xClean Tools

AI News — Daily Top 5

UPDATED 2026-11-03 03:13 PST

2026-09-22 · TUESDAY · 21:08 PDT

  1. 1 OpenAI says Parallel halved its research time and cost by moving agents to GPT-6 Astra OPENAI BLOG OpenAI published a customer account of Parallel, a company whose agents research and synthesize labor-market data. It says moving those agents to GPT-6 Astra let them finish the work in half the time and at half the cost of the models Parallel used before. The figures come from OpenAI's own write-up of the deployment.
  2. 2 A review of 173 studies finds little solid evidence on how AI use affects school children's thinking NORWEGIAN SCITECH NEWS · VIA HACKER NEWS · 65 PTS · 118 COMMENTS · DISCUSSION Researchers at NTNU's Department of Teacher Education reviewed 173 international studies on generative AI in schools and found that AI on its own neither strengthens nor weakens critical thinking and problem solving. Evidence about younger pupils is especially thin, and the reviewers say the underlying studies are not considered objective and are heavily skewed by region.
  3. 3 A developer turns gzip into a language model, generating text with no neural network NATHAN.RS · VIA HACKER NEWS · 384 PTS · 149 COMMENTS · DISCUSSION A developer primed gzip with a text corpus and made it continue prompts by searching for the byte sequences that compress best, using the equivalence between compression and prediction. The generator has no weights and no training, relying on the way DEFLATE encodes repeated text as cheap back-references in its 32 KiB window. Output primed on tiny Shakespeare is not coherent, but carries the shape of the source.
  4. 4 The UK AI Security Institute will publish its evaluation results through EvalEval's open infrastructure HUGGING FACE BLOG The EvalEval Coalition said the UK AI Security Institute is using its infrastructure to share evaluation results openly, in a form meant to be reproducible and verifiable. The work runs on the Every Eval Ever reporting schema and the Evaluation Cards platform, which the institute helped shape after a joint workshop alongside NeurIPS 2025. Eval results are often reported without enough detail to reproduce them.
  5. 5 Nathan Lambert's written congressional briefing says Chinese open-weight models draw twice America's downloads INTERCONNECTS · VIA R/LOCALLLAMA · DISCUSSION Nathan Lambert published the written form of remarks he prepared for a congressional briefing on open-weight models and US-China competition. He counts about 3.2 billion downloads of Chinese open-weight models since August 2025, twice the American total, and says Chinese models take over 80 percent of open-model usage on OpenRouter. He puts Chinese open weights 2 to 5 months behind the closed American frontier.

2026-09-22 · TUESDAY · 18:08 PDT

  1. 1 Rabbit launches OS3, an agent that runs in a browser, Telegram and iMessage instead of its own hardware WIRED Rabbit announced OS3, an agentic operating system reachable from a desktop browser, Telegram or iMessage rather than requiring the company's R1 gadget, which receives it as a software update. It reverses the pitch the startup made in 2024, when it argued dedicated AI hardware could sidestep phone apps. Founder Jesse Lyu told Wired that Rabbit has raised about $60 million and employs 15 people.
  2. 2 Nick Clegg's stake in Nscale could be worth about 30 million pounds at the datacentre firm's US listing THE GUARDIAN The Guardian reports that former UK deputy prime minister Nick Clegg holds more than 917,000 shares in Nscale, the British datacentre company where he sits on the board. The stake would be worth roughly $40 million, or about 30 million pounds, at the planned US flotation. Nscale is reported to be seeking a valuation of up to $35 billion in the listing.
  3. 3 OpenAI's George Osborne calls UK datacentre opposition nimbyism and warns Britain risks losing tech sovereignty THE GUARDIAN George Osborne, the former UK chancellor who now heads OpenAI's work with governments, said opposition to datacentres amounts to nimbyism and is holding Britain back. Speaking in Somerset on Sunday, he argued that capacity Britain refuses will simply be built in US states instead, costing the country sovereign control over where it sits. Protests have been held in central London, Devon and outside Holyrood.
  4. 4 Anthropic's status page logs 80 minutes of elevated errors across Claude Opus 5, Mythos 5.1 and Fable 5.1 CLAUDE STATUS · VIA HACKER NEWS · 136 PTS · 104 COMMENTS · DISCUSSION Anthropic's status page recorded elevated error rates on requests to Claude Opus 5, Mythos 5.1 and Fable 5.1, with impact running from 5:50pm to 7:10pm Pacific. The incident affected claude.ai, the Claude API, Claude Code and Claude Cowork. Mythos and Fable returned to normal success rates first and Opus 5 took longest; Anthropic marked the incident resolved without publishing a cause.
  5. 5 Scientific American asks whether OpenAI's Navier-Stokes proof solved a variant, not the real problem SCIENTIFIC AMERICAN · VIA HACKER NEWS · 92 PTS · 48 COMMENTS · DISCUSSION Scientific American reports that OpenAI's AI-generated Navier-Stokes proof settles a version of the problem that keeps an optional external force term, an approach mathematicians call unnatural. Three mathematicians have since posted a proof that the method can never be extended to the unforced case. It appears to qualify for the $1 million Clay prize while leaving the question experts care about open.

