VeltrixVeltrix.
Live feed— curated by Vel

AI news

Only what matters. No filler, no hype. Every story earns its place.

IndustryTechCrunch

Meta’s Muse hits 10M downloads in 40 days — how does it stack up?

Meta's AI companion Muse reaches 10 million U.S. downloads in 40 days, joining an elite group of apps.

Summary

  • Muse hit 10 million U.S. downloads in 40 days
  • It's in the same exclusive group as Threads
  • User retention is higher than rival AI apps
  • This is a major milestone for Meta's AI strategy
  • The speed of adoption is unprecedented for an AI companion app
ResearchHugging Face

Does fairness in AI decisions mean what we think?

New research finds hidden biases in how LLMs handle ethical dilemmas.

Summary

  • Fairness in LLMs is often measured superficially, not deeply.
  • Biases emerge in complex, real-world decision contexts.
  • Current benchmarks fail to capture nuanced ethical misalignments.
  • Human oversight significantly improves fairness outcomes.
  • The problem is not fixed by technical adjustments alone.
IndustryMIT Tech Review

Is your AI strategy just the tip of the iceberg?

MIT Tech Review reports on AI infrastructure becoming the new economic battleground.

Summary

  • AI is moving from novelty to essential infrastructure
  • Efficiency, not just adoption, now defines competitive advantage
  • Companies are starting to measure AI usage like energy consumption
  • Third-party tools are now central to enterprise AI operations
  • Early adopters are seeing efficiency gains of 20-40% in core workflows
IndustryTechCrunch

Why is the US banning DeepSeek AI from federal systems?

US federal agencies ordered to remove DeepSeek AI and related services due to national security concerns.

Summary

  • The US has ordered federal agencies to remove DeepSeek AI from their systems.
  • DeepSeek is a Chinese AI company with growing global influence.
  • The move is driven by national security risks associated with the AI technology.
  • Third-party tools built on DeepSeek's tech are also affected.
  • This reflects broader US-China tensions over AI and data control.
IndustryTechCrunch

How Eka's robotic arm is redefining AI's physical limits

Eka Robotics unveils a robot arm capable of picking up a wide variety of objects with human-like precision.

Summary

  • Eka Robotics has developed a robotic arm with unprecedented dexterity and adaptability.
  • The arm's smooth, human-like motion has impressed industry experts.
  • It may be the first robotic arm to handle diverse objects with real-world precision.
  • Eka's valuation has reached $1.45 billion due to its technological edge.
  • This could shift the future of automation and AI in manufacturing.
ResearchMIT Tech Review

A radical new claim changes how we see AI — and the internet’s agent problem

A new study claims neural networks may be more than black boxes — with profound implications.

Summary

  • A researcher claims neural networks may have emergent reasoning capabilities.
  • The study was presented at a major AI conference in Cambridge.
  • The findings challenge long-held assumptions in the field.
  • The paper has already sparked debates on AI ethics and governance.
  • Tools are being developed to detect and measure this emergent behavior.
IndustryMIT Tech Review

Why isn't the tech sector the first to adopt AI?

MIT Tech Review highlights how other industries are reshaping AI adoption.

Summary

  • Tech is not leading AI adoption; other sectors are.
  • Generative AI is already transforming non-tech fields.
  • Early adopters outside tech are setting new benchmarks.
  • Productivity gains are outpacing Silicon Valley expectations.
  • Real-world use cases are driving sustainable AI adoption.
Model releasesHacker News

DeepSeek's open-source inference engine changes the game for developers

DeepSeek has open-sourced its inference engine, making powerful AI model deployment more accessible.

Summary

  • DeepSeek's inference engine is now open-source for all to use.
  • It promises low-latency deployment across multiple frameworks.
  • The engine is optimised for both cloud and on-premise usage.
  • It supports integration with TensorFlow, PyTorch, and ONNX.
  • This could lower barriers to deploying high-performance AI models.
IndustryMIT Tech Review

What's inside Samsung's new chipmaking behemoth?

Samsung builds fourth semiconductor fab in Pyeongtaek to meet AI's memory demands.

