AI digest: Training gets faster, governance lags behind
New training techniques slash compute time whilst most companies still have no AI policies in place.
We’re seeing the gap widen between what’s technically possible and what organisations actually do about it.
Nous Research cuts training time by 60% with token tricks
Nous Research released Token Superposition Training, which speeds up LLM pre-training by up to 2.5x by averaging token embeddings into “bags” during the first phase before switching to normal next-token prediction. The clever bit is it works without changing model architecture or inference behaviour, so you get all the speed benefits with none of the compatibility headaches. This is the kind of optimisation that makes training accessible to smaller teams.
Most companies still flying blind on AI governance
New data shows 63% of organisations have no AI governance policy whilst shadow AI spreads through their tech stacks. Meanwhile, Anthropic now has more business customers than OpenAI according to Ramp’s spending data. Companies are clearly buying AI tools faster than they can figure out how to manage them properly.
Anthropic goes after small business with packaged workflows
Anthropic launched Claude for Small Business with 15 pre-built workflows for tools like QuickBooks and HubSpot. They’re even doing a ten-city workshop tour, which feels very much like the enterprise playbook applied to smaller customers. Smart move given that’s where the growth is, but also shows how the AI platform wars are moving downmarket fast.
DeepMind wants to reinvent your mouse pointer
Google DeepMind built an AI-powered mouse cursor that understands visual and semantic context around where you’re pointing. The idea is you point, speak naturally, and get things done without switching to a separate AI window. Feels like the kind of interface experiment that could either be brilliant or completely annoying in practice.