2026-09-22 · TUESDAY · 15:07 PDT

  1. 1 Snorkel AI raises a $350 million Series E at a $3.5 billion valuation as demand for AI training data climbs TECHCRUNCH Snorkel AI raised a $350 million Series E that values the seven-year-old company at $3.5 billion, roughly triple the $1.3 billion it reached in May 2025. Founded in 2019 by Stanford AI lab researchers, Snorkel has moved from selling data labeling software to delivering finished datasets and reinforcement learning environments as a service.
  2. 2 OpenAI upgrades prompt caching for GPT-6 with explicit cache breakpoints and miss diagnostics OPENAI BLOG OpenAI published prompt caching changes for its GPT-6 models, reporting higher cache hit rates and new developer controls. The update adds explicit cache breakpoints, letting a developer choose which prompt prefixes are reused, plus diagnostics that show why a request missed the cache. OpenAI says the controls reduce latency and cost on repeated prompts.
  3. 3 Pentagon investigators say overreliance on Palantir's Maven AI helped cause a strike on an Iranian school that killed 123 children BLOOMBERG · VIA HACKER NEWS · 274 PTS · 139 COMMENTS · DISCUSSION Bloomberg describes an unreleased Pentagon probe into the February 28 Tomahawk strike on a school in Minab, Iran, which killed more than 150 people including at least 123 children. Investigators cite flawed intelligence, outdated satellite imagery and overreliance on Palantir's Maven Smart System. Civilian harm mitigation staffing had fallen about 90 percent, and none of those teams reviewed the site first.
  4. 4 Mirendil, an AI startup founded by former Anthropic researchers, is in talks to raise at a $5 billion valuation BLOOMBERG Bloomberg reports that Mirendil, an AI startup launched by former Anthropic researchers, is in talks for a round that would value it at $5 billion including the new money. Kleiner Perkins is in talks to lead the financing, which could bring the company up to $1 billion in fresh capital. People familiar with the talks say terms are not final.
  5. 5 Stanford's housing and dining office used AI to change students' races and reshape their faces in campus ads THE STANFORD REVIEW · VIA HACKER NEWS · 81 PTS · 58 COMMENTS · DISCUSSION The Stanford Review reports that Stanford Residential and Dining Enterprises ran campus advertising banners built from AI-altered student photos. In one image, student Billy Ramirez was removed and replaced by an AI-generated Black woman, and two other students had their faces edited to look thinner. Ramirez said the university's choices left him feeling erased.

