Justin

Ex @snap | AI/Eng | ⧉ | ❓

[Justin @justkrup]

Kev hitting 1.0 is great! but I find the hard part of finetuning is getting the training data, so I built a proxy that sits in front of Kev, keeps every request, and fine-tunes from that corpus

On kev's own emotion eval:
kev-27b: 58.75%
finetuned kev-9b: 75% @ < ½ the size

[Jared Palmer @jaredpalmer]

Kev 1.0: an open weight family of decision models you can train and run on your own.

• New Kev-27B + updated Kev-9B
• 64k-token document window on Kev-27B
• TypeSafe SDK compatible
• New fine-tuning + deployment skills so you can train Kev on your own data

[Justin @justkrup]

@ritakozlov ffs… I’m trying to relax on the weekend plz do these drops like Tuesday noon next time

[Justin @justkrup]

@tdinh_me Clef has had image support and @TypeLLM supported reasoning already. Trying to figure out what dimensions this wins on

[Justin @justkrup]

@sashavalegorov @bcdsignature Why can’t you be you and me at the same time?

What science are you referring to? There’s nothing empirical about subjective experience / qualia

[Justin @justkrup]

@tdinh_me Wait till you learn about the retina and occipital lobe

[Justin @justkrup]

@turboderp_ What would showcase the new powers post-Jev most is something that takes in unstructured input along with dynamic options. And the fact that it needs to fit within 36k context window and is uncached for sequential calls has been pretty limiting in that world.

[Justin @justkrup]

@turboderp_ I agree they've been pretty poor demos. Almost all the difficult work is in converting board state into some structured input, which frankly defeats the purpose of these SystemOne models. Because the thing that makes them special is that they do support arbitrary input

[Justin @justkrup]

@ethanCaballero It’s not that I’m certain that current AI models are not conscious, but I do feel much more confident that if they’re conscious so were smaller weights models that existed years ago and so are image gen models

Like idk where you’d draw a line or that there even is one at all.

[Justin @justkrup]

@giovani_ultra @bayeslord Cosmic apotheosis wears off faster than salvia

[Justin @justkrup]

@bayeslord This sounds like something 4o would output when prompted for thoughts on existing.

[bayes @bayeslord]

I've been going to the same place in my mind since I was a kid, and it's never gotten any less mysterious. What is this reality? What is consciousness? How does any of this make any sense at all? What the fuck?

[Justin @justkrup]

@AlexanderKalian Is God conscious?

[Justin @justkrup]

@whyarethis ya know whats funny. If somehow we do manage to observe completely emergent behavior from LLMs that’s unexplainable by the raw math. Then that could be the empirical evidence into free will (which while not necessary for consciousnes, consciousness may be necessary for it)

[Justin @justkrup]

@spicey_lemonade Everyone’s missing the point. I’m not conscious because I have a large clump of neurons (the phrase doesnt even make sense- what’s “I”). I am conscious AND I currently feel things some of which expressed by the action potentials of a large clump of neurons…

[Justin @justkrup]

@maxjendrall @michael_chomsky I mean I’ve been getting tiny finetunes running on my iPhone 15 in 10ms… the round trip time overhead in not self hosting can make a big difference for certain workloads

[Justin @justkrup]

@tonychang430 this is sick! did you finetune laya for your labels or is that zero shot? I've been personally finding Laya to be pretty underwhelming until finetuned

[Justin @justkrup]

@margbaramrikkka @Michael_Druggan I mean you’re also made of atoms…

[Justin @justkrup]

@victormustar hugmez

🙏 🙏

[Andrej Karpathy @karpathy]

Cool eval. Simply ask an LLM “Land or Water?” and give it a latitude and longitude coordinate as text. Ask 16,200 times, plot as image. The models know. From compressing the internet.

[Justin @justkrup]

@jmbollenbacher If you have a smart one it’s easier to ft smaller ones for your usecase. But I do agree 27B is kind of awkward size

[Justin @justkrup]

I open-sourced the proxy: https://t.co/AKzafixq1z

or you can finetune and run your own models with https://t.co/E37KFNSLd7 (GPUs via @modal)

[Justin @justkrup]

Kev hitting 1.0 is great! but I find the hard part of finetuning is getting the training data, so I built a proxy that sits in front of Kev, keeps every request, and fine-tunes from that corpus

On kev's own emotion eval:
kev-27b: 58.75%
finetuned kev-9b: 75% @ < ½ the size

[Jared Palmer @jaredpalmer]

Kev 1.0: an open weight family of decision models you can train and run on your own.

• New Kev-27B + updated Kev-9B
• 64k-token document window on Kev-27B
• TypeSafe SDK compatible
• New fine-tuning + deployment skills so you can train Kev on your own data