The A.I. Daily
Trump Rejects Global Entity for AI Oversight
Trump rejected a global entity for AI oversight, signaling a unilateral US stance on AI governance that could affect international AI safety cooperation.
Industry
Everything · newest firstModels, products and AI startups, near-duplicates collapsed to the most credible source
Simulating fault tolerance with stage skipping in pipeline-parallel training
Templar team simulated stage-skipping fault tolerance in pipeline-parallel training on the Crucible distributed pre-training platform. With a 178M model, eight replicas, and four stages per replica, validation loss stayed close to the no-failure baseline at 1% per-replica failure probability. Fixed projections shared across layers improve robustness, supporting training on unreliable compute.
Siemens and Salesforce Deepen AI Partnership
According to Industrial Equipment News, Siemens and Salesforce are deepening their AI partnership, aiming to redefine industrial sales and service. The report gives no investment amount, deal structure or product details.
Agents Can Speed Up Rust Code 2x-20x, Study Finds
Testing shows modern agentic LLMs with proper guardrails can write Rust code faster than current state-of-the-art approaches, with cumulative 2x-20x speedups across domains as frontier models improve.
LinearSolveBench: new benchmark for linear solvers
LinearSolveBench is a new benchmark measuring the ability of a model or harness to write fast, accurate, and general numerical solvers for large sparse linear systems in C. Its goal is to encourage algorithmic advances in numerical methods for linear systems.
xAI's Grok Bot tops 400,000 weekly users in first month
xAI's Grok Bot AI agent surpassed 400,000 weekly users about a month after launch, a sign of modest early demand for a product key to Elon Musk's AI ambitions.
AstroForge Puts In-House AI in Command of Spacecraft
Asteroid-mining startup AstroForge built an in-house transformer-based autonomy stack, Solo, for its Autonomy-1 spacecraft, set to launch in 2027 on Stoke Space's first rocket with NASA backing. The company has raised $56M in venture funding, and anomalies on its two prototype spacecraft, including the loss of contact with Odin, pushed it toward onboard autonomy.
TechCrunch Disrupt 2026 Hosts Five AI Safety Sessions
TechCrunch Disrupt 2026 will host five AI safety sessions across its AI Stage and Real World AI Stage, featuring speakers from Anthropic, NVIDIA, AWS and Waabi. The sessions focus on how enterprises can safely deploy AI agents into production, with passes discounted up to $200 before Sept 25.
Analysis: OpenAI may fast-follow TypeSafe's Jev model
An Arcturus Labs piece argues TypeSafe's Jev model uses an approach close to conventional LLMs, so OpenAI is positioned to fast-follow and fold the capability into its models and agents. The key question is whether TypeSafe's training data and processes form a moat.
European Commission publishes two reports on AI in education
The European Commission published two reports on generative AI's impact on teaching and how European education systems are handling digital transformation. GenAI is becoming mainstream, but its impact depends on use and long-term skill effects remain unclear; nearly two-thirds of systems have AI policy frameworks.
OpenAI Urges US-China Coordination on AI Safety
OpenAI is urging US-led global standards for the most powerful AI systems, including shared safety tests and US-China coordination on serious incidents. The proposal comes just before Trump's talks with Xi, with AI risks and technological competition high on the agenda.
Maryland Gov. Moore Proposes AI Rules
Maryland Gov. Wes Moore proposed a series of AI regulations to protect residents' rights, privacy, and dignity, joining the national debate. The move contrasts with Trump's opposition to AI guardrails.
Jacksonville AI Adoption Expands Amid Security Risks
The University of Florida is investing $300 million in a new Jacksonville campus offering graduate programs in cybersecurity, AI, and data analytics to meet local finance, healthcare, and defense employer demand. US information security analyst jobs are projected to grow 21% through 2035, as local firms tighten fraud controls and privacy protections.
Stanford R&DE used AI to race-swap students in ads
Stanford Residential & Dining Enterprises used AI to change students' races and alter faces and body shapes in ads, according to The Review. Student Billy Ramirez said he was removed from a photo and replaced by an AI-generated Black woman, feeling erased from a representation meant to include him.
GPT-6 Astra cracks Enigma message unsolved since 2005
On 15 September 2026, Carter Leffer submitted a break of German Army Enigma message MVUEH of 10 July 1941, unsolved since 2005. The key and plaintext were validated; the key differs entirely from others that day, using wheel order 253 instead of 512.