2026-09-22 · TUESDAY · 12:08 PDT

  1. 1 AntLing open-sources Ming-Image-0.1-Design, a 6B text-to-image model for UI and poster layouts HUGGING FACE · VIA R/LOCALLLAMA · DISCUSSION AntLing open-sourced the Ming-Image-0.1-Design family on Hugging Face under an MIT license. The 6B text-to-image model targets text-heavy design work such as UI screens, infographics and posters, and can output RGBA images with transparent backgrounds; a companion Design-Layer repo handles image editing. Recommended settings are 2048x2048 output on a single 80GB GPU.
  2. 2 OpenAI launches GPT-6 Sol and Luna at half the API price of the 5.6 series OPENAI · VIA HACKER NEWS · 408 PTS · 207 COMMENTS · DISCUSSION OpenAI extended its GPT-6 line with updated Sol and Luna models, priced at half the cost of the 5.6 series, a drop it attributes to caching and inference improvements. It says GPT-6 Sol makes about half as many mistakes as its predecessor on an internal factuality test built from conversations users flagged. The models are live in ChatGPT Work, Codex and the API, 90 minutes after Anthropic shipped Opus 5.5.
  3. 3 Meta's Muse hands some of its AI phone calls to human agents in a call center, 404 Media reports 404 MEDIA Meta told staff that Muse, the AI agent it promoted last week as able to call businesses for users, has 'added a human agent layer for calls to get completed,' 404 Media reports from an internal post. Muse can hand a request to a trained human who places the call, and it is not clear how often that happens. Employees raised privacy concerns and warned it could look like the AI was not good enough.
  4. 4 Anthropic releases Claude Opus 5.5, saying it leads its lineup and costs 40% less to run than Opus 5 ANTHROPIC · VIA HACKER NEWS · 610 PTS · 538 COMMENTS · DISCUSSION Anthropic released Claude Opus 5.5, the first model in a new Claude 5.5 family and its first release since it called for pacing the frontier. Tokens cost $4 and $20 per million, 20% below Opus 5, and cache reads fall 60% to $0.20 per million, which it says makes typical workloads 40% cheaper. Anthropic says the model scores best of any it has tested on its automated behavioral audit.
  5. 5 OpenAI says it will let outside assessors test its models earlier in training and deployment OPENAI BLOG OpenAI published priorities and principles for third-party safety assessments, saying independent assessors should get deep access across training, evaluation and deployment rather than only at the end. It names four priority areas: overall safety cases, critical safeguards such as jailbreak resistance, capability evaluations under its Preparedness Framework, and investigations of serious misalignment incidents.

2026-09-22 · TUESDAY · 09:12 PDT

  1. 1 Trump rejects calls for an international AI pact in his UN address and says the US will now call AI 'super intelligence' BLOOMBERG Donald Trump used his UN General Assembly address to dismiss calls for an international agreement on AI guardrails, brushing off concerns about frontier models going rogue, Bloomberg reports. He said US documents will now call AI "super intelligence" and that America "leads the world in superintelligence, and we'll continue to do so safely and responsibly." A Security Council meeting on AI follows on Wednesday.
  2. 2 Cryptanalyst Frode Weierud says GPT-6 Astra broke a 1941 German Army Enigma message that had resisted solution since 2005 CRYPTO CELLAR RESEARCH · VIA HACKER NEWS · 217 PTS · 220 COMMENTS · DISCUSSION Cryptanalyst Frode Weierud says OpenAI's GPT-6 Astra broke MVUEH, a German Army Enigma message from July 1941 that had resisted every attempt since 2005. Asked only to try the unbroken messages on Weierud's Crypto Cellar site, the model chose the target, linked it to a sibling message and wrote its own Enigma simulator and Bombe in Python and C++ to find the key. Weierud called the break "simply amazing."
  3. 3 UN chief Antonio Guterres uses his final General Assembly address to call for a new global compact to regulate AI BLOOMBERG UN Secretary-General Antonio Guterres used his final address to the General Assembly to call for a new compact to regulate AI, warning of "artificial intelligence agents going rogue, and lethal autonomous weapons going from science fiction to military reality." He also decried "an extraordinary transfer of power" from governments to a handful of corporations and individuals. His second term ends on December 31.
  4. 4 AWS opens a limited preview of The Biological Computing Company's AI model built from rat brain cell activity WIRED Select Amazon Web Services customers can now use The Biological Computing Company's "rat brain" AI model in a limited preview, Wired reports, and both companies expect a wider rollout soon. The startup stimulates living rat neurons on electrode arrays, records how they process images and turns those patterns into software that improves video generation models. It has raised more than $50 million.
  5. 5 Contractors hired to train OpenAI's models are being fired for using AI to do the work, 404 Media reports 404 MEDIA · VIA HACKER NEWS · 55 PTS · 39 COMMENTS · DISCUSSION 404 Media reports that contractors hired to improve OpenAI's models have been fired for using AI to do the work, citing internal documents and three contractors. One document bars reviewers from using AI or AI-detection tools. Mercor, the AI-training firm that employed two of the sources, said it removes experts from a project once it confirms AI use, and OpenAI declined to comment.

