FYI I use my Codex models with Claude code and they work pretty great. It can even pick up on existing conversations w/ Opus and then resume w/ OAI models.
They're weighted differently. Currently it's set to: reposts 2.0x, replies 1.5x, bookmarks 1.2x, likes 1.0x, and clicks 0.5x. These are guestimates based on public info.
It's real embeddings (MiniLM, truncated to 128-dim), so the cosine similarities between tweets are genuine, but it's obviously not X's actual trained model. The relative distances are meaningful, though.
It should technically work (everything is being routed via LiteLLM under the wraps) with Mistral, but it's untested haha. I don't think Mistral's lost the race; it's just that it doesn't seem to be that popular relative to Gemini/Anthropic/OpenAI, so I didn't bother testing it.
The 'Fast'/embedding model, unless the input is absolutely clear, snaps to the closest alternative.
Here's the song it created when I asked Gemini Flash to do it!
It called it, "Cobalt Kinetic" and explained that, "A high-velocity Mixolydian engine fueled by syncopated electronic pulses and glitchy accents, evoking the iridescent blur of hidden magic weaving through a midnight skyscraper canyon."
It's just that the 'Fast' model picks up from a library of existing tracks using a small embedding model.
If you're really curious, you could try the 'Custom LLM' feature - specifically 'Gemini Flash'. Here's the track it created when I gave it the same input.
It titled it, "In the mood for country cowboy-ish music played for someone like John Wick bleeding out on a cold, snow-covered park bench." and here's the justification it came up with, "A weary, grit-laden Dorian lament for a fallen gunslinger, blending the cold isolation of vast reverb with the sharp, intimate pluck of a final, fading heartbeat."
It's just that the 'Fast' model picks up from a library of existing tracks using a small embedding model.
If you're really curious, you could try the 'Custom LLM' feature - specifically 'Gemini Flash'. Here's the track it created when I gave it the same input.
It titled it, "Throne of the Eternal Night" and here's the justification it came up with, "A glacial, minor-key procession of heavy brass hits and deep sub-pulses, evoking the suffocating grandeur of an ancient, blood-soaked dynasty ending in shadow."
BTW - we used Gemma 270m model - not the 1B model. It's sheerly the size - I wanted to see if I could get a really, really tiny LLM to generate coherent music. Tbh, it didn't quite work as well as I expected. It barely beats a randomly generated track.
In fact, the 'fast' model (literally embedding lookup over a pre-generated library of music ... generated using Best-of-N on Gemini Flash) beats nearly everything - including Gemini Flash, Claude Opus, Gemma models.
Yeah ... that's a bit of an issue - it's because I built a custom simple synthesizer that's not quite as good as a full blown one. Thank you for the feedback! This is something I intend to tackle in upcoming releases.
Yeah! That's one of the reasons I've exposed the Python SDK ... you could, in theory, even attach some sort of DAW that manipulates the music in real-time.
This is amazing. I paid for Phind many months ago for coding QnA. However, this is a much better product and even beats Perplexity, ChatGPT premium, etc. Congratulations to the Phind team!