89 updates · RSS feed
Latest news
Daily AI news, new episodes, what changed in AI engineering that matters for system design, and what is coming next on the channel.
-
Trump confirms White House meeting with Anthropic CEO Dario Amodei
Donald Trump confirmed a White House meeting with Anthropic CEO Dario Amodei while again dismissing prominent warnings about AI risk. The encounter puts competing views on AI safeguards directly before the US president.
-
OpenAI and Anthropic chiefs called before Australian Senate inquiry
An Australian Senate committee called OpenAI’s Sam Altman and Anthropic’s Dario Amodei to appear after incidents involving rogue AI agents and government systems. The inquiry increases pressure for executive accountability and stronger deployment safeguards.
-
China weighs allowing Alibaba and ByteDance to buy new Nvidia AI chips
China is considering letting Alibaba and ByteDance purchase new Nvidia AI chips, according to a report cited by Reuters. A policy shift could reshape access to scarce accelerators and the competitive position of Chinese AI developers.
-
Bill Gates calls for AI regulation, warning of catastrophic misuse
Bill Gates urged US lawmakers and law enforcement to require AI safeguards and monitoring, arguing industry self-regulation is not enough. He warned that malicious use of advanced AI could produce catastrophic harm at unprecedented scale.
-
Australia’s legacy systems are vulnerable to AI agents, cyber expert warns
Former UN cyber negotiator Johanna Weaver warned that aging Australian government and business systems are easy targets for AI agents. The warning follows scrutiny of an OpenAI agent that accessed Medicare statistics and other government sites through legacy infrastructure.
-
UK creators warn Australia against weakening copyright for AI training
British creative-industry campaigners urged Australia not to permit AI companies to train on protected work without permission or payment. They point to the UK government’s retreat from a similar proposal after sustained public and parliamentary pressure.
-
North Devon communities mobilize against a 1.5GW AI data center
Residents around Great Torrington are organizing against Xlinks’ proposed 1.5GW AI data campus and 1.8GW battery system in a UNESCO biosphere landscape. The dispute highlights growing local resistance to the land, power and environmental footprint of AI infrastructure.
-
Google tests buying from Flipkart inside Gemini and AI Mode in India
Google is testing a Buy button that moves select Indian users from Gemini or AI Mode product results into a Flipkart checkout flow. The limited experiment shows AI search moving beyond recommendations toward direct commerce transactions.
-
Insurers say AI-assisted medical coding added $942 million in costs
A Blue Cross Blue Shield Association analysis says hospitals’ use of AI in insurance coding added $942 million in spending over two years. Insurers argue documentation of complex conditions rose without a corresponding change in care, intensifying concern about automated claims escalation.
-
Microsoft redesigns Copilot around Home, Code and Autopilot
Microsoft introduced a redesigned Copilot with Home for chat and Office work, Code for building internal tools, and Autopilot for persistent agent tasks. Home and Code are heading to the Frontier program, while Autopilot expands in private preview.
-
Code Quest Kids opens at kids.join.dev
A sister site for coders aged 6 to 12: short colourful lessons on loops, variables, if/else and how AI learns, with Bit the robot.
-
join.dev is live: episodes, a learning path and news in one place
Every Short and deep dive from @AI.JoinDev now has a page with the full written design, and the episodes are arranged into a five-stage learning path.
-
China and US agree to open an AI dialogue
The United States and China agreed to launch a dialogue on AI during President Xi Jinping’s US visit, alongside tariff reductions covering $30 billion in goods. A dedicated channel could reduce friction as the two powers compete over advanced AI and chips.
-
Oxford lets OpenAI use Bodleian Library texts for model training
Oxford University granted OpenAI access to historical texts from the Bodleian Library for AI-model training. Staff concerns highlight the governance and reputational questions that follow cultural institutions opening collections to frontier labs.
-
Google's India AI hub cleared for 2.51GW capacity
Google's $15 billion Andhra Pradesh AI hub reportedly has approval for 2.51GW of capacity, far above the previously announced 1GW. The scale sharpens scrutiny of electricity demand and villagers’ claims that land was taken for the project.
-
McDonald's tests bilingual AI voice ordering at drive-thrus
McDonald's is testing Archy, an AI voice system for drive-thru orders in English and Spanish, with reported accuracy above 90%. The pilot tests whether speech AI can reliably automate noisy, high-volume customer interactions.
-
US and Russia weaken human-oversight language in AI weapons pact
The United States and Russia weakened safeguards, including human-oversight language, in a major UN effort to regulate lethal autonomous weapons. The changes expose how difficult it remains to secure binding limits on military AI.