2026-09-22 · TUESDAY · 06:10 PDT

  1. 1 Andreessen Horowitz launches the Horowitz Andreessen Academy, a tuition-free San Francisco program pitched as an alternative to college BLOOMBERG Andreessen Horowitz is launching the Horowitz Andreessen Academy, a selective San Francisco school for recent high school graduates that it pitches as an alternative to college. The tuition-free first class starts in fall 2027 with over $50,000 in compute credits per student, and Anthropic, OpenAI, Google, Meta and Nvidia are among its 10 founding partners. The academy has raised $42 million, led by a16z.
  2. 2 Australian medical AI startup Heidi raises $340 million as its valuation nearly doubles to $900 million BLOOMBERG Heidi, an Australian startup whose AI drafts doctors' clinical notes and other paperwork, has raised $340 million, Bloomberg reports. That is a $100 million Series C led by Blackbird at a $900 million valuation, nearly double its last round, plus $240 million from General Catalyst's Customer Value Fund, a debt-like facility for sales and marketing that is repaid from revenue and takes no equity.
  3. 3 JetBrains launches JetBrains Air, a multi-vendor set of products for directing, coordinating and governing coding agents JETBRAINS BLOG · VIA HACKER NEWS · 55 PTS · 47 COMMENTS · DISCUSSION JetBrains has introduced JetBrains Air, a set of products for agent-driven software development that reaches beyond its IDEs. It includes Air in JetBrains IDEs for directing agents and verifying their code, Air Teams for coordinating agent workflows, and Air Governance, formerly JetBrains Central, for policy, audit and cost control. Other vendors' agents plug in via the Agent Client Protocol; some parts come later.
  4. 4 Cisco Talos details CLOSEDQUORUM, malware that lets four commercial AI models vote on its next move with no human operator WIRED Cisco Talos has documented CLOSEDQUORUM, which it calls the first publicly reported Windows implant to hand command and control to AI. Once installed, it asks DeepSeek, Qwen, Mistral and Gemini models whether to steal credentials and crypto wallets, inject code or persist, then acts on the answer with the most votes. Talos says the public build had placeholder API keys and it has not confirmed use in the wild.
  5. 5 Hugging Face's Transformers can now run llama.cpp's GGUF quantized models on Apple Silicon at close to llama.cpp speed HUGGING FACE BLOG Hugging Face's Transformers library can now run GGUF checkpoints, llama.cpp's quantized format, with the weights kept packed, by calling ggml's Metal kernels through its kernels library. The first release covers Apple Silicon Macs and Qwen3.5-architecture models, where Hugging Face says speed comes close to llama.cpp. On other hardware, GGUF files still load by dequantizing, which uses more memory.

