Got back from AI Engineer World's Fair in SF and I'm still buzzing.
Here's what blew my mind 🧵
Here's what blew my mind 🧵
AI capabilities are doubling every 70 days for coding complexity.
Every 7 months for task length.
That's 16-64x capability increase per year.
We're not in gradual improvement territory anymore.
(h/t @ScottWu46 from @cognition_labs)
Every 7 months for task length.
That's 16-64x capability increase per year.
We're not in gradual improvement territory anymore.
(h/t @ScottWu46 from @cognition_labs)
50% of engineers surveyed at the conference are already using LLMs for both internal and external work.
Not "experimenting with" not "testing out" using.
The future arrived quietly while we were still debating if it would come.
Not "experimenting with" not "testing out" using.
The future arrived quietly while we were still debating if it would come.
According to that same survey, for customization: RAG is eating the world.
70% use among engineers looking to customize.
Fine-tuning at 40%.
70% use among engineers looking to customize.
Fine-tuning at 40%.
Remember when GPT-3 cost $60 per million tokens?
Now we're at ~$0.10 for distilled versions.
That's a 600x cost reduction in 3 years.
Compute is becoming free.
Creativity is becoming expensive.
Now we're at ~$0.10 for distilled versions.
That's a 600x cost reduction in 3 years.
Compute is becoming free.
Creativity is becoming expensive.
Models are getting smaller but stronger.
405B → 70B → 24B parameters offer same capability, fraction of the size.
Your phone is about to become a supercomputer.
405B → 70B → 24B parameters offer same capability, fraction of the size.
Your phone is about to become a supercomputer.
Gemini went from 2% to 35% market share in 12 months according to @OpenRouterAI.
That's not growth, that's conquest.
Meanwhile AI inference processing increased 50x year-over-year.
That's not growth, that's conquest.
Meanwhile AI inference processing increased 50x year-over-year.
If you're still thinking about AI as a tool to make you 10% more efficient...
you're thinking too small.
The companies being built right now are assuming AI can do 80% of the work.
They're designing for that world...
not retrofitting for this one.
you're thinking too small.
The companies being built right now are assuming AI can do 80% of the work.
They're designing for that world...
not retrofitting for this one.
Shout out to @swyx for organizing a phenomenal conference.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
