Liu Gong’s Weekly Review: The 10 Signals That Mattered in AI This Week (Oct 31-Nov 6)

Liu Gong believes the week’s real theme was not another parameter race but a simultaneous re-pricing of entry points and billing units. ChatGPT’s search query volume was put at one-fifth of Google’s; publishers took the fight to browser summaries; Nvidia moved the definition of a rendering generation onto its cloud service. Antigravity, ElectroChemLLM, StateBench and SRPI point to the same engineering line: agents enter production only when intermediate results are verifiable, state is reversible, and failure samples are amortizable. Liu Gong’s call: in the next six months, competition shifts from “can it be done” to “who owns the error, and what unit to bill by.”

Signal 1: OpenAI Puts ChatGPT Search at One-Fifth of Google’s Query Volume (11月3日). Twenty percent of queries is not 20% of information retrieval: Google still indexes hundreds of billions of documents, while ChatGPT runs a hybrid of licensed indexes and its own crawler, with visible recall gaps on obscure entities and breaking-news queries. Conversational search also splits one intent into multiple turns—traditional search averages one or two queries per need, while a chat product can trigger five or more rewrites and clarifications. Liu Gong thinks Google will disclose an “AI Mode” query share within six months; once advertisers accept the 20% figure, the negotiating anchor moves from cost-per-click to in-conversation exposure.

Signal 2: New York Times and Other Publishers Sue Brave Over Scraping and Paywall Circumvention (11月2日). This case will not end with Brave shutting off its crawler within 12 months; the more likely landing zone is a compromise built on content licensing plus output-side technical limits. robots.txt and RSL live on the server, but Brave’s summaries happen on the user’s device, where crawl requests are mixed into real reader sessions and are hard to separate from readers. The enforceable boundary is summary length, clickable links to the original, and the share of an article that can be quoted. Liu Gong expects a second major browser vendor to sign summary-attribution or licensing terms within six months, turning compliance cost into a competitive moat.

Signal 3: Google’s Antigravity Adds Chemistry Skills, Moving Agents Into Drug Discovery (11月3日). Within 12 months, fewer than five top pharma companies will actually plug a general-purpose agent into lead-compound optimization. The chemistry tasks agents do well share one trait: intermediate results are verifiable—can RDKit parse the SMILES, does the substructure match, does LogP fall inside the target range—so errors surface immediately and the loop closes. Antigravity’s real contribution is turning those deterministic functions into a service layer that natural language can schedule. Liu Gong thinks the “skill pack” will degrade, within 18 months, into version-controlled tool calls with unit tests; R&D wants reproducibility, not conversation.

Signal 4: ElectroChemLLM Targets Numerical Hallucination in Electrochemistry With a Multimodal Pipeline (11月4日). The winner here is not parameter count but whether extracted results can be audited. The paper pairs “context fragmentation” with “numerical hallucination” because they share a root: once a number leaves its table cell or curve coordinate, it loses its constraint and the model falls back on linguistic priors. The real point of multimodality is not reading another file format but tracing every extracted value back to a specific location in the source. Liu Gong believes that within 12 months, chart channels will move from a bonus to an admission requirement in materials and electrochemistry extraction pipelines; text-only architectures will be eliminated in the first acceptance round of property-database projects.

Signal 5: Nvidia Brings DLSS 5 to GeForce NOW, With Hogwarts Legacy at Launch (11月5日). DLSS 5 landing first on GeForce NOW rather than on a new local GPU shows Nvidia moving the power to define a rendering generation from the hardware launch to the cloud service. DLSS 4’s multi-frame generation inserts up to three frames around one real frame, with errors mostly along the time axis; once lighting and materials are also reconstructed by the network, errors spread into spatial dimensions and need priors such as depth, normals and material IDs. Liu Gong predicts 3D-Guided Neural Rendering will reach local RTX drivers within 12 months, but the cloud will keep a three-to-six-month exclusive window—and that window is itself a subscription-conversion tool.