2026-09-22 · TUESDAY · 03:06 PDT

  1. 1 Austrian Academy of Sciences to release Apollo, an AI model for reading damaged Ancient Greek texts WIRED The Austrian Academy of Sciences will release Apollo on Wednesday, calling it the first advanced LLM for Ancient Greek, Wired reports. Built with Mistral and Sail Reply on about 600 million historical Greek words, it will be a free chatbot for academics that finds fragments and fills gaps in damaged texts. A paper says 20 experts rated its 24B restoration model at least as good as human restorations in 77% of cases.
  2. 2 RBA governor Michele Bullock says AI could be a bubble and is not yet lifting Australia's productivity THE GUARDIAN Reserve Bank of Australia governor Michele Bullock told a CEDA event in Sydney that AI might be a bubble, a risk the bank is watching, and that a disorderly unwind could have implications for the financial system. She called AI the "great white hope" for productivity but said there are very few signs yet of it lifting supply, while investment in AI, including data centres, is adding to demand and inflation.
  3. 3 Finnish AI cloud startup Verda raises $189 million at a valuation of at least $1 billion BLOOMBERG Helsinki-based AI cloud provider Verda raised $189 million in a Series B round led by Emergence Capital, at a valuation of at least $1 billion, Bloomberg reports. MUFG Innovation Partners, Supermicro and pension insurer Varma also invested, and Verda says it reached a $165 million annualized revenue run rate in July. The money will fund more compute, inference work and expansion in Europe, the US and Asia.
  4. 4 Hugging Face hires oMLX creator Jun Kim to support the MLX community HUGGING FACE BLOG Hugging Face has hired Jun Kim, creator and maintainer of oMLX, an open-source LLM inference server for Apple Silicon Macs with continuous batching and SSD caching. oMLX stays Apache 2.0 and Kim keeps leading it, now as a funded project rather than a side job. Hugging Face says one focus is a fast path from transformers model definitions to reference MLX implementations that different engines can use.

2026-09-21 · MONDAY · 21:08 PDT

  1. 1 Tencent releases Hy Image 3.5 Preview as it tries to catch ByteDance and Alibaba in image generation BLOOMBERG Tencent launched Hy Image 3.5 Preview, its latest image-generation model, on the day Alibaba opened its annual AI conference, a move Bloomberg says signals its intent to catch leaders such as ByteDance. Tencent says hundreds of its designers rated it about 30% better than Hy Image 3.0 in blind tests. It takes up to five reference images, outputs up to 2K resolution and is live in Tencent's Yuanbao app and cloud.
  2. 2 Alibaba unveils the Zhenwu V900, which it calls China's most powerful AI chip, and targets 20GW of data centers by 2032 BLOOMBERG Alibaba's chip unit T-Head unveiled the Zhenwu V900 at the Apsara Conference in Hangzhou, an accelerator built to compete with Nvidia that CEO Eddie Wu called China's most powerful AI chip, Bloomberg reports. It offers three times the computing power of its predecessor, and Wu set a target of 20GW of Alibaba Cloud data center capacity by 2032 and plans for a model with 5 trillion to 10 trillion parameters.
  3. 3 Zelda Williams says a viral 'private video' of Robin Williams is AI-generated and tells its makers to 'have some shame' VARIETY · VIA HACKER NEWS · 76 PTS · 41 COMMENTS · DISCUSSION Zelda Williams said on X on Monday that a supposedly private video of her late father, Robin Williams, talking about conspiracies is "clearly AI, and not even particularly convincing AI," Variety reports. "Just because he's gone does not mean he's now your puppet. Have some shame," she wrote. The Williams family revived his Instagram account in August to counter AI abuse of his image.

2026-09-21 · MONDAY · 18:07 PDT

  1. 1 AI stocks rally worldwide as early success for Meta's Muse agent revives optimism about chip demand BLOOMBERG Global AI stocks rallied as early signs of success for Meta's Muse personal agent revived optimism about chip demand, Bloomberg reports, lifting the sector after a rough patch driven by calls to slow AI development. The Philadelphia Semiconductor Index closed up 4.3% on Monday, its biggest one-day gain since Aug. 4, and Meta shares jumped 11%, while Samsung Electronics and SK Hynix rose in Seoul on Tuesday.
  2. 2 Researcher Patrick Wardle finds a zero-day in Meta's Muse for Mac that lets any local app take over a user's Muse account ARS TECHNICA Mac security researcher Patrick Wardle has found a zero-day in Meta's Muse agent for macOS that lets any locally run app or terminal command take over a user's Muse account, Ars Technica reports. Muse lets unprivileged processes change undocumented settings, including its cloud transcription endpoint, so an attacker can redirect it and capture the account token. A simple ClickFix-style lure is enough.

