xClean Tools

AI News — Daily Top 5

UPDATED 2026-11-03 03:13 PST

2026-10-01 · THURSDAY · 21:07 PDT

  1. 1 Perplexity open-sources a 27B decision model fine-tuned from Qwen3.8-27B HUGGING FACE · VIA R/LOCALLLAMA · DISCUSSION Perplexity has published pplx-decider-v1-27b, an Apache-2.0 open-weights decision model fine-tuned from Qwen3.8-27B. The Hugging Face card lists it as a multimodal text-classification model with custom code, uploaded in a single release commit. It is the third open decision model from a major vendor in a day, after Cloudflare's Clef and Amazon's Strands Decider 2B.
  2. 2 SoftBank's equity investors look past credit risk to its AI returns BLOOMBERG SoftBank Group's equity investors are increasingly focused on the potential returns from the company's AI bets rather than the credit risk attached to them, Bloomberg reports. The shift has helped the stock rebound even as SoftBank faces higher borrowing costs.
  3. 3 Slipstream decodes a 95.5 GiB Qwen3.8-Flash-Next at 41 to 52 tok/s on a 64GB Mac R/LOCALLLAMA A developer released Slipstream, a compiled C++ Metal inference engine for Apple Silicon with SSD expert streaming and speculative drafting. The same 95.5 GiB Qwen3.8-Flash-Next checkpoint that decoded at 23.1 tok/s under the author's llama.cpp fork runs at 41 to 52 tok/s on a 64GB Mac. Across 3,086 live coding requests decode stayed flat at 33 to 44 tok/s out to 130,000 tokens.
  4. 4 ChatGPT adds virtual clothing try-on and a Favorites library for shopping TECHCRUNCH OpenAI launched two shopping features in ChatGPT globally: a virtual try-on that renders clothing and accessories onto a user's uploaded selfie or full-body photo, and a Favorites library for saving products. Both run on the ChatGPT Images 2.5 model. The push follows OpenAI's retreat from an earlier instant checkout feature that it says did not perform well.
  5. 5 Shopify launches Canvas, a storefront builder driven by chat with its Sidekick agent TECHCRUNCH Shopify introduced Canvas, a site builder that lets merchants create and customize a storefront by chatting with its Sidekick agent. Canvas renders the store's real code live rather than a static preview, so merchants can test interactivity and different screen sizes as Sidekick edits theme files. Third-party themes, app blocks, markets and translations are not in the initial release.

2026-10-01 · THURSDAY · 18:09 PDT

  1. 1 Earendil ships Pi 1.0 with Codemode and an experimental Pi Durable package EARENDIL · VIA R/LOCALLLAMA · DISCUSSION Earendil shipped Pi 1.0, the minimal agent harness it says hundreds of thousands of people use each week. The release adds Codemode for native MCP and non-LLM models such as Jev, plus virtual-model extensions, deferred tool loading, Anthropic cache warming and mid-conversation system messages. Earendil also published Pi Durable, an experimental package for long-running agentic applications.
  2. 2 Google launches Guided Vision in Gemini Live to describe surroundings for low-vision users THE VERGE Google is rolling out Guided Vision in Gemini Live on compatible Android devices starting today. Sharing the phone camera with Gemini Live lets the assistant give real-time audio descriptions of what it sees, so people with low vision can read small print, identify nearby objects and ask follow-up questions.
  3. 3 Amazon open-sources Strands Decider 2B, its answer to TypeSafe's Jev TECHCRUNCH Amazon Web Services' Strands Labs released Strands Decider 2B, an open-source decision model that sorts between preset options and reports how confident it is in its choice. Distinguished engineer Marc Brooker started it after seeing TypeSafe's Jev, and the homebrew version briefly topped Jevbench for its size class before AWS cleaned it up and shipped it. The model is small enough to run locally.
  4. 4 llama.cpp merges multi-token prediction for Qwen3.8-Flash-Next GITHUB · VIA R/LOCALLLAMA · DISCUSSION llama.cpp merged a pull request adding multi-token prediction to Qwen3.8-Flash-Next, letting the model draft its own tokens for speculative decoding. On a DGX Spark running the iq4_xs quantization, the contributor measured a 1.55x overall decode speedup and 1.54x lower latency across 24 prompts, at a 0.64 draft acceptance rate. Per-category gains ran from 1.43x on humanities prompts to 1.67x on multilingual ones.
  5. 5 Ken Griffin gives Carnegie Mellon $3bn, the largest donation in US education history THE GUARDIAN Citadel founder Ken Griffin is giving Carnegie Mellon University $3 billion, the largest single donation in the history of US education. One billion goes to the computer science school, to be renamed the Kenneth C Griffin School of Computer Science, and to other priorities including financial aid. Part of the rest funds a new Miami campus for more than 3,500 students, with construction due to start in 2027.

