Attack Oranges

[Attack Oranges @iattackoranges]

ygmi

[Ivan Fioravanti @ivanfioravanti]

The more I play with Claude Sonnet 3.5 the more I think: “What is the role of Human Intelligence in 2-3 years from now?” 🤔

[Attack Oranges @iattackoranges]

it's a spectrum between 'ai systems can literally help us in no way' to 'ai systems literally develop their hardware/software entirely', and we're inching closer to the latter

[Trevor Levin @trevposts]

I feel like the term "recursive self-improvement" has grown from a truly dangerous thing -- an AI system that is sufficiently smart and well-equipped that it can autonomously improve *itself* -- to "any feedback loop where any AI system is useful for building future AI systems"?

[Mario Cannistrà @Blueyatagarasu]

Explains a lot.

Of course I'd want to accelerate if I didn't think superintelligent AI was even possible.

[Jing Yu Koh ✈️ COLM'26 @kohjingyu]

LLM agents have demonstrated promise in their ability to automate computer tasks, but face challenges with multi-step reasoning and planning. Towards addressing this, we propose an inference-time tree search algorithm for LLM agents to explicitly perform exploration and multi-step planning in interactive web environments.

It is the first tree search algorithm for LLM agents that shows effectiveness on realistic and complex web environments: on the challenging VisualWebArena benchmark, applying our search algorithm on top of a GPT-4o agent yields a 39.7% relative increase in success rate compared to the same baseline without search, setting a state-of-the-art success rate of 26.4%. On WebArena, search also yields a 28.0% relative improvement over a baseline agent, setting a competitive success rate of 19.2%.

[Jan Leike @janleike]

I like the new Sonnet. I'm frequently asking it to explain ML papers to me. Doesn't always get everything right, but probably better than my skim reading, and way faster.

Automated alignment research is getting closer...

[Alexander Kruel @XiXiDu]

Nat Friedman interviewed the Sora creators: https://t.co/XYG3KBfJCK

[Aryaman Arora @aryaman2020]

bruh why do the 3 biggest llm companies have their interp teams working on the same thing independently

[Attack Oranges @iattackoranges]

...holy fuck.

[Google DeepMind @GoogleDeepMind]

With @Harvard, we built a ‘virtual rodent’ powered by AI to help us better understand how the brain controls movement. 🧠

With deep RL, it learned to operate a biomechanically accurate rat model - allowing us to compare real & virtual neural activity. → https://t.co/GaToq3AWTQ

[Attack Oranges @iattackoranges]

holy fuck

[Sam Ro 📈 @SamRo]

don’t think I’ve ever seen a chart like this

[Cody Fenwick @codytfenwick]

You used to hear critics of AI hype say "Human-level AI is always supposedly 20 years away." They don't say this anymore.

[Attack Oranges @iattackoranges]

appears to be the case

[fofr @fofrAI]

Runway is gonna release sora-grade video tools before OpenAI

[serena shahidi @glamdemon2004]

A lot of you guys are miserable because you don’t do arts and crafts. You should be having arts and crafts time at least once a week

[Attack Oranges @iattackoranges]

i have gotten this vibe from a lot of e/accs

[ControlAI @ControlAI]

Google DeepMind CEO Demis Hassabis on AI accelerationists:
“They don’t actually fully understand the enormity of what’s coming. Because if they did — I’m very optimistic we can get this right — but only if we do it carefully and take the time needed to do it.”

[alex fazio @alxfazio]

Aravind Srinivas says the role of CEO is highly replaceable and future AI like GPT-5 or 6 could perform CEO tasks better by continuously processing information and making decisions.

[tetsuo @tetsuoai]

George Hotz says everyone he's met who can program well learned it the same way: they had an idea, and then they built it.

[Gytis Daujotas @gytdau]

Model interpretability techniques allow us to understand and control what models do. Here's how Sparse Autoencoders give us general fine grained control on image models:

[Andy Masley @AndyMasley]

I think about this a lot

[Attack Oranges @iattackoranges]

jimmy apples basically confirmed this

[Nicholas Dunzelman @nicdunz]

bro dalle 3 getting wild wtf. why is it so good? these dont even look ai generated they look like theyre from a movie or something. dalle 3 always getting iteratively better, openai just never says anything about it.

[Attack Oranges @iattackoranges]

damn

[Micah Goldblum @micahgoldblum]

Virtually all large models today contain huge matrices, and these dominate their compute. By incorporating structure in these matrices, we can improve the performance/compute tradeoff!

[Attack Oranges @iattackoranges]

ok i might have to check this out now

[tldraw @tldraw]

finally we can do maths