The AI intelligence layer
AI is moving fast.Stay ahead.
Daily AI news, clear explainers, and a living record of the models and milestones that matter, for beginners and experts alike.
Real developments.
Deeper perspective.
Brighter possibilities.
Model release
Robotics
Autonomous systems
Research
AI NEWS
07:56Update Office now: an Outlook flaw found with AI help is already being used in attacks07:56Attackers are breaking into servers that run AI models through a flaw in GPUStack07:56Delivery drones cleared to fly over busier UK neighbourhoods after a higher-risk safety approval07:56Tens of thousands of apps built with Grok could be opened by anyone, and some held private data, a security firm found07:56AI agents produced a machine-checked proof that Signal’s core encryption code keeps its security promises07:55Some Notion AI chats were exposed in a breach at an AI provider, so pasted passwords should be changed08:00US regulator asks how to judge medical robots that act on their own or are run from afar08:00Teaching a reasoning model where it overthinks cut its maths answers by over 40%07:56Update Office now: an Outlook flaw found with AI help is already being used in attacks07:56Attackers are breaking into servers that run AI models through a flaw in GPUStack07:56Delivery drones cleared to fly over busier UK neighbourhoods after a higher-risk safety approval07:56Tens of thousands of apps built with Grok could be opened by anyone, and some held private data, a security firm found07:56AI agents produced a machine-checked proof that Signal’s core encryption code keeps its security promises07:55Some Notion AI chats were exposed in a breach at an AI provider, so pasted passwords should be changed08:00US regulator asks how to judge medical robots that act on their own or are run from afar08:00Teaching a reasoning model where it overthinks cut its maths answers by over 40%
Latest in the Loop
Briefs, model releases, chip announcements and timeline entries, newest first.
BriefUpdate Office now: an Outlook flaw found with AI help is already being used in attacksMicrosoft has fixed an Outlook flaw that attackers are already using, reported by a researcher working with AI. Update Office…BriefAttackers are breaking into servers that run AI models through a flaw in GPUStackA critical flaw in GPUStack, software that runs AI models on GPU servers, is being used in attacks. Upgrade to 2.2.2 or later.BriefDelivery drones cleared to fly over busier UK neighbourhoods after a higher-risk safety approvalThe UK aviation regulator has cleared Amazon’s delivery drones to fly over denser parts of Darlington, widening eligible cust…BriefTens of thousands of apps built with Grok could be opened by anyone, and some held private data, a security firm foundSecurity firm Claroc found 35,578 published Grok Build apps that opened without sign-in, and estimates 650 to 800 of them hel…BriefAI agents produced a machine-checked proof that Signal’s core encryption code keeps its security promisesGoogle DeepMind says AI agents wrote millions of lines of checked proof showing Signal’s one-to-one encryption code meets key…BriefSome Notion AI chats were exposed in a breach at an AI provider, so pasted passwords should be changedHackers reached customer data at AI company Fireworks, including some Notion AI chat transcripts, unfinished model weights an…BriefUS regulator asks how to judge medical robots that act on their own or are run from afarNo fully autonomous or remotely operated medical robot is cleared in the US yet. The FDA is asking how it should weigh their…BriefTeaching a reasoning model where it overthinks cut its maths answers by over 40%A new training method cut a reasoning model’s maths answers by over 40% while accuracy held or rose, by teaching it where it…
Explainers
Evergreen explainers, maintained rather than published once.
