OpenAI expands ChatGPT ads to five more markets and starts testing the price of trust
On August 11, OpenAI took its ChatGPT ads pilot to the UK, Mexico, Brazil, Japan, and South Korea, following earlier tests in the US, Canada, Australia, and New Zealand. The program targets logged-in adult users on free and Go plans — Plus, Pro, Business, Enterprise, and education plans show no ads, and accounts identified as under 18 are excluded. OpenAI restated its four principles: ads are clearly labeled, answers stay independent of advertisers, conversations stay private, and users control the experience. Ads appear below responses, each carrying an advertiser name, favicon, headline, description, landing page, and image.
The buying mechanics now look like a conventional ad platform. Ads Manager Beta reports impressions, clicks, spend, click-through rate, average CPC and CPM; CPC campaigns get a recommended $3–5 max bid and run on a relevance-weighted second-price auction. The pitch is intent — reach people comparing options, planning trips, weighing insurance. Same day, Yelp said ChatGPT users in the US and Canada can book restaurant reservations directly in the chat window, a transaction step that is a different animal from a sponsored card.
The trust question is the real story. A search engine can put sponsored links above organic results inside a familiar model; a chatbot answering "which laptop should I buy" has a blurrier line between recommendation and advertisement. One early academic audit of more than 3,000 simulated US accounts found ads were clearly separated from response text but that lower-income accounts were more likely to receive them, regardless of race. OpenAI says advertisers get no chats, memory, or personal details, and ads won't appear near health, mental-health, or political topics during the test. I don't think the design promises are enough on their own — conversational ads need outside monitoring as formats and targeting mature, and this five-country expansion is where the pressure to grow revenue meets the pressure to keep the free tier trustworthy.
— OpenAI · techscurrent
🔗 OpenAI (Testing ads in ChatGPT) · techscurrent · 极客网
DeepSeek announces a major API price hike, and the price war starts to end
DeepSeek posted a notice on August 6: it plans to raise API prices "overall" and "by a considerable margin," with specifics to follow. The move lands three weeks after it introduced peak/off-peak pricing for the V4 series (9–12 and 14–18 hours at 2x). Under current list pricing, V4-Pro is 3 yuan in / 6 yuan out per million tokens, V4-Flash 1/2 yuan; Goldman estimated the peak mechanism pulled blended averages to about $0.35 for Pro and $0.12 for Flash. The announcement gave no timeline or rate, only a request that developers "arrange usage accordingly."
The demand numbers explain the direction. OpenCode reported V4-Flash's official version burned through 8 trillion tokens in a single day on its platform — 5 trillion from free quotas, 3 trillion from paid plans. OpenRouter, which routes more than 400 models, does about 6.6 trillion tokens a day across the whole platform; DeepSeek's one model on one entry point exceeded that. Vercel's data put DeepSeek first in token volume, with V4-Flash at roughly 5.3 trillion tokens a week. Liang Wenfeng's July framing was that demand at these prices is nearly inelastic — "double the price and token consumption barely changes" — and that pricing is built on recovering hardware costs in about ten months. ARR is estimated at $400–500 million with the V4 gross margin above 50%.
This is the price war ending, and I think it's worth being precise about what that does and doesn't mean. DeepSeek is raising prices, not abandoning cheap inference — at Artificial Analysis's late-July measurements, a single V4-Flash call cost around $0.03, roughly 1/105 of Claude Fable 5's $3.15. The hike signals that buying share at below-cost prices is over, and agents are the reason: one agent task can burn a hundred times the tokens of a chat. The honest risk is that Chinese labs raised prices across the board this year (Zhipu three times, Moonshot's K3 output up over 3.5x), while 71% of enterprises told Stanford they can swap the underlying model in routine scenarios — so the increase has to be paired with capability or lock-in, not just a bigger bill.
— DeepSeek · 新浪财经 (via 金融时报/中国企业家)
🔗 DeepSeek (开放平台) · 新浪财经 · 行业分析 (via 腾讯新闻)
Anthropic confirms it is building an in-house chip team for Claude
Anthropic confirmed on August 5 that it is assembling an internal chip team to design custom silicon for Claude — the first official acknowledgment, given to Business Insider and TechCrunch by a spokesperson. The company says it will co-design hardware and models so Claude runs faster and more efficiently at scale, and stresses this is part of a "multi-chip strategy": AWS, Google, NVIDIA, and AMD hardware remain central to its expansion. The job postings show the scope — chip architecture, front-end design, verification, physical design, foundry and packaging — with salaries of $320,000–$485,000 in San Francisco, New York, and Seattle, and a requirement that candidates have shipped silicon through tape-out and into use.
