I think
@karpathy
is onto something - I've started going straight to the gym then to work, no "small tasks" right after I wake up, and productivity been significantly higher
@corbtt
I think they're finding a balance between customer facing products and hard research; the team is pretty big now and they can definitely afford this kinda thing. the team building gpt5 is linearly independent from the gpt4o API team for sure
Referencing either a tweet or podcast (that I have since lost) where he said he goes straight to the office upon wake up to avoid filling RAM with a bunch of smaller things
LLM response times are too long. Caching hit rate is too low for natural language. Try out some semantic caching! 👀
I built GPTCacheLite tonight to save time, compute, and headache.
Check it out with `pip install gptcachelite` 🧵
I'm back in college moving out of my apartment, watching the homies graduate, saying all my goodbyes - surprisingly now everything feels super real. Like the past year was play mode but now it's hitting things won't be the same. crazy
Daily schedule since moving to SF with
@cerebral_valley
:
9am wake up
10am gym + sauna + shower
12pm lunch at the office
12:45pm code + calls + meetings
11pm DJ practice + music (optional)
2:30am get home and sleep
New morning routine has been popping one technical article/paper off my reading list with some coffee - crazy how reading things makes you learn what a wild concept
Reading a research blog by
@NormalComputing
that
@iporollo
sent me yesterday about "Extended Minds" for LLMs as an alternative to RAG. Pretty cool results, a thread: 🧵
The community asked, and we answered. Excited to publicly launch our Cerebral Valley community token, $CV! Thanks to our ICO participants for making this happen.
There are tons of "chat with postgres" projects. But I'm built different.
Introducing AISQLite: a sqlite3 wrapper with natural language query ability.
Because who needs to use postgres anyway.
How does attention REALLY work? What's the buzz about? I've found it's one of the hardest things in introductory deep learning to gain intuition for, so I wrote about it 🚀
If you're an AI engineer, but want to learn the basics, check out:
I've fallen deep down the RAG rabbit hole lately. In a quest to optimize the process, I've been using the highly customizable VLite by
@sdand
. I just made some upgrades 😈
Introducing Eagle-7B
Based on the RWKV-v5 architecture, bringing into opensource space, the strongest
- multi-lingual model
(beating even mistral)
- attention-free transformer today
(10-100x+ lower inference)
With comparable English performance with the best 1T 7B models
Just wanna throw some thoughts on there on the optimal selection of K in RAG apps. If you work in AI engineering, please comment your thoughts! Always want to learn :)
Retrieve Many: the time to retrieve the top k=10 results with the legal corpus already ingested in the database. VLite2 is still over 2x as fast at retrieval, and scales incredibly well to data (thanks to
@unum_cloud
)!
@jowyang
my thesis is that over time there will be less devs, but the devs that do remain will increase in value over their current value, and will continue to generate out of distribution data to enhance "AI software engineers".
Most furniture today is manufactured, but hand-crafted