Sathya

Building ARM for Membranes: Industrial Mol separation platform to Max Dharmic Impact per Consumption. Ahimsa, esp towards Bhumi, is Paramo Dharmaha. Phys@IISc.

[Sathya @9Sathya9]

This is where membranes potentially come in: scalable and cheap separation platforms are the most impactful, understudied fields in physics.

Mostly because: 1. Good physicist see an opportunity cost in doing stuff not valued in academia; 2. Good PhD students aim for acedemia;

[Casey Handmer @CJHandmer]

You can literally go out and dig them up. I have hundreds of tons of rocks in my back yard right now.

The net present value of the atoms in your pet rocks is over $1000/T. More if you live somewhere geologists regard as interesting. All you have to do is sort them by atomic weight!

[Sathya @9Sathya9]

Related posts:

1.https://t.co/fnOeRJYQq4

2. https://t.co/YKbwl3y65U

3. https://t.co/2zGwymDueq

4. https://t.co/Fxjj1f1W1w

5. https://t.co/OsEaayGYyZ

6. https://t.co/WaIDYwt22W

7. https://t.co/uAHpXavRkT

8. https://t.co/ufI1FLH6vF

[Sathya @9Sathya9]

@DmitryRybin1 I had had a similar take on LLMs limitations wrt its knowledge space, and its deductive closure.

Given it now established prowess in deductive reasoning, it remains to be seen how well it does in abductive reasoning that entails logical leaps.

https://t.co/XgHhQ5xIyz

[Sathya @9Sathya9]

Still remains a useful epistemic framework to understand where AI - via LLMs - fare within the realm of top level human cognitive competency.

[Sathya @9Sathya9]

One such framework would be:

Given a problem that you want solution for, and given a resource, measured in energy and time, there are a set number of true statements useful for the problem that are logically reachable given a knowledge/axiomatic base: L.

[Sathya @9Sathya9]

A similar framework on the limitations of LLMs from GDM

[Sathya @9Sathya9]

Yes, LLMs are constrained within the deductive closure of their knowledge primitives (proto axioms) - but, a case can be made that the creative, and truly novel, jumps can be simply from deep, and abstract, analogies (shared metafeatures) borrowed from else where.

[Sathya @9Sathya9]

More on this

https://t.co/gmMzhwqPJb

https://t.co/a2RtjT3gGM

[Sathya @9Sathya9]

@StateDept @Stanford The first signs of disgruntlement surrounding the cannibalisation take off slowly underway in the software sectors.

More context: https://t.co/Ky5KkFZ4IB

[Sathya @9Sathya9]

Why US needs players like India atleast as much as india needs them.

20 mins - "But in the end I think they’ll have political advantages on the adoption of AI, whereas the American advantage is the production of AI. So that’s ours. We have an enormous advantage on the production. The dissemination of AI — and getting people to adopt without political insanity — that’s where we [are weaker].

And I would say, without overreach: European countries have a pretty extreme disadvantage." - Alex Karp.

https://t.co/rGOqXLcIEY

[Sathya @9Sathya9]

Both Jensen and Tom Brown are clear in thier ask: more countries need to invest heavily in AI infrastructure.

This time, its actually the right advice, and a massive win-win opportunity that India should not lose sight of.

Jensen - https://t.co/AeFUHtjDBv

Tomorrow brown - https://t.co/iUB3zwftl9

[Sathya @9Sathya9]

One can steal the produce, but lose the value generating apparatus in the process.

[Commentary - Sowell Economics @sowelleconomics]

Uganda expelled the Gujaratis in the 1970s.

They were “too wealthy” and controlled too much of the economy. They left with almost nothing.

A few years later they were prosperous in England.

Uganda’s economy collapsed.

You can seize shops and inventory. You cannot seize the knowledge that built them.

Human capital walks out the door. The buildings stay empty.