The honest caveat is the timeline. Anthropic disclosed no architecture, no foundry partner, and no launch date. Reuters reported in April that it was exploring custom chips; The Information said in July it had talked to Samsung about manufacturing. One named early hire is Clive Chan, who worked on OpenAI's chip team. This is a recruitment-stage program, not a product.
I read this as the frontier labs converging on the same conclusion: when inference cost is the constraint on the business model, co-designing silicon with the model beats buying general-purpose GPUs at the margin. OpenAI shipped Jalapeño in June via Broadcom; Meta reportedly has its own chip in production from September. Anthropic's version is distinctive in sequencing — it is staffing the team after the model, not before — which is also why this takes years to matter. The "multi-chip" framing is the thing to watch: if custom silicon stays a supplement rather than a replacement, this story is about negotiating leverage with suppliers, not about leaving NVIDIA.
— Anthropic (发言人 via Business Insider/TechCrunch) · 智东西
🔗 Anthropic (招聘页) · 智东西 (via 腾讯新闻) · TechCrunch (via CCID)
NVIDIA releases open-weight Nemotron 3.5 Lightning, plus a router for agent workloads
NVIDIA released Nemotron 3.5 Lightning on August 11: an open-weight, 30-billion-parameter mixture-of-experts model that activates about 3 billion parameters per token, distilled from its larger Nemotron 3 Ultra and aimed at always-on agent workloads. NVIDIA claims up to 4x faster token generation and roughly 30% faster task completion than comparable open models. It is free for commercial use — weights on Hugging Face and build.nvidia.com with no license fee — and NVIDIA published the training data and techniques subject to licensing constraints. It runs locally on RTX PCs, DGX Spark, GB10-based OEM systems, and Jetson, with day-one support across vLLM, Ollama, llama.cpp, LM Studio, and Unsloth in NVFP4 and GGUF formats.
The same day NVIDIA shipped NeMo Switchyard, an open-source model router that picks the most efficient model for each agent task. The sales pitch is pointed: companies in specialized fields like cybersecurity or material science don't want to send proprietary IP to a model provider that could one day compete with them. NVIDIA also confirmed to the WSJ and Chinese media that Nemotron 4 — a flagship reported at roughly 1 trillion parameters, about double Nemotron 3 Ultra — is in training.
The business logic is naked and I respect it. Jensen Huang said it himself: "Free AI should be great for hardware. Free AI should be great for chips." Open models need GPUs to run, and this release lands at a moment when US open-weight momentum looks thin — Qwen reportedly passed 50% of global open-model downloads, and Hugging Face counts put Chinese open-weight downloads at 41% versus 36.5% for the US. Whether 3.5 Lightning changes that depends on whether developers actually run it locally, but the direction is clear: NVIDIA is competing for the center of gravity of open models because that is where the chips get sold.
— NVIDIA · WSJ Tech News Briefing
🔗 NVIDIA (本地 AI 博客) · WSJ Tech News Briefing · 每日经济新闻 (via 腾讯新闻)
WorldExam asks whether AI-generated worlds actually react — and most models fail the test
A team from CASIA (中科院自动化所), Chinese University of Hong Kong, Shanghai AI Laboratory, AMAP (高德), and Tsinghua published WorldExam on arXiv (August 3): a hierarchical benchmark asking whether controllable video-generation models behave like world models, not just video generators. It spans 1,474 cases across eight tasks in four layers — visual quality, control adherence, spatial consistency, and world reactivity. The last layer is the point: terrain interaction (feet should step up stairs), object interaction (a pushed object should move), social interaction (passersby should avoid a car), physical reaction (a tilted basket should fall), and goal completion (pick up the screwdriver, not the coin).
