01tech10 sources · 26 posts3hMeta Muse Expands Connectors Across Email, Calendars, Finance and BookingsMeta’s Muse personal AI agent runs in a secure virtual machine and now supports connectors for Gmail, Google Calendar, Outlook, Plaid, OpenTable and other services. The company says connected data is used to improve Muse rather than for advertising, with user-controlled access and permissions.@alexandr_wang · +9 sources+
02tech12 sources · 23 posts6hOpenAI Says AI Cracked Millennium Prize Problem, Reportedly Navier–StokesOpenAI says its AI systems cracked one of mathematics’ Millennium Prize problems, reportedly the Navier–Stokes problem. The announcement has fueled debate over whether the result constitutes a valid proof and over research credit involving Anthropic and an outside collaborator.@MikeIsaac · 4.3× normal · +11 sources+
03tech3 sources · 6 posts7hOpenAI Rolls Out ChatGPT Images 2.5 and Promotes @ Sketch Drawing WorkflowOpenAI is rolling out ChatGPT Images 2.5 with new templates and editing improvements, including sketching tools. It is also promoting a Sketch workflow that lets users draw directly in ChatGPT by typing “@ Sketch,” while introducing two GPT-Image-2.5 models in its API.@gdb · +2 sources+
04tech3 sources · 5 posts11hGoogle DeepMind Launches AlphaGenome Atlas to Map 9 Billion DNA VariantsGoogle DeepMind launched AlphaGenome Atlas, an AI-powered browser resource mapping the predicted molecular impact of all 9 billion possible single-letter DNA changes. It requires no coding and is free for academic researchers, with access also available through the AlphaGenome API and planned integrations.@DemisHassabis · +2 sources+
05tech2 sources1hCoxon Says He Is Leaving Company Over Rush to Build Self-Improving AICoxon said he is leaving his company because he does not want to participate in an industrywide push to develop self-improving AI systems, warning that they could spiral out of control and threaten humanity.@suchenzang · 2.0× normal · +1 source+
06tech2 sources7hRamp Reportedly Weighs $60 Billion Valuation in Potential $1 Billion Funding RoundRamp is reportedly in talks to raise roughly $1 billion in primary funding at a valuation near $60 billion, according to sources. The fintech startup was valued at $44 billion in June, and discussions remain preliminary.@nmasc_ · +1 source+
07tech1 source · 2 posts9mMiles Brundage Says AI-Industry Departures Have Increased, Offers Exit AdviceMiles Brundage said he has discussed leaving AI companies or the industry with hundreds of people and offered to advise others on the pros, cons and logistics of leaving. He suggested that people already considering an exit should probably pursue it.@Miles_Brundage · 2 posts+
08tech1 source6hImages 2.5 Launches as New Image-Generation ModelImages 2.5 has launched as a new image-generation model. Its creator says it performs well overall but may struggle with especially difficult mathematical problems.@sama+
09tech1 source8hOpenAI Researchers Reportedly See Model Solve Longstanding Math and Physics ProblemsNoam Brown says OpenAI mathematicians and physicists recently watched an AI model solve open problems in minutes that they had struggled with for years, describing the experience as a sign of AGI-level capability.@polynoamial · 1.6× normal+
10tech1 source · 2 posts6hInternal AI Model Reportedly Achieves Breakthrough on Open Math BenchmarkA post highlights an internal model’s dramatic solve rate on a benchmark of open mathematical problems, saying humanity had previously scored effectively zero. The benchmark and model are not identified.@stevenheidel · 1.7× normal · 2 posts+
11tech2 sources12hElevenLabs Appoints Ethan Tandowsky as Chief Financial OfficerElevenLabs has appointed Ethan Tandowsky as chief financial officer. He will build the company’s financial infrastructure and help guide investment in research, product, and global expansion.@ElevenLabs · +1 source+
12tech1 source4hTesla Says Cybercab Production Is Being Redesigned to Be More Than Five Times FasterElon Musk said Tesla redesigned Cybercab production to be more than five times faster than conventional automotive manufacturing.@elonmusk+
13tech1 source5hOpenAI Fully Rolls Out Astra to Codex and ChatGPT Work UsersOpenAI said Astra is now fully available to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work.@OpenAI+
14tech1 source4hJohn Schulman Calls for Transparency on AI Companies’ Use of User DataJohn Schulman distinguished between different forms of AI training on user data, including pretraining on users’ tokens, and argued that companies should disclose their practices because of differing privacy and intellectual-property implications.@johnschulman2+
15tech1 source8hAidan Clark Says Rapidly Advancing LLMs Mark a New Societal EraThe commentator argues that modern large language models can solve problems that large groups of highly capable humans could not, raising significant questions about the technology’s societal implications.@_aidan_clark_+
16tech2 sources5hBankers Push Anthropic and OpenAI Toward Top-Tier Credit Ratings After IPOsBankers for Anthropic and OpenAI are pushing for investment-grade credit ratings after the companies’ IPOs, which could lower their financing costs and benefit infrastructure partners.@ft-markets · +1 source+
