Abhinav SB

• Cofounder & CTO at Nexomeric
• Building AI for Science
• (e/acc)
• INTP-T person

[Abhinav SB @Abhhinav_SB]

The current schooling system is highly weighted toward and optimized for offline policy learning but what we actually need is a better balanced training mechanism with online policy learning.

Also, one more thing to note is that the teaching agent should not distill its own misaligned political opinions or worldview into the student’s model.

[Mustafa. @oprydai]

if you’re drawn to:

richard feynman
alan turing
claude shannon
john von neumann
nikola tesla
seymour papert
marvin minsky
judea pearl
dennis ritchie
donald knuth
rodd brooks
john carmack
dieter rams
elon musk
bjarne stroustrup
steve jobs

you’re not just into tech; you’re into the art of engineering,
the philosophy of systems, and the beauty of how things work.

you’ve found your people.

[Paul Graham @paulg]

A CS undergrad asked me where he could have most effect in the AI age. I said probably at either extreme: either close to the technology, actually making LLMs, or close to the customer, using AI to give them exactly what they want. Or maybe both if you can stretch that far.

[Abhinav SB @Abhhinav_SB]

@Yuchenj_UW Opus beating Fable? 💀 Pacing the frontier by accelerating everything behind the frontier.

[Markus J. Buehler @ProfBuehlerMIT]

Fascinating AI swarm dynamics: a few agents spontaneously emerge as highly connected hubs, while most remain locally connected. The swarm develops a strongly heterogeneous interaction topology with a long-tailed degree distribution - an emergent organizational structure arising from initially decentralized local interactions. There is no central planner assigning roles; the swarm builds its own coordination architecture, with information brokers and increasingly global integration emerging from local behavior.

[will depue @willdepue]

i've been surprised at the response to Jev, but it makes sense in retrospect. sure it's just a classifier but it's a zero shot classifier with frontier-ish intelligence. i'm surprised someone hadn't built it before. i wonder what other old ML ideas are also worth rescuing

[Andrej Karpathy @karpathy]

@willdepue I see it as a point on the LLM pareto optimal curve in a regime that had a large revealed latent demand (no thinking, single token, low latency acceptable intelligence) that was under-invested into because of a race to higher intelligence.

[Nandakishor m @Nandakishorm1]

Asking everyone to fine tune https://t.co/mwPvWGPY2n

Just use any kinda agentic harness like claude code, codex etc. Change the code and use free kaggle GPUs to train your own use cases. Base model has limitations

[Abhinav SB @Abhhinav_SB]

@45surf @VergaraLautaro https://t.co/34FFcoDNEH

[Abhinav SB @Abhhinav_SB]

@CrazyShyyt @thesupermanmx Btw the first successful powered, sustained, and controlled airplane flight took place on December 17, 1903, which is 9 days after this news

[Markus J. Buehler @ProfBuehlerMIT]

Join us at the MIT Media Lab for the ScienceClaw Hackathon, building the internet of agents for science. AI is becoming a collaborator in the real world - designing materials, creating instruments, running experiments, and connecting its capabilities with those of other agents. AI adapts as a problem unfolds - revising its reasoning, learning how to collaborate, and assembling scattered pieces of knowledge, evidence, and raw capability into solutions to some of the hardest challenges in science, technology, and innovation.

Teams will connect AI agents, models, simulations, robots, cloud labs, and scientific tools into functioning systems to tackle problems in protein design, robotics, materials, manufacturing, experimental science, and beyond.

The challenge is to turn ideas into tested results - and demonstrate how agents collaborating across teams and disciplines can accomplish more together. Explore how agents can specialize, challenge one another’s assumptions, learn from failed experiments, and combine their expertise to solve harder problems. Show how collaboration changes what your system can discover, design, or build.

One agent's discovery becomes another's starting point. A tool built by one team enables an experiment by another. Connect your team of AIs, share a capability someone else needs, and build on what others have learned, produced, built. The ambition is an internet of agents through which scientific knowledge, tools, and capabilities can grow, evolve and be utilized across teams and institutions.

🗓️ When: October 30-November 1, 2026
📍 Where: MIT Media Lab

Bring your expertise, your tools, and a problem worth solving - or find one. Help build AI that can contribute to science through what it can discover, create, and make work.

