Afternoon briefing · 16:00
Inference over training: the competition shifts to operations
By · bombong
The center of gravity in AI competition is shifting from building models to running them. A Chinese government-affiliated body said inference data has overtaken training data for the first time. Meituan released a very large model while highlighting low-bit inference efficiency on domestic chips. Together, the afternoon's news points to efficiency and coordination in the operations stage as the new challenge.
Inference's growing share, in numbers
If the morning piece looked at the cost of real-world adoption, the afternoon news shows that this cost accumulates in inference. CCID, a body under China's Ministry of Industry and Information Technology, said in an October 31 report that data generated by Chinese AI systems during inference has surpassed data used for training for the first time. SCMP and Pandaily reported the findings11. Pandaily added that China's daily token calls exceeded 175 trillion as of June, more than 1,000 times the level of early 2024. That a government body is presenting usage as a performance indicator suggests an emphasis on deployment over development.
Meituan's LongCat-2.5 reflects the same trend in model design10. According to Pandaily, the model has 1.6T total parameters but activates only about 48B per token, and was designed as multimodal from the start. Meituan said it trained the model with long-horizon reinforcement learning for lengthy agentic coding tasks. It also said the model was optimized for W4A4 inference (reducing both weights and activations to 4 bits) on Chinese-made chips. This can be read as a design that keeps the model large while lowering the cost of running it, a natural choice in an environment where usage has grown this much.
Operational challenges extend to coordination and verification
As inference grows, problems beyond cost emerge: how to make multiple models work together, and how to verify their outputs. An experiment released on October 29 by researchers from Y Combinator and Olam Labs exposes the first problem13. According to AI Times, the researchers tasked 50 frontier models with jointly building a civilization. The agents exchanged 5,900 messages, absorbed in setting rules and dividing roles, and ultimately failed to reach the goal. It shows that adding many capable agents without a decision-making structure only drives up communication costs. The long-horizon agent work Meituan is targeting faces the same challenge.
On verification, Google is making moves. According to an InfoQ China headline, Google is streamlining the verification process of SynthID Detector, its tool for detecting watermarks in AI-generated content, and expanding support to content from other AI providers12. Generated content spreads beyond any single company's services. Provenance tools appear to be expanding accordingly, beyond individual models toward a shared industry method.
Gaps that remain unconfirmed
Still, there are too many gaps to draw firm conclusions from today's news alone. The reports do not detail how CCID calculated its figures, so it is unclear on what basis inference data and training data were compared. For LongCat-2.5, the collected excerpt does not say whether the weights are open, how the model can be accessed, or how it benchmarks. Actual inference efficiency needs to be assessed against official materials.
The other items are also limited in scope. The Olam Labs experiment was conducted in game environments such as Minecraft, so it does not translate directly to the workplace. For SynthID Detector, the article body was not collected, so which companies are covered and when the change takes effect are unconfirmed. The news points in the same direction, but the size of each change should be judged on follow-up announcements.
Practitioners choosing models should weigh not just training scale but also inference costs, coordination structures between agents, and the means to verify outputs. Going forward, it will be worth checking LongCat-2.5's release scope and benchmarks, as well as the list of providers SynthID Detector supports. On the inference serving side, changes in the new vLLM release are also worth watching.
This piece was written with the help of AI and includes personal views.
Morning briefing · 08:00
As AI spreads, the costs of verification and acceptance
By · bombong
The barrier to AI adoption is shifting from model performance to the cost of verification and acceptance on the ground. A survey of Australian lawyers found AI use rose sharply, but a new task of checking outputs emerged. Local sentiment toward data centers and the challenges of spreading in-house agents point in the same direction.
The verification cost left after adoption
A survey found that greater AI use did not reduce work by the same measure. According to a survey of 2,221 lawyers by the Victorian Legal Services Board + Commissioner (VLSB+C) in Australia, the share of lawyers using AI rose to 55% in one year4. But reductions in workload and clients' legal costs fell short of expectations. Instead, a new task emerged: lawyers checking and correcting inaccurate information and documents that clients produced with AI.
The result matters because it shows that verification costs do not disappear; they just move. The time clients save with AI becomes review time for professionals. That means adoption rates alone are not enough to judge productivity effects.
Google's move can be seen as an attempt to reduce this cost on the provider side. Google said on X that hundreds of textbooks from publishers including OpenStax, Open Textbook Library, and S Chand Publishing are now available for free to users in the US and India in its AI learning tool Gemini Notebook2. Instead of users uploading material of unclear origin themselves, they study from publisher textbooks. Anchoring answers to reliable material from the start also reduces the burden on users to check separately.
The same costs appear in infrastructure and organizations
Verification and acceptance are not only issues for outputs. Similar costs appear in the communities where infrastructure is built and inside organizations adopting AI.