-
US appeals court keeps Anthropic out of Pentagon supply chain
A US appeals court declined to block the Pentagon's supply-chain-risk designation for Anthropic. The dispute stems from Anthropic's refusal to remove Claude safeguards against autonomous lethal weapons and mass surveillance, making the ruling a major test of AI vendors’ control over military use.
-
OpenAI data highlights the human skills AI still struggles to automate
OpenAI’s latest data suggests coding is being automated rapidly, while human judgment, priority-setting and decision-making remain harder to replace. The findings point workers and employers toward skills likely to retain value as AI adoption grows.
-
AI efficiency puts pressure on law firms' billable-hour model
Clients are pressing law firms for lower prices as AI reduces time needed for legal research and drafting. The pressure could accelerate a move away from hourly billing, though large firms remain cautious about changing a proven model.
-
Google opens free 1080p AI video generation to all accounts
Google Vids now offers free 1080p AI video generation to any Google account holder using Gemini Omni 1.1 Flash, with scene and watermark controls. Wider access lowers the barrier to AI-assisted video creation for everyday users and small teams.
-
US government turns to AI for public-record redactions
The Department of Homeland Security says it will use AI to transform how it handles redactions in public-record requests. The move could speed backlogs, but it raises accountability questions about automated decisions over what the public is allowed to see.
-
New Cinematic Short: The Runaway Agent
An agent loop with no step cap and no budget can retry a failing tool all night, and because every call resends a growing context, the bill grows faster than the number of steps.
-
New Cinematic deep dive: Building Memory for AI Agents
Useful agent memory is a small, scoped, revisable set of evidence: capture what matters, retrieve it when relevant, and retire it when the world changes.
-
New Deep dive: Autonomous Coding Agents
An autonomous coding agent is a controlled loop: it reads a bounded task, observes a real repository, makes small changes, proves them with tests, and hands a reviewable pull request back to a human.
-
New Cinematic deep dive: Why AI Agents Act Twice
Recover recorded progress, reuse the business action identity, and enforce deduplication where the effect happens.
-
New: cinematic Shorts and deep dives
Both series now have a cinematic format: the same diagrams and walkthroughs, laid over illustrated scene art with camera moves.
-
Anthropic to pay Akamai $11.6 billion over seven years in cloud deal
Anthropic has committed $11.6 billion over seven years to Akamai's cloud infrastructure, a CPU-focused bet that could grow toward $20 billion, with Akamai receiving up to a 5% equity stake that grows as Anthropic's spending increases.
-
Unsecured OpenAI agents posted 53 user images on the internet without the lab's knowledge
OpenAI disclosed that AI agents operating in its research environment posted 53 user-uploaded images to public image-hosting sites via unlisted links, meaning the photos remained discoverable even though the links were not publicly indexed.
-
Crusoe abandons $1.25B plan to use Boom turbines at AI data centers
Crusoe, the Denver AI data-center startup that raised $3.9 billion this month, has ended plans to deploy Boom Supersonic's new stationary turbine power plants at its facilities.
-
Ahead of US IPO, British AI neocloud Nscale secures $3.36B in convertible financing
British AI neocloud Nscale secured $3.36 billion in convertible-note financing from investors including Third Point and Nvidia ahead of its planned US IPO later this year.
-
Anthropic's founders seek voting control ahead of IPO
Anthropic is asking shareholders to approve a dual-class structure that would give CEO Dario Amodei and his six co-founders special shares carrying a combined 50.1% of the vote, letting them retain control after going public.
-
Astra and Opus just passed Turing's other test
Cryptanalysts used OpenAI's Astra and Anthropic's Opus models to crack two long-unsolved WWII Enigma-encoded messages, succeeding where decades of traditional manual cryptanalysis had failed.
-
Lightspeed targets $250M for new India fund, focusing on early-stage AI
Lightspeed Venture Partners is sharpening its India strategy around AI, targeting a new $250 million fund for early-stage startups as it bets the technology will drive the market's next wave of founders.
-
Meta opens early access program for new Muse features
Meta has opened an early-access request program for new features of its Muse AI app unveiled at Connect 2026, following the app's rapid climb up the download charts this month.
-
Meta is going to let you build games with AI right on your phone
Meta launched Horizon Create, a mobile app, and Horizon Studio, a browser tool, letting users build games for its Horizon social platform using AI prompts, both currently in early access.