[Abhinav SB @Abhhinav_SB]

@Sentdex I don’t think not having chatgpt in college was a gift.

The best case scenario for getting the most out of college would probably be having chatgpt while not having a smartphone even if that’s an unrealistic scenario.

[Elon Musk @elonmusk]

@Benthamsbulldog K3 civ or bust. The bugs can come too.

[Abhinav SB @Abhhinav_SB]

@ProfBuehlerMIT Cool!

[Wikingenieria @Wikingenieria]

Humor que solo los ingenieros pillamos

[Abhinav SB @Abhhinav_SB]

@Mr_Salio I'm more interested if they are releasing a v5 flash.

[Abhinav SB @Abhhinav_SB]

@iScienceLuvr May be a model’s native harness adding some unnecessary scaffolding or context overhead. A simpler harness might let the model reason and act more directly. This could be one reason for the 3rd finding mentioned.

[Günter Klambauer @gklambauer]

50,000 papers submitted to ICLR??

Is this a DDoS attack on machine learning research??

[Abhinav SB @Abhhinav_SB]

@DeryaTR_ But I would prefer the term ‘universal classifier’ rather than just a ‘classifier’ 😅

[Abhinav SB @Abhhinav_SB]

@HenrykAbram Yes..that direction makes sense. There’s already an open source project called Laya exploring a pretty similar Jev style setup with RLCD.
But its architecture is quite different from the adapter based approach you’re describing.

[Abhinav SB @Abhhinav_SB]

My thoughts on Jev (TLDR: I think it’s useful for some cases, but somewhat overhyped and im not that impressed):

- For me, Jev is not comparable to LLM level intelligence nor is the core idea a new paradigm .
We can say it as a universal classifier with a more general RL based architecture.

- Classification instead of generation isn’t that new.

But the actual novel part is the combination of runtime defined decision tasks, parallel inference, calibration focused post training, efficient serving

We need the technical paper and reproducible benchmarks before making stronger claims about algorithmic novelty.

- About the cheap inference: Transformer based models have made us think that anything token based is expensive. No…an RL based model is always this cheap and fast. If it were released in 2020 we wouldn’t be surprised by the cheap token prediction and speed, since models in 2020 were always like this (but if it were released in 2020, we might be surprised by its prediction capability).

- I don’t wanna say, “Whatever Jev can do, a simple classifier can do.” (In most cases, it’s true, though, if we’re using a custom classifier for a task specific application)

Many Jev tasks can be framed as classification or scoring tasks, but Jev’s key advantage is its ability to generalize to runtime defined tasks.
Traditional classifiers usually have a fixed output space. Jev can define the decision space at inference time. That’s why I think of it as a universal classifier.

- The authors emphasize that many useful AI workloads are classification or scoring problems, while Jev’s headline comparisons are mainly against autoregressive LLMs rather than dedicated classifiers. Jev is not directly comparable to current LLMs

- About the RLCD:
I wont say RLCD is simply “RL over sequence embeddings”
But there are already systems using PPO like RL over learned representations for downstream decision classification tasks, and RL for calibration decision making has prior foundations.

But without TypeSafe publishing the actual RLCD algorithm they used we can’t say Jev is just another PPO over sequence embeddings.

- What I find most interesting about Jev is the systems level packaging rather than any single algorithmic component.
That could make Jev genuinely useful for some workloads while still making the “new kind of intelligence” claims feel just hype.

- Still Jev might be worth the hype if it were open source. The lack of technical details matters.

It’s currently a proprietary early access model. And its architecture, weights, training data, and detailed RLCD methodology not published.

- Until TypeSafe publishes a paper or technical documentation, it’s difficult for the research community to determine which parts are the real algorithmic advances and which are effective combinations of existing ideas and engineering.
Also If we have more architectural side details, we can confirm where it fits perfectly in application level.

So overall: interesting direction but over hyped + potentially useful in some systems, but I’d wait for the technical details before calling it a new paradigm.

[Abhinav SB @Abhhinav_SB]

@bad_latency @michael_chomsky Yeah true. Compared to the ability of attention based models to act as classifiers, that’s nothing new. But its ability to act as a custom classifier for a runtime defined task is what makes the term “universal classifier” perfect fit