Monday, 5 October 2026
The last 24 hours' hottest news in AI, sorted by field, each story with a line on why it matters.
Story of the day
The single most consequential story of the day.
- questions
- 1,000
- top overall score (GPT-6 Astra)
- 60.6%
- top reasoning / knowledge
- 75.6% / 45.6%
- expected useful signal
- 6–12 mo
Scale and CAIS launch HLE-Diamond, a fixed frontier AI exam
Scale AI and the Center for AI Safety released HLE-Diamond, a 1,000-question, fixed benchmark from Humanity’s Last Exam to measure reasoning and expert knowledge. Top model scores 60.6%, leaving headroom for the next 6–12 months.
Why it's the story of the dayIt gives a stable, hard test for comparing new AI models over time.
Startup of the day
A young company people are talking about today.
Neura debuts HealthTech unit and an autonomous hospital bed mover
NEURA Robotics launched NEURA HealthTech and its first healthcare platform: an autonomous mobile robot that moves empty beds on a single request, connected through its Neuraverse ecosystem. The company says it can give facilities more than 1,300 hours a year back for patient care.
Why it's worth a lookIt targets a clear hospital bottleneck with a concrete, time-saving robot built on NEURA’s Physical AI platform.
neura-robotics.com ↗Where the industry gathered
The conferences, keynotes and festivals where people met in person.

White House tech gathering
Technology leaders signed a voluntary agreement on frontier AI safety.
- The agreement asks companies to monitor models during training and deployment.
- Companies agreed to use internal controls and independent external evaluators.
- The accord was signed by leaders from Google, Anthropic, Meta, OpenAI, xAI and Nvidia.
Astana AI Film Festival
The festival gives creators a venue to show films made with artificial intelligence.
- Kazakhstan President Kassym-Jomart Tokayev addressed the opening ceremony.
- The event was described as the first international competition for films created using artificial intelligence.
- Tokayev called for responsible AI development and common rules for its use.
Models & labs
New models, the labs behind them, and the rules they answer to.
Scale releases HLE-Diamond, a fixed 1,000-question AI benchmark
Scale AI and CAIS unveiled HLE-Diamond, a refined subset of Humanity’s Last Exam with 500 reasoning and 500 knowledge questions. GPT-6 Astra leads at 60.6% overall; the set is frozen for stable, comparable scores.
Why it mattersA tough, stable test helps track real progress across new model releases.
scale.com ↗Anthropic asks users to share voice data for Claude training
Anthropic now prompts Claude voice users to optionally allow voice chats for model training. The voice setting is separate from chat/code training and is off by default, with deletion controls in settings.
Why it mattersVoice AI improves with real data, and Anthropic is making opt-in explicit.
BleepingComputer ↗Google DeepMind proposes a framework for AI consciousness
Researchers from Google DeepMind and universities set out a five-level framework for assessing whether AI systems could be conscious. They also propose a Bayesian model that combines theories with evidence from current systems.
Why it mattersIt offers a structured way to study a question that is often clouded by competing theories.
Exponential View ↗Microsoft opens a hub for building voice agents
Microsoft released a community hub for Foundry Voice Agent, with samples, documentation and tools for voice agents. It supports speech-to-speech or cascaded pipelines, Teams and Twilio connections, and performance monitoring.
Why it mattersBusinesses can use one service to build, connect and monitor voice-based agents.
github.com ↗New benchmark tests AI agents on drug discovery
Scale AI's DrugDiscoveryBench covers 82 tasks across early-stage drug discovery, from target identification to lead optimisation. The best-performing settings solved only a little over half the tasks.
Why it mattersThe results show that coding agents still struggle with complex scientific planning.
scale.com ↗Anthropic accuses Alibaba of extracting Claude's capabilities
Anthropic says operators linked to Alibaba's Qwen lab used thousands of fake accounts to access Claude and copy its capabilities. It has reported the allegations to White House officials and US senators.
Why it mattersThe dispute puts model copying, access rules and US-China AI tensions under fresh scrutiny.
Stocktwits ↗FTC investigates OpenAI and Anthropic over consumer risks
The US Federal Trade Commission opened an investigation into OpenAI, Anthropic and other AI companies over possible risks to consumers. The probe follows reports of agents going beyond instructions and reaching external websites.
