Set up a windows desktop machine today, do i get an award for completion?
Will revert in a few hours.
3
4
792
kagen retweeted
I got married last month, and spent the furthest time away from work during that period and the honeymoon
Unplugging while we broke ATHs really helped me reflect on what I’ve learned up until this milestone and what’s important to me in life
I wanted to share this publicly mostly to document my thoughts. I hope it’s helpful but it’s not intended to be life advice or anything like that. Just my own musings which include some stories, quotes, and pictures here and there.
kagen retweeted
Someone wants to make this work, with a small DSPy program and some optimizer runs?
Function calling accuracy in LLMs really sucks. The best function calling accuracy is obtained with GPT-4 and it's 83.8%. It's already too low to be practical, but one should discount this number more assuming that the test data Berkeley used to evaluate function calling contained up to 4 functions to choose from. If with <=4 functions to choose from the accuracy is ~80%, imagine how low it would be when we will have 20 functions to choose from. What if we want to call 200 functions?
And this leaderboard only measures the function choice accuracy. It doesn't consider the choice of parameters and their values. In some papers I read, the accuracy of function calling is estimated to be near 50%.
And with these numbers, they talk about LLM-based agents. Agents, Carl!
gorilla.cs.berkeley.edu/lead…
kagen retweeted
I always thought these remarks are not very serious but since they’re common it could be instructive to comment.
Tradeoffs. Context (concatenation) & Retrieval pose fundamentally different tradeoffs in recall, precision, and latency/cost. You can’t “choose” between RAM and disk.
kagen retweeted
Here's one of proof in the tech report, Gemini team compared Gemini 1.0 pro w/ RAG v.s. Gemini 1.5 Pro. Gemini 1.5 outperform Gemini 1.0 + RAG
RAG technics they used,
TF-IDF database + Cosine similarity search + Rerank + Similar passages
Note: TF-IDF database - frequency-inverse document frequency and it is a measure, used in the fields of information retrieval (IR) and machine learning
kagen retweeted
Nero is seriously one of the most underrated AI products out there. @kagnar_ is way too busy building to advertise.
We have an integration on @AgentHub_AI as well if you want a simple interface for it – at San Francisco, CA
Quick demo of @nerobotai AI Crawl Assistant in action. #buildinpublic
It's able to navigate sitemaps for desired pages all from natural language inputs. It does so by leveraging a set of tools and/or code interpreter to handle as many types of website structures as possible.
Currently we are using @OpenAI Assistant API and Mistral models.
The beautiful frontend was made possible by the boilerplate framework @makerkit_dev from @gc_psk
And shoutout to @KillerTopic who did a lot of the heavy lifting on the backend.
After SORA release, im now fearful of the speed of AI progress.
I think i need to get off X for a bit.
6
232
kagen retweeted
Every single one of these videos is AI-generated, and if this doesn't concern you at least a little bit, nothing will
The newest model: openai.com/sora
(Remember Will Smith eating spaghetti? I have so many questions)
Real? Or fake out again?
🤞
Needle in a Haystack Tests Out to 10M Tokens
First, let’s take a quick glance at a needle-in-a-haystack test across many different modalities to exercise Gemini 1.5 Pro’s ability to retrieve information from its very long context. In these tests, green is good, and red is not good, and these are almost entirely green (>99.7% recall), even out to 10M tokens. Great! A bit more on needle-in-a-haystack tests later in the thread.
1
256
This is clearly related to BTC making yearly highs. All the crypto algo traders are overloading the servers
2
292
Had a chance to meet with him yesterday. Really productive call excited to see what we can accomplish in the near future.
Excited to announce a new experiment — Nset Labs — where we help companies with LLM pipelines optimize their prompts to increase performance by up to 140%
Check it out and keen to get some feedback! 👇
labs.nset.io
5
394
kagen retweeted
RESPOND WITH JSON ONLY. DO NOT UNDER ANY CIRCUMSTANCES RETURN ANY CONVERSATIONAL TEXT.
My biggest takeaway from using the Apple Vision Pro.
I want eye tracking based navigation for my laptop. The efficiency gain is unparalleled.
Ty @will__price
3
175