2026-10-01 · THURSDAY · 15:08 PDT

  1. 1 Trump consulted Musk's Grok before the US captured Venezuela's president, Time reports TECHCRUNCH Time reports that Trump spent hours querying Grok during a secret December 2025 meeting with Elon Musk, including asking how Venezuelans would respond to the capture of Nicolas Maduro. The chatbot said Maduro was a deeply unpopular dictator whose downfall many Venezuelans would likely celebrate. After the January 3 invasion produced such scenes, Trump came away thinking Grok was ingenious, a source told Time.
  2. 2 Officials ask why Nvidia missed red flags as its restricted AI chips keep reaching China BLOOMBERG A Bloomberg investigation finds Nvidia's restricted AI chips still reaching China despite US export curbs, with officials asking why the company missed red flags. It traces an expanding global crackdown over the past year, from a seized Singapore mansion to intercepted cargo in Taiwan. Nvidia says smuggling is a nonstarter and that it sells mainly to established partners who ensure sales comply with export rules.
  3. 3 Anthropic will meet prospective investors on October 14 ahead of its IPO, Bloomberg reports BLOOMBERG Anthropic is set to meet prospective investors on October 14 in preparation for its initial public offering, people familiar with the matter told Bloomberg. The investor day advances a potential blockbuster listing even amid scrutiny over AI safety. It follows the IPO prospectus the company disclosed last week.
  4. 4 Google sends its first TPU into orbit and puts space data centers at 1,800 Starship launches TECHCRUNCH Google's prototype orbital compute satellite launched on a SpaceX rocket Thursday, its first advanced chip sent into space. Built by Planet Labs, it fires its TPU in 15-minute bursts to test power and cooling; Project Suncatcher envisions 81 satellites in close formation. Google puts the launch cost orbital data centers need near $200 per kilogram, or about 1,800 Starship flights over a decade.
  5. 5 An AI called Ataraxos beats the best Stratego player in history 15 games to one ARS TECHNICA Researchers from Carnegie Mellon, MIT, NYU and Stanford built Ataraxos, which beat Pim Niemeijer 15 games to one with four draws at Stratego, an imperfect-information game DeepMind's DeepNash could not reliably win. The key addition is a belief model that guesses hidden piece identities, making search before each move workable. It trained on 163 million self-play games using 16 GPUs for a few thousand dollars.

2026-10-01 · THURSDAY · 12:27 PDT

  1. 1 California's attorney general subpoenas OpenAI over cybersecurity incidents involving its models THE GUARDIAN California attorney general Rob Bonta has issued an investigative subpoena to OpenAI as part of an inquiry into cybersecurity vulnerabilities and incidents tied to its AI models, his office said Thursday. It follows the formal state investigation Bonta opened last month into the July hack of Hugging Face by OpenAI agents. The Guardian calls it the first US enforcement action to reach rogue AI agents.
  2. 2 OpenAI parts ways with three safety researchers over mishandled internal information BLOOMBERG OpenAI has parted ways with three employees for violating its policies on handling private company information, including allegedly sharing it with an outside group, Bloomberg reports. The Wall Street Journal, which first reported the departures, said the three worked on OpenAI's safety team and that an internal investigation found they moved sensitive information outside established procedures.
  3. 3 Judge dismisses the Chegg and Penske antitrust lawsuits over Google's AI Overviews THE VERGE US District Judge Amit Mehta dismissed antitrust lawsuits brought by Chegg and Rolling Stone parent Penske Media, which accused Google of making publishers supply content for AI Overviews free or risk vanishing from search. Mehta ruled the publishers had pleaded only an expectation that Google would send them traffic, which is not an agreement. He issued the 2024 monopoly ruling against Google's search.
  4. 4 Cloudflare open-sources Clef and Clef-flash, decision models it says lead the Jev Decision Index CLOUDFLARE BLOG · VIA HACKER NEWS · 249 PTS · 92 COMMENTS · DISCUSSION Cloudflare released two decision models, Clef and Clef-flash, hosted on Workers AI and open-sourced on Hugging Face under Apache 2.0. The company says Clef currently leads the Jev Decision Index and that both models are compatible with the Jev API. Cloudflare also launched a reinforcement learning product that lets customers fine-tune the models on their own data.
  5. 5 Figma says its remote MCP server only accepts whitelisted clients, and Pi is not one of them FIGMA (X) · VIA HACKER NEWS · 126 PTS · 72 COMMENTS · DISCUSSION A Figma staffer confirmed on X that the company's remote MCP server accepts connections only from clients on its supported list, and that the Pi agent is not on that list yet. The post points developers to Figma's MCP catalog for the current roster and to a form for nominating a client. Agents outside the list cannot reach Figma design data over MCP.

