents have been systematically pushed into a stage where mistakes cost tooling and tape-out money. The vendor's own boundary is clear as well: the model can produce a design within a week, but everything from design to a mold-ready part still falls to people. On distribution, the official post drew 414K views, 4,101 likes and 310 reposts, the highest like count among OpenAI's main account posts in this batch of material.
Key X posts: @OpenAI release post (11-04 02:56, 414K views, 4,101 likes, 310 reposts); an industry retelling is at @Sam_Badawi (11-04 03:18, 5,126 views, 12 likes, 3 reposts).
Verification: The prototype within four weeks, the 23-fold token growth, the two scores of 0.875 and 0.594, and the content of the Chip Ganassi Racing and Mazda partnerships were checked against OpenAI's official article (direct fetching of that page returns 403, so details follow Runtimewire's report and CellCog's item-by-item writeup, the latter citing the public BenchCAD 2.0 leaderboard: Astra 0.59, Claude Opus 5.5 0.57, per-instance cost 11.05 dollars against 27.58 dollars). Two points of wording deserve care: 0.875 is self-reported by OpenAI, the model is unnamed, and it does not appear on the public BenchCAD leaderboard; the official article's wording is that the design files were completed "without direct human modification," which is not the same as completing a mass-producible product with no human involvement.
Assessment: What is worth recording is the boundary, not the score. "A design in a week, a prototype in a month" holds at the level of design files; tolerances, process and assembly, three kinds of engineering knowledge, still have to be supplied by people. Treating 0.875 as evidence of the next generation's capability is premature, since the model does not even have a name.
OpenAI brings outside firms into its next security sprint and raises the top bug bounty to 172,000 dollars¶
What happened: OpenAI published two articles on November 2 (posted on X on November 3 at 10:29), announcing on one side that its next internal security sprint will be run with Cathedral Systems, whose researchers will attack OpenAI's systems together with OpenAI's own teams using OpenAI models; and on the other raising bug bounties: up to 30,000 dollars for critical vulnerabilities in core products such as ChatGPT, Codex and the API platform, and up to 172,000 dollars for "exceptional findings" including bonus awards, with the median of the medium, high and critical bounty ranges raised by an average of more than four times. OpenAI also disclosed how the previous sprint worked: 25 percent of product engineers were assigned to finding and fixing vulnerabilities, using Codex to investigate leads, reproduce issues in isolated environments and draft patches. The bounty programme also adds a quality multiplier, five tiers of loyalty bonuses and event bonuses, and applies retroactive top-ups: pending reports are paid under the new standard, and reports paid in the 30 days before the announcement are topped up to the difference.
Why it matters: The vendor's reasoning is blunt: "traditional cyber defence is no longer enough." Pulling a quarter of product engineers into security is an organisational move, not just more money. On distribution, the official post drew 743K views, 3,351 likes and 324 reposts, second among official posts in this batch by views.
Key X posts: @OpenAI release post (11-03 10:29, 743K views, 3,351 likes, 324 reposts); a security-community retelling is at @alexmichael (11-03 10:37, 1,175 views, 9 likes, 1 repost).
Verification: The 172,000 dollars, the more than fourfold increase, the 25 percent of product engineers and the Cathedral partnership can all be found in the wording of the two official articles, the security sprint announcement and the bounty adjustment announcement; a comparison of figures and examples is also in CellCog's writeup. Two points of precision need stating: "more than four times" refers to the average of the median values across the medium, high and critical bounty tiers, not to the top cap quadrupling; the 172,000 dollars is the result of stacking the quality, event and loyalty multipliers on a 100,000-dollar base for the "exceptional findings" tier, while the critical tier itself is capped at 30,000 dollars. @busenur038's retelling of "25 percent of product engineers last round" matches the official account, but that post has extremely low engagement (7 views, 0 likes).
Assessment: This is not a security-budget story but a signal that the process of "attacking your own systems with your own models" is becoming routine. The fourfold increase buys the attention of outside researchers, and the 25 percent engineer commitment shows the internal judgement is that capability is sufficient and what is missing is coverage.