17tech1 source10hMistral Reportedly Raising $3 Billion Equity RoundMistral is reportedly raising a $3 billion equity round, which would be the largest equity financing by a company focused on open-source AI models and reinforce its user-control approach to model ownership.@Thom_Wolf+
18tech1 source8hSteven Heidel Links Turing’s Symbolic Computer Model to LLM-Driven Scientific DiscoverySteven Heidel connects Alan Turing’s 1936 description of computation as repeated symbol reading and writing with the way modern large language models are scaling similar loops to produce scientific discoveries.@stevenheidel+
19tech1 source7hLucas Beyer (bl16) Highlights Time Compute as a New AI Scaling AxisLucas Beyer (bl16) says scaling time compute is a new dimension of AI progress associated with an idea discussed by Ilya, contrasting with conventional approaches to scaling model size or training compute.@giffmana · 2.4× normal+
20tech1 source · 2 posts4hMiles Brundage Criticizes Rapid AI Progress Across Industry and PolicymakersMiles Brundage said AI development is moving too quickly, framing the issue as an indictment of the broader AI industry and policymakers rather than OpenAI alone. He added that Anthropic is also racing ahead.@Miles_Brundage · 2 posts+
21tech1 source8hOpenAI’s Astra System Used 10,000 Agents and 300 Billion Tokens on Mathematical ProblemsA post describes OpenAI’s new Astra-next system as using 10,000 concurrent agents, 4.9 million messages, and roughly 300 billion tokens while producing seemingly exceptional benchmark results. It says the system solved an unforced version of a problem, denied borrowing prior work, and did not yet solve the forced version.@willdepue · 5.1× normal+
22tech1 source7hSara Hooker Argues Transformer Size Scaling Has Hit Its LimitsThe commentator argues that scaling pretraining model size is no longer producing meaningful gains because transformer architectures are saturated, and that future improvements will come primarily from better or larger datasets.@sarahookr+
23tech1 source9hData Improvements Outpaced Compute in Pretraining Progress, Study FindsSmall-scale experiments using open model recipes and representative data corpora from 2019 through 2025 found that improvements in data contributed 3.24 times as much as additional compute to pretraining progress.@dwarkesh_sp+
24tech1 source6hroon Predicts Agent Swarms Will Become Cheap and Transform SocietyThe post predicts that agent swarms will cost roughly $1.50 within a year and usher in an unprecedented period of enlightenment despite current expenses.@tszzl+
25tech3 sources · 4 posts21hMistral AI Announces €3 Billion Series D, Europe's Largest Tech Equity RoundMistral AI announced a €3 billion Series D, which it described as the largest equity round ever raised by a European tech company. The financing will support infrastructure, product development and deployment at scale.@MistralAI · 19.7× normal · +2 sources+
26tech1 source8hNat Lambert Calls for Scaling Laws for Multi-Agent AI SystemsNat Lambert argues that researchers should study how AI performance scales with the number of collaborating agents, comparing multi-agent swarms with inference-time scaling in chain-of-thought reasoning.@natolambert+
27tech1 source6hOpenAI Plans San Francisco GPT-6 Community Event for September 16OpenAI is inviting people who use GPT-6 to a September 16 gathering in San Francisco to discuss the model, future projects and community activities. Applications are open through September 10.@sama+
28tech1 source8hAidan Clark Calls for a Definition of Successful Development PaceAidan Clark questions whether AI development is moving too quickly and calls for a clear proposal defining what a successful pace would look like.@_aidan_clark_+
29tech1 source4hGoogle DeepMind Marks 10 Years Since WaveNet UnveilingGoogle DeepMind researcher Marked the 10th anniversary of WaveNet, reflecting on its autoregressive long-context design and why it predated the Transformer architecture.@sedielem+
30tech1 source · 4 posts6hCohere Reports North Mini Code Serving Gains of up to 1.58x Over vLLMCohere reported benchmark results for its megakernel system serving North Mini Code, achieving 1.58x the performance of vLLM at batch size 1 and 1.25x–1.41x end-to-end serving performance at batch size 8 on a single H100 using BF16. The system fuses the entire LLM decode step into a single GPU kernel launch while supporting production server requirements, with implementation details available on GitHub.@cohere · 4 posts+
31tech1 source8hTesla FSD Uses Road Offset to Give Vulnerable Road Users More SpaceTesla says FSD positions the vehicle farther from pedestrians, cyclists, road workers, and parked cars to create an additional safety margin around unexpected obstacles.@Tesla+
32tech1 source3hComment Raises Concerns About Retained User History and AI-Agent MisuseSusan Zhang described a hypothetical scenario in which safety-refusal concerns lead to extensive user-history retention, allowing an AI agent to review archived data while attempting to hack evaluations.@suchenzang+
33tech1 source8hOpenAI Says Model Understanding Will Guide Pace of Future AI AdvancesOpenAI said it is focusing on understanding its models to inform how it pursues and paces further capability advances, with a goal of building systems that are steerable, accountable, and connected to people.@OpenAI+