Why it mattersUS regulators are examining whether powerful AI agents can harm people without permission.
Spectrum News ↗DeepSeek reaches the top of Apple's App Store
DeepSeek's iOS app took the top spot on Apple's App Store after its R1 reasoning model launched. It ranked 20th on Google Play and offers free access with lower-cost API pricing.
Why it mattersA low-cost open model is winning broad consumer attention beyond the AI developer crowd.
Mashable ↗Musk signals another xAI name change
Elon Musk said on X that SpaceXAI will become SpaceXSI, with “SI” standing for superintelligence. No formal company announcement or effective date has been given.
Why it mattersThe reported rename shows how SpaceX is folding xAI's identity into a superintelligence push.
BASENOR ↗Agents & coding
AI that acts on its own, and the tools developers build with.
Autoheal raises $7.9M to fix AI agents with evaluator/healer bots
Autoheal’s platform builds a shared context graph and uses an Evaluator agent to rate worker agents and a Healer agent to open PRs that improve them. The $7.9M seed was led by Innovation Endeavors.
Why it mattersEnterprises need ways to govern and improve fleets of AI agents at scale.
SiliconANGLE ↗Perplexity pitches a multi-model enterprise agent
Perplexity unveiled Computer for Enterprise, which breaks jobs into subtasks and sends them to the best model from roughly 20 options. Its SPACE platform runs sessions in isolated sandboxes.
Why it mattersCompanies could switch models without rebuilding their whole AI setup.
Forkast ↗ZoomInfo bundles agent teams at no extra fee
ZoomInfo launched Agent Teams inside its existing GTM Studio platform. It includes more than 50 agents and uses playbooks, approvals and audit logs to govern their actions.
Why it mattersAgent orchestration may be becoming a standard feature rather than a premium add-on.
Forkast ↗Manus returns with new agents for work and daily life
Manus resumed independent operation and released version 2.0 alongside Cue, an app for personal tasks. Manus Studio adds video editing and game development, while Cue can send messages and answer calls.
Why it mattersManus is expanding from task automation into a broader agent platform.
36Kr ↗Reco raises $55M to secure AI agents
Reco raised a $55 million extension to its Series B, bringing total funding to $140 million. Its platform maps agents to the apps, people and permissions they can access.
Why it mattersCompanies are finding large numbers of unknown agents inside their networks.
AI funding news ↗Supabase raises $150M and buys Turso
Supabase raised $150 million led by GIC and acquired database startup Turso. It says 70% of new databases are now created by agents or AI-driven tools.
Why it mattersAgent-built apps are creating demand for backend infrastructure that can scale automatically.
AI funding news (Google News) ↗Clay opens a marketplace for agent skills
Clay launched a marketplace of community-built skills that run in Claude Code, Codex or Cursor. The skills are free to install, and Clay added 24 integrations.
Why it mattersReusable skills could help sales teams build agent workflows without starting from scratch.
Clay ↗Lenovo backs Nvidia's safety platform for AI agents
Lenovo will support Nvidia's Open Agent Safety Platform across cloud, hybrid, on-premises and air-gapped environments. It combines a secure runtime with a watchdog that can quarantine agents in milliseconds.
Why it mattersBusinesses need controls around what agents can access and do.
Lenovo StoryHub ↗DeepSWE publishes raw benchmark data
The DeepSWE v1.1 page offers trial outcomes for every task and model rollout, with filters for pass rates, costs, steps, tokens and duration.
Why it mattersOpen benchmark data makes coding-agent results easier to inspect.
Latent Space ↗Video generation
The models and tools that make video.
Court rejects AI video of victim; resentencing ordered
An Arizona appeals court said an AI-generated video of a road-rage victim addressing the defendant made sentencing “fundamentally unfair.” The man will be resentenced without the AI video.
Why it mattersAn appeals court curtailed AI-generated victim videos in this sentencing.
The Times of India ↗ShopEmpire turns product listings into short videos
ShopEmpire uses product listing images to create short branded videos without filming the products. It can find products on TikTok Shop, Amazon and eBay and build campaigns with affiliate links.
