Finetuning > Frontier?
Will enterprises fine-tune models for more cost-efficient intelligence? Plus Saronic rescue, AI Safety concerns, and Fable
Insane week. Hard to keep up with all this news alongside 400 Frontier applications to review!
Firstly, after I mentioned last week that enterprises would start to fine-tune models to bring down cost without sacrificing accuracy, Microsoft announced:
It’s another sign the market could evolve away from solely OpenAI/Anthropic frontier models serving every use-case, even if it’s only large, early-adopter enterprises in the short term.
Then Saronic’s Corsair helped rescue the US helicopter pilots shot down this week. Such a great example of the potential for autonomy in defence.
And then of course Fable. This is probably my favourite Fable post, because video editing is still so hard with AI.
You can feel the safety concerns rising within the labs. Fable is impossible to use on anything with a hint of biology in the question, recursive self-improvement in research feels around the corner, and the concerned letters are coming thick and fast:
Other Links
I had no idea how quickly prices are declining as a result of the battery rollout:
Unitree is wildly under-rated. This is a detailed breakdown of the company and how it’s building its competitive advantage:
A moment of zen…
Turns out the world vs language models debate is centuries old:
Tracks
this had me bouncing in the office this week