Harrison Kinsley

gpus and tractors.

Director of AI and Engineering @ networks from Scratch book:

[Harrison Kinsley @Sentdex]

Lucky robots fellowship is for researchers who are looking for a simulator that bends to their will, not the other way around.

we're looking to partner with a select few researchers in a paid fellowship to push robotics research into the next frontier

[Harrison Kinsley @Sentdex]

@travisk src: https://t.co/AwlSZyaCbu

Distillation is a training technique. Stop letting fear mongers tell you distillation is somehow a bad/wrong thing.

Especially when those same people are arguably infringing on copyright and IP theft on the greatest scale of all time.

[Harrison Kinsley @Sentdex]

@travisk Thinking Machine's (US AI company) latest release, Inkling:

[Harrison Kinsley @Sentdex]

@tarat_211 hopefully soon!

[Harrison Kinsley @Sentdex]

Good morning!

[Harrison Kinsley @Sentdex]

@Gibson61350 Ooh thatd be a cool one actually. Will have to do that one of the days

[Harrison Kinsley @Sentdex]

@deanwball @xlr8harder you're gonna have to decouple the idea that capex == accel.

They often go hand in hand, but OP said accel and all you can reference is capex. We understand you're feeling OAI contraction, but there's more to accel of AI than how much $$$ we put into your company.

[Harrison Kinsley @Sentdex]

@deanwball If open-weights are decel and decreases capex why is capex in AI going up in China and why are the models accelerating faster than USA?

Capex is also a measure of capex, and not necessarily accelerationism. 2 or 3 companies getting all the capex is decel.

[Eiso Kant @eisokant]

ASAP (As Sparse As Possible 😂)

[Harrison Kinsley @Sentdex]

@ValtteriValo Can i find this env somewhere? Would love to tinker

[Harrison Kinsley @Sentdex]

@TheAhmadOsman you're the best, thanks for sharing this. A while back I remember the nightmare of trying to get gh200s to work. figured gb300 would be similar.

Seems way better off building a 3-4x rtx pro 6k machine + gobs of RAM.

I've got a machine i could potentially do it with.

[Harrison Kinsley @Sentdex]

Such an iconic scene in Breaking IPO

[Lucky Robots @luckyrobots]

Coming soon to the real world!

[Harrison Kinsley @Sentdex]

@moochieball yeah idk. i need more details from dario about what a distillation attack is vs regular distillation vs crawling the entire internet including webpage and books that dont wanna be parsed but I'm sure he's got an explanation

[Harrison Kinsley @Sentdex]

Anthropic's waiting eagerly to see the open weights license to find out if they can just do a regular distillation or if they need to do a spooky distillation attack

[Arena.ai @arena]

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

[Harrison Kinsley @Sentdex]

@TheReapsZ I'm inclined to often agree about that wrt oai, but GPT OSS 120B was a darn good model that stayed useful for quite a while. Even today it's got a good use.

[Harrison Kinsley @Sentdex]

Now would also be another great time to release an opensource successor to GPT-OSS-120B

[Harrison Kinsley @Sentdex]

Now would be a great time to release an open source successor to GPT-OSS-120B.

[Harrison Kinsley @Sentdex]

@meltingdiodes @yacineMTB don't know yet, haven't used. But I will!

[Harrison Kinsley @Sentdex]

@ZixuanLi_ GLM 5.2 will go down in history as the turning point

[Harrison Kinsley @Sentdex]

@ValtteriValo certainly not arguing sol or inferno are simple RL tasks, it just seems like zuk is an inherently harder RL task compared to sol, unless you break it up into HP thresholds.

did you find either to discover novel tactics you would actually consider using?

ironman btw

[Harrison Kinsley @Sentdex]

@_hasanfaesal Not even 8 RTX Pro 6Ks will handle this one. This is like... maybe 16 RTX Pro 6Ks type situation, and even then we'll see what the token rate would be.

Think this model's really gonna require a server node unfort, but this model will absolutely be distilled smaller!