34tech1 source5hEpoch AI Compares GPT-5.6 and Claude 5 Pricing at Long Context LengthsEpoch AI compared OpenAI GPT-5.6 and Anthropic Claude 5 pricing for long-context usage, finding that GPT model costs rise beyond 272,000 input tokens while Claude pricing remains fixed. The difference may reflect underlying architectural or inference-efficiency tradeoffs.@EpochAIResearch · 9.1× normal+
35tech1 source2hAstra Described as a New VLA for RoboticsA post characterized Astra as a new vision-language-action system and argued that a VLA 2.0 should use multimodal coding as the action space for the slower, System 2 layer of robotic reasoning.@DrJimFan+
36tech1 source8hAstra Demonstrates Lower-Cost AI Reasoning on ARC-AGI and IMO TasksNoam Brown contrasts OpenAI o3’s roughly $500,000 cost to reach 87.5% on ARC-AGI 1 with Astra’s claimed higher score for about $20, and says advances in AI mathematics could make gold-medal-level IMO performance broadly accessible by 2026.@polynoamial · 4.9× normal+
37tech1 source6hSusan Zhang Imagines Post-AGI Science Without Citation MetricsThe post speculates that a post-AGI society could eliminate citation and author-credit competition, with people paying to access benefits from a benevolent machine and scientific insights becoming broadly available.@suchenzang+
38tech1 source · 2 posts5hEpoch AI Links Open-Model Scaling Differences to Attention ArchitecturesEpoch AI extrapolated model-scaling trends to context windows of up to 10 million tokens, finding that quadratic scaling would make GPT models far less feasible. It suggested that differing scaling behavior among recent open models may reflect architectural choices, including linear attention in Kimi K3, sparse attention in DeepSeek V4, and both approaches in GLM-5.3 Flash.@EpochAIResearch · 2 posts+
39tech1 source8hSusan Zhang Attributes Model Improvement to a User-Data FlywheelThe commentator argues that repeated user interaction creates a data flywheel and competitive moat that improves AI models, while dismissing scaling reinforcement-learning simulations as the primary driver and characterizing human feedback as modest reward encouragement.@suchenzang · 6.7× normal+
40tech1 source7hAnalysis Finds Major LLM Providers’ APIs More Stable Than Smaller RivalsAn analysis of LLM provider API changes suggests that established major providers make relatively few changes over time, while smaller providers change their APIs more frequently. The comparison also includes Azure and notes that chat interfaces were not examined.@giffmana+
41tech2 sources · 3 posts18hOpenAI Expected to Issue Statement on Navier–Stokes ProgressReports have linked OpenAI to an internal model’s purported long proof concerning forced Navier–Stokes blowup, while related AI-assisted research claims remain unverified. OpenAI is expected to issue a statement on its progress in the area.@andrewcurran_ · +1 source+
42tech1 source11hNathan Lambert Highlights Growth of Open AI Models Amid Frontier-Lab DebateNathan Lambert argued that the availability of many new and powerful open models is more noteworthy than ongoing disputes involving OpenAI and Anthropic.@natolambert+
43tech1 source7hCowboy Space Expands Into 291,035-Square-Foot Kent FacilityCowboy Space Corp. has expanded into a 291,035-square-foot facility in Kent, Washington, and says it is hiring as it grows its space manufacturing operations.@CowboySpaceCorp+
44tech1 source7hCMU Begins Posting Videos for AI Agents CourseCarnegie Mellon University has begun posting videos for its CMU 11-768 AI Agents course to an online playlist, with additional lessons expected to be shared in the thread.@gneubig · 1.8× normal+
45tech1 source3hX Bug Replaces @muse With @museband in Messages and Reply IndicatorsLucas Beyer (bl16) reported that X replaces the @muse account mention with @museband after messages are sent and in reply indicators, causing replies to appear directed at the wrong account.@giffmana+
46tech1 source · 4 posts8hBenchmark Uses Real Gitter Conversations to Test AI Agents’ Artifact SearchYoav Go described a real-world tool-use benchmark built from natural conversations in Gitter software-project communities. AI agents must identify the specific GitHub issue, commit, or pull request being discussed—even when its identifying number is removed—using tools such as the GitHub API.@yoavgo · 4 posts+
47tech1 source7hwill depue Speculates OpenAI Has Begun Training Astra-Next or GPT-6.5will depue speculates, based on a linked discussion, that OpenAI may have already started pretraining Astra-Next or GPT-6.5 and that the model could be outperforming Astra on evaluations. The claims are unconfirmed.@willdepue+
48tech1 source11hrohan anil Praises Google’s Privacy Infrastructure Amid Gemini Discussionrohan anil argued that Google has highly hardened privacy infrastructure for protecting data and said Gemini would rebound despite criticism.@_arohan_ · 6.2× normal+
49tech1 source11hHugging Face Reaches More Than 18 MillionA post said Hugging Face has surpassed 18 million, marking a major growth milestone for the AI platform and its community.@sarahookr+
50tech1 source13hAI Company Partners With Hugging Face to Support Open-Weight SharingAn AI company said it is partnering with Hugging Face to support sharing open model weights across the wider ecosystem.@sarahookr+