The best operators we've worked with ask one question before anything else:
"How will we know which lines the AI wrote?"
If you can't separate AI's work from your team's, you can't measure it. If you can't measure it, you can't trust it.
@qasar@blackstone 4. Korean companies has been tradionally undervalued considering PE (this includes Samsung / Hynix / Hyundai) - wonder if that pattern will materially change next few years
5. Interesting! While Bending Spoons are taking headlines, PE acquisition market itself is slowing down
@saradu Great post! Especially tying it to history of city development. Tokyo is fascinating one as their city crawl is reaching other prefecture, but a lot of their buildings are actually quite old, and their strict building code enforces rebuild
1/ Harvey just published how they fine-tuned Kimi (Chinese open-weight model) and hit frontier-level quality on legal work at lower cost.
A top US AI company. Chinese weights. Rational choice.
That should worry US labs more than any benchmark.
I think it gets tougher every year. Each frontier release absorbs more of what used to need domain training.
I think it gets tougher every year. Each frontier release absorbs more of what used to need domain training.
6/ Related question I keep chewing on: can specialized model makers beat frontier labs on vertical use cases long-term?imbler, right direction.5/ The fix isn't restricting demand. It's supply —