AFL-CIO president: AI threats to workers need no doomsday warning
AFL-CIO President Liz Shuler spoke about AI in the workplace, saying it inflicts very real threats on workers every day and that it shouldn't require doomsday warnings to wake up to them.
Economics of Open-Weight Inference: Older GPUs Retain Value
A paper finds open-weight inference costs about one-fifth of comparable closed models, with self-hosting at $0.12-$0.35 per million output tokens. On gpt-oss-120b, the A100 is cheaper than the H100, and A100 five-year rental prices retain 80% of one-month prices, showing older GPU families keep multi-year earning life.
Qonto releases new FAQ information retrieval benchmark
Qonto open-sourced QontoFAQ, a benchmark for evaluating embedding models on FAQ retrieval. It uses a new metric that better reflects document relevance, aiming to address benchmark gaming, with a write-up and code released.
Dyson's $499 CameraJet toothbrush may have waterproofing problem
Dyson this month unveiled the $499 CameraJet electric toothbrush with a tiny camera to aim rinsing fluid between teeth. Reports suggest the overengineered gadget may have a waterproofing problem.
OpenAI contractors fired for using AI to train AI
404 Media found multiple OpenAI contractors hired to improve its models were fired for using AI to train the AI. The practice could worsen model collapse, and contractors say it is common, with many caught in groups of thousands.
Beware fake sites selling AI assistant subscriptions
Malwarebytes found fake subscription websites impersonating AI products including GPT-6 Astra, DaVinci Resolve and PixAI, charging $10/month up to $2,000/year, with some requesting file uploads. Researchers believe one group runs them all, posing data risks for enterprise shadow IT buyers.
JevBench: a reproducible benchmark for typed decision models
A developer launched JevBench, a reproducible benchmark for typed decision models that return bounded choices and probabilities, covering 534 English decisions and scoring accuracy, latency and cost with configurable weights. Jev leads at 74.4, followed by SemIf, djev, Winnow-12B Q8 and reflex 4B.
Sitefire trains classifier to detect AI content by structure
YC W26 startup Sitefire ported the StoryScope method to commercial web pages: it collected 2,250 pre-ChatGPT B2B blog posts from 268 company sites, had five models including GPT-5.4 rewrite each, and trained a classifier on 214 structural questions, reaching 98% accuracy separating AI from human posts.
Andreessen Horowitz launches AI school as college alternative
Venture capital firm Andreessen Horowitz launched a program for recent high school graduates and college dropouts to identify and train young talent building AI products, positioning AI education as an alternative to the traditional college path.
Report: Data center opposition driven by local experience
Nonprofit Data & Society published a report based on 18 months of ethnographic fieldwork interviewing 44 Pennsylvanians to understand opposition to data centers. Surveys show over 60% of Americans favor limiting new data centers, and $68 billion of projects were disrupted by local opponents in Q2 2026.
After AI doomsday panic, Trump-Xi talks unlikely to yield regulation
Trump and Xi will meet in Washington next week with AI high on the agenda. But both sides have recently stressed no slowdown in AI development is needed, and early signs suggest neither is willing to accept regulatory checks.
AI models built from rat brains edge closer to reality on AWS
The Biological Computing Company is bringing its biological computing AI tools to Amazon Web Services, a major boost for the once-fringe field that aims to marry nature with code.
Viture's Vonder glasses map your mind without cameras
Viture launched Vonder, its first display-free smart glasses. They have no cameras and use bone-conduction microphones to privately record users' daily musings.
Nscale IPO tests Wall Street's appetite for concentrated AI bets
British AI data center developer Nscale is listing on the NYSE at an expected $35B valuation, seeking $3B. It holds over $103B in contracts, but about 85% depends on Microsoft and Anthropic; Anthropic's deal is contingent on financing and can be cancelled if milestones are missed. Revenue hit $140.6M for the six months to June 30, with net losses of $1.02B.
Global X launches AI memory ETF and six other funds
Global X launched seven new ETFs on the Toronto Stock Exchange, led by the Global X Artificial Intelligence Memory Index ETF (DRAM) and the Global X Asia Semiconductor Index ETF (ACHP), giving Canadian investors exposure to AI infrastructure and Asia's semiconductor ecosystem.