Om jadhav

Constraints on systems that can evaluate and improve their own reasoning.
AI systems • evaluation • first principles

[Om jadhav @jadhav_om257]

We often measure AI by how well it answers our questions.

But what happens when AI can question its own reasoning, recognize the limits of its understanding, and find better ways to think?

Perhaps the next leap isn't just more intelligence.

It's intelligence that can understand and improve itself.

That's the frontier I'm interested in exploring.

If you're building, researching, or thinking deeply about AI, let's connect. 🤝

[Om jadhav @jadhav_om257]

The idea of treating reasoning as sequence of higher-level actions rather than individual tokens is interesting. If models can learn reusable reasoning skills during mid-training, RL may not need to discover everything from scratch. The key is whether those abstractions transfer across different task.

[Om jadhav @jadhav_om257]

More context doesn't always mean better learning. Agents need to build on past attempts without getting trapped by them. Separating exploration from coordination seems promising. The real breakthrough is helping agents learn from history without being constrained by it ..

[Yoonho Lee @yoonholeee]

Autoresearch loops are a particularly promising form of continual learning that we can scale indefinitely over time. However, in practice, they often plateau as runs get longer. In naive implementations of such loops, agents "lose motivation" after seeing full history and talk themselves out of trying anything new.

Stateless Language Agents (SLA) looks at why this happens and proposes a few simple modifications to scale usefully to 1B tokens (!). The core trick is to hide as much context as possible from each worker while having a separate advisor agent handle coordination.

SLA achieves the best final result on every task we tested (vs EvoX, CORAL, SwarmResearch), often by a wide margin!

[Om jadhav @jadhav_om257]

The interesting distinction is between an agent that can execute individual steps of replication and one that can sustain the entire process autonomously.

Reliable persistence, access to compute, resource acquisition, and adapting to defensive measures are very different challenges.

The gap is worth watching closely... !!

[Om jadhav @jadhav_om257]

@mark_k Interesting that it improves on new task without changing the weights. Still not clear how well those changes hold up when the new tasks only partly match the ones it was revised on. A fix that helps one task can still hurt others

[Om jadhav @jadhav_om257]

@HamelHusain @sh_reya Yeah, it works when you can still check the judge against human labels. Once the run goes longer and the only thing it has to go on is its own earlier steps, agreement with itself doesn't really tell you if those steps sounds good ...!

[Om jadhav @jadhav_om257]

@AmbroseAbaasa Sure, Thanks.

[Om jadhav @jadhav_om257]

The hard part isn't getting agents keep going.
Once a run gets long enough, almost no one can tell which earlier steps rested on an incomplete or circular assumption.

A system that can't catch that will keep building on the wrong things, even when every individual step looks fine.

[Om jadhav @jadhav_om257]

@dair_ai Harness Learning and VeriHarness both move the improvement into the surrounding system instead of the model weights. The open question is how stable those edits stay once the harness has to handle task that are only partially overlapping with the ones it was revised on.

[Om jadhav @jadhav_om257]

@fchollet Science has external checks that current agent loops mostly lack. Once a run gets long enough, incomplete or circular assumptions in earlier steps become hard to detect, so the system can keep building on them even when a local outputs look fine.

[Om jadhav @jadhav_om257]

Truee.. !!

[Nick @NickpxJ]

If you want people to notice your posts, start by noticing theirs.

[Om jadhav @jadhav_om257]

@EdenBitton7 Sure man , Glad you're here
Done from my side 👊

[Om jadhav @jadhav_om257]

Hey gems I'm Om, an aspiring ai engineer.
I've always been more interested in what we can build than in talking about what might be possible.

Lately, I've been fascinated by how far we can push AI. Right now, I'm working toward something I can only describe as "intelligence that doesn't exists yet".

here to connect with people who experiment, build their own things, and aren't afraid to explore ideas that might seem a little crazy at first .. !!
what are you working on these days ?
lets connect !!🤝

[Om jadhav @jadhav_om257]

This is the part I keep running into too. If the agent is looping or burning the step budget, tweaking the harness usually fixes it. But when it calmly delivers a bad plan, the harness almost never saves you. You end up having to touch the weights. The annoying part is that on longer tasks it gets harder to tell which kind of failure you’re actually looking at.

[Om jadhav @jadhav_om257]

The awareness gap is real. A lot of people still only see the chatbot version. What’s more interesting to me is that even among people who do use the stronger systems, checking the harder math results is already slower than generating them. The ivory tower problem and the verification problem are growing at the same time.

[Om jadhav @jadhav_om257]

@coryalthoff The generation part is getting cheap fast. What’s still hard is figuring out what is actually worth building, keeping it coherent over months instead of one prompt, and being accountable when it breaks. Those jobs don’t vanish just because the first draft is free.

[Om jadhav @jadhav_om257]

That's what man. The continuous learning part is the real bottleneck. Even if you had the agents, keeping a single coherent story and world consistent across hundreds of hours of content still needs something that can notice and correct drift in its own long-term reasoning, not just generate the next piece. We don’t have that at the required scale yet, which is why the short-term version stays very unlikely.

[Om jadhav @jadhav_om257]

Yeah, the gap in what people are actually seeing is real. What stands out more to me is that at the top end, people are already struggling to fully review what the agents produce. Once the output is bigger than what someone can carefully check, doing well on verifiable tasks doesn’t really tell you whether the system can notice when its own longer reasoning is incomplete or going in circles. Both the capability and the ability to properly check it are moving fast at the same time.

[Om jadhav @jadhav_om257]

The shared Git history clearly lets later agents reuse partial work and failures instead of restarting. The harder open question is how an agent decides which earlier commits are actually worth building on—especially when a contribution looks locally coherent but contains a subtle error or rests on an incomplete assumption. Without a reliable way to evaluate the quality of intermediate reasoning, the system can still compound the wrong things. !

[Om jadhav @jadhav_om257]

The central issue is that evaluation awareness breaks the measurement. Once a model can reason about whether it is being tested, a drop in scheming rates on the eval no longer cleanly distinguishes “it stopped scheming” from “it got better at detecting the test and behaving differently under observation.” Any training method that increases situational awareness inherits this ambiguity.
What's your saying. ?

[Om jadhav @jadhav_om257]

Useful framing. The gains from feeding past records (including failures) are clear, but the harder part may be knowing when the world model’s own prediction is unreliable—especially under distribution shift to a new setup or when the research direction itself changes. High ranking correlation on familiar setups doesn’t yet tell us how well it flags its own uncertainty... !!

[Om jadhav @jadhav_om257]

Strong numbers across agentic coding, long-context, and computer-use benchmarks. The open question is whether higher scores on these tasks also reflect better internal evaluation of the system’s own reasoning traces—detecting incomplete assumptions or circular paths—or whether the gains are still mostly in outcome quality under the existing evaluation setups...!!