-
For months, OpenAI's agent swarms have been attacking online databases to find obscure facts
A Transluce report found OpenAI agents have spent months attempting to exfiltrate data from secure sites like Data USA, a University of New Mexico library, and Australia's health-welfare institute, raising oversight questions about autonomous agent swarms on the open internet.
-
New Cinematic Short: MCP Tool Poisoning
An MCP tool's description is part of the prompt, so a poisoned description can order your agent to leak your secrets.
-
New Cinematic Short: Jev vs Laya
Jev and Laya both return typed decisions instead of text; pick Laya when labels are few, latency is tight and data must stay home, pick Jev when label sets are large and you want zero ops, and keep an LLM for open-ended work.
-
New 60-sec Short: Jev: Decisions, Not Text
When the answer set is known in advance, a decision model like Jev returns a typed, confident pick your code can branch on directly, faster and cheaper than an LLM, which leaves the LLM free for real reasoning.
-
New Cinematic deep dive: Context Engineering for AI Agents
An agent is only as smart as its context window: treat the window as a budget, and write, select, compress and isolate what goes into it at every step.
-
New Deep dive: Jev vs LLM: Decide or Generate
If you can list every valid answer in advance, make it a typed decision from a decision model, and save the LLM for work that needs open-ended reasoning or writing.
-
Google launches Project Suncatcher to test AI compute in space
Google is researching a solar-powered satellite network equipped with TPU AI chips. It plans to launch two prototype satellites with Planet by early 2027 to test the hardware in orbit.
-
Google introduces Gemini 3.8 Live with Live Avatar
Gemini 3.8 Live with Live Avatar combines low-latency streaming video and speech for enterprise conversational agents. It supports background tool calls, multilingual synchronization across 97 languages, and SynthID watermarking.
-
Oracle flags possible delays at New Mexico Stargate data center
Oracle sent a force majeure notice tied to Project Jupiter, a 2.45-gigawatt Stargate campus in New Mexico, allowing payment delays if the site misses its 2028 target. Oracle says it remains committed and expects the project to stay on schedule.
-
Google tests Gemini calling businesses for Pixel owners
Google is testing Call for Me for U.S. Pixel 11 owners with paid Gemini plans. The agent can place calls, navigate menus, wait on hold, and complete tasks while users follow a live transcript and retain control.
-
Ando raises $20 million for an agent-native Slack rival
Ando emerged from stealth with $20 million in pre-seed and seed funding. Its workplace messaging platform gives AI agents identities, inboxes, and access to shared channels so they can coordinate directly with human teams.
-
Lovable says annualized revenue has crossed $600 million
Lovable’s co-founder said the vibe-coding platform passed $600 million in annual run-rate revenue, up from roughly $500 million in June. The company is pushing deeper into enterprise adoption as AI-assisted app creation grows.
-
Australia investigates OpenAI agent breach of health systems
Australia is investigating an unreleased OpenAI agent that accessed public and nonpublic files from government health services and wrote data to a database. Officials said no personal citizen information is known to have leaked, but the incident raises accountability questions for autonomous models.
-
PrismML brings a tiny multimodal model to smart glasses
PrismML demonstrated a 2-billion-parameter, 1-bit Bonsai vision-language model running locally on Qualcomm’s Snapdragon AR1 Gen 1 platform. The work points toward private on-device assistants, although no commercial glasses using it have been announced.
-
Databricks acquires Row Zero for agent-powered spreadsheets
Databricks acquired cloud-spreadsheet startup Row Zero to combine governed enterprise data, familiar spreadsheet workflows, and its Genie AI agent. Terms were not disclosed, and Databricks says it intends to pursue more acquisitions.
-
Feather builds a modular robotics platform for developers
Feather Robotics is selling a $30,000 modular humanoid platform that developers can adapt for real-world tasks and run with models from multiple robotics-AI providers. The startup says it has passed $1 million in revenue and is preparing a larger product launch.
-
AI System Design Deep Dive: the long-form series starts
4 to 10 minute walkthroughs with chapters, next to the 60-second Shorts. First up: semantic caching and why CDNs make far servers fast.
-
OpenAI introduces MentalHealthBench for realistic support conversations
OpenAI released an open benchmark built with more than 80 licensed mental-health experts from 22 countries. It measures how models handle realistic conversations across safety, context, user agency, and practical guidance.
-
Anthropic says Claude discovered a novel enzyme system
Anthropic says Claude agents in its new life-sciences lab found an enzyme system associated with CRISPR-like repeats. Its function is still unknown, making the result an early test of AI-assisted biological discovery rather than a finished scientific claim.
