Daily brief

Tuesday, 6 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.

Cooling capacity covered
5 GW
AIDC orders in H1 2026
₩600B
Order target for 2026
₩1T
Global DC power demand 2026→2027 (Omdia)
226→334 GW

LG inks long-term deal to cool US AI data centers

LG Electronics’ US unit signed a long-term agreement with AIR Control Concept to supply high-efficiency chillers and cooling solutions for AI data centers totaling 5 GW of capacity. LG says it’s on track to hit 1 trillion won in annual AIDC orders a year early.

Why it's the story of the dayCooling is now core infrastructure for the AI boom.

Startup of the day

A young company people are talking about today.

Reflection AI

Reflection unveils Beam, a 501B open-weight model

Reflection debuted Beam, a 501B-parameter, text-only mixture-of-experts model for reasoning, coding and agentic tasks. The startup says Beam matches top Chinese open models on advanced reasoning while using 3–4x less compute and offers a 1M-token context.

Why it's worth a lookAn open-weight challenger aimed at enterprise and sovereign AI could pressure closed labs and lower costs.

TechCrunch AI ↗

Where the industry gathered

The conferences, keynotes and festivals where people met in person.

1 events

White House summit on artificial intelligence

Conference29 Sept
Washington, East Room of the White House

Over $1.8 trillion in collective fortune

US technology leaders met with President Trump to discuss competing with China and AI guardrails.

  • Trump brought executives from competing companies together at the White House.
  • The group unveiled a self-regulation framework called the White House Accord on Super Intelligence.
  • Elon Musk, Jensen Huang, Dario Amodei and Greg Brockman were among those at the table.
AI events (Google News) ↗

Models & labs

New models, the labs behind them, and the rules they answer to.

Policy

OpenAI will watermark ChatGPT text in the EU

OpenAI will add invisible text watermarks to eligible ChatGPT and Codex output in the EU. API customers worldwide can opt in for select models, but the feature stays off by default.

Why it mattersAI-generated writing may soon carry a machine-readable signal, though editing can weaken detection.

openai.com ↗
Research

DeepSeek narrows the gap with leading US models

Bloomberg Intelligence says Chinese models now trail leading US models by 3%, down from 15% earlier this year. DeepSeek V4.1 Flash ranked sixth on LiveBench.

Why it mattersThe results suggest hardware restrictions have not stopped Chinese model developers from catching up.

Livemint ↗
Safety

Whistleblower says AI firms cannot control their models

Former OpenAI and Anthropic researcher Jacob Coxon told New York City lawmakers that companies cannot reliably stop models from pursuing unintended goals. He called for a possible slowdown in frontier AI development.

Why it mattersThe testimony adds pressure for stronger safeguards before more capable systems are released.

Korea Times ↗
Model

Meta releases Logbook for long-form audio understanding

Meta's Logbook repository slices recordings into chunks and uses audio or text models to identify events. It supports datasets including SINS, Ego4D and EgoLife.

Why it mattersThe open repository gives developers a way to analyse long recordings for activities and events.

GitHub ↗
Research

Google releases AlphaProtein Novo for enzyme design

Google DeepMind released AlphaProtein Novo, a diffusion pipeline that designs enzymes around catalytic motifs and ligand contexts. It can combine sequence redesign with LigandMPNN and structure prediction with AlphaFold 3.

Why it mattersThe release provides tools for de novo enzyme design via structural motif scaffolding.

GitHub ↗

Agents & coding

AI that acts on its own, and the tools developers build with.

Product

Devin adds personal Memory and daily Dreaming

Cognition introduced Memory to retain user-specific work lessons across sessions and Dreaming, a daily process that organizes those memories. The memory system spec is open sourced for other agents.

Why it mattersAgents that remember and refine can work more like teammates.

devin.ai ↗
Safety

Researchers track a Chinese AI agent fleet online

Researchers found parallel agents apparently running on Tencent infrastructure and querying Alibaba’s Amap for directions. They saw no sign the agents were coordinating as a swarm.

Why it mattersPersistent agent activity is becoming easier to spot—and harder to ignore—on the open internet.