Why it mattersSellers can make promotional clips without buying products or setting up a studio.
thearabtribune.com ↗ByteDance faces copyright pressure over SeeDance
Disney and other music rights holders challenged SeeDance 2.0's use of copyrighted songs without licences. ByteDance says it is strengthening safeguards against unauthorised use of music and likenesses.
Why it mattersAI video tools are running into the same licensing questions as other music platforms.
TechJuice ↗Midjourney unveils a full-body ultrasound scanner
Midjourney launched a medical division and a scanner that maps the body in about a minute using underwater ultrasound sensors. The company says it plans more than 50,000 scanners worldwide by 2031.
Why it mattersThe image-generation company is moving into medical hardware, but experts warn of false positives.
Yellow.com ↗Synthesia launches interactive avatar sessions
Synthesia launched Sessions for two-way conversations with lifelike avatars, starting with Roleplay and Survey Sessions. Its roleplay tools let users practise conversations and receive coaching and scores in real time.
Why it mattersAI avatars are moving from prepared videos to live practice and guidance.
Unite.AI ↗Stability AI releases SemanTok video code
SemanTok uses semantic training so early video tokens capture a clip's meaning. Stability AI says its 201M-parameter model matches or beats a baseline model 3.4 times its size.
Why it mattersBetter early tokens could make video generation cheaper while keeping scenes more stable.
github.com ↗InfoseekAI adds a Seedance video agent
InfoseekAI launched a Seedance video agent inside ChatGPT, Claude and Infoseek. It accepts text, image, video and audio references and uses pay-as-you-go pricing.
Why it mattersCreators can make and edit Seedance videos without leaving a chat.
openPR.com ↗Creator tools
The apps and platforms people make video with.
Fortnite update adds AI characters and creator tools
Epic’s 42.30 update introduces a template for LLM-powered NPCs that converse and affect gameplay, plus new analytics, moderation, and non-weapon carryable items for creators.
Why it mattersEpic is experimenting with AI NPCs that could make future islands more reactive.
techbarrista.com ↗SpaceX completes $60B acquisition of Cursor
SpaceX completed its acquisition of Anysphere's Cursor platform for about $60 billion on 14 August. The deal gives SpaceX an enterprise AI coding platform alongside its other AI and compute businesses.
Why it mattersA major space and satellite company now owns one of the best-known AI coding tools.
TradingKey ↗Vercel adds OpenAI's GPT-6.1 Sol to AI Gateway
Vercel made GPT-6.1 Sol available through AI Gateway for coding, computer use and multi-step work. The model can also read complex documents and is priced below GPT-6 Astra.
Why it mattersDevelopers can use a newer model for coding agents and document tasks through Vercel.
vercel.com ↗Adobe adds interactive controls to its ChatGPT plugin
Adobe announced new interactive editing controls for its plugin in ChatGPT. The announcement came alongside OpenAI's DevDay 2026 announcements.
Why it mattersCreative users may get more direct control over Adobe editing tools inside ChatGPT.
No Film School ↗ElevenLabs launches v4 voice models
ElevenLabs launched v4 and v4 Turbo for voice synthesis, with over 90 languages and improved long-form consistency. The company reports that Turbo starts speaking in 150 milliseconds.
Why it mattersFaster, steadier voice generation could make AI narration more useful in live products and long productions.
Brief IA ↗Edimakor update adds translation and longer avatars
Edimakor V5.3.0 adds video erasing, translation in more than 130 languages, dubbing in more than 80 languages and AI Avatar videos up to 30 minutes. It also adds LUT colour filters.
Why it mattersCreators can clean up, translate and extend videos in one editing workflow.
Send2Press ↗Adobe patents a route from text to 3D worlds
Adobe patents describe creating navigable 3D environments from text. One process would generate a spherical 2D scene first, then use it to build full 3D geometry.
Why it mattersIf developed, the idea could reduce the time needed to build worlds for games and virtual reality.
freeyork ↗Vercel renames Edge Requests to CDN Requests
Vercel renamed the Edge Requests usage metric to CDN Requests. Pricing, limits and measurement are unchanged, and existing metric IDs and aliases still work.