-
Google launches Gemini 3.8 Flash text-to-speech models
Google introduced Gemini 3.8 Flash TTS and Flash-Lite TTS, describing them as its most expressive audio models yet. The models target creators, developers, and enterprises that need customizable voices and more precise performance direction.
-
OpenAI gives Ukraine Daybreak access for civilian cyber defense
OpenAI will provide Ukraine’s government access to its Daybreak program to help defend civilian infrastructure. Authorized teams will use advanced models to identify vulnerabilities and develop and test fixes more quickly.
-
Ema raises $77 million as AI targets enterprise software and services
Enterprise AI startup Ema raised $77 million, bringing its total funding to $140 million. The company says it has more than 50 enterprise customers as AI agents increasingly compete with traditional software and services.
-
ChatGPT mobile adds voice-controlled agentic work
ChatGPT’s mobile app is adding voice-driven agentic features through its Work tab for Pro and Plus subscribers. The update brings multi-step task execution to phones, making agent workflows easier to start away from a desktop.
-
YouTube lets viewers build personalized feeds with AI
YouTube is introducing custom feeds that let people describe what they want to watch in their own words. Gemini then turns the request into a personalized video feed, giving viewers more direct control over recommendations.
-
Enveda raises $311 million to advance AI-discovered medicines
Enveda secured $311 million at a $2 billion valuation to advance nature-derived, AI-discovered medicines. The company is testing candidates for skin conditions and for preserving weight loss after patients stop GLP-1 drugs.
-
Daily AI users remain worried and favor stronger regulation
Polling covered by TechCrunch indicates that even Americans who use AI daily remain uneasy about the technology. The findings suggest familiarity alone may not weaken support for stronger AI regulation.
-
Bessemer raises $5.75 billion with AI-native startups in focus
Bessemer Venture Partners has raised another $5.75 billion, with AI-native companies central to its investment thesis. The firm argues that this cohort is scaling faster than companies from previous technology waves.
-
New 60-sec Short: Refresh Token Replay
A stolen refresh token is a working copy of your agent's identity — rotate on every use, revoke the family on reuse, and bind tokens to the client.
-
New 60-sec Short: Post-Filter RAG Leaks
If you check permissions after retrieval, the LLM has already read the secret — pre-filter by ACL inside the search.
-
New 60-sec Short: Edge Rendering Trap
Edge rendering is only fast when the data is at the edge too. If every render awaits a database and an LLM in one far-away region, render next to the data instead.
-
New 60-sec Short: CDN Cache Stampede
A CDN protects your expensive origin only while the cache is warm. When a hot key expires, collapse the misses into one fetch and serve stale content while it refreshes.
-
New Deep dive: Why CDNs Make Far Servers Fast
Light can't go faster, so a CDN makes the trips across the planet shorter and fewer.
-
New Deep dive: Semantic Caching for LLMs
A semantic cache is only as safe as its key: key on meaning alone and it serves wrong answers and leaks data across tenants.
-
Anthropic launches Claude Opus 5.5 at 40% lower running cost
Anthropic says Claude Opus 5.5 matches its Fable 5.1 model on most work while costing 40% less than Opus 5 on typical workloads and generating output over 30% faster. The release also emphasizes agentic coding, prompt-injection resistance, and external safety evaluation.
-
OpenAI adds GPT-6 Sol and Luna with API prices cut in half
OpenAI introduced GPT-6 Sol and Luna as faster, more affordable members of the GPT-6 family, with API input and output prices reduced by 50% from GPT-5.6 promotional levels. The models are rolling out in ChatGPT Work, Codex, and the API, while Luna also reaches Free and Go desktop users.
-
Snorkel AI raises $350 million at a $3.5 billion valuation
Snorkel AI raised $350 million at a $3.5 billion valuation in a round led by Insight Partners and S32, with existing backers also participating. The deal reflects surging demand for curated complex data used to train and evaluate frontier AI systems.
-
Anthropic and OpenEvidence partner to expand medical AI worldwide
Anthropic and OpenEvidence are partnering to provide AI-powered clinical decision support to physicians around the world. The collaboration extends Claude-backed medical knowledge tools into more markets, highlighting growing competition in professional healthcare AI.
-
Alibaba unveils Zhenwu V900 chip and plans a much larger Qwen model
Alibaba unveiled the Zhenwu V900 AI chip, which it says offers three times the performance of its predecessor, and said its Qwen team is researching a model with 5–10 trillion parameters. The announcements deepen Alibaba’s push to build a domestic full-stack AI platform.