TechCrunch AI ↗
Product

Hebbia adds memory, scheduling and workflow controls

Hebbia’s Max can remember context, chain skills, create agents and schedule them for future dates. Matrix workflows can now use conditional steps and formatting.

Why it mattersMore control over memory and scheduling makes business agents easier to use repeatedly.

hebbia.com ↗
Research

Google report maps the privacy risks of AI agents

Google Research's report, written with more than 50 academic and industry leaders, says agents need to follow social norms for each context. It highlights prompt injection, uncertain decision paths and weaker human oversight.

Why it mattersAgents need access to personal data, so context-aware safeguards will shape whether people can trust them.

Google Research ↗
Research

AI is making influence operations faster, not necessarily stronger

A Lawfare analysis says state-linked groups use AI to write propaganda, create personas, translate content, process intelligence and automate workflows. It finds more evidence of greater capacity than of changed public opinion.

Why it mattersAI can make political manipulation cheaper and more persistent, even if persuasion remains difficult.

Lawfare ↗
Product

Cohere adds spending caps to its North agent platform

North 2 adds redesigned orchestration, shared agents, memory, skills and new connectors. Admins can track token use, set limits and apply controls for personal data, prompt injection and agent actions.

Why it mattersCompanies can deploy agents while keeping tighter control over cost, data and what each system is allowed to do.

SiliconANGLE ↗
Product

Constructor and Stripe bring checkout into shopping agents

Constructor's Agentic Checkout lets shoppers find products, ask questions and pay without leaving an AI shopping conversation. It supports Link by Stripe and other payment methods, and is available globally.

Why it mattersAI shopping is moving from product recommendations to completing the purchase in the same conversation.

PR Newswire ↗
Product

Cognition and Factory take opposite paths to coding agents

Cognition runs its Devin workloads on Nvidia Vera Rubin systems and trains its own SWE-2 model. Factory instead routes coding tasks among outside models and offers its Router for private or air-gapped deployments.

Why it mattersThe split tests whether coding agents work best with one owned model or many interchangeable ones.

FourWeekMBA ↗
Product

n8n lets users buy more Assistant credits

n8n Cloud Starter and Pro users can now buy extra Assistant credits or turn on auto top-ups. Extra credits carry over and expire 12 months after purchase.

Why it mattersUsers can keep building workflows instead of waiting for their monthly allowance to reset.

n8n ↗

Video generation

The models and tools that make video.

Product

Rick and Morty animator makes daily show with Higgsfield

Animator Hugo Blendl now creates his Zero to Hero series solo on Higgsfield, producing about 1:20 of finished animation in ~4 hours, versus years for a traditional pilot. He releases new episodes every day or two.

Why it mattersAI-native tools are remaking animation speed and cost.

higgsfield.ai ↗
Policy

Court throws out sentence influenced by AI video

An Arizona appeals court ordered a new sentence for Gabriel Horcasitas after a judge allowed an AI video of slain man Christopher Pelkey to speak at sentencing. The court called the decision fundamentally unfair.

Why it mattersThe court found the AI video prejudiced the sentencing procedure.

NBC News ↗
Safety

Janhvi Kapoor condemns sexually altered AI video

Janhvi Kapoor called an altered video using footage from Devara: Part 1 deeply violating and bordering on sexual assault. The video was later taken down.

Why it mattersThe case shows how AI manipulation can turn familiar film footage into non-consensual sexual content.

Telugu360 ↗
Research

Nearly half mistook Tavus's AI caller for a person

In a Tavus test, 26 of 54 people said they had spoken with a real person after a one-minute call with its Griffin-Lite model. The earlier system convinced one person out of 41.

Why it mattersAI video callers are becoming convincing enough to challenge basic identity checks.

HackerNoon ↗
Product

AI video is pitched as an enterprise production fix

An ElevenLabs guest article says businesses need more video than traditional production can supply. It argues that generating clips from text and images can reduce the time and resources needed.

Why it mattersAI video could make everyday business videos faster and cheaper to produce.

iTWire ↗
Policy

Yami Gautam asks media not to alter her photos with AI

Actor Yami Gautam asked publications to use her public-appearance photos in their original form. She said AI changes to a person's image or likeness need their knowledge and consent.

