Afternoon briefing · 16:00
Chinese AI firms compete on purpose-built components
By · bombong
This afternoon's announcements from Chinese tech companies focused less on general-purpose model rankings and more on components that plug directly into specific tasks. Xiaomi's small document model, Baidu's search API discount, Horizon Robotics' automotive chip and Tencent Cloud's physical AI stack all target the layer that developers and manufacturers build into their own products.
Small components for developer pipelines
The most direct examples are in documents and search. Xiaomi has reportedly open-sourced Xiaomi-OCR-0, a document model with 0.8B parameters 26. The model handles document parsing, key field extraction and OCR-based question answering. Xiaomi said it scored 96.83 on OmniDocBench v1.6, the highest among models under 1B parameters. At this size, it becomes an option for running document processing on-device or on low-cost servers. It also follows a run of small open OCR models, including MiniCPM-OCR and Baidu Unlimited-OCR 2.
Baidu is going after the same layer on price. Baidu said its search engine retrains itself without human tuning, based on real search queries and user feedback 27. Alongside this, it is selling an enhanced search AI API at 4.9 yuan per 1,000 calls until December 31. Agents (AI that carries out multi-step tasks on a user's behalf) call search frequently, so the per-call price drives operating costs. Both announcements share a common thread: rather than finished services, they offer components to be built into other companies' agent and document systems.
The same approach extends to vehicles and robotics infrastructure
The component-supply approach also shows up in layers beyond software. Horizon Robotics' Journey 6P autonomous driving chip and its HSD driving software will reportedly go into BYD's driver-assistance system God's Eye B 28. The first vehicle with the system arrives in November. Because an automaker is sourcing both Chinese chips and software together for advanced driver assistance, this can be seen as a case of widening options in the automotive chip supply chain.
Tencent Cloud has bundled the robotics and autonomous driving development process into a managed service. It launched an integrated stack covering multimodal data processing and model training, and said it plans to release TCLake XPhysis, a physical AI (AI that operates in the physical world, such as robots and vehicles) platform 29. The pipeline from cleaning sensor data through to training is considered a major bottleneck in the field. That similar announcements came on the same day from two different layers, chips and cloud, shows Chinese companies expanding the layers that run and supply models rather than the models themselves.
What the numbers don't yet tell us
Still, it is too early to judge the results of this component race from today's news alone. Xiaomi's benchmark ranking comes from reports relaying the company's claim, and Baidu's account of self-retraining has not been externally verified. Baidu's pricing is a limited-time offer through year-end, and the launch date and detailed features of Tencent Cloud's new platform have not been disclosed.
How to express performance and terms in numbers is also still unsettled. Horizon Robotics CEO Yu Kai said the company will report compute performance as two separate figures going forward. This ties into an industry debate over whether chip performance can be compared using TOPS (a unit of operations per second) alone. The labeling issue has come up outside China too. Anthropic reportedly apologized for confusion after some users took the "20x" in a plan name to apply to weekly usage as well, and said it will rename the plan 24. According to reports, verified subscribers on the $200-per-month plan will now be subject only to a weekly limit, without the 5-hour limits. For those choosing components, checking the actual terms now matters more than names and headline figures.
Today's news shows one front of AI competition broadening from general-purpose model rankings to the unit price and ease of integration of components with clear use cases. When adopting small models and APIs, practitioners should look beyond company-reported scores to performance on their own data, pricing after discount periods end, and the actual units in which usage limits are measured. Next to watch: whether the system actually ships in BYD's first vehicle in November, the release schedule for TCLake XPhysis, and how Qwen 4, rumored to launch next week, enters this competition.
This piece was written with the help of AI and includes personal views.
Morning briefing · 08:00
A training opt-out standard and the AI cost ledger
By · bombong
The AI industry is moving away from company-by-company pledges toward shared rules that the whole industry follows and toward measuring AI in numbers. Four major frontier companies have agreed to follow a common training opt-out signal, and markets and cloud providers have begun judging AI spending by financial metrics. Security operations, however, have not yet kept pace.
Training data rules converge on a single standard
A common standard is emerging for training data. According to The Verge, Google, Microsoft and Anthropic have followed OpenAI in saying they will adopt a 'policy token' signal that web publishers can use to opt out of collection for AI training1. With this signal, publishers can tell all participating companies at once whether they allow training, instead of configuring each company separately. OpenAI said it will respect a common signal that conveys training preferences to participating providers.
This matters because it greatly reduces the work publishers must do to manage opt-outs. Until now, they had to block each company's crawler one by one. Going forward, a single signal will be enough to state their preference. Still, announcing that a company will adopt a standard is different from actually running it. Scope and timing will have to be confirmed from each company's implementation.
On the same day, OpenAI compiled and published excerpts from court briefs filed by creators, academics, accessibility advocacy groups, legal experts and others9. OpenAI's position is that fair use helps AI training and innovation. So a standard way to opt out and a legal case for training are advancing side by side. The data question is being settled on two tracks at once: technical signals and courtroom arguments.
