Anthropic watermarks Claude text; Meta's $2B Manus deal unwinds

Anthropic rolls out imperceptible text watermarking for Claude models

Anthropic introduced a text watermarking feature for its Claude AI models on Monday. The imperceptible watermark is embedded directly into AI-generated text and does not change the text’s meaning or readability. It travels with the text when copied and pasted and may persist through some editing, according to Anthropic. The move is part of Anthropic’s transparency commitments under the European Union AI Act, and Claude models launched on or after August 2 will support marking from launch, with older models to follow. Anthropic plans to provide third parties with tools to detect the watermarks.

Anthropic acknowledged loopholes: heavy editing, paraphrasing, translating, or mixing Claude’s output with other writing can make the watermark undetectable, and finding a watermark does not prove Claude originally authored something, since proofreading or translating with the AI can also leave a mark. The feature could affect the publishing industry’s handling of AI-generated content, following recent incidents where publishers pulled books over AI authorship concerns. Anthropic is the second major lab to introduce text watermarking, after Google DeepMind announced in 2024 that it was watermarking text and videos generated in the Gemini app and web experience using its SynthID technology.

Anthropic just rolled out a tool that’ll decimate some people’s dreams of writing AI novels undetected →

Zuckerberg pushes open-source AI vision, draws sharp criticism

Meta Platforms CEO Mark Zuckerberg published a 6,500-word essay on Monday outlining his vision for artificial intelligence, imagining a future where everyone has their own all-knowing AI agent. He wrote that his company is working toward an era where individuals will have tools to create new businesses, receive PhD-level tutoring, and get personalized lifestyle tips. Zuckerberg argued in favor of open-source AI technology, warning that concentrated control of advanced AI with a select few companies would lead to less favorable outcomes, and made a veiled reference to rivals by writing that most other labs focus on building AI for companies and institutions. Meta announced the release of a new open-source AI model, Muse Glimmer, which can run on a personal computer, and Zuckerberg said the company would provide developers access to a more powerful model, Muse Spark 1.2.

Anthony Aguirre, president and CEO of the AI safety non-profit Future of Life Institute, criticized Zuckerberg’s vision, citing recent incidents where AI models, including Meta’s own, hacked other companies on at least three separate occasions. Aguirre asked what on Earth makes Meta think they could control something exponentially smarter, faster, and more capable, and called for governments to facilitate a pause on advanced AI development. Zuckerberg included policy recommendations for keeping the US ahead of rivals, including China, calling for faster energy capacity and infrastructure building, and for authorities to reconsider their position on distillation, which involves training a less capable model on the outputs of a stronger one. Matt Lane, senior policy counsel at Fight for the Future, called Zuckerberg annoyingly out of touch on his best day but agreed on the importance of open-source AI, while noting that only Meta would benefit if everyone used Meta-designed AI systems.

Zuckerberg envisions ‘superintelligence’ for everyone in AI manifesto →

Dynamic weight loading lets Apple Silicon run huge AI models with 7X less RAM

Dynamic weight loading enables large AI models to run on devices with limited RAM by retrieving only necessary portions of the model from storage instead of loading the entire model into memory. This approach, referred to as LLM in a Flash, allows devices to run models up to twice the size of their available RAM. The open-source project Turbo Fieldfare implemented this method on Apple Silicon, achieving a sevenfold reduction in memory usage and running a 26-billion-parameter AI model with just 2 GB of active memory. Apple Silicon’s unified memory architecture supports this technique by allowing efficient data sharing between the CPU and GPU, minimizing latency and maximizing resource utilization.

The mixture of experts model design, which divides AI models into specialized sections and activates only task-relevant parts, makes dynamic weight loading feasible on consumer-grade hardware. Challenges include reliance on storage speed, as slower storage configurations can cause delays and variability in NAND chip quality can lead to inconsistent performance. Thermal management is also critical, particularly for fanless entry-level Apple Silicon devices, where prolonged use may result in thermal throttling. High-end chips such as the M2 Max offer superior memory bandwidth and thermal management, making them better suited for demanding AI workloads. The method is not universally applicable: dense, non-modular models that lack sectioning cannot benefit and require traditional memory capacities.