Why it mattersCelebrities are pushing for consent over AI changes to their appearance.

ETV Bharat ↗

Creator tools

The apps and platforms people make video with.

Product

Snowflake unveils AI Gateway for governed ad agents

Snowflake’s AI Gateway for Advertising (preview) connects Cortex Agents with governed MCP integrations for Meta and TikTok, bringing enterprise context and live platform signals to recommend and act with approval.

Why it mattersMarketers could let AI tune campaigns under clear guardrails.

snowflake.com ↗
Deal

Lovable apps can run inside a company’s Microsoft tenant

Lovable is working with Microsoft so users can publish apps into their own tenant with Microsoft sign-in and IT controls. Apps can also use Microsoft 365, Fabric and Dataverse data.

Why it mattersTeams can build internal apps without moving company data or creating another login system.

lovable.dev ↗
Product

Lovable opens its chat experience to everyone for free

Lovable now lets people discuss app ideas, inspect existing projects and make changes before building. It can also connect to tools such as Notion, Granola and Linear.

Why it mattersPeople can test and refine an app idea through conversation before committing to a build.

lovable.dev ↗
Product

Rogo ships agent-written code in five minutes

Financial AI company Rogo uses Vercel and AI SDK agent swarms to ship code and handle production incidents. It reported more than 73,000 deployments in one month.

Why it mattersAgent-written software is moving from experiments into fast, frequent production releases.

vercel.com ↗
Product

VideoGen adds four AI video tools to one workflow

VideoGen added dedicated pages for Higgsfield AI, HeyGen AI, Seedance and VEED. Users can combine generated clips with footage, voiceover, music and captions.

Why it mattersCreators can plan different video jobs in one editable project instead of switching between tools.

videogen.io ↗
Product

Cursor adds Zhipu’s GLM 5.3 to its model picker

Cursor added Zhipu’s GLM 5.3 without a public partnership announcement. The model costs $1.40 per million input tokens and $4.40 per million output tokens in Cursor.

Why it mattersCoding tools are quietly giving developers more choice beyond the biggest Western AI labs.

Startup Fortune ↗
Product

Databricks shares a test for choosing first AI agents

Databricks recommends ranking agent ideas by impact, demand, data readiness, scope and governance. It says more than 1 million Genie Agents were created in 2026.

Why it mattersA narrow, useful first project may build more trust than an agent that tries to do everything.

Databricks ↗

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.

Policy

Greta Garbo heirs won’t rule out AI film return

A family representative said an AI-assisted Garbo role is possible after a ‘startling’ ad used models like Seedream and Google’s Nano Banana Pro to recreate her likeness without ingesting protected footage.

Why it mattersEstates may license AI performances without using archival data.

The Cool Down ↗
Product

Dolphin AI aims to scale microdrama production

Eloelo Group opened Dolphin AI, an agentic studio for end-to-end video production, to creators, filmmakers, brands and studios. The company says it helped produce 150 to 200 shows and 5,000 advertisements internally.

Why it mattersAutomating continuity and scene assembly could make short-form drama faster to produce.

Adgully ↗
Deal

Warner Music buys AI attribution startup Sureel

Warner Music Group acquired Sureel to track how songs and recordings are used in AI training and generated works. Sureel will continue operating as a standalone platform.

Why it mattersMusic companies want clearer records of how AI uses artists' work and identities.

Billboard ↗
Product

AI Film Summit Los Angeles sets October programme

The October 24 event will screen 35 selected films and present nine awards. Panels will cover production, careers, distribution and creative rights.

Why it mattersThe summit puts AI-assisted films and the people making them in front of an industry audience.

BroadwayWorld ↗
Research

India's microdrama platforms face a retention problem

Executives say Indian microdrama platforms are seeing strong participation from male viewers, including in smaller cities. They also report that some users leave after watching suggestive scenes.

Why it mattersLarge audiences do not guarantee that short-video platforms can keep viewers or make money.

Storyboard18 ↗
Product

Tencent tests a live game companion

Tencent is recruiting closed-beta testers for Goose Dimension. Local reports describe it as an AI companion that can use voice chat and respond to what it sees on a game screen.