OpenAI builds its own Git backend, Pando: push latency down about 80 percent, nearly 1 PB a day¶
What happened: OpenAI's developer account posted about Pando at 08:09 on November 3, a newly written Git backend for its monorepo, with no change to how engineers and agents use Git. The result is roughly 80 percent lower push latency, and the system now carries close to 1 PB of Git data a day; the official line is that it can "scale Git capacity alongside agent usage." The official long-form article adds detail: written in Rust, 10 weeks from project start to launch (into production in April), with the core designed by one engineer and the implementation done by a team of Codex agents; reliability of the backend had previously fallen below 90 percent; Pando drops two conventional assumptions, that all branches of the whole repository are consistent at a single instant and that a packfile is built on the fly for each client, and instead coordinates per branch and streams verification while writing blobs during a push, with the cache service hitting more than 99 percent; quarterly commit volume equals 1,378 percent of the first quarter of 2024, and the fourth quarter of 2026 is projected at 2,442 percent. Pando itself is not open source, and only some Git client improvements have been contributed upstream.
Why it matters: This is the first concrete case of agents overwhelming a piece of basic tooling, and more informative than the commit growth is where the bottleneck sits: not the Git protocol but the way conventional Git servers serve requests. On distribution, this post drew 1.33M views, 6,298 likes and 494 reposts, the highest engagement among these nine items, with almost no contentious discussion.
Key X post: @OpenAIDevs release post (11-03 08:09, 1.33M views, 6,298 likes, 494 reposts).
Verification: The roughly 80 percent lower latency (about 5 times faster), nearly 1 PB a day, Rust, 10 weeks, one engineer designing the core, production in April, reliability falling below 90 percent, the Express cache hit rate above 99 percent and the two commit figures of 1,378 percent and 2,442 percent were checked against OpenAI's official article (direct fetching is blocked with a 403, so the content follows The New Stack's report and Interesting Engineering's report). The two retellings agree: Pando is not open source; OpenAI says it is "studying ways to share the underlying technology with users."
Assessment: OpenAI's choice was to keep Git compatibility and replace only the server side, which shows that the branches and commits generated by agents have already exceeded the design scale of conventional Git servers, with the bottleneck appearing first in infrastructure rather than in the protocol.
OpenRouter lists stealth model Crouton Alpha, and OpenCode carries another free one, Pulse Alpha¶
What happened: OpenRouter listed the stealth model Crouton Alpha at 04:17 on November 4: free, a 1-million-token context, support for text, image, audio, video and file input, and a statement that user data is not used for training; in OpenRouter's API metadata the context is 1,048,576, with reasoning forced on and defaulting to the highest level. opencode then posted additional details: the model is free for a week, multimodal, and limited to certain regions. In the same week there is another free stealth model on OpenCode Zen, Pulse Alpha Free (1-million context, text plus image input), whose model page states it is unavailable in the European Union, Iceland, Liechtenstein, Norway, Cuba, Iran, North Korea, Russia, Ukraine and Venezuela, and notes that data collected during the free period may be retained to improve the model.
Why it matters: A stealth launch is a common move before a major release: it collects real traffic and feedback with a model that carries no vendor name while committing to nothing. Two free models with million-scale context going live at the same time turns "whose model is this" into an open guessing game: in the Chinese-language community some people compared fingerprints using vocabulary lists, the token formula for vision encoders and error message wording, claiming to rule out every known vendor; others judged it not to be a flagship model by generating low-polygon ponies as a test. These conclusions contradict one another and all rest on the model's own statements and error messages. On distribution, the opencode post drew 100K views and 2,046 likes, and the OpenRouter post 58K views and 826 likes.
Key X posts: @OpenRouter listing post (11-04 04:17, 58K views, 826 likes, 37 reposts), @opencode supplementary post (11-04 04:37, 100K views, 2,046 likes, 60 reposts); the Chinese-language fingerprint analysis is at @mylifcc (11-04 07:01, 465 views, 2 likes).
Verification: The OpenRouter model page and public API metadata confirm that it is free, that the context is 1,048,576, that input modalities are text plus image, video, audio and file, that reasoning is forced on, and that "EU routing is currently unavailable"; the OpenCode model documentation confirms that Pulse Alpha Free is a time-limited free stealth model, the regional restriction list, and that data from the free period may be retained for improvement. What could not be verified: who owns Crouton Alpha, its parameter count and actual output limit (X carries a claim of a 270K output limit, while the API's max_completion_tokens field is 943,718, which does not match), and whether it shares an origin with Pulse Alpha; the identifications of specific vendors on X have no official or reported support.
