←
Dylan Patel14,271 posts
Dylan Patel avatar✓

Dylan Patel ✓

@dylan522p

SemiAnalysis AI Infrastructure Research and Consulting DMs are open for consulting, quotes, or to talk shop, Opinions my own

◷ Joined April 2018▧ 1,511 Photos and videos
Dylan Patel avatar
Dylan Patel@dylan522p·Nov 17

Airlines freezing Banks down We have created a technology that you have to pay to protect yourself It's roman fire brigades all over again

Dylan Patel avatar
Dylan Patel@dylan522p·Nov 15

I have been informed that Venus and Mercury retrograde have ended. It's time for a return of the king

Dylan Patel avatar
Dylan Patel@dylan522p·Nov 14

The entirety of the semiconductor supply chain, neocloud complex, and open source ecosystem

Post attachment
Dylan Patel avatar
Dylan Patel@dylan522p·Nov 12

Nvidia Board needs to chase down Charlie Kawwas or Johny Srouji. They are the only somewhat viable CEO replacements.

Dylan Patel avatar
Dylan Patel@dylan522p·Nov 11

What I love about Perplexity is that they're very fast on any feature request. Even though they did not invent the category, they are fast followers and implement any change that's needed in less than 24 hours. They even got the name from us back in February.

Post attachment
Dylan Patel avatar
Dylan Patel@dylan522p·Nov 11

The funny thing is that Perplexity is primarily using Anthropic models for our use cases, but Anthropic sucks at product, and they're not really receptive to any feedback on their product Same with OpenAI

Dylan Patel avatar
Dylan Patel@dylan522p·Nov 11

Persistent Agents are the next unlock of AI usage Our spend on AI is now 45% persistent agents Unfortunately for Anthropic, Claude Tags sucks OpenAI's Slack agent is even worse We primarily use Perplexity Computer, which uses a lot of different models Perplexity's Harness mogs

Post attachment
SemiAnalysis avatar
⟳ Dylan Patel retweeted
SemiAnalysis@SemiAnalysis_·Nov 10

Fable 5.2 is very token inefficient for certain tasks due to its increased persistence and proclivity for spawning subagents. For example, on GDPx materials science tasks, Fable 5.2 costs 89% more than Astra and uses 4.3x more output tokens to achieve effectively the same score. The price difference vs Gemini 4 Carbon is even more startling at 7.5x. (1/3)🧵

Post attachment
Dylan Patel avatar
Dylan Patel@dylan522p·Nov 10

Fable 5.2 is significantly better at chip design, software engineering, biology, and material science. GDPx is the only eval that test these categories at this level of detail. True experts make these tasks, unlike other benchmarks. Notably it is not better at finance, legal, quantative trading, RSI, or hardware design. Anthropic explicitly nerfed this model in RSI tasks and it's interesting seeing that reflected in the score.

Post attachment
Dylan Patel avatar
Dylan Patel@dylan522p·Nov 6

Why is Baseten worth more than Nebius? This makes 0 sense Venture market valuations are nuts

Dylan Patel avatar
Dylan Patel@dylan522p·Nov 4

Great work that pushes our understanding of how agents and experts can collaborate on optimizing system performance Beyond the strong results, the team also found some surprising cases where our human inductive biases end up limiting what agents can achieve