@thsottiaux FYI, the codex status and Status lines thingy seems to be broken. In one term it says 88% left on 5h and in another it says 2% :)
Haven’t gotten to dig through the codex code yet though so no “analysis” yet
Chris retweeted
cloudflare tunnel and a raspi you own can serve more traffic than most people's "apps" and can scale easily into 1000s of concurrent users (which is more than 99+% of all apps)
Chris retweeted
my biggest inspiration as an engineer
Readers added context they thought people might want to know
The picture has been altered.
x.com/ID_AA_Carmack/…
Economy: @FlySWISS > @Lufthansa_DE
Business: @Lufthansa_DE > @FlySWISS (by a looooooot)
First: ?
@FlySWISS : Economy > Business
Chris retweeted
I have very mixed feelings when reading these recent search + rl papers.
I understand the LLM folks are too young to read the 2020 RAG papers. But claims like "RAG lacks the flexibility for multi-turn, multi-query retrieval" and "We are the SoTA by achieving NQ = 41%, HotpotQA=38%" really baffle me.
1. Iterative RAG has been studied for quite a long time. There are plenty of great work done in this space.
2. The SoTA for NQ and HotpotQA are both around 70%!
There are some good curated RAG paper list in github.com/coree/awesome-rag and github.com/hymie122/RAG-Surv….
Chris retweeted
New research shows that LLMs don't perform well on long context.
Perfect needle-in-the-haystack scores are easy—attention mechanisms can match the word. When you require 1-hop of reasoning, performance degrades quickly.
This is why guaranteeing correctness for agents is hard.
Chris retweeted
o3-mini is the most feature complete + developer friendly o-series model we've released to date: function calling, structured outputs, streaming, batch, assistants!
It's also:
1. It's 90+% cheaper than o1
2. It's _much_ faster
3. And on things like coding, it's better than o1
I hope you like it! Please send any model or API feedback my way ❤️