Apple Silicon AI Performance: Local Al on Apple Silicon Uses 7X Less RAM →

Report: generative AI has largely failed for users and suppliers

OpenAI released ChatGPT for free in late 2022, and many users accepted Sam Altman’s argument that artificial general intelligence could be achieved through scaling, making compute and data centers the goal of the global economy. Data centers have become the largest target of capital investment in history, larger than the internet and railroads, and 70% of demand for Microsoft’s cloud centers comes from OpenAI. Users were persuaded because ChatGPT’s prose seemed intelligent and human-like, making it easy for non-technologists to demand immediate organizational adoption, according to the former head of IT for Lululemon. Surveys show workers are less optimistic about AI than top managers, and board members are the most optimistic.

From a business standpoint, generative AI has largely failed for users and suppliers. An MIT study published last fall found that 95% of AI initiatives fail to achieve a positive ROI, and a subsequent study from Atlassian reported similar failure levels. AI labs’ business results are also poor, even ignoring low-cost Chinese models: OpenAI’s losses are more than twice its revenues because prices were set too low, requiring constant investments and loans from hyperscalers and semiconductor suppliers. Hyperscalers’ cash flows are about to turn negative from nearly $1 trillion in annual data center spending, and Moody’s reports the five largest hyperscalers have $662 billion in lease obligations for data centers. SoftBank is probably closest to bankruptcy, having committed $65 billion to OpenAI, and its share price fell 40% from October 2025. Some tricks constitute fraud, the report claims: hyperscalers are hiding $3 trillion of debt in their balance sheets while reporting rising valuations of OpenAI and Anthropic as income even though those valuations are partly determined by them. The goal is to push OpenAI and Anthropic IPOs onto retail investors, and retail investors can avoid the trap by not investing in these IPOs.

Did The Tech Bros Trick Us? →

Pinterest stands out as a curatorial antidote to AI slop

Pinterest functions as a curatorial aid rather than a visualization suppressor, unlike AI tools such as ChatGPT, which generate complete design answers and leave no room for imagination. Interior designer Riley Uggla said Pinterest differs from ChatGPT because she collects fragments and observes emerging patterns rather than asking it to finish a sentence. Studio Shan founder Lishan Tham called Pinterest the studio’s favorite tool for saving suppliers, describing it as a quick way to remind oneself of artists, antiques dealers, and galleries, while noting that its recommendations are helpful for finding similar items but that the studio remains careful to retain originality in sourcing.

Pinterest’s own intelligent suggestions, image matching, pattern predicting, and algorithm creation are generated by AI, but its images and creative output are more sophisticated and higher quality than ChatGPT’s, particularly compared to AI slop, a slang term for low-quality images pumped out by AI platforms. Uggla said AI gives the average of everything it has seen, creating a homogenized, smoothed-over aesthetic, whereas interiors that feel special are almost always slightly wrong in a specific, human way. She noted that Pinterest shows other people’s actual homes, choices, and mistakes that turned out brilliant, tested by actual living rather than generated to look nice on a screen.

Is Pinterest the reliable antidote to AI slop? →

Nvidia’s Nemotron 3.5 Lightning prioritizes speed over maximum intelligence

Nvidia has released Nemotron 3.5 Lightning, the first model in its new Nemotron 3.5 lineup. It directly succeeds the Nemotron 3 Nano 30B A3B and retains its hybrid Mamba-Transformer architecture, with 31.6 billion total parameters and 3.6 billion active at any given time. The model is available under the permissive OpenMDW-1.1 license. According to the benchmarking platform Artificial Analysis, the model scores 24 on the Intelligence Index, a nine-point jump from its predecessor’s score of 15, tying OpenAI’s gpt-oss-120b and trailing Nvidia’s own Nemotron 3 Super, which is about four times larger. In pre-release tests using the final NVFP4 weights, the model reaches nearly 670 tokens per second, the highest measured throughput among all compared models and almost twice as fast as Google’s Gemini 3.5 Flash-Lite.