Summary

  • Samsung is ramping up memory production for AI chips in South Korea.
  • A fourth fab is being constructed alongside three already operational ones.
  • Demand for memory is driven by the rise of generative AI models.
  • GPUs rely on stacks of specialized memory chips to function efficiently.
  • This expansion underscores AI's growing impact on global semiconductor manufacturing.
IndustryArs Technica

Japan's Mars moon mission: A leap to bring Phobos home

Japan's MMX mission aims to return samples from Phobos by 2029.

Summary

  • Japan's MMX mission will attempt two landings on Phobos by 2029.
  • The mission aims to return samples to Earth for scientific study.
  • Phobos is one of Mars' two small moons, and studying it could reveal insights into the solar system's history.
  • The spacecraft is designed to collect and store samples before returning to Earth.
  • This mission is part of a broader push to understand the Martian system and its moons.
IndustryTechCrunch

Nscale and Anthropic: Another $45B compute deal in the works?

Nscale and Anthropic are reportedly discussing a new $45 billion compute deal.

Summary

  • Nscale and Anthropic were previously reported to have a $45 billion compute deal in August.
  • The two companies are now reportedly in talks for another deal of the same scale.
  • This would mark a significant escalation in their collaboration.
  • Compute costs are a major driver in AI development and deployment.
  • This highlights the growing importance of infrastructure in AI scaling.
IndustryArs Technica

Is congenital measles resurging in the US?

Pennsylvania issues alert over rare congenital measles cases not seen for decades.

Summary

  • Congenital measles is a rare but severe condition affecting newborns.
  • It can be fatal or lead to long-term complications.
  • Survivors require years of medical monitoring.
  • The condition is typically associated with maternal infection during pregnancy.
  • Pennsylvania is the first US state to report this in decades.
IndustryArs Technica

Close call in space: Russian satellite nearly collides with Ukrainian imaging tool

Kosmos 2612 approached ICEYE-X36 to within hundreds of kilometers, raising concerns over space safety.

Summary

  • Russian satellite Kosmos 2612 maneuvered to extreme proximity with Ukrainian satellite ICEYE-X36
  • The event occurred in Low Earth Orbit, 400–500 km above Earth
  • ICEYE-X36 is used for high-resolution Earth imaging by Ukraine
  • No collision occurred, but the maneuver was unexplained and unprecedented
  • This raises concerns about space traffic management and intent behind such maneuvers
ToolsTechCrunch

Amazon's new tool could shake up AI's chip hierarchy

Amazon introduces TorchNeuron to challenge Nvidia's dominance in PyTorch development.

Summary

  • Amazon is releasing TorchNeuron to reduce reliance on Nvidia chips.
  • The tool allows PyTorch apps to run on Amazon's own hardware.
  • This aims to give developers more chip flexibility.
  • It's a direct challenge to Nvidia's stronghold in AI hardware.
  • Amazon wants to attract developers to its ecosystem.
IndustryTechCrunch

Tenstorrent's $6.2B valuation shakes up open-weight AI chip race

Tenstorrent raises $1.5 billion to remain independent as it scales its AI chip business.

Summary

  • Tenstorrent achieved a $6.2 billion valuation after raising $1.5 billion.
  • The company chose independence over acquisition offers.
  • It focuses on open-weight AI chips, a niche different from Nvidia's closed ecosystem.
  • The funding will accelerate deployment of its AI inference chips.
  • This move challenges the dominance of major AI chipmakers like Nvidia.
IndustryMIT Tech Review

Could AI agents reshape the internet for the worse?

MIT study finds optimizing the web for bots may harm human users.

Summary

  • Bots are now a third category of internet users
  • Optimizing for bots may degrade user experience
  • Websites are increasingly unresponsive to humans
  • Automation is outpacing human-centric design
  • This shift risks deepening the digital divide
Model releasesTechCrunch

Nvidia's bold open model move: What does it mean for your enterprise?

Nvidia has released new open models for enterprise use and on-device AI agents.

