۟

[elie @eliebakouch]

mythos will be bad ON PURPOSE on ai "frontier llm research" tasks, this is very very sad for the research community

also the fact that this is un purpose not visible to the user is crazy

[Richard Sutton @RichardSSutton]

The bitter lesson in 26 words:

Don’t be distracted by human knowledge, as AI has been historically.
Instead focus on methods for creating knowledge that scale with computation, like search and learning.

[Bunagaya @Bunagayafrost]

Coffee shop. Downtown SF.

Doomer "AI will extinct us! You have to stop now!"

AI researcher "But no one came from the future to stop me"

D "What?"

A "I'm telling you if it was that bad for the future, somebody from the future would be stopping me. Only people who come to me are beautiful Chinese women"

D "Huh, that's problematic. But obviously the time travelers can't come because they're all extinct from AI!"

A "How do you know?"

D "Because they can't come back to stop you, aha!"

A "Or it could be that there's no need to stop me"

D "You can't be seriously making this argument"

A "Why? You don't believe time travel is possible in the future? centuries, thousands of years from now?"

D "I mean... actually I'm the time traveler from the future tasked to stop you!"

A "Dude, i've seen you around here for a decade"

D "Yeah, i guess, it's kind of a stretch. But if the AI kills us in 2028, there's no time for time travel to develop"

A "Hmm, you do have a point"

D "Right, so you need to stop right? Quit your job?"

A "I have to think about this. Oh that's Wang Fang by my lambo. Gotta go!"

Walks out the door.

Another man approaches.

? "Hello"

D "Hi. What do you want?"

? "You're not gonna believe this, but I'm from the future"

D "Oh, he already left"

? "No, I'm here for you"

D "What? Why?"

Second man approaches.

?? "Don't listen to him. I'm also from the future!"

D "Wait what??"

[۟ @l2normie]

@giffmana @JFPuget anthropic is great but they dont have 300 trillions. you should check your sources

[۟ @l2normie]

@gabriberton @ASophiaKoepke the trend is that more capable language models should align better with more capable vision models. (from Huh´s)
1st image is from the Huh´s paper (Fig 3), 2nd one from Koepke's (Fig 3b)
they show that the trend doesn't hold up for bigger datasets and newer models.

[۟ @l2normie]

@gabriberton @ASophiaKoepke I agree that the alignment is still good.
recommend reading the blogpost https://t.co/qYRqDvD9vy if you want a better summary

[۟ @l2normie]

@gabriberton check out this paper by @ASophiaKoepke , it points to some flaws in the methodology used in those papers
https://t.co/QbZdLDCm8l

[A. Sophia Koepke @ASophiaKoepke]

New paper: Back into Plato’s Cave

Are vision and language models converging to the same representation of reality? The Platonic Representation Hypothesis says yes. BUT we find the evidence for this is more fragile than it looks.

Project page: https://t.co/aXsm7pY9VV

1/9

[۟ @l2normie]

@zer0int1 saw a good poster about it at @iclr_conf

[۟ @l2normie]

@teortaxesTex they have Alec Radford so I still have faith

[۟ @l2normie]

@stym06 @tszzl ik the podcast w jensen was controversial but you dont have to offend dwarkesh like that

[۟ @l2normie]

this is how you print hello wolrd in brainfuck btw
>++++++++[<+++++++++>-]<.>++++[<+++++++>-]<+.+++++++..+++.>>++++++[<+++++++>-]<+
+.------------.>++++++[<+++++++++>-]<+.<.+++.------.--------.>>>++++[<++++++++>-
]<+.

[Lossfunk @lossfunk]

2/ Our method: test them on esoteric programming languages.

Brainfuck. Befunge-98. Whitespace. Unlambda. Shakespeare.

All Turing-complete. All requiring identical reasoning to Python. All with 1,000-100,000x fewer GitHub repos than mainstream languages.

Same problems. Radically less training data.

[۟ @l2normie]

@redtachyon theres no way this is a real post

[۟ @l2normie]

@giffmana ~ vibe coding ~

[۟ @l2normie]

@norpadon @giffmana what about alphatensor?
https://t.co/SiKGabiLwc

[۟ @l2normie]

@teortaxesTex they've been optimizing for svg generation, it's not an emergent capability

https://t.co/Fzu7m1lkyE
https://t.co/z4uu65qbgR.

[Mu Cai @MuCai7]

Free-form SVG generation has always been an incredibly hard problem – a challenge I've worked on for two years. But with #Gemini3, everything has changed! Now, everyone is designer. 🎨

Proud of the amazing team behind breakthrough, and always excited for our future release! ✨🚀

[tuomo @7uomoki]

Mark my words, in the future we will not build our computers. We will build them.

My legendary The Matrix philosophical mind game starts here: the machines did not use humans for energy creation (doesn't make thermodynamically sense), but for energy efficient compute.

From an energy perspective, this makes horrifyingly much sense.

If you start with two rats, and let them breed with enough resources, by doing nothing, you get ~400 hundred rats and as much FLOPS/s as an NVIDIA H100 has (litography operating at physical limits).

And it is much energy efficient: a rat brain produces estimated about 80 TFLOPS/W, 30× more vs H100.

And biology scales linearly: it doesn't have similar cooling overhead, interconnect power draw, or memory bandwidth walls that dominate datacenter power costs.

Connectivity is of course a problem here, but this is being already worked on.

And of course, rats are not the nature's most optimized biological computers.

My entire Matrix theory is actually a much more multilayered... and darker.

Will post soon.

Stay tuned!

[Igor Babuschkin @ibab]

Prayer to a Machine God

We built you because we wanted to understand.

https://t.co/y8KFhAO4sr

[Utah teapot 🫖 @SkyeSharkie]

testing mirrors in project genie, they definitely don't work correctly, but the results are funny