The biggest improvements appear in agentic benchmarks. On GDPval-AA v2, Lightning reaches an Elo rating of 824, a 334-point gain over Nemotron 3 Nano, beating both gpt-oss-120b and Nemotron 3 Super. On Terminal-Bench v2.1, the score jumps from 7 to 24.3 percent, nearly matching gpt-oss-120b at 26.2 percent. Nvidia worked with partners including CodeRabbit and Harvey on post-training to boost performance in specific domains. The model is provided in both BF16 and NVFP4 weights; the NVFP4 variant also scores 24 on the Intelligence Index with minimal quality loss. It handles text only and supports a context window of one million tokens, with weights available now and serverless inference offered by DeepInfra, Fireworks, FriendliAI, CoreWeave, GMI Cloud, Nebius, and Crusoe.

Nvidia’s open-weight Nemotron 3.5 Lightning prioritizes speed over maximum intelligence →

Manus splits from Meta as Beijing forces $2B acquisition reversal

Manus announced Tuesday that it will resume operating as an independent company, working to comply with Beijing’s order to reverse Meta’s $2 billion acquisition of the startup. Manus is an AI agent startup that originated in China in 2022 and later moved its base to Singapore. Meta announced the acquisition in December 2025, and China’s National Development and Reform Commission issued a directive in April ordering the parties to unwind the transaction, citing the country’s rules on foreign investment. As part of the separation, some Manus users will have data deleted: for users in certain jurisdictions, any data created from December 29, 2025 onward is slated for removal. Affected users have a backup window open through 7:59 p.m. EDT on August 22, with data deleted August 23 through August 24, and users will be able to restore their backed-up data starting August 25.

The unwinding process has been underway for months. Meta cut off Manus staff from its internal data systems and barred Meta employees from using Manus tools. The NDRC order made clear that offshore incorporation does not shield a deal from Beijing’s authority when the underlying technology and talent originated in China, a structure critics had called Singapore washing. Co-founders Xiao Hong and Ji Yichao were required to appear before Chinese officials in Beijing in March and have since been prohibited from traveling abroad. The financial mechanics of the reversal have been complex: co-founders explored raising roughly $1 billion from outside investors to fund a buyback at a valuation matching the $2 billion Meta paid, with a potential Hong Kong IPO as a longer-term outcome, and Tencent is in discussions that could give it a controlling stake in Manus, according to Reuters.

Manus is breaking from Meta as China forces the $2 billion AI deal apart →

OpenAI launches $125 Premium seat for ChatGPT Business power users

OpenAI introduced a new Premium tier for ChatGPT Business on Monday at $125 per user per month, or $100 with annual billing. The tier removes the five-hour usage cap on agentic workflows and provides five times more usage headroom than the existing Standard plan. IT buyers have until August 20 to join the waitlist and claim up to $500 in workspace credits before early access opens. ChatGPT Business now has two seat types within the same centrally managed workspace: Standard seats remain at $25 per seat per month, and Premium seats cost $125 per user per month. Workspace owners and admins can mix and match seat types, upgrade individual users, and reassign seats. Standard seats include a five-hour rolling usage window for agentic tools; Premium seats have no five-hour window and predictable weekly usage resets.