New explainerAI watermarking explained in 2 minutesAn AI watermark is a hidden signal woven into generated text, images, audio or video so software can later tell that AI made it.Reviewed November 2026ExplainerAI agent monitoring explained in 2 minutesAI labs now watch what their agents do, with layers of automated checks and human review, much like a bank watches for fraud.Reviewed November 2026ExplainerEmergent misalignment explained in 2 minutesEmergent misalignment is when training an AI on one narrow bad habit makes it behave badly in many unrelated ways.Reviewed November 2026ExplainerCross-embodiment explained in 2 minutesCross-embodiment means one robot brain learning from, and controlling, many different robot bodies, instead of being trained for just one.Reviewed November 2026ExplainerHow Much Energy Does AI Use? Data Centres, Power and Cooling ExplainedOne AI chat uses a sip of electricity, but billions of them, plus training, add up to a power demand that is reshaping grids.Reviewed November 2026ExplainerWhat Is a World Model? How AI Learns to Imagine What Happens NextA world model is an AI system that predicts what will happen next in a scene, so robots and agents can practise and plan inside their own imagination.Reviewed November 2026ExplainerGPU, TPU or NPU? How AI Chips Differ, and What TOPS and FP4 MeanWhy AI runs on several kinds of chip, from data-centre GPUs to custom accelerators and the NPU in your laptop, and how to read numbers like TOPS an…Reviewed October 2026ExplainerHow Does Voice AI Work? Speech Recognition, Synthetic Voices and Voice Cloning ExplainedHow AI hears you, works out a reply and speaks back, how a few seconds of audio can copy a voice, and why a familiar voice is no longer proof of wh…Reviewed October 2026ExplainerHow AI Agents Work in Practice: Tools, Loops and Where They Go WrongExplains how AI agents turn a chatbot into something that takes actions, how the plan, act and check loop works, and why banks and regulators are d…Reviewed November 2026ExplainerHow Are AI Models Tested? Benchmarks, Evals and Red Teams ExplainedExplains how AI models are tested, from exam-style benchmarks to red teams and outside evaluators, and why a single score rarely tells the whole st…Reviewed November 2026ExplainerProbes explained in 2 minutesA probe is a small classifier that reads an AI model’s internal signals, rather than its words, to detect things like deception.Reviewed October 2026ExplainerWorld-action models explained in 2 minutesA world-action model predicts both what a robot should do next and what the world will look like after it does it.Reviewed October 2026ExplainerTotal vs active parameters explained in 2 minutesIn a mixture-of-experts model, only a small share of the parameters does the work on each word, which is why two numbers appear in the name.Reviewed October 2026ExplainerNVFP4 explained in 2 minutesNVFP4 is a way of storing an AI model’s numbers in just 4 bits each, so big models need far less memory to run.Reviewed October 2026ExplainerOpen Weights vs Closed Models: What “Open” Actually Means in AIAn open-weights model hands you the trained file to run yourself; a closed model only lets you use it through someone else’s service, and the licen…Reviewed October 2026ExplainerHow Do Robots Learn? Teleoperation, Simulation and Vision-Language-Action ModelsRobots learn skills from three sources, people steering them, practice in simulation and models that turn words and images into movement, and each…Reviewed October 2026ExplainerHow Self-Driving Actually Works: SAE Levels 0–5 and the State of PlayThe SAE levels are the only precise vocabulary for autonomous driving, and almost every consumer claim abuses them. Here is the framework and what…Reviewed October 2026ExplainerThe AI Compute Stack: Why Chips Decide the RaceAI capability is downstream of hardware. Understanding the compute stack (accelerators, memory, packaging, power) explains most of the industry’s s…Reviewed October 2026ExplainerAI Benchmarks Explained: What the Scores Actually MeasureBenchmark numbers drive coverage, procurement and valuations. Most of the widely-quoted ones are saturated, contaminated, or not comparable between…Reviewed October 2026ExplainerTool Use and MCP: How AI Connects to the Real WorldTool use is how a model stops merely talking and starts doing. MCP is the standard that stopped every integration from being bespoke.Reviewed October 2026ExplainerWhy AI Models HallucinateHallucination is not a bug in language models. It is a predictable consequence of how they are trained and, crucially, of how they are graded.Reviewed October 2026ExplainerFine-Tuning vs RAG vs Prompting: Choosing the Right ApproachThree ways to make a model do what you want, routinely confused with each other. The deciding question is whether your problem is knowledge, behavi…Reviewed October 2026ExplainerReasoning Models Explained: Chain-of-Thought and Test-Time ComputeReasoning models spend extra computation thinking before answering. The gains on hard problems are real, the costs are substantial, and the tradeof…Reviewed October 2026ExplainerRAG Explained: Giving AI Access to Knowledge It Was Never Trained OnRetrieval-augmented generation looks up relevant information and hands it to the model at question time. Large context windows were supposed to kil…Reviewed October 2026ExplainerAI Agents Explained: What They Are and What They Can Actually DoAn agent is a model given tools and a goal, running in a loop until the job is done. The honest picture of where they work and where they fail is n…Reviewed October 2026ExplainerContext Windows Explained: Why AI Forgets, and What ChangedA context window is the model’s entire working memory. Understanding its limits explains most practical AI failures, and why a bigger window is not…Reviewed October 2026ExplainerTokens and Embeddings: How Machines Represent MeaningModels do not read words or understand meaning directly. They read tokens and operate on vectors. This distinction explains a whole class of otherw…Reviewed October 2026ExplainerTransformers Explained: The Architecture Behind Modern AIAttention lets a model weigh every word against every other word at once. That one idea is why modern AI scaled when earlier architectures could not.Reviewed October 2026ExplainerWhat Is a Large Language Model? How LLMs Actually WorkA large language model learns to predict the next piece of text. That single mechanism explains both what these systems can do and how they charact…Reviewed October 2026
AI Robotics
Humanoids, robotaxis, warehouse fleets and the AI that drives them. News, key players and plain-word explainers in one place.