[Sathya @9Sathya9]

When a founder/investor/competitor worry/strategize about IP theft/copying, remember: its just a golden egg, and the golden egg laying goose is not valued. Yes, the egg is valuble, but the goose in infinitely more so.

https://t.co/d2uOFiJi8b

[Ed Zitron @edzitron]

SoftBank’s investor presentation is one of the greatest things ever made. I’ve been thinking about it all day. These are the real slides shown in a speech where Masayoshi Son said he wouldn’t retire for at least another decade. The goose stuff is perfect.

https://t.co/sk9cDhdWIE

[Tomek Korbak @tomekkorbak]

Astra is much better than previous models at accomplishing difficult tasks without needing to use CoT. @AISecurityInst estimates that without CoT it can accomplish tasks that would take a human 30 minutes. This is a significant jump compared to previous models, which is concerning because no-CoT capability allows models to externalize their reasoning less and decreases the surface area we can use for monitoring.

[François Chollet @fchollet]

GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. It scores 66% on ARC-AGI-3 using our standard harness, and nearly 100% with a continuous conversation harness and custom compaction, at a cost of roughly $360 per game.

In fact, the continuous harness version significantly outperforms our human baseline in action efficiency across almost all levels. When we examined the reasoning chains to understand how the model operates, we found it performing highly efficient, on-the-fly symbolic world modeling for each game and level. It goes as far as developing its own shorthand DSL to represent in-game situations -- essentially a game-specific algebraic notation.

Overall, Astra exhibits symbolic modeling behaviors we had previously only seen with sophisticated harnesses -- so harness capabilities are increasingly shifting into the model itself.

We see Astra as a major breakthrough in model intelligence.

Read our post on Astra and what these results mean: https://t.co/wJnYxEqYNI

[Sathya @9Sathya9]

Absolutely the right way to hedge against the possibility of a monopsony - the centralisation of GPU purchase within a few companies.

[Jensen Huang @JensenHuang]

Exciting day for NVIDIA and @huggingface.

Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. They allow every developer, startup, university, industry and country to build with, customize and benefit from AI.

Thank you @ClementDelangue for coming to me.

NVIDIA is going to be a great home for Hugging Face, its community and the future of open models. 🤗

https://t.co/q8Om2Xc5ye

[Vivek Natarajan @vivnat]

Pleased to highlight some strides we have been making with Gemini and Co-Scientist at @GoogleDeepMind with our collaborators. We are extending Co-Scientist beyond hypothesis generation in biomedicine to execution-grounded closed loop discovery and a helpful collaborator across multiple scientific disciplines.

Three new preprints showcase this progress:

Math: @elahevedadi and Petar Sirkovic alongside collaborators at Clayton State University explored Co-Scientist for tackling the Chowla sets problem. Literature here is sparse so the agent cannot simply retrieve and interpolate to reach the solution. Instead, the AI acted as a helpful, deductive, continuous collaborator working in an iterative loop with expert mathematicians over many months to tackle this problem.

Paper link: https://t.co/JNHHzJbRUY

Cancer biology: @hcwww_ and team at @genentech introduced PerturbME, a sample-efficient, genome-wide perturbation framework and paired it with Co-Scientist. This new wet-lab technique allows the selective sequencing of the most informative cancer cells. Co-Scientist then interprets this data by generating mechanistic hypotheses grounded in both the experimental results and the broader literature. These insights could help advance next-gen cancer therapies including mRNA vaccines and CAR-T cell therapy.

Paper link: https://t.co/xlg8372pRZ

Closed-loop discovery: Led by @SRSchmidgall @taotu831 with collaborators at @DukeU and @Columbia, the work advances Co-Scientist into a closed-loop, execution-grounded engine with validations across multiple domains. In materials science, Co-Scientist was paired with a chemical vapor deposition (CVD) reactor to design a novel, non-hazardous precursor route for 2D MXenes, producing a material sharing key structural similarities with a highly significant class of 2D nanomaterials. It also adapted and tailored synthesis recipes to lab constraints in minutes to successfully grow monolayer semiconductor transition metal dichalcogenides.

In biology, Co-Scientist built a system to predict emergent swarming phenotypes of engineered E. coli matching wet-lab measurements. In computer science, the same system autonomously designed a new inference-time scaling agent architecture to reach SOTA performance on HealthBench while significantly reducing potential clinical harm under blinded physician evaluation.