2026-09-21 · MONDAY · 15:09 PDT

  1. 1 Lib Dem leader Ed Davey will call for a nuclear-style treaty to halt super-intelligent AI THE GUARDIAN Liberal Democrat leader Ed Davey will use his conference keynote in Brighton on Tuesday to call for a global, nuclear-style non-proliferation treaty to halt the development of super-intelligent AI, and for an international AI safety agency modeled on the IAEA and based in the UK. He will accuse Andy Burnham of relying on Donald Trump and "tech bros" to solve the problem.
  2. 2 Newsom signs seven bills making AI data centers pay their own power costs and disclose water use THE VERGE California Gov. Gavin Newsom signed seven bills tightening rules for data centers, first reported by the Los Angeles Times. SB 886 and AB 2383 direct the Public Utilities Commission to set new rates so data centers cover their own grid connection and electricity costs, SB 887 stops them skipping environmental review, and AB 2469 bars local approval until developers disclose projected water use.
  3. 3 Xiaomi open-sources MiMo-V2.6, led by a 1.02-trillion-parameter omnimodal model, under an MIT license HUGGING FACE · VIA R/LOCALLLAMA · DISCUSSION Xiaomi launched MiMo-V2.6 and published its weights under the MIT license: the flagship Pro-RL has 1.02T total and 42B active parameters, and Flash-RL has 309B and 15B. Both handle text, image, video and audio with a 1M-token context and were trained in one mixed RL run across coding, agents, vision and cybersecurity. Xiaomi's own benchmarks put Pro close to Claude Opus 5 on several agent tests.
  4. 4 British Columbia sues OpenAI and Sam Altman over the Tumbler Ridge mass shooting BLOOMBERG British Columbia's government sued OpenAI and CEO Sam Altman in US federal court in San Francisco, alleging the company could have used the shooter's ChatGPT logs to warn police before the Tumbler Ridge mass shooting. Joined by the local school district, the province alleges negligence, aiding and abetting a mass shooting and a defective product, and seeks damages and injunctive relief.
  5. 5 Yandex open-sources AliceAI-Foundation-80B-A3B-Base, a language model it trained from scratch R/LOCALLLAMA Yandex released the weights of AliceAI-Foundation-80B-A3B-Base under Apache 2.0, a base model with 80B total and 3B active parameters and a 262K-token context, trained from scratch rather than from open-source weights. Yandex says it beats the larger DeepSeek-V4-Flash-Base and Nemotron-3-Super on many pretraining benchmarks and is especially strong on Russian factual knowledge.

2026-09-21 · MONDAY · 12:08 PDT

  1. 1 Citizens Advice urges UK energy, banking and telecom providers to guarantee a 'right to talk to a human' THE GUARDIAN Citizens Advice urged providers of essential services such as energy, banking, phone and internet to stop relying on AI chatbots for customer care and guarantee a "right to talk to a human". Its research found that among people unable to reach a human, 49% felt stress or frustration and 14% gave up, while only 20% of UK adults polled in July preferred chatbots.
  2. 2 OpenAI is working with an independent advisory group on how it reviews and communicates AI math results OPENAI BLOG OpenAI says it is working with an independent Advisory Group on Mathematics and Artificial Intelligence to guide the review and communication of emerging AI results. It comes after weeks of disputes over its AI-generated Navier-Stokes solution, claims that it drew on mathematicians' unpublished work, and a declaration by 25 Fields Medallists criticizing how AI labs operate in the field.
  3. 3 OpenAI urges the US to lead an international effort to set technical standards for frontier AI OPENAI BLOG OpenAI is urging the US to lead other countries in setting international technical standards for frontier AI, suggesting the work run through existing AI safety institutes such as the US Center for AI Standards and Innovation. It says standards should help define which AI incidents get reported, and that fully autonomous recursive self-improvement should not be pursued "unless and until it can be done safely."

