Sabitlenmiş Tweet
Deep learning is not “the thing”, it is “a thing” that takes us to “the thing”.
English
Lakshay Sagar Rana
7.4K posts
Lakshay Sagar Rana
@lsrspeakstocomp
abstraction engineer. interested in complexity theory. playing with computers :)
Katılım Şubat 2018
409 Takip Edilen461 Takipçiler
@IterIntellectus keep hand designing things so that you can file for patents.
English
Solving a problem is not the reason for working on it.
[Even] if one never solves it, one realizes approaches that looked right and then understands why they were wrong.
~Conjecture Institute Ambassador @DavidDeutschOxf with @GadSaad
English
Lakshay Sagar Rana retweetledi
Our paper on optimize_anything has been accepted to CAIS 2026, and is out on Arxiv with expanded experiments and details!
A unified API to optimize agents (with architecture), CUDA kernels, cloud scheduling policies, or even graphics!
x.com/LakshyAAAgrawa…
Lakshya A Agrawal@LakshyAAAgrawal
Excited to release @gepa_ai's optimize_anything: a universal API for optimizing any text parameter. It consistently matches or outperforms domain-specific tools optimizing code, prompts, agent harnesses, cloud policies, even visuals! If you can measure it, you can optimize it.
English
@goodhunt @lateinteraction @nrehiew_ @DimitrisPapail do you think this could help you improve your teaching with kids too?
English
@lateinteraction @nrehiew_ @DimitrisPapail bruh thank you I am finally not going insane using my teacher brain looking at ai research
good teachers scaffold material and use pedagogy
the method is as important as the answer
English
It feels like the next breakthrough on a scalable training algorithm is close, likely on top of GRPO with denser credit assignment beyond outcome rewards, but done so with lower bias.
- ECHO does this by limiting credit to environment responses
- Composer2/Self distillation/OPD
wh@nrehiew_
Feels like the vibe has shifted. The current wave of research feels very similar in spirit to the reasoning/r1/grpo period
English
dont let anyone confuse you with “agentic/llm search”. Its just a keyword search for entities but without their ids.
English
English
@lsrspeakstocomp @pfau @curiouswavefn theory scales perfectly! If the correct hypothesis class is included at the start
English
Cosmo Shalizi: "I like to think I am not a stupid man, and I have been reading about, and coding up, neural networks since the early 1990s. But I read Vaswani et al. (2017) ["Attention Is All You Need"] multiple times, carefully, and was quite unable to grasp what "attention" was supposed to be doing. (I could follow the math.) I also read multiple tutorials, for multiple intended audiences, and got nothing from them...the sheer opacity of this literature is I think a real problem."
bactra.org/notebooks/nn-a…
English
because our civilisation was always in an abundant environment. Mother india had all the fruits. Our ancestors knew problem isnt in resources its in people hence the spiritual awakenings.
you can look at all the teachings, they relate somehow on how to have a prosperous society. How to think of a society like a system.
hopefully we find a solution using these teachings, the problem of mindset.
English
the more i read up on indian philosophy be it hindu or buddhist i realise it was made for a post abundance society where struggle has lost all meaning all you’re left with is you and your thoughts. its post absurd post overman its after there’s nothing left to do.
what led to this tho
English
@pfau @gleech @curiouswavefn do you think its because statistics does not have tools to think of the effects of scale ?
English
@gleech @curiouswavefn If I had a dime for every time a statistician declared deep learning was just kernel learning in a trenchcoat...
English
Here's a better lesson, don't fall for bitter lesson.
Richard Sutton@RichardSSutton
The bitter lesson in 26 words: Don’t be distracted by human knowledge, as AI has been historically. Instead focus on methods for creating knowledge that scale with computation, like search and learning.
English
Lakshay Sagar Rana retweetledi
arxiv.org/abs/2605.18607
Measuring the capability of the model using token-level statistics with weight on expert trajectories. The metric behaves much more smoothly with high predictive power. This could be super useful.
English
1. Computational Complexity Theory is actually just the study of the large-scale structure of the Logos.
2. Human ideas/engineering are so concentrated in easier complexity classes that the rest of it is likely a useful map of many kinds of unknown unknowns in science and math.
English
there is a reason JESUS said “Love thy Neighbour”
Pratyush Ranjan Tiwari@PratyushRT
I believe a 1-person research lab is possible now for specific enough problems: If someone has incredible range, ability to dig deep into a problem + the mindset to take complete ownership Everything is more fun with collaborators though!
English
tbh it was very difficult to remember what is row and what is column. fml.
English
English
One line of code is all it takes to prevent LLM agent delusions, instead of post-training patches like RL.
love4all.ai/blog/why-it-is…
❤️ 4 ∀
github.com/nandodef/love4…
English
@PratyushRT you get to know specific enough problems only with people 🤷
English
I believe a 1-person research lab is possible now for specific enough problems: If someone has incredible range, ability to dig deep into a problem + the mindset to take complete ownership
Everything is more fun with collaborators though!
English
@alexabelonix thanks!! Please sure to connect with you too!
English
@lsrspeakstocomp keep shipping like this. 🤝 pleasure to connect with you!
English
i keep coming back to Karakauer’s idea of intelligence.
maybe the permanent underclass are the one’s with a LLM sub.
English
Unpopular opinion but to properly use agentic AI you need to be able to think and communicate in pseudo-code.
Imagine you are actually writing (big picture, strategic) code with procedures and boundaries but using English.
English
@profleonn @Rafa_Schwinger Hahah no. Just read essence of software.
English