Across 20 representative models the results show a clean capability split. Camera-driven models dominate camera control — NeoVerse at 97.33 — but their interface cannot do dynamic interaction. Action-driven models control the subject precisely (WorldPlay at 92.74 in action tasks) yet often leave the world unresponsive; WorldPlay's terrain-interaction score collapses to 27.49. Language-driven models handle interaction better but follow complex controls less faithfully, with the best, Hailuo 2.3, at 63.29 in camera tasks and only a 50.5% success rate returning to the starting viewpoint. The grading pipeline is notable in itself: VGGT-Ω reconstructs 3D trajectories from frames, and GPT-5.5 acts as an AI grader against per-case checklists, correlating with human raters at Spearman 0.86 across 800 samples.
The headline takeaway is that visual quality and explicit instruction-following don't predict whether a model builds a world that behaves — which is exactly the gap that matters for robotics and simulation, where the video has to be right in ways a human viewer might not notice. A foot that phases through a step is invisible in a single frame but fatal in training data. The benchmark is also honest about its limits: no model covers all four layers well, and the fact that "world model" is still mostly aspiration is the most useful result in the paper.
— arXiv · 腾讯新闻
🔗 arXiv:2608.02603 · 腾讯新闻 (WorldExam 深度报道)
ParamBench: the tool-call parameter is the least-examined part of agent work
A Zhejiang University-led group of 16 authors published ParamBench on arXiv (August 4), targeting the least-examined part of tool use: filling a tool call's parameters correctly. Most agent research focuses on which tool to pick and in what order; the paper notes that in domains like cloud networking, even frontier models correctly complete fewer than half of tool calls. The authors' discovery is a hidden-state signal — while a model generates a parameter value, a simple linear probe on its hidden states can predict whether that value will be correct.
They turn that signal into two methods: probe-filtered bootstrapped training (PBT), which uses the probe to filter reliable self-generated calls for fine-tuning, and probe-guided reranking (PGR), which selects better candidates at inference. ParamBench itself is built from real cloud-network APIs and grades every instance into five difficulty levels by nesting depth, cross-parameter dependencies, and how much reasoning is needed to derive values from earlier calls. Across five open models and six external benchmarks, average exact match rose from 19.7% to 59.6%.
I keep coming back to the numbers because they expose how shallow the agentic hype can be — a model can name the right tool and still fail half the calls because the parameters are wrong. The probe idea is the interesting part: if hidden states carry a correctness signal, that is a cheap supervision channel that needs no labels. The honest caveat is scope: ParamBench is cloud-network APIs, and whether the method transfers to messier, real-world tool schemas is untested. Still, this is the kind of work that quietly moves the agent stack forward, because parameter correctness is where a lot of real-world agent failures actually live.
— arXiv
🔗 arXiv:2608.03071
Oxide Computer raises $445M for on-premises cloud infrastructure
Oxide Computer disclosed a $444,999,052 Series D in an SEC Form D filing signed August 4 — 15 investors under Rule 506(b) — following a $200M Series C in January led by Thomas Tull's US Innovative Technology Fund with Eclipse, Riot Ventures, and Jane Street. The Emeryville, California company builds rack-scale systems that fuse servers, networking, and control software into a single box, so enterprises can run cloud-style infrastructure on their own premises. For an infrastructure startup the raise is unusually large; Dealroom calls it among the largest in US enterprise software.
The thesis is that public cloud alone doesn't fit every AI workload. Oxide's pitch is data sovereignty, lower latency, and full control for organizations running sensitive data or customized configurations — the same reasons enterprises keep standing up private GPU clusters. The company has a strong engineering reputation, but Hacker News commenters were split, with some questioning whether Oxide has shipped hardware at real volume and one engineer complaining about not getting sales callbacks after spending heavily on AWS.
I'd read the round as a bet on the direction of enterprise AI infrastructure rather than on Oxide's current revenue — $445M is a lot of runway for a company whose products are still making their way into production. The signal is that investors are pouring capital into on-prem infrastructure at a moment when the default answer to "where do we run AI" is the public cloud. If the market splits between public cloud for training and on-prem for sensitive inference, Oxide's timing looks good; if enterprises stay with the hyperscalers, this is a very expensive bet on a niche.
— SEC Form D · Dealroom
🔗 SEC Form D (via Dealroom) · The AI Brief · dev.to (分析)
Top comments (0)