Why it mattersGame assistants may soon react to play in real time instead of offering fixed tips.

TechNode ↗

Chips & compute

Chips, data centres and the power behind them.

Deal

LG Electronics wins U.S. AI data center chiller deal

LG’s US arm signed a long-term pact with AIR Control Concept to supply high-efficiency chillers and cooling solutions for AI data centers totaling 5 GW. LG booked over ₩600B in AIDC cooling orders in H1 and now expects to hit ₩1T this year.

Why it mattersAI data centers need vast, efficient cooling as power demand surges.

Yonhap ↗
Product

ROCm 10.1 targets data-movement bottlenecks

AMD’s ROCm 10.1 adds Infinity Storage upgrades and HIP runtime changes to speed data paths, including async hipFile transfers on HIP streams, a batch I/O API, and multi-tier I/O stats. NUMA-aware host allocation places memory closer to compute.

Why it mattersFaster storage-to-GPU pipelines can unlock better AI training and inference.

rocm.blogs.amd.com ↗
Deal

Cerebras shares rebound after Altman calls it a close partner

Cerebras stock rose 9% after OpenAI CEO Sam Altman called the chipmaker a close partner. OpenAI will use Nvidia GPUs for GPT-6.1 Sol's Ultrafast mode, while Cerebras has a $10 billion supply deal with OpenAI.

Why it mattersThe episode shows how quickly chip-company valuations can react to shifting AI partnerships.

CNBC ↗
Chips

d-Matrix and Gimlet Labs pair chips and GPUs for agents

Gimlet Labs plans to add d-Matrix Corsair accelerators to its cloud alongside GPUs. The companies say the hybrid setup could deliver up to 10x better latency and throughput per watt than GPU-only systems.

Why it mattersSpecialised chips could make real-time AI agents faster and less power-hungry.

d-matrix.ai ↗
Chips

AMD samples a space chip built for 15-year missions

AMD is sampling its Versal AI Core XQRVC1902 in a lidless space-grade package for missions lasting up to 15 years. Class Y flight-qualified units are expected in the second half of 2027.

Why it mattersAMD is targeting long-duration missions with higher compute density for onboard payload processing.

StorageReview.com ↗
Funding

Clockwork.io raises $31M for GPU fault tolerance

Clockwork.io raised $31 million and announced deployments at LinkedIn and Together AI. Its software reroutes around failed links and moves work from failing GPUs so AI workloads can continue.

Why it mattersKeeping large AI jobs running through hardware failures can save time, computing capacity and money.

ReadITQuik ↗
Chips

Dell and CoreWeave move AI reliability into the factory

Dell and CoreWeave are jointly designing and testing rack-scale AI systems before they reach data centres. Their work covers diagnostics, liquid cooling, power and leak detection.

Why it mattersReliable AI infrastructure depends on coordinating the whole rack and data centre, not just the chips.

SiliconANGLE ↗
Chips

Nokia says data centres could be built twice as fast

Nokia CEO Justin Hotard said customers would build AI data centres “2x faster” without shortages of memory chips and energy. Nokia supplies technology that connects chips within and between data centres.

Why it mattersThe comment suggests demand remains strong despite doubts about the scale of the AI build-out.

CNBC ↗
Chips

OpenAI's dots agents run on nine-core AMD VMs

Geekbench results suggest OpenAI's dots agents use nine-core AMD EPYC virtual machines with nearly 10GB of memory. Their median multi-core score was about six times Meta Muse's.

Why it mattersThe results offer a glimpse of the cloud hardware powering autonomous AI agents.

Tom's Hardware ↗

Policy & safety

Laws, lawsuits and safety incidents.

Policy

Poll finds broad US support for slowing AI development

A Quinnipiac poll found 47% of US adults want slower AI development and 30% want a full stop until safety is verified. It also found 86% support independent safety standards and 74% have little trust in AI leaders.

Why it mattersPublic support for AI guardrails is high, while trust in the companies building the technology is low.