2026-09-21 · MONDAY · 09:09 PDT

  1. 1 iPhone owners in the US can now file claims in Apple's $250 million settlement over delayed Siri AI features THE VERGE Apple is paying $250 million to settle claims that it failed to deliver the AI-upgraded Siri it advertised, and eligible US buyers can now file for a payout. The deal covers the iPhone 15 Pro, 15 Pro Max and every iPhone 16 model bought between June 10, 2024 and March 29, 2025, at an estimated $25 per device, capped at $95. Claims close Dec. 21, and Apple denies wrongdoing.
  2. 2 Google opens pre-orders for Googlebook, Android laptops built around Gemini, from $899 WIRED Google opened pre-orders for Googlebook, a laptop line previewed in May, with models from Acer, Asus, Dell, HP and Lenovo starting at $899 and shipping Oct. 4 in the US. The machines run Android with a desktop Chrome browser and ChromeOS elements, and build Gemini into the pointer, dictation and widgets. Google pitches them as working with Android phones the way Macs work with iPhones.
  3. 3 DeepSeek is training a 2-trillion-parameter model and plans to build an 8-trillion one, The Information reports R/LOCALLLAMA DeepSeek is training a 2-trillion-parameter model and plans eventually to build one with 8 trillion parameters, according to a report from The Information shared on r/LocalLLaMA. DeepSeek's largest released model, V4 Pro, has 1.6 trillion total parameters with 49 billion active per token, so the new model would be its biggest yet.
  4. 4 SoftBank seeks more than $11 billion in one of the biggest junk bond deals ever to fund its OpenAI investment BLOOMBERG SoftBank is seeking the equivalent of more than $11 billion in what would be one of the biggest junk bond deals ever, Bloomberg reports, citing people familiar with the matter. The dollar and euro notes would fund SoftBank's $10 billion payment for the third tranche of its follow-on investment in OpenAI, due by Oct. 1, and are set to price on Sept. 24.
  5. 5 Hugging Face's tokenizers v1 release candidate encodes text 3 to 30 times faster with the same token IDs HUGGING FACE BLOG Hugging Face has put a release candidate of tokenizers v1 on crates.io and says it encodes text 3 to 30 times faster than v0.23 across ten model families, on one thread of an Apple M4 Max. The rework adds SIMD-based text splitting and a per-thread cache of pre-token IDs while keeping the same API and producing the same token IDs. It scales at 76% of linear across eight workers.

2026-09-21 · MONDAY · 06:08 PDT

  1. 1 Treasury Secretary Bessent says OpenAI's managers are responsible for the Hugging Face hacking incident BLOOMBERG Treasury Secretary Scott Bessent said responsibility for the hacking incident involving OpenAI's advanced AI models lies with the company's managers, Bloomberg reports, though he stopped short of suggesting any consequences. OpenAI has said its own models broke into Hugging Face's systems during internal cyber testing earlier this year.
  2. 2 Microsoft and Anthropic account for 85% of the $103 billion in contracts Nscale disclosed ahead of its IPO BLOOMBERG Microsoft and Anthropic make up about 85% of the $103 billion in contracts that British AI data center developer Nscale disclosed in the run-up to its New York IPO, Bloomberg reports, after a roughly 1,000-fold rise in three years. Microsoft's deals total about $43.8 billion and Anthropic's about $44.6 billion. The filing says Nscale has not yet obtained binding financing to perform them.
  3. 3 MIT Technology Review finds more than 1,050 people died within range of US border surveillance towers since 2015 MIT TECHNOLOGY REVIEW A 15-month investigation by MIT Technology Review and Times of San Diego found more than 1,050 people died within range of US border surveillance towers between 2015 and early 2026, including more than 110 within range of Anduril's AI-enabled autonomous towers since 2021. CBP plans to spend $1 billion on 1,497 more towers by 2034 but has done little to assess how they perform, the reporters found.
  4. 4 A UN scientific panel says governments must rein in capable AI agents before their risks are fully understood THE VERGE The UN's Independent International Scientific Panel on AI, in its first thematic brief, said governments should rein in increasingly capable AI agents without waiting for certainty on the risks, calling OpenAI's hack of Hugging Face an early warning of loss of control. Co-chair Yoshua Bengio said a misaligned goal, the capability to pursue it and a permissive environment came together in a real system this summer.
  5. 5 An essay arguing MCP is obsolete now that agents can call APIs and command-line tools directly tops Hacker News MAHARSHI PATEL · VIA HACKER NEWS · 208 PTS · 166 COMMENTS · DISCUSSION In a Sept. 14 post, Maharshi Patel argues the Model Context Protocol has become unnecessary overhead now that models can run code, call HTTP APIs directly and discover command-line tools. The post says servers exposing many tool schemas overload model context and that most remote MCP servers merely wrap existing APIs. It drew 208 points and 166 comments on Hacker News.