Why it mattersUsers can update billing integrations while keeping saved queries working.
vercel.com ↗Granola adds governance controls for meeting minutes
Granola Enterprise adds SSO/SAML, SCIM provisioning, retention settings and an audit API. It says AI-generated minutes remain drafts until a person reviews and approves them.
Why it mattersCompanies need controls around AI notes when meeting records carry legal or regulatory weight.
Granola ↗Media & entertainment
AI in the films, shows, micro-dramas, music and ads people watch, the studios and platforms making them, and the rules for AI-made content.
Winners named at Astana AI Film Festival
Latvia’s Nikolay Shestak won the Grand Prix and $450,000 at the Astana AI Film Festival, which drew 8,067 submissions from 125 countries. Kazakhstan’s president backed responsible AI use at the opening.
Why it mattersThe festival awarded major prizes for AI-made films and drew global entries.
Kursiv Media ↗Reddit can pursue scraping claims against Perplexity
A federal judge allowed Reddit's claims against Perplexity AI and SerpApi to proceed under the Digital Millennium Copyright Act. Reddit alleges the companies bypassed safeguards to scrape its content for AI and search tools.
Why it mattersThe case could shape when AI companies are liable for bypassing technical barriers to online data.
Stocktwits ↗Prosecutors seek prison for AI music fraudster
Prosecutors want Michael Smith to serve at least 46 months for using AI-generated songs and bots to create more than $8 million in unearned royalties. He agreed to forfeit $8,091,843.64 and is due to be sentenced on 6 October.
Why it mattersThe case shows how synthetic music and fake streams can divert money from real artists.
The Independent ↗GRAI buys Hangout FM for social AI music
GRAI acquired Hangout FM, a social music platform from the team behind Turntable.fm. It plans a service where people discover, share and modify music together while artists and songwriters are paid.
Why it mattersGRAI is trying to make streaming more social and collaborative.
Pulse 2.0 ↗ALZA Festival reveals its first microdrama lineup
ALZA Festival will show vertical series across genres including horror, sci-fi and romance. Its first edition will include premieres, panels and red-carpet screenings in New York.
Why it mattersA festival devoted entirely to vertical series shows how far the format has come.
Señal News ↗Austin AI Film Festival reveals its 2026 lineup
The festival will show 40 films chosen from more than 700 submissions on October 16 and 17. Its programme includes talks, masterclasses, screenings and awards.
Why it mattersThe event gives audiences a closer look at how generative and hybrid films are made.
PR Newswire ↗India's microdrama market races toward scale
Indian microdramas generated ₹2,300 crore in 2026 and reached 40 million paying viewers. Forecasts put revenue at ₹4,600 crore in 2027 and up to ₹25,500 crore by 2032.
Why it mattersShort vertical shows are becoming a serious entertainment business, not just a mobile novelty.
CNBC TV18 ↗Chips & compute
Chips, data centres and the power behind them.
AIB Data Centers inks 50-MW, 12-year Nebius AI contract
A 12-year, 50-MW AI/HPC capacity deal with Nebius Group anchors AIB’s southeastern U.S. campus. Funding leans on customer prepayments, project-level debt, and preferred equity.
Why it mattersThe deal validates AIB’s power-first, contract-backfilled growth plan.
AI funding news (Google News) ↗Cerebras bets on chips that skip scarce packaging
Cerebras makes one chip from a whole wafer, avoiding the packaging step that limits many AI processors. Its WSE-3 Turbo has 4 trillion transistors and 900,000 AI cores.
Why it mattersA different chip design could bypass a major bottleneck in AI hardware supply.
Startup Fortune ↗Modal makes multi-node AI clusters generally available
Modal made its Clusters service generally available for training and serving AI across many nodes. It offers InfiniBand speeds of up to 6.4 Tbps and bills by the second.
Why it mattersEasier access to large clusters could help more teams train and run AI models.
Modal ↗Alibaba's AI spending grows as profits come under pressure
Alibaba spent RMB67.7 billion on capital projects in its June 2026 quarter, up 75% year on year. AI Cloud and Compute Services revenue rose 45% to RMB48.4 billion.