Assessment: The value of this item is the observation method, not the conclusion. The identity of a stealth model can only be settled when a vendor claims it; elimination by vocabulary lists and error messages can serve as a lead but is not enough to name it, especially when a vendor can switch backends by region and by batch.
Pruna turns MiniMax's video model into the open image model P-H3-Image, cutting the VRAM bar to 21 GB¶
What happened: Pruna open-sourced P-H3-Image at 01:01 on November 4, an image generation and editing model derived from the MiniMax-H3 video model. On efficiency it uses 8 sampling steps instead of 50: a 4MP image takes 2.6 seconds on a single B200 (down from 12.4 seconds), and 8MP takes 5.8 seconds (down from 30.6 seconds); the minimum weight set is 20 GB (down from 143 GB), and at 8MP the 4-bit peak VRAM is about 21 GB. The changes span the whole stack: a prompt amplifier based on Qwen3.5-9B, a text encoder distilled from 32B to 4B (from 67 GB to 9 GB), a diffusion schedule retuned for images, a transformer backbone compressed from 33B to 20B, and a retrained single-frame decoder. On capability it supports rendering from 2 to 16 MP and compositing and editing with up to 8 reference images (another passage of the blog says 9), with strengths in documentary and editorial photography, portrait skin texture, film stills and graphic design with short quotations; weaknesses are hands, small background figures and dense small text. The licence is the MiniMax H3 Community License, which explicitly excludes the European Union, the United Kingdom, South Korea and the United States.
Why it matters: Using a video model as an image base is a rare route among this round of open image models, and the reasoning holds up, since video models must learn materials, lighting and object rotation to keep frames consistent. The real change is the bar: from 143 GB down to 20 GB, with 8MP running inside 21 GB of VRAM, 24 GB consumer cards enter the usable range for the first time. On distribution, the official post drew 44K views, 784 likes and 89 reposts; ComfyUI announced native support the same day.
Key X posts: @PrunaAI release post (11-04 01:01, 44K views, 784 likes, 89 reposts), @ComfyUI support announcement (11-04 05:53, 5,477 views, 128 likes, 8 reposts).
Verification: 2.6 seconds against 12.4 seconds (4MP, B200, 8 steps against 50), 5.8 seconds against 30.6 seconds (8MP), 20 GB against 143 GB, 21 GB at 8MP in 4-bit, the 4B text encoder, the licence excluding the EU, UK, South Korea and the United States, and the requirement to label output "Powered by MiniMax H3" were each checked against Pruna's official technical article and the Hugging Face model card; the model card also states that bf16 at 8MP needs about 101 GB of VRAM (H200/B200 class) and about 80 GB with offloading. The number of reference images differs between two places (the official post and one passage of the blog say 8, another passage of the blog says 9). Among Chinese-language retellings, @ShenfuX's "about 76 GB for a standard diffusers pipeline" (11-04 06:04, 43 views) is close to but not the same as the model card's 80 GB offloaded figure, and the model card takes precedence; "5 times faster" holds only on a single B200 at 4MP with 8 steps against 50, and cannot be carried over directly to consumer graphics cards.
Assessment: This is a typical case of compressing a large model, and what is worth recording is the compression route rather than the speed multiple: distilling a 32B text encoder to 4B and replacing the timestep modules that account for four tenths of the transformer do not depend on specific hardware. The licence's exclusion of four major markets will materially limit commercial adoption in those regions.
2. AI compute and industry¶
The Information: Anthropic pitches "designing chips with Claude" to IPO investors¶
What happened: The Information reported on November 3 under the headline "Anthropic Touts Claude's Chip Design Skills to IPO Investors." According to the report, in recent communications with potential IPO investors Anthropic said a newly formed 15-person chip design team was already using Claude for engineering and design work on custom chips; the chips are for training and running Claude. The report cites four people familiar with the matter and says engineers estimated time savings by reference to the conventional chip design flows of Nvidia and AMD. Retellings on X went further, writing that it is "faster than Nvidia or AMD." As background, OpenAI took another, already executed path: on June 24 it unveiled a custom inference chip codenamed Jalapeño jointly with Broadcom.