Signal 6: Auditors Find AI Rewriters Hit Evasion Targets but Lose Factual Fidelity (11月4日). Within six months, at least one mainstream humanizer leaderboard will put factual fidelity on the main board rather than in a footnote; once buyers ask whether their numbers survived the rewrite, a ranking that only compares evasion rates loses its bargaining power. The paper’s audit standard is the useful part: numbers, dates, people and institutions, and quotations are four enumerable segment types that regex and entity recognition can catch with high confidence, and they are also the most costly errors. But hedges (“about,” “more than”) and reversed causal links fall outside the checker’s view. Liu Gong, drawing on industrial inspection practice, insists false positives and false negatives be reported separately; a precision range of 0.84–1.00 still leaves plenty slipping through in the worst case.

Signal 7: Anthropic Updates Claude Code Prompts to Ban Replicating Existing Games (11月2日). This will not stop anyone determined to clone a game, but it will become an industry norm for coding agents within 12 months. A prompt-level ban is a soft gate: Claude’s 200K-token context makes a few dozen tokens of constraint less than 0.02% of the window, with almost no effect on inference cost or latency, but there is no verification mechanism—if a user claims to own the original cartridge and ROM, the model may well let it pass. Liu Gong expects at least one major coding agent to add a similar clause within six months and to disclose prompt differences as a compliance asset; reliable blocking has to sit on the output side.

Signal 8: Hugging Face Releases StateBench to Test Long-Document State Tracking (11月3日). StateBench is not really testing whether a model remembers; it is testing whether it revises correctly. State tracking requires maintaining a variable table that updates as the document advances and rolling back old values when it hits revocation language such as “this clause is void” or “the supplemental agreement prevails.” Errors cluster around “revoked revisions” because models tend to treat any later text as an update, while real documents are full of mentions that never take effect. Liu Gong believes top vendors will add a “state-tracking accuracy” field to model cards within six months, replacing context-window length as the headline spec—after windows grew from 8K to 256K, localization tasks neared saturation while state tasks got worse.

Signal 9: SRPI Uses Failure Samples to Self-Refine Streaming Text-to-SQL Prompts (11月4日). Within 12 months, “prompt assetization” will become a standard part of enterprise SQL-assistant stacks; methods like SRPI, which reverse-engineer instructions from failure samples, will land first in mid-sized teams held back by fine-tuning costs, because they turn adaptation from a fixed expense into an amortizable one. The interesting part is not the Refiner but the design of three LLM judges plus reference-conditioned feedback: using execution errors alone as a signal teaches the model dialect approximations that merely run, while still missing window boundaries. Reference answers give the reviewer a ruler and separate “syntactically legal” from “semantically correct,” at the cost of more offline calls—at least three independent reviews per round, plus the Generator and Refiner.

Signal 10: Starbucks Explores Personalized Drive-Thru Menus and Frictionless Payment (10月30日). This one falls just before the window but is worth including. Starbucks’ personalized-menu business will most likely appear within six months as opt-in pilots in three to five metro areas, not as license-plate recognition at the speaker box; whoever turns recognition into a customer-raised hand will avoid the privacy minefield. Drive-thru is one of retail’s most latency-sensitive interfaces: a driver has only a dozen seconds or so from entering the order zone to speaking, so recognition, retrieval, ranking and rendering must fit into a few hundred milliseconds or cars back up into the street. The hard part is not the recommendation model but recognizing people reliably in backlight, rain and tinted windows.

Republishing: feel free to quote or repost with a link back to the original article on Wenfeng Radar. · Business & partnership: see our media kit
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private
Get next Sunday’s review in your inbox:
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private

Leave a Reply

Your email address will not be published. Required fields are marked *

中文EN
Contact · Privacy · Disclaimer · AI Disclosure · Weekly Review · Media & Business
Email: wenfengleida@wenfengleida.cn