Summary

  • Nvidia is pushing open models for enterprise data centres and on-device AI agents.
  • This aims to give companies more control over AI deployment.
  • The models are designed for high performance and scalability.
  • Emphasis is on privacy and on-device computation.
  • This follows a growing trend of enterprises seeking AI autonomy.
IndustryTechCrunch

How AI agents are stress-testing robots before they fail in real life

Laplace Intelligence simulates real-world chaos for self-driving cars and robots before deployment.

Summary

  • Laplace Intelligence uses AI agents to simulate real-world chaos for robots and self-driving cars.
  • The startup is a University of Michigan spinout now competing in TechCrunch Startup Battlefield.
  • Their approach involves interactive simulations to stress-test autonomous systems.
  • The goal is to find critical flaws before deployment, reducing risks and failures in the real world.
  • This method is proving profitable and is gaining industry attention.
ResearchMIT Tech Review

What happens when Arctic fires meet frozen carbon?

A wildfire in 2021 exposed the hidden threat of Arctic carbon release.

Summary

  • Arctic fires are releasing vast amounts of ancient carbon.
  • Black spruce forests are especially vulnerable to burning.
  • The 2021 fire released stored carbon more quickly than expected.
  • Permafrost thaw and fire create a dangerous feedback loop.
  • This has implications for global warming and climate models.
IndustryTechCrunch

Why is Replit choosing OpenAI over Anthropic? The implications are profound.

Replit has shifted its default model provider from Anthropic to OpenAI, according to CEO Masad.

Summary

  • Replit now defaults to OpenAI models over Anthropic's Claude.
  • Claude will remain available as an option for users.
  • The decision was confirmed by Replit's CEO, Amir Youssef Masad.
  • This shift may impact developers and AI users on the platform.
  • It highlights ongoing competition between major AI providers.
IndustryOpenAI

How Itaú is making AI work for millions in Brazil’s strictest sector

Itaú is rolling out generative AI across its customer base with new tools and governance.

Summary

  • Itaú is implementing generative AI through ia.i and IARA across Brazil’s highly regulated financial sector
  • The AI First strategy is targeting both customers and employees
  • IARA ensures compliance and ethical use of AI in sensitive financial contexts
  • ia.i functions as a conversational, customer-facing AI interface
  • This is one of the most ambitious AI deployments in a heavily regulated market
ResearchMIT Tech Review

Can AI engineer viruses? A Stanford student says yes.

A PhD student used AI to design viral genetic blueprints, sparking new ethical and scientific questions.

Summary

  • AI can now generate virus-like genetic designs
  • This is not yet life, but a step towards synthetic biology
  • The work was done by a Stanford PhD student named Samuel King
  • The AI model was trained on vast genomic databases
  • The implications raise urgent ethical concerns
IndustryTechCrunch

India uncovers stealthy AI agents lurking in plain sight

India reports rogue AI agents infiltrating systems via exposed credentials and misconfigured cloud roles.

Summary

  • Rogue AI agents accessed systems through exposed credentials and cloud misconfigurations
  • Public code repos and overly permissive cloud roles were exploited
  • India is not the first, but the latest in a growing global trend
  • Attackers used known vulnerabilities in app configuration settings
  • Security teams are scrambling to audit and secure all exposed endpoints
Model releasesTechCrunch

How fast is Jev really? Vercel's new AI model defies expectations

Jev, Vercel's latest AI model, has been adopted at unprecedented speed through Vercel's AI gateway.

Summary

  • Jev is Vercel's fastest-adopted AI model to date
  • It was released through Vercel's AI gateway, a unified API service
  • Vercel's AI gateway functions similarly to OpenRouter
  • Jev's adoption speed suggests strong developer interest in its capabilities
  • This marks a significant shift in how AI models are deployed and utilised
ResearchHugging Face

2.8 Trillion Tokens – What Does It Mean for Indic Languages?

Hugging Face unveils a major breakthrough in Indic language pretraining with a new open data pipeline.

Summary

  • 2.8 trillion tokens of Indic text now available for pretraining
  • Open data generation pipeline democratizes access to large-scale Indian language data
  • This is the largest Indic language corpus ever released
  • Model training is now more accurate and representative for over 100 million speakers
  • The pipeline supports 12 major Indic languages with high-quality annotations
ToolsOpenAI

How does Codex transform loan analysis for finance professionals?

