Actualités IA Quotidiennes

Amazon launched Alexa for Shopping today — an AI assistant built directly into the Amazon search bar. It works with your voice or touch, runs on Alexa+, and works across the Amazon mobile app, desktop, and Echo Show displays.

The assistant gives personalized product recommendations and can handle parts of the shopping experience automatically: searching across Amazon and other online stores, comparing options, and filtering results based on what you tell it.

This is Amazon taking the AI assistant race into its own backyard. Google Assistant and Apple Siri have been in your phone's search bar for years. Amazon's version lives where the buying happens — which gives it a clear purpose that general-purpose assistants have always struggled to nail.

WhatsApp is testing a new incognito mode for its Meta AI chat feature. When you switch it on, the conversation is not saved. Messages disappear by default once you close the chat — similar to disappearing messages, but for your AI conversations.

Meta built this after feedback that people wanted to ask the AI sensitive questions without leaving a permanent record. The concerns ranged from personal health questions to financial queries to things people just did not want stored.

It is one of the first major AI chat features designed around privacy as the default rather than an afterthought. The AI industry has largely built retention and memory into its products. WhatsApp is testing whether users will actually engage more with an AI that forgets.

Adaption launched AutoScientist today — a tool that lets AI models improve themselves by running their own training experiments automatically. Instead of a team of engineers designing each training update, AutoScientist runs tests, measures the results, and adjusts the model on its own.

The technology targets a specific and expensive bottleneck: getting a general AI model to work well on a narrow task usually takes weeks of specialized work. AutoScientist is designed to cut that down by automating the experimentation loop that humans normally run by hand.

This is generating attention in the AI research community today because it points toward a future where models do not just respond to users — they actively improve themselves between sessions.

A new report from Ramp — a company that tracks how businesses spend money on software — shows that 34.4% of its clients are now paying for Anthropic, compared to 32.3% paying for OpenAI. That makes Anthropic the top-paid AI provider among Ramp's customers today.

This is the first time any independent spending data has shown Anthropic ahead of OpenAI in business adoption. OpenAI invented the market. Anthropic is winning the bills.

The shift is being discussed across the AI industry today, with analysts pointing to Anthropic's push into legal, healthcare, and now small business as the driver. When a company that is years younger and less well-known than its rival starts taking the lead in paying customers, it changes what the next chapter of the AI market looks like.

Anthropic is making a direct play for small business owners — the 36 million companies that make up the backbone of the U.S. economy but have largely been ignored by AI labs focused on big enterprise deals.

The new offering is designed to be simple enough for a non-technical owner to set up and use. Anthropic is pricing and packaging it for people who run a restaurant, a law firm, or a local service company — not for a Fortune 500 IT department.

For investors and founders, this signals that the AI market is expanding downmarket in a serious way. The next wave of AI customers will not come from corporate procurement. They will come from millions of small businesses that want a simpler way to handle the admin work that eats their day.

Google and SpaceX are in early talks to build data centers in orbit. The pitch: space removes the problems of land, power, and cooling that are slowing down AI infrastructure on the ground. SpaceX would launch and operate the hardware; Google would use the compute.

The costs today are far higher than ground-based options. But both companies are treating this as a long-term bet — that demand for compute will outpace what can be built on Earth fast enough.

This is a generating real discussion today about where AI's physical infrastructure goes next. Most people think about AI in terms of software. The companies building it are increasingly thinking about it in terms of real estate, energy, and now, orbital mechanics.

Cactus, a research team, released Needle — a 26-million-parameter model trained specifically to call tools and run functions. It processes input at 6,000 tokens per second and generates output at 1,200 tokens per second on a regular consumer device. That is fast enough to use in real products, not just demos.

The model was distilled from Gemini, meaning the team took a much larger model's knowledge of tool-calling and compressed it into something tiny. The result is a model that does one thing — deciding when and how to call external tools — at a speed and size that can run locally.

This is generating significant buzz among developers building AI systems that chain tools together, because the expensive part of those systems is usually the decision-making layer. Needle makes that layer very cheap and very fast.

Hypercubic launched Hopper today — an agentic development environment built for IBM mainframes. Mainframes still run the core of banking, insurance, airlines, and government systems, but they use TN3270 terminals, ISPF panels, JCL job queues, and COBOL — an environment that no existing AI tool was designed to navigate.

Hopper gives an AI agent the ability to operate inside the mainframe environment directly: browsing datasets, submitting jobs, reading spool output, and patching COBOL source files — the same loop a human developer would run, but automated. Sensitive operations require human approval, and the terminal stays visible at all times.

The project is drawing attention from developers doing legacy modernization who have been stuck trying to apply modern AI tools to infrastructure that predates the internet.

Gigacatalyst lets non-technical users build custom workflows inside an existing SaaS product by describing what they need in plain language. The AI connects to the product's APIs, learns its data model, and generates working apps — governed and sandboxed — without any engineering involvement.

The use cases shown at launch are practical: a maintenance manager who built a parts stockout alert, a facilities team that automated invoice OCR from phone photos, a pizza chain that triaged maintenance requests by priority. All built by the people who needed the tool, not the engineers.

With 2,000 daily users and 900 apps already built before today's public launch, the product has been validated quietly before opening to the public.