Anthropic puts an invisible watermark on every Claude output, everywhere
Anthropic confirmed on August 11 that every Claude model launched on or after August 2 now carries a marking at launch: an invisible watermark woven directly into generated text, plus digitally signed provenance metadata attached to supported files (.svg, .png, .jpg). The trigger is regulatory — Anthropic signed the EU AI Act's Article 50(2) Code of Practice on transparency of AI-generated content, one of roughly 190 organizations to do so — but the rollout is global, not European. The watermark is applied at the model level, which means it shows up no matter which surface produced the text: the API, the Claude app, Claude Code, Claude Cowork, or Claude Tag, including access through AWS, Google Cloud, and Microsoft Foundry.
The two techniques answer different questions. The text watermark is designed to survive copy-and-paste and may persist through some editing, so it travels with the passage rather than sitting in metadata. The file manifest, built on the C2PA open standard, signals that a file passed through Claude and lets a C2PA-aware tool detect whether it has been tampered with. The company is explicit about the limits: a detected mark is a signal, not proof — someone using Claude to proofread or translate their own document gets a watermarked result, and heavy editing, short passages, or old models can all strip the signal. Metadata can be lost entirely through format conversion or screenshots.
What's notable is the direction, not the novelty. Google DeepMind already ships SynthID for text, image, and audio; OpenAI has talked about watermarking but moved more slowly. Anthropic is the one making it model-level and default-on everywhere. I find the honesty about limits more interesting than the watermark itself — the company essentially admits this is a compliance floor, and that detection tooling it promised to publish is still forthcoming. Enforcement sits with national market-surveillance authorities, with fines up to €15 million or 3% of global turnover, and the European AI Office plans task forces in September where signatories compare what they actually shipped. The real test comes when the first false-positive prosecution happens — a human author whose text trips a Claude detector.
— Anthropic · euronews
🔗 Anthropic (How Claude marks AI-generated content) · euronews · Unite.AI · 界面新闻 (via 腾讯新闻)
DeepSeek opens a "Harness Team" and publicly targets Claude Code
DeepSeek has made its challenge to Anthropic's Claude Code explicit. The Hangzhou lab set up a WeChat account for its "DeepSeek Harness Team" — verified by Tencent, registered to a Beijing entity that Chinese corporate data shows is wholly owned by DeepSeek, and branded with a black-whale logo — and is posting job listings for agentic product work. The team has been assembling since May, led by Cui Tianyi, a Zhejiang University CS graduate with six ACM regional gold medals, nine years at Jane Street, and a stint co-founding quant firm TSY Capital. The internal formula is blunt: "Model + Harness = Agent," where the harness is everything wrapped around the model — tool connection, context management, task decomposition, execution feedback, result verification.
The moves go beyond hiring. On August 12 DeepSeek announced that its flagship V4 Pro got better agentic capabilities, and its API documentation now includes an integration guide for calling DeepSeek models from inside Claude Code. The team lead opened an internal beta recruiting developers of open-source Agent Harness projects on August 1. The strategy is the reverse of Anthropic's: DeepSeek's models are already the cheap workhorse everyone routes to (V4-Flash's official version alone burned 8 trillion tokens in a day on one platform), so the missing piece is the harness that turns raw capability into a usable agent product.
This is the real front of the China-US model war now. OpenAI has Codex, Anthropic has Claude Code with AutoMode going default on August 14, Google is in talks to license Mechanize's tech, and Meta shipped Muse Code in beta on August 5. DeepSeek joining with a price advantage is the pattern that keeps repeating — the models are commoditizing, and the harness, the terminal UX, and the approval/verification loop are where the competition moved. I'd watch the pricing question more than the demos: Claude Code and Codex charge real money for agent runs, and a V4-backed agent at a fraction of the cost is the one thing that could actually dent the incumbent terminals. The unknown is whether DeepSeek can build the reliability layer — harnesses fail in the verification, not the generation.
— DeepSeek (via Bloomberg) · 21经济
🔗 Bloomberg (via Yahoo Finance) · 21财经 (黑鲸出水) · 新浪财经 (Harness 团队单设公众号)
Mistral turns European AI sovereignty into a capacity contract
Mistral announced on August 11 a three-part infrastructure push: Regional Endpoints are generally available so customers choose whether inference runs in Europe or the US; a Priority Tier (public preview) adds custom rate limits and an uptime SLA for mission-critical workloads; and the platform will host third-party open models, starting with GLM-5.2 from China's Z.ai. The bigger move is the compute coalition. Mistral is aggregating multi-year commitments from an anchor group of European enterprises — ASML, CMA CGM, Amadeus, and Caisse des Dépôts are named — into "European Compute Units" (ECUs), claims on Mistral-built capacity, to underwrite 200 megawatts of infrastructure by the end of 2027 and a full gigawatt by the end of 2030.
