Co-founder of @mesosphere/@D2iQ (acq. @nutanix) · Early @airbnb engineer · Angel investor in cloud & AI startups · Sharing what I learned building at scale.
I'm still not sold on the idea that these single uber models that are experts at everything will be the winning architecture long-term. They are very expensive to run and it seems that everyone is hitting capacity limits right now with no end in sight.
Now that frontier models are great at writing tasks (including writing code), next we will see them rapidly improve on a number of highly specialized tasks that require deep domain expertise. We went from 0% to 52%+ on Frontier Math in < 2 years and we'll accelerate from here.
These are 100% real concepts that were coming out of Detroit during the Space Age days
We were so close to (now) seemingly unreachable levels of American excellence
China just open-sourced a trillion-parameter model that burns fewer tokens than your favorite "efficient" US model.
Ling-2.6-1T is now public, inspectable, and benchmarkable.
The closed-model moat just got smaller.
@tetrisgm Kimi K2.6 is right behind GPT/Claude/Gemini in this chart. I'm using it for coding and for most tasks it's as good as them. I'm still using GPT or Opus for the most complex tasks though.
@tetrisgm They're not there yet, but very close. At the rate at which they're improving, I predict parity will happen this year. Lots of numbers here: https://t.co/LsonUDma5r
When future historians write about Silicon Valley, they’ll have an entire chapter dedicated to the Ron Conway way: how he turned generosity, warmth, and showing up for founders into a winning strategy.
The rollout process that @AnthropicAI is using for Mythos will become the industry standard for frontier models. It's similar to the responsible disclosure process that's been a standard for a long time.