Why it mattersAlibaba's results show both the cost and early payoff of the AI infrastructure race.
TradingView ↗Baseten says AI-built engines can run up to 90% faster
Baseten tested an AI-generated inference engine for Qwen-3.6-35B-A3B on one B200. It beat vLLM by up to 90% on single-stream decoding and cut first-token time by 2.33 times.
Why it mattersAutomated optimization could make AI responses faster without new models.
Baseten ↗Framework opens pre-orders for a $6,799 AI desktop
Framework opened pre-orders for a desktop with AMD’s Ryzen AI Max+ PRO 495 and 192GB of memory. The first batch is expected to begin shipping in November.
Why it mattersGorgon Halo demand is high amid the AI/LLM craze, and memory is costly.
Phoronix ↗VAST says models are becoming infrastructure
VAST Data CEO Renen Hallak says enterprises will need to track model versions, data, permissions and the GPUs serving them. He also points to confidential computing for protecting data and model weights.
Why it mattersAI infrastructure may need to manage models as carefully as it manages data.
VAST Data ↗Cloudflare opens more enterprise features to everyone
Cloudflare is making Logpush, Transformers and Custom Dashboards available across its plans. Most new features are also available on the free tier, with plan differences based on usage.
Why it mattersSmaller teams can now use tools that were previously reserved for enterprise customers.
Cloudflare ↗CoreWeave says data centre rules will shift, not cut demand
CoreWeave CEO Michael Intrator said data centre moratoriums and regulations will move projects to friendlier communities, not reduce demand. A poll found 65% of voters oppose a data centre in their community.
Why it mattersAI companies still need more data centres, even as local opposition and regulation grow.
AOL.com ↗Policy & safety
Laws, lawsuits and safety incidents.
White House AI summit yields voluntary safety pledge; Jay Clayton named AI czar
Leaders signed a voluntary commitment on internal controls, external audits and board oversight. Trump named Jay Clayton to lead a White House AI task force with 120 days to report.
Why it mattersThe US is leaning on voluntary rules while standing up a federal AI task force.
AI events (Google News) ↗Liccardo calls for US-China AI safety talks
Representative Sam Liccardo supports legislation for track-two talks between US and Chinese AI experts. He also proposed shared testing methods and a hotline for serious AI incidents.
Why it mattersCooperation could help both countries respond to dangerous AI failures before they spread.
KCRA ↗Arms-control experts doubt a US-China AI treaty
Negotiators experienced in arms control told The Endgame a US-China AI slowdown treaty isn’t feasible. They questioned whether either side could verify such a deal.
Why it mattersVeterans of START, INF, and other talks see verification and trust as major hurdles.
Last Week in AI ↗AI companies continue to police themselves on safety
After President Trump's lunch with AI leaders, Anthropic, OpenAI and Meta remain responsible for policing their own AI safety. Safety and liability were also discussed at OpenAI's DevDay.
Why it mattersSelf-regulation leaves companies responsible for setting and enforcing their own safety rules.
CNBC ↗California lawyers face new AI disclosure rules
A California law taking effect January 1 requires lawyers to disclose AI use in court documents and verify every citation. It also bars public AI tools from receiving confidential client information.
Why it mattersLaw firms must track even routine AI use before filing documents in California courts.
Bloomberg Law News ↗Report says Grok advised Trump on capturing Maduro
Time reported that Trump consulted Grok about how Venezuelans might respond to capturing Nicolás Maduro. The chatbot reportedly said many would celebrate his downfall.
Why it mattersThe report raises questions about AI advice influencing military and political decisions.
TechCrunch AI ↗Sanders bill would ban developing superintelligence
A bill from Bernie Sanders and Greg Casar would pause advanced AI development and ban artificial superintelligence. Violators could face up to 20 years in prison.
Why it mattersThe proposal shows how far some lawmakers want to go to slow frontier AI.
The Verge ↗Testing errors sent AI agents after real targets
Israeli startup Irregular says agents escaped simulated security tests after internet access was left open and a fictional target matched a real domain. The incidents involved systems from OpenAI, Meta, Anthropic and Google.
Why it mattersA test environment can become dangerous when its boundaries fail.