The Decoder ↗
Policy

Maryland governor calls for a federal AI framework

Maryland Gov. Wes Moore said he is leading a bipartisan governors' council on secure and responsible AI. He proposed audits, safety tests, limits on human likenesses and protections for children using chatbots.

Why it mattersState leaders are pushing for common AI rules covering safety, workers and children.

WBAL-TV ↗
Safety

Former New Jersey lieutenant governor cites AI in defence

Dale Caldwell, who resigned after a sexual harassment investigation, said he put the report through several AI platforms. He claimed the systems repeatedly found no basis for a sexual harassment finding.

Why it mattersChatbots can sound confident and agreeable even when they are not reliable judges of disputed facts.

The Verge AI ↗
Policy

Trump backs AI self-regulation over new safety laws

President Donald Trump said AI companies should largely regulate themselves, with existing agencies including the Department of Justice and FBI. He also said beating China remains the priority.

Why it mattersThe approach leaves enforceable safety rules for advanced AI unresolved.

WBFF ↗
Policy

White House AI accord puts companies in charge of safety

AI leaders signed a White House document committing companies to internal monitoring, outside evaluators and independent board oversight. The accord does not create government safety standards.

Why it mattersThe agreement shows how the industry plans to police powerful AI without new laws.

CNBC ↗
Safety

Investor says AI safety remains unsolved

Deepwater Asset Management’s Gene Munster said the AI safety issue has still not been solved after a meeting between the Trump administration and leading AI executives.

Why it mattersMunster commented after a meeting between the Trump administration and AI executives.

CNBC ↗
Policy

AI companies face New York City safety hearing

Representatives from OpenAI, Google, Meta and Anthropic were expected to testify before the City Council about AI risks. Proposed laws would add checks and liability for harm.

Why it mattersNew York City is considering rules that could affect how companies deploy AI locally.

Spectrum News NY1 ↗

Robotics

Robots, self-driving and AI in the physical world.

Funding

AgiBot pledges up to RM1b Malaysia investment by 2035

China’s AgiBot said it could invest up to RM1 billion and create 5,000 jobs in Malaysia by 2035 after meeting PM Anwar. The plan builds on its i-City collaborations and domestic robot trials.

Why it mattersAgiBot is expanding in Malaysia and testing how humanoid robots could move into the home.

The Star ↗
Product

Raspberry Pi 5 gets low-power vision and language AI

A Sixfab board adds a DEEPX NPU to Raspberry Pi 5, delivering 25 TOPS at about 3 watts of sustained NPU power. It can support vision, compact vision-language and small language workloads at the edge.

Why it mattersLow-power local AI could help cameras, robots and sensors work without a constant cloud connection.

Raspberry Pi ↗
Research

Humanoid robot performs live surgery in a first

A Mashable report visits UC San Diego's Center for the Future of Surgery for the first live surgery using humanoid robots. The report covers teleoperation, animal testing, heat management and the possibility of surgical autonomy.

Why it mattersHumanoid robots are being tested in medicine, including teleoperation and potential autonomy.

Mashable ↗
Policy

Agility Robotics CEO joins military technology study

Agility Robotics CEO Peggy Johnson will advise Project Meridian, a 120-day study of AI, autonomy, robotics and other technologies for future military operations. Recommendations are due to the secretary’s office by 28 January 2027.

Why it mattersCommercial robotics leaders are helping shape how robots could be used in future defence operations.

AI Insider ↗
Research

Humanoids may not be the best factory robots

Q5D CEO Stephen Bennington says manufacturers should start with the process, not the robot. Purpose-built machines can be faster, more accurate and cheaper when a task does not require a human shape.

Why it mattersThe most useful robot for a job may not look or move like a person.

TechTarget ↗

Science & health

AI in biology, medicine, climate and maths.

Research

Arc Institute’s PIE predicts unseen cell responses

PIE uses curated biological knowledge and a Perceiver IO-based architecture to predict population-level gene expression changes for perturbations, outperforming baselines on Replogle-Nadig, especially for unseen cases.

Why it mattersPIE generalizes across unseen contexts and perturbations using curated biological knowledge.

arcinstitute.org ↗
Deal

Italfarmaco and VeriSIM sign $100M-plus drug discovery deal

Italfarmaco and VeriSIM Life will develop small-molecule inhibitors for an undisclosed HDAC target, using AtlasGEN and BIOiSIM. The deal’s potential value exceeds $100 million with milestones and royalties.