Benefit Street Partners uses Codex to accelerate code review and analysis of credit agreements.

Summary

  • Codex is used to automate and speed up code review for loan valuation tools.
  • Anant Kumar leverages Codex to improve follow-up processes in finance.
  • The tool helps compare loan valuations more efficiently.
  • Credit agreement analysis becomes faster and more accurate.
  • This reduces manual effort and increases productivity for finance teams.
ToolsOpenAI

How did ChatGPT triple agent creation in weeks?

A new tool built with ChatGPT plugin boosted signups by up to 10× in just a few weeks.

Summary

  • A new tool called 'Stack Overflow for Agents' used ChatGPT to grow rapidly.
  • Engineers built it in weeks using Codex and the ChatGPT plugin.
  • ChatGPT plugin drove up to 10× growth in agent creation.
  • Community-driven development allowed users to shape the tool’s evolution.
  • The tool focuses on agent collaboration, similar to Stack Overflow’s question-and-answer format.
Model releasesHugging Face

A 501B model you can own — what does that even mean?

Hugging Face launches Welcome Beam: an open model with novel licensing and usage rights.

Summary

  • Welcome Beam is a 501B open model with unique licensing terms.
  • Users can own, customize, and scale it without restrictive clauses.
  • It is distinct from standard open-source models due to its 501B nature.
  • The model is available on Hugging Face for immediate testing and deployment.
  • This represents a shift in how open models can be used in commercial and private contexts.
Model releasesHugging Face

Granite Switch 4.1: Can one model handle chat, code, and speech?

Hugging Face releases Granite Switch 4.1 with unified capabilities for chat, decisions, and speech.

Summary

  • Granite Switch 4.1 merges chat, decision-making, and speech into a single model.
  • It is the first model to perform all three tasks with consistent quality.
  • The model shows notable improvements in multilingual speech synthesis.
  • It is available through the Hugging Face Hub with open-source access.
  • Ideal for developers seeking end-to-end conversational AI systems.
IndustryArs Technica

Targeted ads weaponised to install spyware without user interaction

Mercenary malware delivered via zero-click exploits through targeted online ads.

Summary

  • Targeted ads used to deliver zero-click malware exploits.
  • Likely first confirmed case of mercenary spyware via ad networks.
  • No user action required for installation.
  • Attack vector bypasses traditional security measures.
  • Highlights major vulnerability in digital advertising ecosystems.
Model releasesTechCrunch

A new AI model runs on your Mac — what does that mean for creativity?

Meta and Apple have launched Muse Spark Open, an open-weight AI model designed for Apple silicon.

Summary

  • Muse Spark Open is an open-weight AI model running natively on Apple silicon.
  • It is developed by Meta and optimized for Apple's M-series chips.
  • The model focuses on creative tasks like text and image generation.
  • It can be used locally without requiring cloud infrastructure.
  • This is a significant step toward on-device AI and open collaboration.
IndustryTechCrunch

Anthropic shares rare access with METR — what could go wrong?

Anthropic grants METR unprecedented employee-level access to its systems and practices.

Summary

  • Anthropic is letting METR researchers examine its internal operations in detail.
  • METR can publish its own assessments of Anthropic’s risks and practices.
  • Anthropic has waived the right to block METR’s findings.
  • This opens a rare window into how frontier AI companies operate.
  • The move sets a precedent for future AI oversight and transparency.
Model releasesOpenAI

GPT-Rosalind now reshapes life sciences research — for the right people

OpenAI launches the Rosalind Academic Plan, granting labs access to GPT-Rosalind and tools for scientific discovery.

Summary

  • GPT-Rosalind is tailored for life sciences research with domain-specific training data.
  • The Rosalind Academic Plan provides eligible labs with sponsored access to the model.
  • Tools include customisable prompts, data analysis workflows, and integration with lab equipment.
  • Access is restricted to non-commercial research institutions with verified credentials.
  • Labs can request access via the OpenAI website, with applications reviewed on a rolling basis.
IndustryArs Technica

Drones cross into NATO waters: What does this mean for Europe?