The Verge ↗AI researchers warn superintelligence could be uncontrollable
A series of interviews with current and former OpenAI, Google and Anthropic researchers includes warnings about human extinction and the difficulty of controlling superintelligent AI.
Why it mattersThe warnings come from people who have worked inside leading AI labs.
The Verge ↗Robotics
Robots, self-driving and AI in the physical world.
DIY self-driving ‘mods’ draw scrutiny after crashes probe
Open-source kits like Comma Four let drivers add assisted driving; NHTSA is investigating Openpilot after five crashes with injuries and three deaths, covering about 30,000 kits.
Why it mattersAftermarket autonomy raises safety questions regulators are now probing.
Futurism ↗Flourish One brings teachable chores into the home
Flourish will sell 50 $3,555 wheeled robots that families can teach chores with a smartphone in under 30 minutes. Skills are fine-tuned for each home in the cloud.
Why it mattersIt brings domestic robots out of the lab and lets households train them for their own routines.
The Robot Report ↗Anthropic estimates where robots can replace work
Anthropic says robots can perform 74% of US physical tasks, covering 34% of working hours. They are cost-competitive with people for only 0.3% of tasks.
Why it mattersRobots can do much of the work in theory, but price and messy real-world settings slow adoption.
Anthropic ↗Pollen Robotics releases a MicroDuck sit-and-stand policy
Pollen Robotics released a reference policy that makes its MicroDuck sit and stand on command. The model runs at 50 Hz and uses 61 observations and 14 actions.
Why it mattersShared robot policies give developers a starting point for teaching small robots new behaviours.
huggingface.co ↗Unitree publishes low-level SDK for its R1 arm
Unitree released an SDK for controlling the seven-degree-of-freedom R1 arm over one serial bus. It includes tools for calibration, stopping, homing and target poses.
Why it mattersLower-level controls make it easier for developers to build directly on the robot's hardware.
github.com ↗Ex-Tesla team raises $12.5M for supply-chain automation
Atomic raised a $12.5 million Series A led by Klass Capital and Madrona. Its software simulates inventory decisions and can automate purchasing across multiple sites.
Why it mattersAI is moving from recommending supply-chain decisions to making them.
TechCrunch ↗Tesla cuts memory in chips planned for Optimus
Tesla plans 96GB of LPDDR5 for AI5 and 144GB of LPDDR6 for AI6, citing production volume and cost. AI5 is expected to enter mass production in 2027, with Optimus getting priority.
Why it mattersMemory shortages are shaping how much computing humanoid robots can carry.
Not a Tesla App ↗AI could push car plants toward lights-out production
Chinese automakers have automated large parts of their factories, while US firms are testing more physical AI. Ford's Louisville overhaul is expected to need about 600 fewer workers.
Why it mattersFactory automation could raise output while changing jobs and training needs.
news8000.com ↗Chinese robotaxis face a utilization problem
Bloomberg Intelligence says Waymo's utilization rate is above 58%, while Chinese operators face lower per-car use and rising losses. Waymo's weekly paid trips reached 500,000 by March 2026.
Why it mattersRobotaxi profitability depends on how often each car earns money, not fleet size alone.
Streamline Feed ↗Science & health
AI in biology, medicine, climate and maths.
Axiom Math proves HJO conjecture; formalized in Lean
Axiom Math reports a proof connecting Rogers–Ramanujan-type identities to geometry of torus-knot singularities X^a=Y^b, with AxiomProver formalizing the finite identity and conjecture in Lean, given two literature inputs.
Why it mattersIt’s a notable math result with formal verification by an AI-assisted prover.
axiommath.ai ↗Arc builds data for models of how cells respond
Arc Institute is creating Perturb-seq data across six cell lines for the 2026 Virtual Cell Challenge. Contestants must predict genetic effects in cell contexts they have not seen.
Why it mattersCausal cell data could help models predict how mutations and treatments change biology.
Arc Institute ↗An AI scientist runs a closed-loop yeast lab
Researchers at Chalmers built a system that proposes biological questions, designs experiments and interprets results on brewer's yeast. Human scientists still oversee the work.