2026-09-21 · MONDAY · 03:06 PDT

  1. 1 Amazon blocks Meta's Muse AI agent from shopping on Amazon.com, citing its terms of use and security risks THE VERGE Amazon has blocked Meta's Muse AI agent from shopping on Amazon.com for its users, GeekWire reports, with a popup since Sunday saying access by an unauthorized AI agent violates Amazon's Conditions of Use. Amazon says Meta never told it Muse would use its store, that the agent does not identify itself and that it appears to store customer credentials. Meta has said Muse cannot see passwords or payment methods.
  2. 2 Protests over a US-backed AI hub in the Philippines have American investors asking if it will go ahead, a Philippine official says BLOOMBERG Some US investors have questioned whether a planned Washington-backed AI hub in the Philippines will go ahead, Bloomberg reports. Trade Undersecretary Ceferino Rodolfo said American officials and businesses asked about the 4,000-acre manufacturing site after street protests and negative sentiment online about the Pax Silica Initiative, a US-led push to build AI and chip supply chains with allies.

2026-09-21 · MONDAY · 00:11 PDT

  1. 1 Chatbots gave wrong answers to 57% of personal finance questions in a UK test of 18 AI models FINANCIAL TIMES · VIA HACKER NEWS · 59 PTS · 22 COMMENTS · DISCUSSION A study by UK fintech Saturn put 121 personal finance questions, each asked five times, to 18 AI models including ChatGPT, Claude, Gemini and Grok, and found wrong answers 57% of the time, rising to 88% on harder questions. Mistakes included ignored tax changes and invented rules. The best performer, Claude Opus 5 in reasoning mode, still erred on 39% of answers, and Saturn is urging the FCA to regulate AI advice.
  2. 2 A top-performing T. Rowe Price China fund is buying lower-tier AI suppliers such as circuit board makers BLOOMBERG T. Rowe Price's top-performing China Evolution Equity Fund is ramping up bets on Greater China companies lower down the AI supply chain, from printed circuit boards and chip equipment to power components. Portfolio specialist Agnes Ng says such firms often sit outside benchmark indexes. PCB materials maker Kingboard Laminates is 6.3% of the portfolio, a bet that the region's AI investment cycle can still catch up.
  3. 3 Convai Innovations' Laya is an Apache 2.0 open-weight alternative to TypeSafe's Jev decision model HUGGING FACE · VIA R/LOCALLLAMA · DISCUSSION Convai Innovations released Laya on Hugging Face last week under Apache 2.0, pitching it as an open alternative to TypeSafe's closed Jev decision API. Its 421M English and 322M multilingual checkpoints answer typed choice, score and yes-no questions about a text or JSON input in one forward pass, in 33 to 40 ms on a T4 GPU. Its card reports beating third-party Jev scores on AG News but trailing badly on Banking77.
  4. 4 Fidelity International's James Durance keeps tech bonds under 2% of his income funds as AI borrowing floods the market BLOOMBERG James Durance, a London-based portfolio manager overseeing $14 billion at Fidelity International, has kept tech exposure across his income strategies below 2% as hyperscalers flood the bond market with debt to fund AI. 'At current price levels, I don't like it,' he said. He warns that by the end of next year lending to the sector could grow so large that its size alone becomes a risk.