L'IA pour les CTOs.

★ Analyse

Google’s game generator is a distribution test

Google opened Playground today to US users aged 18 or over, while Switzerland has no access; prompt-built games can stay private or be shared through links and a public gallery. The prompt box is bait. Behind it, Google controls login, sharing, discovery and ratings, while Google One membership sets creation limits; Unity Spark is meant to add richer 3D and professional game systems later in 2026. Google calls Playground an early experiment; it publishes no quality or delay figures and gives no API, export format, cost model or commercial terms, while Spark still awaits a closed beta. Our pilot gate would be simple: assets and usage data must leave cleanly, or the work stays in the experiment budget.

★ 3 actualités

HEALTH AI

Healthleap raises $38M to flag patients for review

Healthleap disclosed an $8 million seed round and a $30 million Series A for AI that flags hospital patients who may need a closer look. That changes who gets seen. Missed patients and alert overload are the deployment risks because a flag can redirect clinical attention before any diagnosis is made.

Lire →

OPEN SOURCE

Strands releases a 2B decision model

Strands released Decider 2B today, a small open-source model built to choose among actions inside software workflows. Two billion parameters are the point. At that size, it gives us a compact candidate for routing requests or ranking an agent’s possible next actions, jobs where a general chat model can add needless cost and delay.

Lire →

WEB PLATFORM

Chrome ships JPEG XL

Google’s browser just reopened an image-pipeline decision. JPEG XL support is shipping in Chrome, so web teams can test the format there. Your own images and delivery costs can now decide whether it earns a production path.

Lire →

★ Aperçu du secteur

• →OPENAI API: OpenAI put its Decisions API into public beta on 7 October, opening the service to public testing. →
• →AUTONOMOUS RAIL: SpaceX alumni at Parallel Systems raised $100 million to scale autonomous electric rail vehicles that can carry thousands of pounds for up to 500 miles. →
• →AI DEVICES: Tony Fadell argues that first-wave AI gadgets failed to solve concrete problems and that the next wave must earn consumer trust. →
• →SOFTWARE RIGHTS: Photopea’s developer says GitHub left AI-assisted cracked copies of the browser editor online after a month of removal requests. →

★ Pendant ce temps en Suisse

AI PRIVACY

Some chatbot vendors share entire conversations

Netzwoche reports that some AI chatbot operators pass full chat transcripts to third parties. More companies then hold the text. For us in Switzerland, each extra copy adds another retention rule and another place the conversation can leak.

Lire →

CUSTOMER SERVICE

AI agents gain ground in Swiss customer service

According to Netzwoche, AI agents are gaining ground in Swiss customer service; these systems can take actions within a case. Refunds expose the boundary. An automated account change can become a costly customer error when escalation fails.

Lire →

IDENTITY

ADN adds Swissbit security keys to its catalogue

ADN now carries Swissbit’s hardware login keys. For us, that adds another channel for buying devices and replacements through approved suppliers.

Lire →

★ L'académie IA

The review bottleneck behind OpenAI’s 722 math manuscripts

OpenAI published 722 AI-produced mathematical manuscripts this week. The release exposes a new bottleneck: candidate proofs can arrive far faster than experts can judge which claims are sound and worth attention.

1. OpenAI gave an unreleased internal model about 4,000 open problems. Most retained results followed one common procedure and used, on average, the equivalent of roughly three hours of ChatGPT Pro thinking time per result.
2. The counts differ. The release contains 722 manuscripts across 372 result families. One family can bundle a main theorem with several related or alternate proofs, which is why the manuscript total is larger.
3. In Lean, a proof is written as machine-checkable steps against stated definitions and starting rules. Fluent prose earns no credit there: one invalid formal step breaks the chain.
4. Lean’s boundary matters here. It checks the proof chain that was encoded, while humans still judge whether the assumptions capture the intended problem and whether the result matters.
5. Only some manuscripts have Lean proofs, and OpenAI says the collection spans different verification stages. The internal model remains unreleased. Its processing number is an estimate, and independent review has yet to match the production rate.
6. The CTO lesson for us is capacity planning: more agent output needs matching review capacity. OpenAI’s version and correction process offers a useful pattern: every generated item keeps its check status and correction history.