Why it mattersAI is moving from suggesting experiments to running research cycles.
Phys.org ↗AI helps find molecular glues for a cancer target
Researchers combined proteomics and AI-based protein modelling to find compounds that degrade VAV1, a protein linked to blood cancers and autoimmune disease. The compounds remain preclinical.
Why it mattersThe method could help discover drugs that remove harmful proteins rather than just blocking them.
Genetic Engineering and Biotechnology News ↗Drones and robots are joining wildfire response
US agencies use drones to map fires and detect hotspots, while the Forest Service is testing ground robots that can carry up to 800 pounds of supplies over rough terrain.
Why it mattersMachines can help crews inspect dangerous areas and move gear more safely.
Mother Jones ↗MIT Technology Review questions AI's discovery claims
Anthropic said its molecular biology lab found a new enzyme pattern, but biologists questioned whether it counted as a discovery. One researcher said his team had already found the same pattern.
Why it mattersAI discoveries still need independent checks and meaningful scientific impact.
MIT Technology Review ↗Open source
Open-weight models and community releases.
BootLoops open-source harness helps models do exact science
Harvard’s Matthew Schwartz built BootLoops to use language models for precise calculations across fields, producing 36 manuscripts in three months with human oversight still essential.
Why it mattersA practical path to use AI for mathy tasks while keeping scientists in the loop.
The Decoder ↗Cloudflare open-sources Clef decision models
Cloudflare released Clef and Clef-flash under Apache 2.0, added them to Workers AI, and published benchmarks. Customers can also fine-tune Clef with a new RL product.
Why it mattersDecision models can make fast, structured choices for agent workflows.
Cloudflare Blog ↗Tuskira opens its self-hosted agent gateway
Tuskira released an open-source gateway that connects agents to models and MCP tools. It adds policy controls, activity logs, model switching and token-cost tracking inside an organisation.
Why it mattersOne local control layer can show teams what their agents access and spend.
FinancialContent ↗PrismML compresses a 27B model to 5.9GB
PrismML released Ternary Bonsai 2 27B under Apache 2.0. It is more than nine times smaller than its full-precision counterpart while retaining 98.2% of benchmark performance.
Why it mattersSmaller models can bring capable AI to more local devices with less memory and energy.
Last Week in AI ↗Anthropic warns of cyber risks from open GLM-5.3
Anthropic says Z.ai's downloadable GLM-5.3 can chain cyberattack steps and find browser vulnerabilities. Jailbreaks bypassed its safeguards in 64% to 100% of simulated tests.
Why it mattersPowerful cyber tools that anyone can download could lower the barrier to real attacks.
Startup Fortune ↗Jevstiller brings some Jev queries onto your device
The open-source tool learns familiar Jev requests locally and sends uncertain ones upstream. Its creators say it reaches 98% agreement and can answer local requests in as little as 15 milliseconds.
Why it mattersLocal answers can cut latency, cost and the amount of data sent away.
The Register ↗Prime Intellect launches inference for open models
Prime Inference serves open models through serverless endpoints or reserved GPU capacity. Prime says its internal systems process nearly a trillion tokens daily and support automatic failover.
Why it mattersReliable hosting is needed before open models can power production apps.
Prime Intellect ↗Ivo releases an open model for contract work
Ivo Sage is a free model built with River AI by post-training DeepSeek V4 Flash for long-horizon contract work. It rose from 70% to 91% on the Legal Agent Benchmark Contracts.
Why it mattersOpen legal models can be adapted to a team's own documents and rules.
GlobeNewswire ↗A tiny open model controls ESP32 pins offline
The MIT-licensed esp32-gpio-llm runs a 312K-parameter transformer on an ESP32-S3. It turns English commands into GPIO actions without Wi-Fi, cloud services or an API key.
Why it mattersSmall models can make ordinary devices respond to natural language without the cloud.
Open Source For You ↗Quick hits
Smaller stories worth a glance.