Why it matters: Model companies have begun using "compressing chip design cycles with our own models" as a selling point to investors, which shows custom silicon has reached the stage where it can carry a story; it is also one of the few public cases of model capability being converted directly into hardware R&D efficiency. On distribution, @kimmonismus's retelling drew the most engagement, while @bearlyai's long post also lists the internal usage described by Nvidia chief scientist Bill Dally at GTC, which is background rather than content of this report.
Key X posts: @kimmonismus's retelling (11-04 02:23, 49K views, 754 likes, 25 reposts), a post carrying the original headline (11-03 22:04, 37K views, 17 likes, 2 reposts), a background long post (11-04 02:44, 19K views, 52 likes, 7 reposts).
Verification: The headline, the 15-person team, the four sources and the IPO context were checked against the original page at The Information (paywalled, with only the headline and summary visible) and against secondhand retellings: AI Weekly's summary and Investing.com's report. What must be flagged: the body text was not read, so "estimating time savings by reference to Nvidia and AMD" comes from secondhand retellings, and "faster than Nvidia or AMD" is a further reworking of those retellings that the original report did not write; Anthropic has published no material on this and has given no quantitative figures. The Jalapeño codename and timing were checked against CNBC's report.
Assessment: The information value here lies in "the audience for the story is investors," not in "a 15-person team." Fifteen people cannot build an advanced-node chip; what they can build is architectural exploration and validation of a custom accelerator. Reading it as "Anthropic wants to make GPUs" overestimates it; reading it as "custom silicon has entered the IPO narrative" is more accurate.
3. Chinese AI ecosystem¶
Tencent Hunyuan Hy4 general release still in teaser stage: the official account posted only "4 Real Productivity", and specs and vision capability rest on a leaked document¶
What happened: Tencent Hunyuan's official account posted a teaser at 20:15 on November 3 containing only "4 Real Productivity" and a video, drawing 229K views, 2,021 likes and 127 reposts. As of 07:50 Beijing time on November 4, the Hy4 general release had not shipped: no official launch post, no updated model page, and Tencent's line remains "coming soon." The verifiable official timeline is this: Hy4 preview was released and open-sourced on August 28; on October 29, at Tencent's Global Digital Ecosystem Summit, senior executive vice president Dowson Tong said the Hy4 general release was imminent, would introduce vision understanding to the series for the first time, would focus on strengthening long-horizon tasks and real productivity, would improve chain-of-thought efficiency, and would be followed by the cost-focused Hy4-Flash. Speculation in the Chinese-language community about the specifications comes from a closed documentation pull request: on November 2 someone submitted "Add Hy4 deployment guide" to the vLLM-Ascend repository, where the document states 770B total parameters, 49B active per token, 78 layers, 256 routed experts per layer with top-8, plus a native MTP layer of about 10B (about 0.7B active), notes that the parameter count "excludes vision components," and attaches a HallusionBench multimodal evaluation (1,129 questions, 178 text-only, 951 with images). The pull request was closed and not merged at 11:09 UTC on November 3.
Why it matters: The Wuzhen summit runs from November 2 to 5 in Wuzhen and opened on November 3, with AI-related sub-forums scheduled across the whole of November 4, and community expectations for a release are concentrated on these days. Vision understanding is the first modality this Hunyuan line has added, and the key step as it moves from the first tier of open models toward multimodality.
Key X posts: @TencentHunyuan teaser post (11-03 20:15, 229K views, 2,021 likes, 127 reposts); leaked writeups are at @NFT_Chen (11-03 02:38, 14K views, 54 likes, 7 reposts) and a follow-up post (11-03 20:44, 10K views, 81 likes, 6 reposts), with @leo114119 speculating about summit timing (11-03 20:51, 3,101 views, 35 likes).