Costs land on the ledger, too
The same shift seen in data is also showing up on the cost side. According to the BBC, U.S. tech stocks fell for a second straight day on views that results fell short of expectations relative to AI investment, and the Nasdaq dropped more than 3.5% on Friday5. Intel fell as much as 30% intraday. TechCrunch reported that the drop cut the value of the U.S. government's stake in Intel by $12 billion. It also said the AI industry's heavy absorption of memory and other components has pushed up PC prices. Fujifilm is increasing magnetic tape production fivefold to meet rising AI-driven demand for data storage14. AI demand now affects prices and production plans even in component and storage media supply chains.
Inside companies, systems for controlling spending and measuring results in numbers are taking shape. AWS laid out a five-step operating model for managing Bedrock costs20: assigning cost owners, observing every request, enforcing budgets, optimizing platform defaults and holding regular reviews. Logistics company Worldwide Express, featured by AWS, saw pending exceptions fall 75% within six months of deploying a Bedrock-based multi-agent system7. The company said the time to collect payment was shortened by four weeks and that it secured $10 million in free cash flow.
The market is asking how much of the money spent on AI can be recouped. Cloud providers released cost management guidance and financial case studies during the same period, which can be read as the providers' answer to that question.
Security operations still have gaps
Rules and cost tracking are being put in order, but security operations still have large gaps. Ars Technica reported that Anthropic's VS Code extension on Cursor was backdoored for nearly 12 hours4. It is still unknown how the attacker got into the verified account. Deployment platform Northflank disclosed a security incident and told all users to assume their environment variables had been decrypted and to rotate their keys11. Several users said their AI API keys had been used without authorization. The two cases show that verification badges or storing secrets on a platform are not enough on their own to protect development tools and credentials. Leaked keys quickly turn into usage charges, so these security gaps also feed into the cost problem described in the previous section.
In agent execution environments, companies keep trying to place controls at runtime (the execution stage). NVIDIA introduced NemoClaw as a way to run specialized agents on a policy-enforcing runtime8. However, its specific feature scope and availability status could not be confirmed. AWS introduced Session endpoints for AgentCore Runtime6. The feature attaches an authenticated URL to each session so that web apps built by coding agents can be viewed directly in a browser. For now, each company is offering its own approach. It remains to be seen whether these will converge on a common industry standard, as the training opt-out signal did.
For practitioners, today's news means AI has to be managed as part of operations, not just as a feature. Content publishers should check when, and how broadly, each company will apply the common training opt-out signal. Companies adopting AI should review who is accountable for costs and how API keys are managed, together. Next, watch for each of the four companies' timelines for implementing the signal, whether the intrusion path in the extension marketplace breach is disclosed, and whether cloud providers' cost guidance leads to real budget control.
This piece was written with the help of AI and includes personal views.
Unconfirmed reports 6
Not yet confirmed by an official source. Check the original report.
- "Anthropic to extend invisible text watermark to Mythos 5, Opus 4.7 and other models from Nov. 3; paid users can't opt out" (unverified) Original
- 'Soft-MTP' paper claims keeping only the KL self-distillation signal in MTP gradients improves speculative decoding and next-token prediction Original
- Google unveils 'EnvHarness' to help self-evolving agents learn in dynamic environments (report) Original
- DeepSeek website code points to 'Cyber' security model with personal and enterprise versions, ID checks and a 'Cyber Key' (report, unconfirmed) Original
- Google is developing an offline mode for Gemini Live, spotted in Android app code (report) Original
- "Anthropic adjusts Claude to stop helping clone games" after weeks of low-quality AI games flooding Steam (community report) Original
Community buzz 5
Topics widely discussed on Reddit, Hacker News, GeekNews and elsewhere. These are community reactions and may include unverified claims.
- Claude policy fatigue Users posted a run of complaints and questions: a claim that text watermarks will extend to existing models from Nov. 3, complaints that the security classifier blocks even legitimate firmware porting work, suspicions that game recreation is being blocked, and questions about the 'non-commercial use' clause in the EU consumer terms.
- AI and deskilling A post arguing that SREs lose the instincts to handle complex outages as AI takes over routine incident response scored over 300 points on HN. News that Proton had canceled its AI-use re-interviews was discussed alongside it.
- DeepSeek V4.1 Flash quantization dispute Unsloth said it could not reproduce ISTA's GSQ-RCO GGUF results (recovering 100% AIME25 accuracy). ISTA countered that the gap came from differences in evaluation conditions, such as a 32K cap on the reasoning token budget. Separately, users shared measurements showing that llama.cpp's new DFlash2 support made Qwen3.8-Flash-Next generation 17% faster on average.
- AI cost worries and tech selloff Discussion centered on a BBC article and a GeekNews summary post. Users pointed to the five big tech companies' 2026 spending plans, which total $800.6 billion, and a reported remark by Larry Page that he would not lose this race even if it meant going bankrupt.
- NES emulator inside model weights Drew Breunig built an NES 6502 CPU emulator into the weights of Qwen3.5-4B and ran Super Mario on it (very slowly). A weekend project that used Opus 5.5 to bring Mario 64 to mixed reality also drew attention.