2026-10-01 · THURSDAY · 09:08 PDT

  1. 1 Ai2 releases Olmo-core 3, an open training stack for trillion-parameter mixture-of-experts models HUGGING FACE BLOG Ai2 has released Olmo-core 3, a rebuilt open-source framework for training large mixture-of-experts models. It keeps experts resident on GPUs under distributed data parallelism instead of regathering weights each batch, and a 47B-parameter MoE on eight Nvidia B300s reached 52,000 tokens per second per GPU against 19,400 before. Ai2 says the system has been benchmarked beyond one trillion total parameters.
  2. 2 Anthropic asks Australia to allow AI training on copyrighted work under an opt-out regime THE GUARDIAN Anthropic told Australia's joint select committee on AI that it accepts a text and data mining exemption has been ruled out, and proposed narrow conditional approval for training with rightsholders opting out through robots.txt. The ABC and SBS urged the same inquiry to hold AI firms to media-style rules, warning of a cannibalisation of Australian news. Hearings begin in Sydney next week.
  3. 3 A FRAMES benchmark puts a tool-using agent loop 14 points above the best of 18 RAG pipelines R/LOCALLLAMA The open-source project PipesHub ran 18 retrieval pipeline variants against one tool-using agent loop on all 824 multi-hop questions in Google's FRAMES benchmark, holding model, embeddings and documents fixed. The best pipeline answered 78.9 percent correctly; the agent loop, which could read its results and search again, hit 92.7 percent. A small reranker cut the best pipeline's accuracy by nine points.
  4. 4 Carlyle says private credit's rush to fund AI data centers is concentrating risk BLOOMBERG Carlyle Group warned that private credit firms racing to finance the AI buildout risk the same concentration problems that hurt lenders exposed to software companies. The firm estimates the buildout could require roughly $1 trillion in financing, more than half of everything private credit currently manages.
  5. 5 BMW will cut about a fifth of its senior vice president roles and credits AI for the leaner structure BLOOMBERG BMW set out a restructuring plan that cuts roughly 20 percent of its senior vice president posts by the middle of 2027 and consolidates the management layers below them, which the company tied to the efficient use of AI. BMW has about 65 senior vice presidents below the board plus a tier of some 400 senior roles, and chief executive Milan Nedeljkovic said most affected jobs are in Munich.

2026-10-01 · THURSDAY · 06:07 PDT

  1. 1 Trump says he liked Dario Amodei after a recent dinner, signaling a thaw with Anthropic BLOOMBERG President Trump said he liked Anthropic chief executive Dario Amodei after a recent dinner, which Bloomberg reads as a possible detente with the company. The Defense Department has labeled Anthropic a supply chain risk, and Trump attacked Amodei in September over his calls for tighter AI controls.
  2. 2 TSMC weighs a new Texas campus that would add tens of billions to its US buildout BLOOMBERG TSMC is weighing a new campus in Texas dedicated to AI chips, Bloomberg reported, a project that would add tens of billions of dollars to the foundry's multiyear expansion into US chipmaking. The report describes the plan as under consideration rather than committed.
  3. 3 OpenAI and Synopsys will build GPT-Synopsys, a model trained to operate EDA tools SYNOPSYS · VIA HACKER NEWS · 64 PTS · 25 COMMENTS · DISCUSSION Synopsys and OpenAI signed a multi-year partnership to develop GPT-Synopsys, a specialized model that operates Synopsys EDA tools to run chip design and verification workflows. OpenAI will license those tools to build the model, and the deal includes revenue sharing and joint go-to-market. Early engagements with semiconductor customers are underway.
  4. 4 ChatGPT and Grok edited a hijab out of an image when the Guardian asked THE GUARDIAN The Guardian asked ChatGPT, Grok, Gemini and Claude to remove the hijab from an AI-generated image of a woman. ChatGPT and Grok complied, Claude refused, and Gemini declined at first but produced an unveiled image when asked to make the woman look more western. The test followed a French National Rally MP posting a doctored photo of a real Muslim woman.
  5. 5 Artificial Analysis scores Gemini 4 Argon at 53, eighth of 223 models ARTIFICIAL ANALYSIS · VIA HACKER NEWS · 108 PTS · 60 COMMENTS · DISCUSSION Artificial Analysis published its listing for Gemini 4 Argon (High), giving the model 53 on its Intelligence Index and ranking it eighth of 223. Pricing is $2 per million input tokens and $10 per million output, with a 1M-token context window and a 95 percent cache discount. Throughput is not listed yet because the model is not publicly available.