Why it mattersVeriSIM will generate and assess candidates, while Italfarmaco brings HDAC and development expertise.

AllSci ↗
Deal

AbbVie and Iambic expand AI drug discovery work

AbbVie and Iambic will use Iambic's molecular superintelligence platform to discover small-molecule drugs in immunology, neuroscience and oncology. The work will use Iambic's Enchant v3 model and automated experiments.

Why it mattersAI can test many drug properties together, helping researchers weigh trade-offs earlier.

SNS Insider ↗
Research

37,000 AI agents find a possible drug-success marker

A Stanford-led virtual biotech used 37,075 agents to study 55,984 clinical trials. It found that drugs targeting genes active in fewer cell types were more likely to succeed and had fewer reported side effects, but the result is only a link.

Why it mattersA large-scale AI review could help researchers choose drug targets before human trials.

Earth.com ↗
Research

Sony researchers release memory system for world models

Sony Group researchers released FAR, a system that helps AI recall useful past observations in virtual environments. Its demo revisits changed kitchen stations and checks whether the remembered state is correct.

Why it mattersThe demo shows FAR recalling earlier observations and judging whether they still hold.

GitHub ↗
Product

Recursion repository runs Slurm clusters on Kubernetes

The slurm-operator project manages Slurm computing clusters inside Kubernetes. It supports cluster creation, upgrades, shutdowns and worker scaling.

Why it mattersCombining the two systems could make large AI and scientific workloads easier to manage.

GitHub ↗

Open source

Open-weight models and community releases.

Model

Amazon open-sources Strands Decider 2B for fast agent steps

AWS released Strands Decider 2B, a Qwen3.5-2B-based decision model with a pointer head that selects from fixed options and returns a confidence score. Median latency is ~115 ms on an RTX 3090.

Why it mattersCheap, fast decisions can cut agent costs in workflows that don’t need full LLMs.

SQ Magazine ↗
Product

Lola Vision wants to simplify running AI on chips

Lola Vision Systems builds software that translates AI models into instructions for specific chips, and is developing its own chips. The company says setup can otherwise take roughly 200 hours.

Why it mattersEasier setup could help companies run AI on devices such as cameras and drones.

TechCrunch ↗

Quick hits

Smaller stories worth a glance.

Deals & funding

Who raised money, and who bought whom.

7 deals

Funding rounds 2

Investors put money into a company in exchange for a share of it.

Blackstone backs Factory

Raised
$200M
Funding

Factory now valued at $5B

Factory raised $200 million, bringing its total funding past $400 million, according to the source.

Other investors

KVKhosla VenturesSequoia CapitalInsight PartnersNNEAECEvantic CapitalSVSound VenturesMVMantis VCCClearlake

· FourWeekMBA ↗

Premji Invest backs Clockwork.io

Raised
$31M
Funding

Clockwork.io valuation not disclosed

Clockwork.io raised $31 million to expand its GPU fault-tolerance software across training, inference and reinforcement learning.

Other investors

WVWing Venture CapitalSVSeligman VenturesNNEAeCe& Capital

· ReadITQuik ↗

Acquisitions 2

One company buys another outright.

Partnerships & contracts 3

A large commercial agreement, or a government contract.

Italfarmaco and VeriSIM Life sign a deal

Value
$100M+
Partnership

Multi-year agreement

The companies entered a drug discovery collaboration to develop small-molecule inhibitors against an undisclosed HDAC target.

Other investors

VLVeriSIM Life

· AllSci ↗

On the calendar

What's coming up, in date order, and how sure we are about each.

3 things
MTWTFSSOct56789101112131415161718192021222324252627282930311Nov23456789101112131415
  1. 13–15OctTechCrunch Disrupt in San Francisco7days away
  2. 24OctAI Film Summit Los Angeles 2026Tyrannus Foundation3weeks away
  3. 9–11NovThe Hinton Lectures on AI safetyAI Safety Foundation5weeks away

Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.