← All editions

Wed Oct 07 2026 02:00:00 GMT+0200 (South Africa Standard Time)

AI Trends: 7 October 2026

Mistral goes trillion-scale with Large 4

Mistral released Large 4, nicknamed "le Chonk": a natively multimodal mixture-of-experts model with 1 trillion total parameters and 49 billion active per token. It was trained from scratch in about two months on 4,000 Nvidia Grace Blackwell GPUs in European data centers. API access starts at $1.36 per million input tokens, and the weights are due at the end of the month, giving European enterprises a sovereign alternative to US and Chinese frontier models.

Labs decline to guarantee safety under oath

At a New York City Council hearing on 5 October, representatives from OpenAI, Anthropic, Google and Meta were asked by Speaker Julie Menin to guarantee under oath that their AI agents would always follow safeguards. All four declined. Anthropic's Logan Graham called the science "fundamentally hard and unsettled," and OpenAI's Morgan Dwyer said no technology can be guaranteed risk-free. Local governments are now questioning the labs directly while federal rules remain thin.

White House launches the "Super Intelligence Force"

President Trump announced the task force on 4 October. It is led by Director of National Intelligence Jay Clayton as AI czar, with FTC chair Andrew Ferguson, Pentagon research chief Emil Michael and OPM director Scott Kupor. Clayton said "the risk of not being first is high." The announcement gave no detail on oversight or regulatory powers. Meanwhile UN human rights chief Volker Türk called for mandatory AI safeguards, saying governments can no longer rely on goodwill.

Security programs and agent standards

Anthropic folded Project Glasswing into a three-tier Cyber Verification Program covering defense, red team and safety-critical access. Partners report more than 129,000 verified software flaws found between April and July, over 33,000 of them rated critical or high. Separately, Meta and Sierra published an open OAuth-based protocol that lets personal AI agents authenticate with businesses, with Walmart, Shopify, Stripe and Genesys as founding partners.

Open-weight and model releases

Reflection, the Nvidia-backed lab, released its first open-weight model, Beam, aimed at rivaling Chinese models at lower compute cost. Z.ai's GLM 5.3 also landed on Amazon Bedrock. Both arrived on 5 October, alongside Mistral's launch.

Power and national investment

Google signed a 3.6 gigawatt deal with Constellation Energy, about a quarter of it from new nuclear capacity across 11 reactor upgrades in Illinois, Pennsylvania and New Jersey, with deliveries starting in 2028. South Korea committed 4.7 trillion won (about $3.5 billion) to homegrown frontier AI from March 2027, backed by roughly 29,000 GPUs. Crunchbase data shows AI startup acquisitions by other AI startups reached 195 through 29 September, up 14% on all of 2025, with OpenAI the most active buyer.


Compiled from public reporting for the 48 hours ending 7 October 2026.

Published Wed Oct 07 2026 02:00:00 GMT+0200 (South Africa Standard Time)

← View all editions