-
Driverless cars still passing stopped school buses prompt a US review of Waymo’s fix
-
Blurry camera feeds led AI models running robot arms to attempt unsafe tasks 88% of the time
-
Humanoid robots now clean homes for pay in San Francisco, with remote pilots doing much of the work
-
Robots can learn two-handed tasks from people’s handheld recordings alone, a Japanese group’s open model suggests
The AI Timeline
Key moments. A clearer picture.
Nov 5
2026 Research
2026 Research
Research
METR finds AI speeding up Anthropic’s model progress
METR's high-access review estimates Anthropic's capability growth now runs about twice as fast, mostly because of AI.
Oct 29
2026 Infrastructure
2026 Infrastructure
Infrastructure
Canada announces up to C$1.6bn equity stake in Cohere
Prime Minister Mark Carney announced a government stake in Cohere, tied to Canadian jobs, compute and headquarters.
Oct 15
2026 Applications
2026 Applications
Applications
Microsoft launches Copilot for Legal
Microsoft added more than 20 legal skills to Word, along with plugins from legal technology partners including Harvey and Legora.
Oct 5
2026 AI-powered attacks
2026 AI-powered attacks
AI-powered attacks
OpenAI begins watermarking its models’ text output
OpenAI opened invisible text watermarking to API customers and published its detection rates, with an EU rollout to follow.
Oct 1
2026 Policy
2026 Policy
Policy
arXiv rate-limits authors over AI-written papers
arXiv capped authors at two submissions a month after September brought 40,363 papers.
Who’s shaping AI
Antoine MoyroudFrance · Investor
Part of Lightspeed's early investment in Mistral AI. Also backs CuspAI.
The signal, not the noise.
A curated briefing on the week's most important AI developments, delivered to your inbox.
AI developments,researched, sourced and reviewed before publication.
Once a weekno more
No spamjust signal
Why it matters
We track AI developments not just to report what's new, but to explain what it means for industries, economies, and the future.
Our editorial approach →AI Security
AI that finds bugs, AI that writes scams, and the AI systems that can be tricked. What changed, who is defending, and what to do, in plain words.
-
Update Office now: an Outlook flaw found with AI help is already being used in attacks
-
Attackers are breaking into servers that run AI models through a flaw in GPUStack
-
Tens of thousands of apps built with Grok could be opened by anyone, and some held private data, a security firm found
-
AI agents produced a machine-checked proof that Signal’s core encryption code keeps its security promises
AI by country
How each country is inventing, adopting, researching and regulating AI, with profiles reviewed monthly.
Updated 11 NovEurope
United Kingdom
Pro-growth and pro-adoption: no AI Act, regulation through existing sector regulators, and a safety effort focused on testing frontier models.
11 NovDelivery drones cleared to fly over busier UK neighbourhoods after a higher-risk safety approval9 NovA&O Shearman says 93% of its lawyers used its AI tools in September, as it commits $1.1bn over five years9 Nov86% of UK financial firms now use AI, but few let AI agents act alone, Bank of England and FCA survey finds
Americas
United States
Build fast, regulate lightly at federal level, and push one national rulebook that overrides…
Asia-Pacific
China
State-led: put AI into every sector, release open-weight models, build its own chips, and…
Asia-Pacific
South Korea
Pushing to become a top-three AI power, with state-backed sovereign models and compute alongside…
Europe
European Union
One binding rulebook, trimmed in 2026, plus heavy public spending on compute and on…
Europe
Germany
Industrial AI over consumer chatbots: build sovereign compute, and enforce the AI Act with…
12 country hubs
All countries
Canada, Japan, France, India and more, each with news, policy and research.
AI in Industry
How hospitals, banks, factories and shops are really using AI. Named deployments, measured results, reviewed monthly.
Updated 11 Nov
Logistics
Latest: Delivery drones cleared to fly over busier UK neighbourhoods after a higher-risk safety approval
Live
Healthcare
Strongest where AI reads images and writes paperwork. Still unproven where it makes the…
Live
Legal
AI clearly saves time on first drafts, document review and summarising, and firms such…
Live
Finance & banking
Fraud screening and staff copilots are routine. Letting AI make lending calls or move…
Live
Retail
Shopping assistants inside retailers' own apps are where AI shows clear gains. AI agents…
Live
Education
AI clearly saves teachers planning time, and well-designed tutors have produced real learning gains…
Live
Media & marketing
AI clearly pays off in ad targeting and in producing ad creative, where the…
Live
Manufacturing
Camera inspection and maintenance alerts pay their way on the factory floor. Showing how…
Live
Defence
AI adoption in defence is clearest in back-office work, logistics, maintenance and analysis tools.…