Eden Gibson

CTO @ Derive Notes

[Eden Gibson @edengib]

@iykshani Since it reuses the kv cache i assume it would be negligible

[Eden Gibson @edengib]

@findfaraj

[Eden Gibson @edengib]

@skeptrune wouldnt this be a lossy approximaition though?

since jev outputs logprobs and your method outputs a tokenised prediction, it would weight towards 0.9, 0.65, ect. "human" numbers

Still looks useful for 90% of jev's use cases

[Eden Gibson @edengib]

@doodlestein this is sick as

how granular are the commits?

[Eden Gibson @edengib]

A fun use case of jev is adding it as a input parameter to your local models. allows for a sort of tailored intelligence, while deferring the annoying stuff. Its a nice upgrade from what i've seen so far!

[Eden Gibson @edengib]

I recently wrote a post about a version of reward models becoming a powerful part of the future, with intelligent and cheap routing deployed at scale.

Jev is this! (though not a reward model)

The future is here, and faster than i thought (as always)

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Eden Gibson @edengib]

@notkevinzhang Intelligent feature discovery nudging + classifer for canvas content

[Eden Gibson @edengib]

@VictorTaelin Get codex i run 10 accounts 8 claude 2 codex

[Eden Gibson @edengib]

@levelsio I'm in uluwatu building right now and its dead as

Dont see it!

[Eden Gibson @edengib]

@JalenDNelson @onlinedopamine @levelsio @marckohlbrugge yeah plus they're kind of annoying to get to at least where we are

[Eden Gibson @edengib]

@bentlegen I love WASM pre ai it was so annoying now i just use it wherever i can

[Eden Gibson @edengib]

@Martinsmora3 @thsottiaux Oh really? mine doesn’t

[Eden Gibson @edengib]

@ryanirl How did you approach the training process?

[Eden Gibson @edengib]

This is the perfect example of what i’m so passionate about

tiny capable models for tiny (but important!) tasks

[Shu @shuding]

I trained a small model to do syntax highlighting in the browser with GPU.

Meet gpu-lexer from Vercel Labs: Small (27.5KB), fast (runs on WebGPU), and language-agnostic (model guesses the syntax).

https://t.co/adphvt2Lv4

It is experimental and built for learning!

[Eden Gibson @edengib]

yep, in my startup data labelling is fully autonomous.
I have a gold calibration corpus, i run an autoresearch loop to optimise the prompt (usually hits 1.0 F1 if i label consistently) and then apply the model at scale to label a ridiculous amout of data.

Allows for some cool and capable internal models!

[Eden Gibson @edengib]

No better dapamine hit than waking up to a model hill climbing amazingly

[Eden Gibson @edengib]

@Abhhinav_SB @_rockt Dlss 5 uses transformer blocks