Verification: The official teaser post and Tong's October 29 wording were checked against Technode's report; Hy4 preview was released and open-sourced on August 28, per IT Home's report and the official Hy4 preview page; the Wuzhen summit runs from November 2 to 5 and opened on November 3, per pre-event releases from People's Daily Online and Xinhua. The existence, content and closed status of the documentation pull request were confirmed through GitHub PR #20671 and the repository API as "submitted on November 2, closed on November 3, not merged," and the document does contain passages on 770B, 49B, 78 layers, 256 experts, the 10B MTP and HallusionBench. What must be flagged: the pull request was submitted by a community developer and is not official Tencent documentation, so the parameters and vision capability are claims from that source and Tencent has not confirmed them; the claim on X that "the PR was hastily closed by administrators to bury it" has no basis, and no reason for the closure was given.
Assessment: The teaser is real; the specifications currently have no official backing. Until the general release, 770B with 49B active, vision capability and the release date can only be treated as leaks and retellings; whether it ships during the summit is something no party has publicly promised.
QuestMobile autumn report: Doubao wins on frequency, DeepSeek on duration, and the two move in opposite directions month on month¶
What happened: QuestMobile published its China Mobile Internet Autumn 2026 Report on November 3. The AI application section uses per-user figures: in September 2026, Doubao was used 72.0 times and 132.5 minutes per user per month, and DeepSeek 65.9 times and 136.7 minutes, clearly ahead of other products; month on month the two diverged, with Doubao's sessions and duration down 6.1 percent and 7.8 percent respectively, and DeepSeek's up 5.8 percent and 9.5 percent. In the same data set, Yuanbao recorded 38.7 times and 78.6 minutes, Kimi 26.2 times and Wenxin 16.2 times. The report's own judgement is that "the battle for the entry point is nearing its end," with competition shifting from land-grabbing to managing high-frequency users. The site-wide figures are 1.281 billion monthly active users (up 0.9 percent year on year) and 189.8 hours per user per month (up 6.5 percent). The same report on the same day carried a related result: the free short-drama app Hongguo overtook Doubao with 430 million deduplicated users.
Why it matters: Frequency and duration measure how often an app is opened and how long people stay, which is a different calculation from installs and monthly actives and does not reflect acquisition ability, a point the report stresses. Only by reading the two together can the split between "entry-point" and "tool" products be seen: Doubao relies on frequent opening, DeepSeek on longer single sessions. On distribution, interpretation posts in the Chinese-language X community drew little engagement, with @NFT_Chen's writeup (11-03 19:20) at 5,487 views, 24 likes and 4 reposts, and @chenYijl's reading on Qwen's monthly actives (11-04 03:20) at 206 views and 3 likes.
Key X posts: @NFT_Chen's data writeup (11-03 19:20, 5,487 views, 24 likes, 4 reposts), @xiax0603's addition (11-03 20:52, 924 views, 11 likes, 3 reposts).
Verification: The 72.0 and 65.9 sessions, the 132.5 and 136.7 minutes and the four month-on-month figures were checked against 21st Century Business Herald's report citing Cailianshe (published 11-03 18:16), and "the battle for the entry point is nearing its end" is the report's wording as quoted there; Yuanbao's 38.7 sessions and 78.6 minutes, Kimi's 26.2 and Wenxin's 16.2 are in Beijing Business Today's report; the 1.281 billion monthly actives, 189.8 hours and Hongguo's 430 million deduplicated users overtaking Doubao are in The Paper's report. Direct fetching of the report page on QuestMobile's site returns a 403, so the original report could not be read and the above rests on reprints. Two things need separating: the figures of 17.8 sessions and 22.7 minutes for Qwen and 42.8 minutes for Kimi appear only in X retellings and were not found in public reporting, so they could not be verified; the September ranking of AI-native apps at 419 million, 158 million and 123 million monthly actives is a separate leaderboard QuestMobile published on October 20 and is not part of the November 3 report, and @chenYijl's cited 5,077 percent year-on-year growth for Qwen comes from that leaderboard (public reporting gives 5,076.6 percent).
Assessment: Looking only at session counts suggests Doubao wins, and looking only at duration suggests DeepSeek wins; both readings are right, and neither is an answer to "who is stronger." More noteworthy is the direction month on month: the largest player is falling and the second is rising. As for explanations such as "a rollback of subsidies caused the decline," those are the retellers' inference, and the report draws no such causal link.
Coverage note: The Information article is paywalled, the QuestMobile report page could not be read directly, and OpenAI blog posts were cross-checked against multiple reprints; the affected items are flagged.