The numbers behind the plan deserve scrutiny. Mistral currently operates under 200 MW — a 44 MW site near Paris (online Q2), a 23 MW site in Sweden with EcoDataCenter, a 10 MW site in Les Ulis (online Q3). Epoch AI puts a typical 1 GW AI data center at roughly $38 billion in upfront capex; Goldman Sachs estimates $15–20 million per megawatt before chips. CTO Timothée Lacroix didn't dispute the scale, telling VentureBeat that a gigawatt "is a large investment that requires also a lot of scaling and revenue behind it," and framed the urgency as a supply crunch — demand exceeding supply around 2027–28, especially in Europe. Mistral raised €830 million in debt earlier this year to fund its Paris data center.
The structure is what makes this different from a cloud roadmap. An ECU reads more like a power-purchase agreement than a hosting contract — buyers commit before the capacity exists, converting sovereignty talk into a financing instrument that de-risks construction. That also means the open questions are contractual: Mistral published no ECU price, no minimum commitment, no delivery schedule, and no clarity on whether allocations transfer. Europe's compute gap is real — Alice Labs counts 3.6 GW on the continent against 31 GW in the US — and the EU is running a €20 billion AI Gigafactories program in parallel. I like the mechanism, but it's worth remembering the commitments are declarations, not orders: the coalition announcement names who signed up, not how much anyone actually bought.
— Mistral (官方博客) · VentureBeat
🔗 Mistral (Regional inference, open models, new European infrastructure) · VentureBeat · BankInfoSecurity
Anthropic signs a $9.1B, 20-year compute deal with bitcoin miner Riot
Riot Platforms disclosed a $9.1 billion, 20-year agreement to supply 191 megawatts of computing capacity from its Rockdale, Texas campus to a "leading frontier AI lab" — which sources confirm is Anthropic. The contract runs through June 2048 with two five-year extension options, which would push the value to $16.1 billion. Capacity comes online in stages: 96 MW by December 2027, the full 191 MW by June 2028. Riot arranged a $573 million interim financing facility through Morgan Stanley to fund early construction. The announcement came alongside Riot's Q2 results, where revenue rose 14% to $174.2 million but the company swung to a net loss of $237.2 million; Riot's shares jumped about 25% in after-hours trading before giving back most of the gain.
This is Anthropic's third mega-deal in a few months — $10 billion with Volta Infra in Norway, roughly $45 billion of compute from xAI in May, and now this — on top of the >$100 billion, 10-year AWS commitment announced in April for up to five gigawatts. CEO Dario Amodei has said run-rate revenue crossed $47 billion by May. Riot, meanwhile, is now a two-tenant campus: AMD is the other tenant, and CEO Jason Les says the leases total 241 MW and roughly $9.8 billion in contracted revenue, all signed in six months.
The deal is a clean illustration of how the compute market is reorganizing. Bitcoin miners hold exactly what AI labs are short on — grid-connected power and existing infrastructure — and the crypto downturn plus the 2028 halving cycle pushed them to repurpose. Cipher, Hut 8, and TeraWulf already run hybrid models; Riot was one of the last pure-play holdouts. The numbers are the interesting part: the $9.1 billion contract is roughly $455 million a year, which at the simple average rivals Riot's entire annualized bitcoin mining revenue. That's how fast the identity flipped. What none of these 20-year leases can tell you is what AI compute demand actually looks like in 2048 — the contracts price capacity, not certainty.
— Riot Platforms · The Next Web (via Bloomberg)
🔗 Riot Platforms (Q2 2026 财报公告) · The Next Web · Quartz · CNBC TV18
CodeRabbit raises $143M and rebrands review as "Agentic Change Management"
CodeRabbit closed a $143 million Series C at a $1.5 billion valuation on August 12, less than a year after its $60 million Series B. Atomico and Smash Capital co-led, with BMW i Ventures, Datadog, Hirtle Callaghan, SineWave, Scenic Management, and existing investors CRV and Scale Venture Partners participating. Alongside the round the company launched "Agentic Change Management," positioning its AI code review as the control layer for software written by both humans and agents: validation with repository-wide context, triage that scores incoming pull requests by value and risk, automated fix-and-re-review loops, and continuous post-merge monitoring. Revenue grew more than 5x year-over-year, it runs over 2 million code reviews a week, and it counts Adyen, BMW, Indeed, JFrog, NVIDIA, Trivago, and Campfire among 17,000 customers. It also committed over $10 million to keep AI code review free for open source over the next year.
The pitch captures the shift in software delivery. "Writing code has gotten dramatically easier. Validating it hasn't," said Blacksmith CEO JP Jayaprakash in a separate but same-day announcement — Blacksmith raised a $45 million Series B led by Peak XV, growing from 800 to 6,000 customers as CI became the bottleneck. CodeRabbit's framing is that the backlog is moving from tickets to proposed code: agents and non-technical staff open pull requests continuously, so the pull request becomes the auditable decision point where quality, risk, and human attention get assigned.