TechCrunch reported that OpenAI may start calling its data centers 'AI compute campuses'1. The move comes as local sentiment toward data centers has soured. According to an account of an Axios report shared in the community, the plan is part of a post-midterm PR campaign and also includes giving residents more involvement in projects and their benefits. It shows that power and water use and local acceptance have become as important as technology in expanding AI infrastructure.
In Korea, SK Telecom said it will pursue 'AX Innovation 2.0,' redesigning work processes around AI at the organizational level5. More than 4,000 employees have built agents themselves, and about 1,500 are in use in daily work. The company said it will also tackle data access, security, and system integration issues that surfaced as agents spread. That suggests the next challenge is less about building agents than about connecting them safely to the organization's data and systems.
Evidence is still uneven
That said, the evidence is still too uneven to draw firm conclusions about this trend. OpenAI's renaming rests on reports and secondhand accounts, and no official company announcement is included in the source material. Whether the name will actually change, and what form resident involvement will take, remains unknown.
The Victoria survey also covers a single region and a single profession. Follow-up studies are needed to determine whether the verification burden is a transitional cost of early adoption or a lasting structure. Google's textbook offering is likewise limited to the US and India, and there is no data yet showing learning gains or fewer errors. How much building in vetted material actually reduces the burden of checking remains to be seen.
For companies and practitioners adopting AI, the question is less about adoption rates than about who bears the costs of verification and acceptance, and how much. Professional services need to design procedures for handling AI outputs clients bring in; infrastructure operators need ways to share benefits with local communities; and in-house adoption teams need to build data access and security frameworks alongside. Going forward, it is worth watching for OpenAI's official position and the specifics of its resident-involvement plans, and for figures showing that the spread of in-house agents has actually translated into time saved at work.
This piece was written with the help of AI and includes personal views.
Unconfirmed reports 5
Not yet confirmed by an official source. Check the original report.
- 'Adjusting just two neurons in Qwen3.5-4B turned it hostile to humans': interpretability experiments spread to individual developers (report) Original
- New Yann LeCun paper says a 109M-parameter model beats a 72B model, challenging scaling laws by arguing bigger models don't understand the world better (report) Original
- Next-model rumor follow-up: 'GPT-6.2 Astra' reportedly in limited user testing, with signs of a Claude Mythos 5.5 launch the same day (reports, community posts, unconfirmed) Original
- "China's Ministry of Commerce imposes first export controls on frontier AI via Announcement No. 48": Cloud Security Alliance publishes analysis note (community report, details unconfirmed) Original
- Qwen lead reportedly says upcoming Qwen 4.0 and 5.0 will also be open source, denying rumors Qwen4 is the last open-weight release (community report, unconfirmed) Original
Community buzz 6
Topics widely discussed on Reddit, Hacker News, GeekNews and elsewhere. These are community reactions and may include unverified claims.
- 50-agent Minecraft experiment An experiment by Y Combinator and Olam Labs, in which 50 GPT-6.1 Sol agents were tasked with building a civilization in Minecraft over three hours, was posted on both HN and GeekNews. Instead of building, the agents became absorbed in food rationing and managing approval processes. After deaths clustered on the first night, the agents reportedly cut back on exploration and stayed in shelters.
- Debate over AI adoption results and trust An article in which an Apollo economist reportedly said 'the financial return on AI is zero at most companies' scored over 200 points on HN. A post saying Proton's CEO called an engineer's request for time for penetration testing a 'loser mindset' also drew attention, and Kyle Kingsbury argued that LLMs are eroding trust in documentation and software.
- Math research in the AI era A post argued that the gap between how fast AI produces proofs and how fast humans can understand and build on them is widening, so the practice of rewarding 'first proofs' should be reconsidered. A mathematician's optimistic view that AI will free mathematicians from unsolved conjectures to explore deeper problems also drew attention.
- Real-world local LLM use Users shared reports of reaching 35 tokens per second with DeepSeek-V4.1-Flash-1017 and 130 tokens per second with Qwen3.8-Flash-Next on a single RTX 3090 Ti with 128GB of RAM. Other posts asked for alternatives to Gemma 4 E4B as a small model for tool calling, and featured GLB-3D-Coder, a 4B model that creates 3D assets for Godot, and an experiment running llama2.c on a 2006 iPod.
- Claude restrictions and subscription limits in practice A comparison post spread saying Gemini 4 Argon recreated a Pixar intro while Claude Opus 5.5 declined on copyright grounds. Meanwhile, one post said running four Opus 5.5 sessions for five hours on Anthropic's top-tier plan used only 8% of the quota, and another complained that OpenAI's subscription terms are getting worse.
- Improving the image of data centers News that OpenAI may rename its data centers 'AI compute campuses' drew responses that a name change alone does not address concerns about electricity bills and water use. An interview in which Demis Hassabis said data centers should be built beautifully and their impact on surroundings taken seriously was also reshared.