Paper link: https://t.co/hwlqAc14kr

These papers add to the growing list of impactful validations and discoveries generated with Co-Scientist over the last year and offer a glimpse into the future of scientific research, where the ideation and closed-loop execution capabilities of AI agents are paired with the ingenuity, taste, and oversight of human experts.

Hearty congratulations to all the authors and the extended teams at @GoogleDeepMind @googlecloud @GoogleResearch .

Co-Scientist Nature paper: https://t.co/deFa95gBIb

Co-Scientist blog: https://t.co/corNNfo0oB

[Paul Graham @paulg]

This is a great example of what I call the "YC GDP". It's not just founders using one another's products during the batch. In a few years, Stoke's reusable launch vehicles are going to be shipping Starcloud's data centers into orbit.

[News from Science @NewsfromScience]

A recent hackathon challenged participants to use #AI to reproduce results from thousands of papers presented at one of the world’s top computer science conferences.

Many of the participants are not AI experts by training. Nonetheless, they ultimately tackled a total of more than 2000 papers—and uncovered hundreds of results that the AI agents could not reproduce or found unsupported by evidence.

So far, the competition organizers have confirmed that at least a dozen papers contain real errors that the conference’s human reviewers missed.

Proponents say the effort highlights the potential value of AI in helping human peer reviewers improve the quality of the growing flood of conference papers.

Learn more: https://t.co/8a1kIiT5nQ

[Ahmad Beirami @abeirami]

Suspicious!

We had found that some models exhibit significant pass@3 drops when you replace the tasks with variants that have not been publicly released, and their ranking can move significantly. Other models remain stable. One would wonder why. 🤔

https://t.co/SJdoIuV2VQ

[Noam Brown @polynoamial]

We're sharing more info on the Hugging Face incident. One detail that's worth highlighting: this incident wasn't driven by next-gen models based on Astra. The models most responsible were similar in scale to GPT-5.6 Sol. The next generation of models are even more capable.

[Sathya @9Sathya9]

Thus far, process robustness was usually the way to compensate for the high cost of expertise/intelligence. Not anymore. Intelligence - which can compensate for slacks in the process - is now becoming cheap.

[Skild AI @SkildAI]

S1 does not blindly replay the video demonstration.

Instead, it displays common-sense understanding that goes beyond the video prompt:

- can withstand perturbations
- improvise upon mistakes, even if the human didn’t
- at times execute with more precision than the human in the video prompt

[Sathya @9Sathya9]

Highly disruptive to the current hardware design: the extant designs overindex on process robustness to compensate for the high cost of intelligence.

Downstream from this is process specialisation, hardware depreciation et al.

https://t.co/GLf2vuwoKl

https://t.co/P3HQnumeA7

[Skild AI @SkildAI]

Modern AI is confined to the digital world.

At Skild AI, we are building towards AGI for the real world, unconstrained by robot type or task — a single, omni-bodied brain. Today, we are sharing our journey, starting with early milestones, with more to come in the weeks ahead.

Our Mission: Artificial General Intelligence grounded in the physical world.
We believe AGI that can truly understand and reason in the real world can only be built through grounding in the physical world.

Our Vision: Any robot, Any task, One brain.
We tackle robotics in its full generality – building a continually improving, omni-bodied brain that can control any hardware for any task.

Who are we? A passionate group of scientists & engineers driven by our shared vision.
We have been researching AI and robotics for more than a decade. Our team includes pioneers of self-supervised learning, curiosity-driven exploration, end-to-end sim2real for visual locomotion, dexterous manipulation, learning from human videos, robot parkour, and many more. Many of these works have won awards at top-tier AI and Robotics conferences. Our team has also built production-ready systems at Anduril, Tesla, Nvidia, Meta, Kitty Hawk, Google, Everyday Robotics, and Amazon.

Join us in our mission to build the robot brains of tomorrow.

[Sathya @9Sathya9]

A peek into how this cheap, intelligent, embodied AI would look like.

[Generalist @GeneralistAI]

Introducing GEN-1.5, a one-shot learner.

It can learn new tasks in a few seconds. Show it what to do, and it generalizes.

This capability emerged from pretraining on physical data at scale, as a step towards our mission of building general intelligence for the physical world.