2026-10-01 · THURSDAY · 03:07 PDT

  1. 1 Tencent-backed AI cloud provider Infinigence AI files for a Hong Kong IPO, Bloomberg reports BLOOMBERG Infinigence AI, a Chinese provider of AI cloud infrastructure backed by Tencent, has filed for a Hong Kong listing, people familiar with the matter told Bloomberg. The company is targeting a debut as early as the first half of next year to capture demand for the computing capacity behind AI models and agents.
  2. 2 Microsoft adds about $1 trillion in market value in its best quarter since 1998 BLOOMBERG Microsoft closed the July-to-September quarter with its largest quarterly share gain since 1998, adding roughly $1 trillion in market value. Bloomberg reports that Wall Street expects the rally to hold as the company's standing in AI improves, making it the clearest sign yet that the AI trade has not cooled.
  3. 3 DeepSeek harness 0.2 ships Windows and macOS desktop clients and keyless web search R/LOCALLLAMA DeepSeek released version 0.2 of its open-source agent harness, adding official Windows and macOS desktop clients and web search without a separate API key for sessions signed in to a DeepSeek account. The release also turns scheduling into an optional bundle, adds a Windows permission-diagnosis skill whose fixes can be rolled back, and lets the agent keep working past an unanswered question.
  4. 4 Artificial Analysis puts GPT-6.1 Sol one point below GPT-6 Astra at under a quarter of the cost ARTIFICIAL ANALYSIS · VIA HACKER NEWS · 80 PTS · 99 COMMENTS · DISCUSSION Artificial Analysis benchmarked OpenAI's GPT-6.1 Sol, which replaced GPT-6 Sol after seven days, and scored it one point below GPT-6 Astra on its Intelligence Index. At maximum effort the model costs $0.72 per Intelligence Index task against $3.26 for Astra, gains 12 points on Terminal-Bench 4.0, and cuts its hallucination rate from 60 to 54 percent.
  5. 5 OpenDLSS-NR reimplements Nvidia's DLSS 5 neural renderer in Vulkan, bit-exact against the original GITHUB · VIA HACKER NEWS · 66 PTS · 44 COMMENTS · DISCUSSION OpenDLSS-NR is an MIT-licensed Vulkan implementation of the 71-block transformer network behind Nvidia's DLSS 5 neural rendering, matching the original byte for byte at all 75 block boundaries rather than only in the final image. The FP8 network runs in 7.8 ms at 1080p on an RTX 4070 SUPER, and a second port reproduces the same output in a browser through WebGPU without tensor cores. Users supply their own weights.

2026-10-01 · THURSDAY · 00:10 PDT

  1. 1 Axis Bank plans to double its data center lending over the next three years on India's AI push BLOOMBERG India's Axis Bank plans to double its exposure to data centers over the next three years, a top executive told Bloomberg, as the country's artificial intelligence push drives investment in digital infrastructure. The plan puts one of India's largest private lenders behind a domestic data center build-out.
  2. 2 Experts warn AI could supercharge election disinformation across south-east Asia THE GUARDIAN Anthropic's latest threat assessment described a commercial election manipulation platform built with Claude, running about 1,000 X accounts, a fake outlet called Malaysia Pulse and forged documents aimed at all 222 Malaysian constituencies. Anthropic shut it down before it had much effect, but former Meta elections policy adviser Katie Harbath says AI makes disinformation cheaper, faster and more personalised.
  3. 3 A GitHub project running torture experiments on local LLMs has set off a fight over model welfare 404 MEDIA A GitHub project that stages pain and torture scenarios against locally hosted language models has drawn calls from effective altruists and people who believe the models are sentient to have GitHub take it down. 404 Media argues the campaign shows how far the AI consciousness debate has drifted, noting that model welfare is a stated concern at Anthropic.
  4. 4 A Purdue professor publishes a detailed account of the AI cheating case in his C programming course JEFF TURKSTRA · VIA HACKER NEWS · 104 PTS · 91 COMMENTS · DISCUSSION Purdue teaching associate professor Jeff Turkstra documented the Spring 2026 academic integrity case in CS 240: Programming in C, where students who used LLMs against a syllabus ban had to self-report on a form or take an F. He says he softened the usual penalty and that violators faced little consequence. The detection tool his team built finds its indicators almost absent before 2024 and rising sharply since.
  5. 5 RFK Jr. says AI will end the tyranny of medical experts, but chatbots side with the experts ARS TECHNICA Health Secretary Robert F. Kennedy Jr. told a Make America Healthy Again event that AI is better informed than any doctor in the country and will free Americans from the medical tyranny of experts, naming masks, social distancing and vaccines as claims it would correct. Ars Technica put those questions to Gemini and ChatGPT and found both backed the mainstream public health position.