Two things strike me about this round. One is the thesis — independent, model-agnostic governance of agent-written code is now a venture category worth $1.5 billion, which says more about how much code agents are generating than about any single vendor. The other is the investor list: Datadog and BMW i Ventures joining is a signal that observability and industrial players see code governance as infrastructure, not a tool. The honest caveat is that "agentic change management" is a marketing frame around features that are still maturing — the actual test is whether teams keep trusting AI review to gate agent PRs once those agents start generating thousands of changes a week, which is exactly the scale this round is betting on.
— CodeRabbit (Business Wire) · The Agent Times
🔗 CodeRabbit (via FT Markets / Business Wire) · PRNewswire (Blacksmith) · The Agent Times
OpenAI completes a $7B employee tender at $852B, inching toward IPO
OpenAI completed a roughly $7 billion secondary tender offer on August 10, letting current and former employees sell shares at the company's $852 billion valuation — unchanged from the record $122 billion funding round that closed in March. The notable detail: OpenAI bought the shares itself rather than bringing in outside investors, so this is a liquidity event, not fresh capital. Bloomberg first reported the size, and CNBC confirmed the tender had been in the works since the March round. It follows a $6.6 billion tender at a $500 billion valuation in October and a $1.5 billion tender in 2024, bringing total secondary sales to over $15 billion. OpenAI confidentially filed its S-1 with the SEC in June, though no listing timeline has been announced and reports suggest the IPO could slip into next year.
The tender does two jobs at once. For employees it converts paper wealth into cash without waiting for an IPO that keeps getting pushed; for the company it removes some of the pressure to go public before it's ready. At $852 billion, OpenAI would arrive as one of the most closely watched listings in history, and the March round's investor list — SoftBank, Amazon, NVIDIA, a16z, Sequoia, Thrive — sets the expectations. The rival context matters too: Anthropic was most recently valued at $965 billion after its $65 billion raise in May, and media reports say its IPO could come as soon as next month. Two of the defining private AI companies are now both publicly preparing for the public market, with the largest remaining question being sequencing, not valuation.
I read the "company buys its own stock back" structure as the tell. OpenAI doesn't need the money — it needs employees to stay while the IPO clock keeps moving. Tenders are how you manage that: give people liquidity now, keep the option of listing later. The counterintuitive risk is that repeated 9-figure tenders normalize a private market valuation that the public market may not match on day one. At $852 billion, the gap between private round pricing and what public markets will pay for AI revenue is the single biggest unknown in the company's story right now, and no tender structure resolves it.
— OpenAI (via Bloomberg/CNBC) · 华尔街见闻
🔗 Bloomberg (via Yahoo Finance) · 华尔街见闻 (via 网易) · TechStartups
LLaDA MoE v2: diffusion language models finally get scaling laws
A team from Renmin University's Gaoling School of AI (Li Chongxuan, Wen Jirong) and Ant Group published LLaDA MoE v2 on arXiv (August 11), giving diffusion language models something they never had: proper scaling laws. The paper characterizes how batch size and learning rate should scale for masked-denoising MoE models, then trains a 30B-parameter model with 3B active parameters from scratch under those laws. The result is striking on efficiency: LLaDA MoE v2 uses roughly 65% of the pre-training tokens of Qwen3 30B-A3B and approaches its performance on knowledge, reasoning, and code benchmarks, and after supervised fine-tuning alone it beats existing diffusion LMs on 7 of 8 reasoning and code benchmarks.
The scaling-law detail explains why diffusion models behave differently. Under masked denoising, a nominal token only contributes supervision when it's masked, which lowers effective supervision strength and changes the gradient-noise profile. So the optimal batch size grows steeper than in autoregressive models — at one FLOPs level, the new law says 3.43 million tokens versus 1.02 million for the autoregressive baseline — and the optimal learning rate decays faster. The IsoFLOP analysis puts the split near balance with a slight data tilt: model-side compute scales at the 0.475 power, tokens at 0.525. The authors say inference code and weights will be open-sourced.
This is the kind of paper that doesn't make a product announcement but quietly moves a research direction. Diffusion LMs have been the "maybe cheaper than autoregressive" bet for a while — the masked objective is parallel-friendly and the generation process is reversible — but they were trained empirically, model by model. With a law, you can allocate compute before spending it, which is exactly how autoregressive models scaled up so reliably. The honest caveat is that "approaches Qwen3" is not "beats Qwen3" — the gap is real, and the benchmark list is short. What I find interesting is that Ant Group is co-funding this line of research at all; if the efficiency holds at 100B+ scale, the diffusion path stops being a curiosity and becomes a cost story, and that's where Chinese labs with compute constraints would push it first.
— arXiv · 腾讯新闻
🔗 arXiv:2608.03457 · 腾讯新闻 (LLaDA MoE v2 深度报道)
Top comments (0)