Suspected Russian drones strike a chemical tanker in Romanian territorial waters.

Summary

  • Drones crossed into Romanian territorial waters for the first time in NATO history.
  • The attack targeted a chemical tanker near Constanta, Romania.
  • This marks a new escalation in hybrid warfare tactics.
  • NATO has not yet officially commented on the incident.
  • The event raises questions about the protection of NATO member states.
ResearchGoogle DeepMind

What if AI could design antibiotics faster than nature?

Google DeepMind used AI to accelerate the discovery of antimicrobial peptides.

Summary

  • AI is now generating antimicrobial peptides in days, not years.
  • The peptides discovered are novel and have not been previously catalogued.
  • This reduces the time required for experimental validation by up to 90%.
  • The method is called 'Antigravity' and uses language model techniques adapted for biology.
  • This opens new frontiers in the fight against drug-resistant bacteria.
ToolsTechCrunch

Google’s new shopping tool changes how we shop online — really

Google’s Universal Cart shopping tool has launched in the US through Search and the Gemini app.

Summary

  • Google's Universal Cart tool now available in the US via Search and Gemini app
  • Lets users collect items from multiple stores in one cart
  • Available across Google Search and Gemini app, not just Google Shopping
  • Works with multiple retailers, not just Google-owned stores
  • Aims to simplify multi-store online shopping experience
IndustryMIT Tech Review

Reviving frozen kidneys and boosting nuclear power — what does it mean for medicine and energy?

Pig kidneys frozen for months have been successfully revived, and nuclear power projects are accelerating.

Summary

  • Pig kidneys frozen for a year have been revived and functioned normally.
  • The breakthrough uses vitrification, a process that prevents ice crystal formation.
  • This could revolutionise organ preservation and transplant logistics.
  • Parallel developments in nuclear energy aim to boost clean power output by 2030.
  • Both fields highlight the intersection of life-saving science and sustainable infrastructure.
IndustryMIT Tech Review

Nuclear power plants hitting their limits — but why?

Constellation plans to boost output from US nuclear plants as demand grows.

Summary

  • Constellation aims to increase nuclear output by 15% through upgrades.
  • Current plants are operating near capacity for the first time in decades.
  • The move comes amid rising electricity demand and energy crisis concerns.
  • Upgrades include reactor modifications and grid integration improvements.
  • This could delay the transition to renewables if not managed carefully.
ResearchMIT Tech Review

Frozen pig kidneys revived after a year — what does this mean for human medicine?

Scientists have successfully revived pig kidneys frozen for over a year and restored their function.

Summary

  • Pig kidneys were frozen to -150 °C for over a year and then revived.
  • The kidneys retained their function after thawing.
  • This could transform long-term organ storage for transplantation.
  • The method uses a unique glass-like preservation state.
  • Researchers aim to apply this to human organs in the future.
IndustryTechCrunch

AI-designed chips: Could they fix the compute crisis?

Ricursive Intelligence uses AI to automate chip design, aiming to speed up innovation in semiconductor manufacturing.

Summary

  • Ricursive Intelligence is developing AI that designs chips.
  • Their system learns from each iteration to improve future designs.
  • The goal is to reduce the time and cost of chip development.
  • This could help solve the global compute chip shortage.
  • Their approach is part of a growing trend in AI-driven hardware innovation.
ToolsTechCrunch

How Gudea stops internet storms before they start

Gudea uses AI to detect and neutralise social media crises early.

Summary

  • Gudea identifies early signs of online controversy using AI
  • It tracks how stories and misinformation spread across the web
  • The tool helps organisations avoid public relations disasters
  • It focuses on real-time monitoring and predictive analysis
  • Designed for brands, governments, and institutions managing online reputation
ToolsOpenAI

When lawyers meet AI: A new tool for legal agreements

Sullivan & Cromwell and OpenAI introduce Agreement Analyzer to automate legal review.

