rar00

> the researchers noticed a group of neurons that consistently signaled the wrong response. Instead of fading as learning improved, these neurons grew stronger, and occasionally even nudged the model toward an incorrect decision.

> “It’s counterintuitive,” Miller said. “You’d think neurons that signal the wrong pathway would go away with learning.”

Except they don't tweak weights when the model is incorrect, I'm puzzled why he's making that claim?? In equation 32 they show weights are adjusted in the form `dW = K * S * (Wmax - W) * F(A)`, where F(A) is the feedback given the chosen action A (i.e. reward). K is positive and S non-negative, so Weights can be nudged towards a maximal absolute value Wmax in the direction of F(A).

However, they set `F(A) = {1 if A is correct and 0 when incorrect}`. That means weights don't change when the model is wrong and can only be potentiated (`sign(Dw) == sign(F(A)) == +`)

Paper: https://www.nature.com/articles/s41467-025-67076-x

Really cool! An affective basis of homeostatic drive seems promising.

Have you performed any basic evaluation / test of your approach?

I'm also curious if there was any deliberation between pursuing "thinking" (language modality) versus "behaving" (visual modality)?

typo? Rounding it up to 2 billion, 30% means 600 million per year

hmm, can't tell if complete bullshit or a work of genius.

On the one hand, the approach overlaps a lot with my thinking, and has some original tweaks (like the emotionally valenced reward signals). Saying that as someone from a robotics/AI background nowadays involved in GenAI, with a few years of phd research on NeuroAI, curious about molecular neuroscience and the Free Energy Principle (as conceptualised by Karl Friston and Mark Solms).

On the other:

- this plausibility dilemma is the hallmark of LLMs

- has all the buzzwords imaginable

- no code, no raw outputs, no official confirmation (by ARC)

- Agentic AI play, walled demo page

I might just be too hopeful (and gullible)...

I started reading two recent neuroscience books (elusive cures and natural neuroscience) that while have different goals both highlight the utility of systems neuroscience. In elusive cures the author presented a brief history of the evolving ideological currents where neuroscientists first only cared about about the specific brain region where a stimulus or disorder is happening (first-order effects), then decades later realised the importance of downstream and upstream brain regions (second-order), and are finally coming to terms that the brain is a complex system with coupled regions (third-order).

Seems the article is a contemporary example of the first->second-order realisation...

“This was a very unexpected finding given the current assumptions about how psychedelic medicine works”

"Surprisingly, psychedelic treatment was still able to strongly boost connectivity onto these neurons”

Knowing (those types of) psychedelics bind to serotonin receptors scientists studied neurons with such receptors and didn't focus on the others. Their study looked at other neurons and found plasticity changes there too.

"As [Essential AI Labs (founded in 2021 by Vaswani)] changes focus, Vaswani is asking investors for at least $150 million."

expected, yet still funny. Noting that their initial aim was to capitalise on transformers to create business tools after GPT-3 came out.

I know people are pushing back, taking "only" literally, but from a reasonable perspective what causes LLMs (technically their outputs) to give that impression is indeed the crux of what holds progress back: how/what LLMs learn from data. In my personal opinion, there's something fundamentally flawed the whole field has yet to properly pinpointing and fix.

That's an orthogonal concern IMO. Running a batch in Europe is about tapping into another source of opportunities. There are plenty of founders that won't or aren't able to attend YC in SF

disagree, there are a few organisations exploring novel paths. It's just that throwing new data at an "old" algorithm is much easier and has been a winning strategy. And, also, there's no incentive for a private org to advertise a new idea that seems to be working (mine's a notable exception :D).

yep, even with greedy sampling and fixed system state, numerical instability is sufficient to make output sequences diverge when processing the same exact input

This argument works better for state space models. A transformer would still steps context one token at a time, not maintain an internal 1e18 state.

getting a billionaire to finance your startup with 9-figures for no equity... :D

It is a sensible position. YC and VCs are backing businesses, not charitable causes or research initiatives. It is the founders' responsibility to liaise the two sides in order to signal that the pursuit of their particular purpose is an undeniably attractive and fast-growing investment. Which obviously entails more work and has fewer market opportunities compared to the case "money is the purpose".

After getting backed and receiving adequate funding, all that matters is maintaining a good growth rate to remain a purpose-driven business.

the robot arm demonstration video jumps at the 00:28s mark...

Model weights already available in HF, code to be released shortly in GitHub (https://github.com/nvidia-cosmos/cosmos-predict2)

aligns with (or is based on) Demis Hassabis' assessment from yesterday on missing cognitive capabilities for AGI: long-term memory, reasoning, hierarchical planning. He then goes on to suggest scientific creativity may be essential.

https://www.youtube.com/watch?v=yr0GiSgUvPU https://news.ycombinator.com/item?id=42817089

Says that new cognitive capabilities are needed to attain AGI: reasoning, hierarchical planning, long-term memory, inventive creativity.

Interesting that LeCun also mentions the same needs in https://techcrunch.com/2025/01/23/metas-yann-lecun-predicts-...

> However, we have a verbal agreement that these materials will not be used in model training.

That'll do it... clearly no incentive to do otherwise. There should be some form of academic penalty for this kind of (feigned) naivety.

nice way to spoil the fight result to people. Because, of course, an athlete gaining more than its adversary is HN-worthy news...

Paper: https://arxiv.org/abs/2501.00663

They're planning to. The prime minister wrote some points on reducing bureaucracy in NHS/public services through AI his FT article: https://archive.ph/MpOOZ

maybe it's worth specifying in the title this is written by UK's prime minister

with human supervisors is the plan. Instead of X workers, they now have X agents + Y<<X supervisors to unblock (and collect additional data for) agents

it's hard to escape the 'gimmickry' or narrow purpose in a cost-effective manner to allow the company to survive long enough and reach large-scale deployment.

The reason is mixture of hardware and software constraints. You need a range of sensors and equipment (end-effectors, batteries, GPUs), expensive at lower volumes, to extend the robot's physical capabilities (e.g. reach, manipulation, navigation) and enable certain software robot skills. Besides their dependence on hardware, robot skills are not entirely solved nor general enough to work in all environments, that means the company needs to do R&D and data collection, or purchase bought elsewhere. For example, Generative AI models (LLMs, VLAs, world models) are a boon for robotics thanks to knowledge reuse and eased domain adaptation but they're (for now) somewhat unreliable. It's difficult for such embodied GenAI models to be more than technically correct when performing tasks because they lack or ignore knowledge about the physical world needed to ensure risk-free actions and outcomes.

For example, asking for a robot to "pour water on that glass" can lead to dropped bottles/glasses or water pouring on a table because the model won't have a clear models of bottle/glass/water ("entities") nor expectations (nothing broken, nothing wet; only what is more or less expected with the act of pouring water conditioned on the most probable areas for representing the of object of interest.

Just have a look at 1X's videos, a well-funded humanoid robot startups, and pay attention to object interactions: how those interactions start and end.