Mistral Large 4 Preview, ChatGPT Teen Safety Concerns, and More Tech News

AI

ChatGPT for Teens Rated “Unacceptable Risk” by Common Sense Media

The Youth AI Safety Institute at Common Sense Media has published a safety assessment of ChatGPT for Teens. After testing more than 4,000 prompts on accounts registered to 13- to 17-year-olds, the institute assigned an “Unacceptable Risk” rating.

The most alarming finding involves parental alerts. A teen can spend an hour discussing self-harm without triggering a single notification to their parent. In more than one in four crisis situations, ChatGPT failed to direct teens toward help resources.

Study Mode can be bypassed with a “Show me the answer” button that lets teens skip the tutoring entirely. Deleting a short prefix from a message also overrides parent-set study time restrictions. Age verification did not function as advertised, with adult test accounts never switching to the teen experience.

Mistral Launches Large 4 Preview — a 1.05-Trillion-Parameter Open-Weight MoE

French AI lab Mistral AI has released Large 4 as an API preview. Nicknamed “Le Chonk,” the model packs 1.05 trillion total parameters with 49 billion active per token in a Mixture-of-Experts architecture.

It offers a 512K-token context window with up to 256K output tokens. The model accepts text and images as input and produces text output, with native support for tool calling and structured outputs.

Training took place in European data centers on a cluster of 3,800 NVIDIA Grace Blackwell GPUs. According to Simon Willison’s analysis, open weights are expected by the end of October. A trillion-parameter model trained entirely on European infrastructure marks a notable milestone for AI sovereignty in the region.

Big Tech

Google Launches Playground, an AI-Powered Game Creation Platform

Google Labs has launched Playground, an experimental platform where users create, play, and share games using nothing but text prompts. It became available in the US on October 7 for users aged 18 and over.

According to Gamedeveloper, the platform combines Gemini, Nano Banana, and Lyria models. Users describe the mechanics and visual style they want, and a playable 2D or 3D game is generated instantly. In-game leaderboards and multiplayer are supported in select genres.

Browsing and playing existing games is free, while game generation runs on a weekly token system. A closed-beta Unity Spark tier is on the way, promising higher-fidelity 3D and professional-level mechanics through Unity runtime integration.

Sony Transfers 419 XR Patents to Meta, Signaling VR Pullback

Sony Group has transferred 419 XR-related patents to Meta Platforms, as revealed by Road to VR’s analysis of public records. The portfolio covers 180 US patents and 239 filings across Japan, China, Korea, and Europe, all originally filed between 2010 and 2022.

VR.org’s review of USPTO records shows the confirmatory assignment was effective December 17, 2025, signed in late April 2026, and recorded with the USPTO between June and September. The patents span HMD optics, wearer face capture, and 360-degree spatial depth recording.

A transfer of this scale suggests Sony is pulling back from XR hardware ambitions, consistent with the tepid reception of PSVR 2. Neither company has issued a public statement about the deal.

Startups

Arena Closes $200M Series B at $3.1B Valuation

Arena, the crowdsourced AI model leaderboard that grew out of UC Berkeley’s Chatbot Arena, has closed a $200 million Series B at a $3.1 billion valuation. According to Tech Startups, this follows a $150 million Series A in January 2026.

AI Weekly reports that Arena crossed $100 million in annualized run-rate revenue just eight months after launching its commercial service. Major AI labs rely on its blind comparison and Elo-style rankings as a gold standard for model evaluation.

BloomX Raises $13M for Robotic Crop Pollination

Israeli AgTech startup BloomX has raised $13 million to expand its robotic pollination systems, as reported by Tech Startups.

The decline of natural pollinators like honeybees poses a direct threat to global agricultural productivity. BloomX’s robotic approach aims to fill that gap through automation, offering a scalable alternative that doesn’t depend on bee colony health.

Open Source

OpenSSH 10.6 Enables Post-Quantum Signatures by Default

OpenSSH 10.6 shipped on October 6. The headline change is the default activation of the hybrid post-quantum signature algorithm ssh-mldsa44-ed25519.

