Liu Gong’s Weekly Review: The 10 Signals That Mattered in AI This Week (Nov 07-13)

Liu Gong thinks the week’s real storyline wasn’t a single launch but the split between two ways of scaling intelligence: Google unbundles a hard question into sub-tasks and hands them to a cluster of agents, while IBM pushes structural symbols into the tokenizer so small models stop failing on format. Both point the same direction — reliability is being engineered into the pipeline rather than hoped for from a bigger checkpoint. Money confirms the mood: OpenAI raises $11 billion at a $1.25 trillion valuation, with Anthropic’s IPO queueing up behind it. Meanwhile Nvidia’s board drama turns a personal story into a governance question, and Adobe opens its project files to third-party agents.

Signal 1: Google ships Gemini 4 Hydrogen, breaking complex prompts into sub-questions (11月12日). Liu Gong believes the deciding factor isn’t breadth of knowledge but the ability to turn one messy sentence into executable sub-tasks. Read the spec, not the adjectives: multi-part questions, typos and highly specific requests all point to structure-first parsing and knowledge-second retrieval. Within six months Google will most likely make sub-question decomposition the default path of its search stack and sell it as the difference against ChatGPT’s search entry point. Splitting first also moves error correction from rewriting a whole answer to patching one branch.

Signal 2: Google opens Gemini 4 Carbon to the public, betting on agent clusters (11月10日). Clusters are not a model copied N times and run in parallel; the real bottleneck sits in merging and verification. At eight parallel sub-agents, the theoretical eightfold cut in wall-clock time typically loses thirty to fifty percent to scheduling, merging and conflict rollback, landing closer to a halving — which is why clusters demo well and deploy badly. Within six months, if Google does not expose parallel-agent count, failure rollback rate and per-task cost to developers, Carbon’s enterprise adoption will lag well behind its consumer noise.

Signal 3: Adobe Max opens creative workflows to third-party AI agents (11月10日). Adobe’s moat was never the generation model; it is the executable context — project files, layer structures, timelines, collaboration permissions. Letting agents read and write that turns the editor into a toolbox for agents. The hard part is permission granularity and audit trails, not plugging in more models: an agent that can change layers but cannot roll back or tag provenance will not be trusted with brand assets. Within twelve months Liu Gong expects the integration to move from creative apps into enterprise content supply chains and brand review.

Signal 4: Hugging Face open-sources ToyBlocks for controllable robot learning and evaluation (11月6日). The controlled variables are the actual product: colour, shape, initial pose, camera angle and lighting each become independently switchable settings, so evaluation yields decay curves against distribution shift instead of one blended score. Traditional benchmarks fold variation into task difficulty, which is why imitation-learning failures cannot be attributed. ToyBlocks splits difficulty from shift and hands researchers a readable ruler. Within twelve months manipulation benchmarks will compete on how well variables are held rather than how many tasks are listed, and the hundred-dollar hardware anchor changes who can reproduce results.

Signal 5: Ring’s indoor flying camera lands January 2027 at $249.99 (11月10日). The gain is viewpoint, not resolution. In industrial inspection Liu Gong keeps seeing the same pattern: on most surface-defect tasks an extra viewpoint beats doubling resolution, and a corner hidden behind a sofa looks identical at 2K and 4K — flying there is the difference. Flight also imports constraints wall-mounted cameras never faced: battery, noise, obstacle avoidance and crash probability. It will not be a hit in the first half of 2027, but within twelve months it becomes the most effective hook in Ring’s subscription bundle, and most likely evolves through software rather than hardware.

Signal 6: A second allegation against Jensen Huang puts Nvidia’s succession vacuum on the record (11月12日). The damage is not the rumour; it is that the equation “Nvidia equals Jensen Huang” is now on the public table. The moat is organisational — CUDA, migration costs and operator coverage are institutionalised and will not vanish overnight. But the annual architecture cadence from Blackwell to Rubin is set personally at GTC, and any wobble in the roadmap disturbs ODM and cloud scheduling downstream. Within twelve months Nvidia will most likely be pushed to formalise successor disclosure in its annual report or proxy statement: not a CEO change, but the succession bench finally written down.

Signal 7: IBM bakes LLMON into Granite 4.1 3B/8B (11月12日). The point is not shorter syntax; it is that structural symbols enter the tokenizer. Most JSON failures are character-legal failures, and grammar-constrained decoding can mask illegal tokens at sampling time, which line-oriented structure suits naturally. Within six months at least one mainstream inference framework will ship line markup in its default grammar templates, because in-vocabulary format beats post-hoc retries by an order of magnitude. Within twelve months it becomes a factory setting for small models without displacing JSON — it eats the edge and on-device share.

Signal 8: OpenAI raises $11B at a $1.25T valuation as Anthropic edges toward IPO (11月6日). The $1.25 trillion number matters less than the fact that it took only $11 billion of new money. By 2026 the capability gap is no longer an order of magnitude; the difference is where the stack sits. OpenAI ties multimodality to consumer distribution, so iteration shows up in retention and call volume; Anthropic ties long context to enterprise workflows, so it shows up in API renewal and customer lifetime. Liu Gong expects Anthropic’s IPO, before May 2027, to price at a revenue multiple below the one implied by OpenAI’s private round, carrying the valuation gap from private markets into public ones.

Signal 9: NYT Cooking’s AI saves dinner but ruins a cake

(11月7日). Dinner succeeding while cake fails is about task type, not model luck. Weeknight cooking is open-ended constraint satisfaction with wide tolerance; baking is a closed numerical problem where leavening, liquid, sugar-to-fat ratio and oven curve usually live inside a plus-or-minus five percent window, so one plausible-sounding rewrite pushes the result out of the edible range. NYT Cooking’s library of over twenty thousand recipes is a ready, verifiable answer pool. Within six months, leading recipe products will most likely switch baking, fermentation and sugar work to retrieval-only, with ratios and temperature curves locked.

Signal 10: Nvidia’s board reportedly prepares to oust Jensen Huang (11月11日). Within six months the board most likely will not complete a handover through a public dismissal; it will first install an executive chairman or interim CEO and cut Huang out of day-to-day management. The trigger for a formal removal is verifiable legal or safety evidence in the second allegation — contain the information, then the seat, then equity and reputation is the usual crisis cadence. Technically, CUDA, NVLink and each GPU generation have been signed off by him personally: H100, Blackwell and Rubin will not stall immediately, but roadmap reviews and custom-accelerator co-design with top cloud buyers will slow, because those projects need CEO-level commitment.

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