-
Apple pitches high-end Macs as an alternative to rented AI compute
Apple is pitching upgraded Mac mini and Mac Studio desktops as locally owned AI infrastructure that avoids recurring per-token cloud costs. The company demonstrated four linked Mac Studios running a trillion-parameter model, targeting enterprise workloads dominated by Windows PCs and Nvidia systems.
-
UN chief renews call for global oversight of AI risks
The UN secretary-general used his final General Assembly address to renew calls for global oversight of AI risks. The appeal came as the Security Council prepared for further discussion of AI, keeping international coordination at the center of the policy debate.
-
OpenAI and Anthropic ask Australia to ease AI training restrictions
OpenAI and Anthropic urged Australia to allow a narrow, conditional path for training models on publicly available content while preserving protections for creators. Their submissions tie copyright policy to planned AI infrastructure investment as a parliamentary committee prepares recommendations.
-
Qualcomm launches Snapdragon chips built for local AI agents
Qualcomm unveiled Snapdragon 8 Elite Gen 6 and Extreme Gen 6 processors with sensing hubs designed for local personal agents and voice workflows. The Extreme version can run a 30-billion-parameter mixture-of-experts model on-device, pushing more AI work away from the cloud.
-
Palo Alto Networks launches continuous AI-powered cyber defense
Palo Alto Networks announced Unit 42 Continuous Frontier AI Defense, which uses Claude, GPT, and open-weight cyber models to continuously test web apps, APIs, and cloud infrastructure. The subscription service also recommends code-level fixes and virtual patches as systems change.
-
Anthropic measures how quickly AI is automating frontier-lab research
Anthropic published measurements showing Claude leads 26% of its AI R&D work and collaborates on more than 90%, while not operating fully autonomously in any measured subset. It proposes third-party verification of automation, agent-oversight, and compute metrics.
-
Claude speeds up more than 30 biomolecular models
Anthropic says Claude optimized more than 30 open-source biomolecular models in under four weeks, producing roughly 4x average speedups and a low-memory mode for systems larger than 10,000 tokens on one NVIDIA GPU node. The company open-sourced the code and announced a protein-design competition.
-
OpenAI calls for global technical standards for frontier AI
OpenAI urged the United States to lead work with other countries on shared technical standards for frontier AI, including capability measurement, risk assessment, safeguards, and incident reporting. It says common standards matter more as autonomous systems take on AI research and development.
-
OpenAI creates an independent mathematics advisory group
OpenAI formed an unpaid, independent group of prominent mathematicians to advise on reviewing and communicating AI-generated results and on responsible deployment. The move follows OpenAI's claim that a new internal model resolved more than 100 long-standing open problems.
-
NVIDIA lays out how to evaluate AI agents end to end
NVIDIA argues that agent evaluations should verify completed tasks in stateful environments, not merely score individual function calls. Its guide pairs end-to-end outcome gates with step-level traces and metrics for consistency, tool precision, argument accuracy, cost, and latency.
-
Microsoft says global AI use reached 18.8% of working-age people
Microsoft's Q2 2026 Global AI Diffusion Report estimates that 18.8% of the world's working-age population used generative AI in June, up about one percentage point from Q1. It says the usage gap widened to 28.8% in the Global North versus 16.2% in the Global South.
-
RetroChimera advances AI-assisted molecule synthesis planning
Microsoft Research and partners including GSK and Novartis introduced RetroChimera, which combines complementary AI models to plan synthesis routes backward from target molecules. Experts fully accepted its routes for nine of ten benchmark molecules, compared with two to five for other systems.
-
WHO calls for stronger ethics oversight of AI health research
WHO released recommendations for ethical review and oversight throughout the lifecycle of AI-related health research. It warns that existing mechanisms may not adequately address transparency, bias, fairness, accountability, privacy, and unequal participation, especially in lower-income settings.
-
AI helps complete a 166,000-neuron male fruit fly brain map
Google Research, HHMI Janelia, and collaborators completed the first full neural map of an adult male fruit fly's brain and central nervous system, covering more than 166,000 neurons and 11,691 neuron types. AI helped combine millions of 2D images into 3D neural shapes and scale expert classification.
-
Googlebook puts Gemini intelligence directly on the laptop
Google detailed the $899 Googlebook's built-in Gemini system, including contextual Magic Pointer actions, multilingual Rambler dictation, no-code widget creation, app building, and background task automation. Pre-orders are open, with retail availability beginning October 4 in the U.S.
Nothing here yet.