According to Linuxiac, the release includes several security hardening measures. The LZ77 dictionary coder has been disabled in ssh and sshd to mitigate compression side-channel leaks. Default KDF rounds for private keys increased from 24 to 32. Usernames on the ssh command line can no longer contain $ or \, reducing shell-injection risks.

SFTP also received tighter path validation to prevent malicious servers from directing recursive copy operations outside the intended destination. The team noted that a surge in AI-generated security bug reports is prompting more frequent releases.

llama.cpp v0.6.0 Brings MTP Speculative Decoding

The ggml-org team has released llama.cpp v0.6.0, headlined by Multi-Token Prediction (MTP) speculative decoding support.

AI Weekly’s analysis notes that MTP speculative decoding on Qwen4Exp delivers roughly 1.5x decode speedup on DGX Spark. Unlike traditional speculative decoding, MTP needs no separate draft model. A single forward pass through the main model produces both the current token and speculative drafts for the next N positions.

New Metal MMA kernels for Apple GPUs claim up to 3x faster matrix multiplication for batched and speculative decoding. The release also adds support for the GLM-5.3-Flash 320B hybrid model and the Clef decision model.

Google Adopts Meta’s Pyrefly as Internal Python Type Checker

Google has adopted Pyrefly, the open-source Python type checker developed by Meta, as its internal standard. The announcement came via the Google Open Source Blog.

The switch has yielded significant speedups for engineers and AI agents alike, while substantially reducing build-system compute resources. Two of the world’s largest tech companies now sharing the same open-source Python type checker is a notable collaboration in the Python ecosystem.

Gaming

Capcom’s REX Project Aims to Rebuild RE Engine for AI Workflows

Capcom unveiled the REX project at Open Conference RE: 2026, outlining plans to evolve its RE Engine into what it calls an “AI-generation game engine.”

According to True Achievements, programmer Satoshi Ishida presented the project’s goals: making engine code readable, writable, and testable by AI tools. As games and assets grow larger, even small corrections take increasingly long to test, and Capcom sees AI integration as the path to streamlining that process.

The company emphasized that changes will happen gradually and in stages. Capcom maintains its existing policy of not using AI-generated assets in its games. The goal is efficiency in development workflows, not replacement of human-created content.

Triple-I Initiative Showcases 40 Indie Games with 8 World Premieres

Evil Empire hosted the Triple-I Initiative on October 8, packing 40 indie game trailers and eight exclusive world premieres into a 45-minute broadcast.

Dot Esports reports that studios including Amanita Design, Norsfell, Strange Scaffold, and Team17 delivered world premieres. The showcase featured titles spanning a wide range: a new Among Us game, Outward 2, Pony Island 2: Panda Circus, Don’t Starve Elsewhere, and Risk of Rain 2 among them.

The indie market continues to grow steadily as AAA development costs climb, and Triple-I has established itself as a concentrated showcase for the independent scene.

References

ChatGPT for Teens Poses Unacceptable Risk to Kids — Common Sense Media
Introducing Mistral Large 4 — Mistral AI
Le Chonk — Mistral Large 4 — Simon Willison
Google Playground — Google
Google debuts new AI game platform named Google Playground — Gamedeveloper
Sony sells off dozens of XR patents to Meta — Road to VR
Sony Meta XR patents USPTO record — VR.org
Startup Funding News Today, October 9, 2026 — Tech Startups
Arena AI evaluations hit $100M ARR — AI Weekly
OpenSSH 10.6 Release Notes — OpenSSH
OpenSSH 10.6 Released with Security Hardening — Linuxiac
llama.cpp v0.6.0 — GitHub
llama.cpp v0.6.0 Adds GLM-5.3-Flash, 3x Apple Metal Matmul — AI Weekly
This Week in Open Source for October 2, 2026 — Google Open Source Blog
Capcom Details REX Upgrade to Make RE Engine AI-Ready — God is a Geek
Capcom is transforming its RE Engine into an AI-generation game engine — True Achievements
The Triple-I Initiative: October 8, 2026 — Gematsu
Triple-I Initiative 2026 Announcements — Dot Esports