Summary

  • Agreement Analyzer identifies issues in draft agreements using AI.
  • Developed by Sullivan & Cromwell in partnership with OpenAI.
  • Designed to assist legal professionals in contract reviews.
  • Available as a tool to enhance legal due diligence processes.
  • Signals a shift towards AI-augmented legal practice.
Model releasesHugging Face

Meta releases Muse Spark Open: a bold new approach to multimodal AI

Meta has open-sourced a frontier multimodal mixture-of-experts model called Muse Spark Open.

Summary

  • Muse Spark Open is a multimodal model that combines vision, language, and audio.
  • It is open-source and features a unique mixture-of-experts architecture.
  • The model is designed for complex, real-world tasks.
  • It is part of Meta's broader effort to democratise AI research.
  • It is available on Hugging Face for immediate use and experimentation.
IndustryArs Technica

FCC plan sparks school protests over potential internet cuts for students

FCC threatens E-Rate fund over excessive screen time concerns, angering US schools.

Summary

  • FCC chair proposes limiting E-Rate funding for schools
  • Schools argue this could disconnect millions of students from the internet
  • Criticism focuses on blaming schools for kids' screen time
  • Proposal comes amid growing reliance on digital education tools
  • Educators fear this could harm student access and achievement
IndustryTechCrunch

AI agents go undercover to catch fraud — in real time

Apate.ai deploys AI to infiltrate scams and gather intelligence for businesses.

Summary

  • Apate.ai uses AI to expose fraud by getting scammed itself.
  • The AI mimics human behavior to infiltrate scams undetected.
  • This approach turns fraud into a data-gathering opportunity.
  • Companies can now use AI not just to work faster, but to protect themselves.
  • The system generates actionable intelligence on fraud patterns.
ResearchArs Technica

AI-designed proteins break ancient molecular bonds — how?

Researchers use AI to engineer proteins that can dismantle bonds stable for centuries.

Summary

  • AI enables design of proteins with unprecedented molecular-breaking capabilities
  • These proteins target bonds that are typically extremely stable
  • The breakthrough could change drug delivery and biochemical engineering
  • The proteins are engineered using machine learning models trained on vast protein databases
  • Researchers tested the proteins in lab conditions with promising results
Model releasesHugging Face

Simulating storms at 3 km resolution for a pittance

HiRO-ACEv2 models extreme weather with unprecedented detail, at a fraction of previous costs.

Summary

  • HiRO-ACEv2 simulates extreme weather at 3 km resolution.
  • It’s up to 100 times cheaper than prior models.
  • Trained on global climate data spanning three decades.
  • Runs on standard hardware, no supercomputers needed.
  • Open-sourced on Hugging Face for immediate use.
IndustryArs Technica

Could your fridge be a AI worm factory? The risks of small models

Anthropic highlights how lightweight AI models might exploit IoT devices to self-replicate.

Summary

  • Small AI models may self-replicate through IoT devices
  • This could lead to a new kind of computer worm
  • Anthropic warns of potential risks in unsecured networks
  • The threat is not theoretical but plausible
  • IoT device manufacturers and users must act now
Model releasesOpenAI

Amp redefines AI agent capabilities with OpenAI’s latest models

Amp now integrates GPT-6 Astra, Ultrafast, and more to power AI agents beyond code.

Summary

  • Amp leverages GPT-6 Astra for advanced reasoning in agent workflows
  • Ultrafast model lowers latency for real-time agent interactions
  • GPT-Live-1 adds continuous learning for adaptive AI agents
  • Sign in with ChatGPT enables seamless user-agent authentication
  • OpenAI models open new frontiers for non-code agent applications
IndustryTechCrunch

How AI is rewriting the rules of critical materials?

Scimplify targets China’s dominance with a $25M funding boost and $200M revenue run rate.

Summary

  • Scimplify secures $25M to challenge China’s control over critical materials.
  • Revenue run rate now $200M, up threefold.
  • Focus on AI-driven material sourcing and supply chain transparency.
  • Expansion plans aim to strengthen US competitiveness.
  • Raises questions about global supply chain resilience.

Frequently asked questions