Kenneth Goodman

@corbtt What happens if you have hundreds of tools? Is the right approach to RL every model after every checkpoint for every new tool?

Announcing MCP•RL: teach your model how to use any MCP server automatically using reinforcement learning! Just connect any MCP server, and your model will start playing with it and (using RL) "learn from experience" how to use its tools most effectively!

@TicTocTick Has to be like 18% interest for the payment to make sense

@mcuban @KamalaHarris Doesn’t that assume there isn’t another way to make the product without COG that are being tariffed?

@rakyll Ya I meant why don’t we do all the logic the docker/proxy is doing behind the API so migration is simple. This feels complicated enough to just use a different model

@pythonprimes What is the custom URL? Our server doesn’t support this protocol yet. This is a temporary solution and is built as a sidecar so any OpenAI client in any language can use it.

Open sourced proxy-to-gemini, a sidecar, for those who want to talk to Google DeepMind models via the OpenAI protocol. Grab an API key and run the proxy. Let me know if you are interested in OpenAI API compatibility in Gemini API more officially. github.com/google-gemini/…

@iamgingertrash @levelsio Please. I have M3 96gb

If you're a local model user, comment under this thread for access. We've made the UI dead simple. You dont need to worry about quantization, inference logic, or anything of the sort. Hit install, and your model is exposed on an OpenAI spec locally. @levelsio, we listened.

We're releasing Look Mom No Cloud today — a Linux + Mac desktop GUI + compiler to run local models We're initially rolling this out as a private beta to collect user feedback and suggestions, and converge on the ideal UX for developers to build local AI apps Threadboii (🧵 1/6)

@TheStalwart The market learned to correctly price in the discounted cash flow of future earnings

So the 'Google is too woke now' was a total fade huh

@spectate_or Some tokenization bug during some training?

llama3-120b randomly creates a new word "prefaceate" that has 0 search results on google

@damengchen How do you know Adobe is doing a branded search term and not a keywordless ad?

Here is what the $230 billion company did for an indie business. I call it a win ✌️

@TheStalwart I mostly buy it based on the expected discounted sum of future cash flows. It feels fairly priced right now

All jokes aside, I do have one question about these memecoin valuations. When you're buying a coin like Jeo Boden, how are you pricing in network risks, and whether fee income for Solana validators will rise enough to compensate for the eventual decline in SOL issuance? Thanks.

@josephsemrai App Store/github link?

inference on your head mistral 7b (4bit quantized) running locally on apple vision pro

@c_gro There is no high watermark for Ackmans fund?

These charts are crazy. $1,000,000 invested in funds led by Cathie Woods & Bill Ackman versus just putting into an S&P500 index fund over the last 8 years. No clue why anyone would trust these funds versus just betting on the market.

This holds true for Cos(emb(t2), emb(t1+t2)) < cos(emb(t2), emb(t2+t1)) as well. So the second string doesn’t matter.

OpenAIs new embedding seems to really prefer earlier text. Cos(emb(t1), emb(t1+t2)) > Cos(emb(t1), emb(t2+t1)) ~97% of the time of the 1000 examples I tried. Interesting retrieval consideration

@jrysana Say more

10 microseconds

Wtf? What is this unknown limit @OpenAIDevs

@EMostaque Stable-tral coming soon?

I mean wouldn’t a 12bn parameter version of mistral 7b trained for like 8 trillion tokens perform like Mixtral, be just as fast and use 4x less vram?

@rohindhar Definitely getting interesting but doesn’t seem insane yet

301 homes with price cuts In the last seven days in West Palm Beach, Florida area

@EMostaque Haha, originally I had 950ms but then changed to s.

@pythonprimes 1s! Give us a bit for 1 ms…

You should get < 1 second an image on a M3 Max If only there was a super lightweight upscaler coming 🤓 Or you could do hi-res tile based upscaling with… stuff Some crazy UIs are incoming that work on your laptops all