Many people in the AI industry believe that harnesses don’t matter, models have become powerful enough to solve problems with minimal scaffolding. Well, that’s true only if problems are fully specified upfront like they often see in benchmarks. Most real world work is ill-specified and humans discover what they want as they go along. There models do require harnesses that optimally help the human explore and discover what she wants AI to do. So harness design must be now thought from human cognitive system point of view, not from the model point of view.

This long weekend has crazy long queues at popular places in India. I suspect this: almost everyone is now asking ChatGPT where to go, and it is sending the masses to similar places via similar routes at similar times. Correlated human behavior is a second-order effect of AI!

New paper release from @lossfunk We find that something as simple as saying information is from a "verified source" flips the answer of an LLM. It has implications on how the LLM answers if information it fetches online simply asserts it is a verified source.

We offloaded physical labor to machines long ago, and now we are doing the same for intellectual labor. There’s no reason to expect we will retain cognitive agency. Thinking is effortful, so if it can be outsourced, we will happily do that. Post-work society seems inevitable.

I asked Astra to work on whatever it felt like and show me the results. It ended up iterating over cellular automata :) This pattern of agents working on whatever they feel like is going to catch on, as we humans would love getting surprises from our agents.

“How can a man so genius be so stubborn!” This line from the biography of Gödel perfectly summarises the personality. The man shook the world of math by proving how formal systems contain truths that can never be proven. Yet, he refused medical treatment consistently and died in the end of starvation. He also believed a world beyond the physical world that human intuition directly sensed, perhaps which is what partly drove his delusions that doctors were trying to poison him. The #book# is extremely well written. The author weaves stories of Austro-Hungarian empire, Vienna and World War 2 alongside. You get feel the infectious energy of the Vienna Circle and get to know famous philosophers, mathematicians and scientists along the way. I strongly recommend reading this biography, if you can get hands on a copy. It’s a beautiful book detailing life of a tragic genius!

It's a minor trend, but I'm definitely noticing a growing trend of 16-20 year old founders who're skipping college and directly building a startup. I'm honestly excited about this. College has its value in terms of human-connect but I think we'll find better alternatives for it.

I think the easiest growth hack for any business right now might be to expose endpoints to agents. So many businesses have their utility locked behind contact us or human sign up. Just expose an endpoint for AI agents and let people derive value programmatically.

These sorts of findings should make us want to massively accelerate the science of consciousness (something we're doing at @lossfunk). We need validated theories to answer the question: what systems are capable of having an experience, and what sort of qualia they have.

The first scaling axis was data & model size. The second one was RL environments & compute. I think model companies are now scaling on third axis: agent swarms. If you notice all the recent eye-popping behaviours are an outcome of 10,000+ agents working together intelligently.

AI increases the average quality of output but collapses the variance of it, making everything looking and feeling very similar. Implication: contribution of a human ought to be to judged by how much difference can they make in an output v/s the default AI-driven output.

Cool approach, i think less latency and cost is the way to go. Here's how this would have likely been made. - Tiny transformer or equivalent model (maybe a few bn or so?), explaining latency and cost - Questions are sent in parallel to multiple copies of it (I'm sure they're edge located) - The model is post-trained for calibration in a wide variety of data (the recipe is relatively simple, and likely targeted on distillation of logprobs / confidence of a bigger model) Notice how cost is ONLY for input tokens as output is merely numbers (few tokens) because input could be huge (questions and options). At 0.042-per-million price they have, Astra estimates the model to be 3bn parameters. One could replicate this by post training Qwen 3.5 2Bn. I expect people to do so soon!

There's something magical about self-modifying software. I'm creating my own harness, which allows customisation just with prompting. I simply point my agent to its own code and tell it what change I want, and it gets it done. It runs regression tests to ensure nothing breaks. I can build plugins for it, roll back to previous versions if I don't like something. I started with codex, but now I use my own harness to improve itself. Makes me think about the future of open source - everyone running their custom version of a software, and agents doing intelligent merges whenever updates need to be installed. (I know pi and dsh are also self-modifying harnesses, but they are on terminal. I'm creating a web based harness as I believe text on terminal is not conducive for deep human feedback, which is what I want to provide in order to meaningfully contribute to the final output.)