AI digest: Models get faster, companies get desperate
Google releases a diffusion text model, Anthropic ships dual safety tiers, and companies burn serious cash chasing AI.
Three stories this week show how the AI race is heating up. Models are getting genuinely faster, safety debates are splitting into different products, and the money being thrown around is getting silly.
Google’s DiffusionGemma generates text like images
Google released DiffusionGemma, a 26B parameter model that generates text through diffusion instead of word-by-word prediction. The approach promises 4x faster generation on GPUs, which could actually matter for production systems. This feels like the first genuinely different text generation architecture we’ve seen gain traction since transformers.
Anthropic splits safety into two products
Anthropic shipped Claude Fable 5 and Mythos 5, same underlying model but with different safety guardrails. Fable 5 gets the full safety treatment, whilst Mythos 5 has cyber safeguards lifted through their Project Glasswing programme. Cybersecurity researchers are already complaining that Fable’s guardrails are too strict for real security work. Smart move to avoid the impossible middle ground.
Companies burning £7.5k per employee monthly on AI
The most AI-obsessed firms are spending roughly £7,500 monthly per employee on AI tools and infrastructure. Meanwhile, Amazon just borrowed £17.5B from banks to fund AI spending. These numbers suggest either transformative productivity gains are coming, or we’re watching the most expensive corporate FOMO in history.