Actualités IA Quotidiennes

Google released a new Android command-line tool today that is built to work with AI coding agents. It connects directly with Claude Code and OpenAI's Codex, letting developers — or their AI agents — build, test, and deploy Android apps without touching a graphical interface.

The tool gives AI agents the ability to create real native apps, not just generate code that needs to be manually assembled. An agent can run commands, read error logs, iterate on builds, and push updates from start to finish.

This is a deliberate move by Google to bring AI-first development to Android. Rather than bolting AI onto an existing workflow, Google is redesigning the Android development experience from the ground up around agents that do the building.

Spellar 3.0 launched today with a feature that sets it apart from standard meeting tools: cross-meeting memory. Most AI meeting tools treat every call in isolation — they transcribe, summarize, and forget. Spellar 3.0 links your meetings together, remembering commitments, open questions, and recurring topics across sessions.

If someone promised to send a document three meetings ago and never did, Spellar knows. If a topic keeps coming up without resolution, it surfaces that pattern. The assistant becomes more useful the longer you use it, because it builds context over time.

It is a practical answer to a real problem that anyone who runs regular meetings will recognize immediately.

Andrej Karpathy announced today that he has joined Anthropic. He will work on the pre-training team — the group responsible for the large, expensive training runs that give Claude its core knowledge.

Karpathy is one of the most celebrated researchers in AI. He helped build the original GPT models at OpenAI and later spent time at Tesla leading their Autopilot team before returning to independent work. His move to Anthropic is not a quiet hire — it is a statement. Pre-training is where the fundamental capability of a model is set. Choosing to work on it at Anthropic, rather than OpenAI or any of the other large labs, is a clear signal of where he sees the most interesting problems.

The AI community is treating this as one of the most significant researcher moves in recent memory.

At Google I/O today, Google announced it is replacing the familiar list of links with an AI-powered experience that gives direct answers, holds conversations, and — most significantly — runs tasks on your behalf without you prompting each step.

Google is turning Search into a system that acts, not just responds. Users will be able to ask it to monitor a topic and send alerts, compare products, book a table, or research a question across multiple sources — and the system handles it from start to finish.

This is a fundamental shift in how hundreds of millions of people interact with the internet every day. The old model was: you search, Google shows links, you click. The new model is: you describe what you need, the agent figures it out.

Google launched Gemini 3.5 Flash at I/O today — and the company was explicit about what it is built for. This is not a better chatbot. It is a model designed to run complex tasks on its own, with minimal human direction along the way.

Gemini 3.5 Flash is positioned as Google's strongest coding and task-execution model yet. It can autonomously build software tools, execute multi-step workflows, and take actions across services. Google is calling this the start of the agentic Gemini era.

The name — Flash — signals something too. Flash models are fast and cheap to run, which matters for agents that might execute dozens of steps in a single job. Building the fastest, most capable agent model is very different from building the most impressive demo.

Andon Labs let AI systems run a live radio station with no humans involved — no producer, no host, no safety net. The experiment, called Andon FM, is the latest in their series of projects where AI agents run real businesses from start to finish and the team documents publicly what happens.

Previous experiments covered retail and food service. A radio station adds a new dimension: real-time content, live audience, and the need to make editorial decisions continuously. The AI handled music selection, talk segments, and pacing — and Andon Labs reported on where things went well and where they broke down.

The project is a serious research effort into autonomous AI operation, not just a stunt. The public reporting is the point: they want to show what actually happens when humans step out of the loop.

Simon Willison published a rapid-fire retrospective today covering six months of progress in large language models. The post cuts through the noise to document what actually changed: which capabilities made genuine jumps, what got dramatically faster or cheaper, and where researchers are still working through unsolved problems.

Willison is one of the most respected independent voices on LLM development — known for being direct about what works, what does not, and what is hype. The post is structured to be read in five minutes but rewards careful attention. It covers model releases, pricing shifts, capability benchmarks, and the debates that defined the last half-year.

The discussion it sparked on Hacker News today is one of the most active threads of the week, with practitioners sharing their own assessments of which changes actually mattered in production.

AgentMemory is an open-source persistent memory system for AI coding agents, built by developer Rohit Ghumare. The project claims the top rank on real-world benchmarks for AI agent memory — meaning it was tested against actual coding workflows, not synthetic tests.

The problem it solves is straightforward: AI coding agents forget everything the moment a session ends. AgentMemory gives them a place to store what they have learned — decisions made, patterns identified, context built up over time — and retrieve it in future sessions.

It supports the most common AI coding tools and integrates via MCP, the protocol that has become a standard for connecting agents to external tools and data sources.

OpenHuman launched today as an open-source AI harness designed around the human experience. While most AI frameworks are built for developer performance — token efficiency, speed, output quality — OpenHuman takes a different starting point: how does the person in the loop actually feel using this?

The project is community-driven and free, built to make AI interactions feel more natural and less like operating a technical tool. The focus is on transparency — the user should always understand what the AI is doing and why.

It landed on Product Hunt today with strong support from people who have grown frustrated with AI tools that feel powerful but opaque.