Mistral launches a huge new model as Altman and Anthropic clash on AI risk
Mistral released a massive new model called Large 4, OpenAI's CEO publicly broke with Anthropic over how much regulation AI needs, and LibreOffice is marketing the fact that it refuses to add AI at all.
Every line links to a primary source in the full briefing.
01Mistral releases 'Le Chonk', a very large new model, with open weights coming soon
French AI lab Mistral released a preview of Large 4, a model with 1 trillion total parameters (49 billion actually used per query), trained on its own cluster of Nvidia chips. It's available now through Mistral's API, and the company says it will release the full open weights version by the end of October.
A credible European alternative to US models matters for businesses worried about data residency or vendor lock-in, and an open weights release means companies could eventually run it on their own infrastructure instead of paying per API call.
02OpenAI's Altman publicly breaks with Anthropic over AI regulation
Sam Altman said publicly that the world needs to accept that AI will cause some harm, arguing the benefits outweigh the downsides, a position that puts him at odds with Anthropic's more cautious public stance on regulation.
The two biggest AI labs disagreeing this openly on how much oversight is needed is a sign that regulation is still unsettled, so businesses shouldn't assume today's light-touch environment will last.
03OpenAI expands deals to put AI agents into real business workflows
OpenAI announced an expanded partnership with Atlassian to connect its models directly to company knowledge and project tools, and a separate collaboration with contract software company Ironclad to train AI agents on real-world contract review and negotiation work.
These are concrete test cases for AI agents doing actual office work such as planning projects or reviewing contracts, which is the direction this technology is heading for any business that manages documents or workflows.
04Google releases a small, open model for building search and retrieval tools
Google DeepMind released EmbeddingGemma 2, an open and lightweight model that converts text and images into a format computers can search and compare, rather than one that generates text or chats.
This is the kind of model developers use cheaply to build internal search over a company's documents, emails or product catalogue, so it lowers the cost of custom search tools even for smaller teams.
05LibreOffice turns its refusal to add AI into a marketing pitch
The maker of the free, open source office suite LibreOffice says it has no plans to build AI features into its default software, citing user privacy as the reason.
It signals a real, if small, market of users and organisations who want guaranteed AI-free software for privacy or compliance reasons, worth knowing if your own customers ever ask the same question about your tools.
06Google highlights AI tools for defending against cyberattacks
Google DeepMind and Google AI published material on using AI proactively for cyber defense, aimed at governments and large enterprises, covering how models can help spot and respond to threats before damage is done.
Attackers are already using AI to scale up attacks, so AI-assisted defense tools are likely to trickle down to the security software smaller businesses rely on, even if this particular effort targets large organisations first.