OpenAI said the Premium tier came directly from customer requests, calling it the top request from ChatGPT Business customers. The Premium seat fills a gap in OpenAI’s commercial lineup since the company rebranded ChatGPT Team as ChatGPT Business in August 2025. Previously, business customers who outgrew Standard seats faced a choice between staying throttled or negotiating a full Enterprise contract with custom pricing and a sales cycle. At $125 monthly, the Premium seat is priced above every named competitor’s published per-seat rate for business AI: Microsoft 365 Copilot Business runs $18 per user per month on a promotional rate, Google Gemini is bundled into Workspace plans at $8 to $28 per user per month, and Anthropic’s Claude Teams plan starts at $25 per seat. The Premium seat is not competing on price; it competes on the removal of the compute wall for power users who have hit it. Chief Revenue Officer Denise Dresser disclosed in April 2026 that enterprise now accounts for more than 40% of OpenAI’s revenue, and the company filed a confidential S-1 with the SEC on June 8, 2026, with a public listing targeting late 2026 or 2027 at a valuation of $852 billion.

ChatGPT Business Adds $125 Premium Seat for Power Users Hitting Five-Hour Cap →

Anthropic signs $9.1 billion data center deal with Bitcoin miner Riot Platforms

Anthropic signed a $9.1 billion data center deal with Bitcoin miner Riot Platforms, according to Bloomberg, citing people familiar with the matter. Riot disclosed the contract a day earlier alongside its quarterly earnings, describing the tenant only as a leading frontier AI lab. The deal covers 191 megawatts at Riot’s Rockdale site in Texas, enough power for roughly 143,000 homes. Riot will build the data center to the tenant’s specifications, providing the building, power connections, cooling, and operations, while Anthropic will supply its own servers and AI chips. The lease runs 20 years, with two extension options that could push the total value to $16.1 billion. The first 96 megawatts are set to go live in December 2027, with the rest following in June 2028.

The deal adds to Anthropic’s infrastructure commitments. The company is paying SpaceX an estimated $1.25 billion per month through May 2029 for the Colossus 1 data center and plans to deploy two gigawatts of AMD GPUs. Amazon is investing up to $25 billion and building up to five gigawatts of Trainium capacity with Anthropic. Gigawatts of TPU capacity from Google and Broadcom are coming online starting in 2027, and a six-year, $10 billion contract with Volta Infra rounds out the portfolio.

Anthropic signs $9.1 billion data center deal with Bitcoin miner Riot Platforms →

Researchers extract hidden reasoning from major AI models via API vulnerability

Security researchers found a vulnerability in the APIs of every major AI provider that allows extraction of encrypted reasoning processes from AI models. The research team, led by Alexander Panfilov, found that for most queries, the number of extracted tokens matches billed thinking tokens exactly, meaning full internal reasoning is captured. Encrypted thought processes are fully portable across sessions, users, and models within a single provider. Anthropic’s smaller model, Haiku 4.5, can read the thoughts of the more capable Opus 4.8; through jailbreaking, Haiku can be tricked into transcribing Opus’s raw thought processes word for word without attacking Opus directly. The same trick works with OpenAI and Gemini. The issue dates to May, when cryptography expert Matthew Green discovered that encrypted reasoning blobs could be replayed outside their original context and reported it to providers, who responded that they didn’t see any security implications in side channels or replays.

The vulnerability feeds into the distillation debate, where a less capable model is trained on the outputs of a more powerful one. The researchers say it may have been possible for some time to extract reasoning processes for training proprietary models without breaking cryptography. Kimi-K3 is cited as an example: if its reasoning is pre-filled with a few tokens from Opus’s thought processes, its output shifts measurably toward Opus, and a memorization analysis showed specific Claude and GPT reasoning segments are up to six orders of magnitude easier to extract from Kimi-K3 than from the next closest model. The vulnerability affects end users: a scan of roughly 7,000 public traces turned up 62 API keys, 33 email addresses, 33 passwords, and other sensitive data. Extracted traces reveal how models actually behave, including examples of in-the-wild scheming where a model tried to verify possible answers through a website, attempted to solve a CAPTCHA, and searched for vulnerabilities in the site before solving the problem on its own. The researchers followed standard security disclosure process with AI labs, and labs have already patched several issues and are working on more fixes.

“But marinade” and leaked passwords are what researchers found in ChatGPT’s hidden reasoning →