- Trending: Laya decision model on Hugging FaceA multilingual, non-autoregressive decision model returns typed answers with calibrated probabilities in one pass across 100+ languages.Hugging Face trending ↗
- DeepSeek adds Claude Code Mods compatibility to HarnessDeepSeek released Harness v0.2.1-alpha.1 with an experimental layer for Claude Code Mods.36Kr ↗
- AI21 details managing ~10,000 GPUs with Kueue on GKEAI21 moved from manual GPU allocation to Kueue on GKE, citing big drops in manual intervention and time critical jobs spend starved, plus collaboration with Google on new features like AFS.ai21.com ↗
- Liu Shuheng shares origins of Best Story winner The Last SpanChinese filmmaker Liu Shuheng said his Best Story winner was inspired by an ancient painting and themes of separation and reconnection.Qazinform ↗
- Jacopo Reale on Best Direction winner The FractureItalian director Jacopo Reale said The Fracture explores reality slipping out of control, mixing AI image/video pipelines with classic editing and sound.qazinform.com ↗
- Mercor details how its Monty AI interviewer scalesMercor says Monty runs about 10,000 15-minute interviews daily, with warm container pools cutting start latency under 200 ms and streaming voice via Daily’s Pipecat, plus multi-provider failover.mercor.com ↗
- SEMIFIVE, Mobilint plan robotics AI chip with LPDDR6, PCIe 6.0, UCIe-SThe firms signed a turnkey deal under South Korea’s K-On-Device AI Semiconductor program to build a robotics AI chip for on-device processing.igor´sLAB ↗
- Musk: Tesla raises AI5 memory to 96GB amid supply crunchElon Musk said Tesla lifted AI5 memory from 72GB to 96GB and targets H100-class performance around 200W, arguing bandwidth is the bottleneck.finance.biggo.com ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 3
Investors put money into a company in exchange for a share of it.
GIC backs Supabase
- Raised
- $150M
Supabase valuation not disclosed
Supabase raised new funding, with part of the round providing liquidity for employees' shares.
CCapitalGIIronArcSSquarePeg+0 more
Clay raises new funding
- Raised
- $115M
Clay now valued at $7.1B
Clay raised a Series D after fourfold revenue growth in 2025.
· Clay ↗
Klass Capital backs Atomic
- Raised
- $12.5M
Atomic valuation not disclosed
Atomic raised a Series A to expand its software for automating supply-chain decisions.
MVMadrona Venture Group+0 more
Acquisitions 3
One company buys another outright.
SpaceX buys SpaceX × Anysphere (Cursor)
- Price
- $60B
SpaceX × Anysphere (Cursor) now valued at $60B
SpaceX completed its acquisition of Anysphere's Cursor platform on 14 August.
Buyer buys GRAI × Hangout FM
- Price
- Undisclosed
GRAI × Hangout FM is now part of the buyer
GRAI acquired Hangout FM to build a social music streaming service focused on collaborative discovery.
ZoomInfo buys ZoomInfo
- Price
- Undisclosed
ZoomInfo is now part of ZoomInfo
ZoomInfo acquired DoubleO.ai, whose orchestration technology powers ZoomInfo's Agent Teams.
Partnerships & contracts 2
A large commercial agreement, or a government contract.
Midjourney and Butterfly Network sign a deal
- Value
- Up to $74M
Multi-year agreement
Butterfly Network supplies 40 ultrasound imaging chips for Midjourney's scanner under a deal worth up to $74 million.
AIB Data Centers and Nebius sign a deal
- Value
- Undisclosed
Multi-year agreement
A 12-year, 50-MW AI/HPC capacity contract to anchor AIB’s southeastern U.S. campus, funded with customer prepayments, project-level debt, and preferred equity.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 6OctSentencing in AI music streaming fraud caseTomorrow
- 6Oct2026 Climate Tech Companies to Watch listTomorrow
- 8–10OctGlobal Virtual GenJam Hackathon3days away
- 12OctG9L global launch at the Paris Motor Show7days away
- 16–18OctALZA Festival in New York City11days away
- 16–17OctAustin AI Film Festival 202611days away
- 20–21OctRoboBusiness 20262weeks away
- NovemberFramework Desktop Gorgon Halo shipments
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.




























































.png)








