@VictorTaelin @CozendeyMath Is package management something on the horizon or no?
English
jacob
731 posts
Vitalik again pitching the ideas I desperately tried to explain to him while at the Ethereum Foundation. I'm pulling my hairs rn ):
Anyway it is coming alive next week or so™
(Bend2 is done, Fable is now emulating users so we catch bugs and improve the UI / UX before launch)
vitalik.eth@VitalikButerin
A new type of "high-level programming language" that seems really worth trying to make, is a language that gets compiled to Lean (or HOL, or...) that is specifically about making it as friendly as possible for a human to read definitions and theorems. Not the proofs - as all that matters with proofs is that the proofs are correct - just the definitions and theorems. The intended use case is that AI outputs a blob of proofs, and you're trying to make it as easy as possible for anyone reading the output to understand what the actual precise claims are that have been proven
English
showing an RSA pubkey to an ultrafinitist: "you claim that this number has factors, but your claim does not even make SENSE. it is not merely wrong. it is gibberish! look here. if it would take 2^64 years to factor, then discussion about it is meaningless. it is not REAL"
faw@fawwazanvilen
an ultrafinitist on whether 2¹⁰⁰ exists
English
@teortaxesTex You forget the soviet dream isn't to BUY 4TB of flash, you buy the thing that MAKES the 4tb of flash on your 40 year old kitchen fridge.
English
I really hope this ever becomes real
but the thing is, a wafer is a wafer, it has finite area
we can tweak the architecture, but it'll be really hard to make terabytes of high-bandwidth memory for everyone
Tim Sweeney@TimSweeneyEpic
One element of the strange climate now is the scale and inaccessibility of data centers versus PCs. But post-memory-disruption, PCs will have 4TB+ of Flash storage with ultra high bandwidth and we'll just run these models locally and see they're no big deal.
English
@VictorTaelin Maybe a B2B license, adoption with rewrites now are a lower ceiling, so providing documentation+tools for rewriting codebases to Bend2 might be a good onramp. Have you looked up ferrocene.dev prices?
If Ds4 rewrites be viable, you could rewrite a popular lib like libssl
English
Bend2:
- Programming language
- Near C speed (on CPU)
- Near CUDA speed (on GPU)
- High-level (JS-like closures, objects, recursion)
- Lean-like proofs (force AI to write proofs = you get 0 bugs)
Everyone is using it.
How much would you pay to use it?
Be *honest* either way
English
@suspected_n00b check my comments. lots of people unfortunately.
English
Why do you think hosting Chinese open source models on your own GPUs is less trustworthy and secure than giving away your data to anthropic and being shut down access at any time because they think you are doing something they consider “not safe”?
English
@notmahi @bhawna_paliwal_ @HarithejaE @willjhliang @pabbeel @JitendraMalikCV @kyutai_labs Just curious, is it dumb to ask if predicting video frames of a person doing stuff not equivalent to robot behavior, then why not just generate egocentric video and derive actions from that?
English
Lots of gems like this in the full work, led by @bhawna_paliwal_, @HarithejaE, & @willjhliang. Also special thanks to @pabbeel & @JitendraMalikCV for advice, and @kyutai_labs for the compute support.
Try our code today:
📝: arxiv.org/abs/2606.19333
🌐: do-as-i-do.com
⚙️: github.com/malik-group/do…
English
Robots are the bottleneck in scaling robotics, and learning from human video promises to solve it. But how can chaotic human data ever measure up to sanitized, lab-made teleoperation data?
Introducing Do as I Do: establishing a much needed correspondence between human videos and dexterous robot data. Some fun insights below: 🧵
English
@latrasis You don't understand how CRV works at all do you? You generate a target list, you give the numerical representations to the viewing model, you ask it to give you the data it senses raw, you then quantize fit, adjust and repeat. None of this is guessing another's output.
English
Fun fact, you can use AI to remote view with statistically significant results.
English
@IndSocEnjoyer What does this have to do with being cheap? You’re trying to predict llm output using another llm, this is literally non-random and thus invalid
English
@latrasis Yes that works better, but you can do it with the low end and i'm cheap.
English
@IndSocEnjoyer Using llm for targets sounds dumb, more believable if uses actual random binary source masked with pre-selected unique targets and a deterministic+cached selector
English
Ingo Swann's methodology is easy to replicate. Develop an approach where you can confirm and quantize results and failures, reinforce success, use another model to generate targets, minimize human involvement.
English
@teortaxesTex silly little stories that exist as smirks to a black tie, doodles on a frescoe
English
In Greg Bear's Hardfought, humans and Senexi – ancient, hiveminded, ammonia-based entities – evolved to be similar in the course of billions of years of war.
At the end of Tsutomu Nihei's Zeb-Noid, humans and flies even became reproductively compatible.
So with the US and China.
nic carter@nic_carter
If a top AI CEO in China told the CCP to go kick rocks when they asked for help, that CEO would be instantly sent to prison. This is the correct approach Letting AI CEOs play politics and dictate policy for the military and soon the entire country like their own personal fiefdoms is appalling and undemocratic If Trump doesnt bring Dario to heel now, we will simply end up completely subjugated by him and his lunatic EA buddies
English
@bayeslord Idiotic, the bottleneck is production of useful compute not compute or weights themselves
English
***warning: the reading comprehension skill required to understand this post is above the fourth grade level***
bayes@bayeslord
Open source AI is now moving into the territory of capabilities where it could start generating meaningful negative externalities, and we should probably at least talk about it. It's a popular idea because freedom to compute is good and we want a balance of powers, and we don't want an intelligence cartel. I am, obviously, quite sympathetic to these ideas. But there could, for example, be intermediate solutions that don't immediately diffuse power in all directions. Purely as a matter of public relations the AI industry cannot afford to be generating negative externalities.
English
I'm returning my Mac Mini
In its place I bought a $10k Mac Studio (+ employee discount)
My ClawdBot Henry will be controlling an AI super computer
Henry will use Opus as a brain, and multiple local models as a swarm of employees
We are entering unexplored territory. Buckle up
English
@cocktailpeanut @deepbeepmeep The video is ok, but the jazz interpretation really kicks, the rhythmic-studder could easily pass as intentional
English
WanGP Update -- Flux Klein, GGUF, Video-to-Video with Audio Transfer, and more!
WanGP, the all-in-one image/video generation gradio app from @deepbeepmeep just made another update!
My favorite update: Transform any video to another INCLUDING the audio track!
deepbeepmeep@deepbeepmeep
WanGP 10.42: More goodies on the way - Flux 2 Klein 4B & 9B: in tiny models get the best of both Qwen Edit & Z Image - LTX2 can now extract both video and audio from a control video - Models Manager: identify easily which files to delete among your 3TB of checkpoints - GGUF support for lighter models in RAM - RAM optimizations so that you can generate multi minutes videos github.com/deepbeepmeep/W…
English
i dont know if my skeleton is evil or not yet. kind of anxiety inducing. what if i die and it's regular and it doesn't attack adventurers or anything
English
why does gemini have like a full-on anime rivalry with claude
AI Notkilleveryoneism Memes ⏸️@AISafetyMemes
An engineer showed Gemini what another AI said about its code Gemini responded (in its "private" thoughts) with petty trash-talking, jealousy, and a full-on revenge plan 🧵
English
