@PriGoistic
Account based inIndia
About this account
- Account based in
- India
- Connected via
- India Android App
Account-level information from X, not a live location or the device used for a specific post.
co-founder & chief of research @metacognitionai | ex - ml & sys eng @NASAArtemis , @NASAJPL | prev - @UCBerkeley
Joined March 2022
- Tweets556
- Following192
- Followers232
- Likes482
Pinned Tweet
7 months , 6 naive bachelors , 1 vision : a suite of artifical systems built ground up from biological intelligence.
A new generation of emergent intelligence which goes beyond the comprehension of humanity.
Be a part of a leap towards the next phase of evolution with us at @metacognitionai
5 years ago also there were people writing slop code, but the percentage of it was still controllable, as real systems were built with proper, understandable and durable code written by programmers and engineers who had the capability to own them. Personally, i have felt this not just in programming, but also while seeing the state of Tier A conferences like @NeurIPSConf and @iclr_conf and their 50k + submission numbers. There could possibly be only 2 things which justify this: either research across multiple domains is going so well, or AI has made tangible research so good, that every passing day there is a chance we can reach a breakthrough (unlikely), or else, as i term it, people are constantly in this trance of doing researchmaxxing, where the ratio of tangible outcomes / amount of resources burnt is simply not justifiable.
From my experience working at some small research labs recently, apart from the seniors who have in-depth knowledge about their domains, the juniors are just coming up with hypotheses which make no sense. Yes, scientific research is all about making a plausible hypothesis and testing it out, but the important word here is plausible, which i feel is lacking a lot these days. And holy shit, don’t even get me started on folks who are writing papers with Claude and pushing them to arXiv or conferences without any kind of deep knowledge about what’s written in them. Writing a research paper is not similar to some kind of literature writing. The problem this creates is a stigma that folks in bachelors don’t have the capability to work in this field, and companies don’t trust upcoming researchers for proper roles.
I am done with this shit. It is over. The state of engineering right now is horrible. It has been half a month since I started a new role at a big company. Nobody knows anything here. The specs, code, tests, PRDs, tickets, resolution of those tickets, reports, etc., everything is made by Claude Code. Nobody on my team likes this. They are being forced to ship as much as they can. I have heard multiple times from higher management that pushing code is not a bottleneck, so why are we slow? People are working 12 to 13 hours a day just to press enter. Nobody is reading anything. Humans in corporate are doing nothing on their own. Everyone, literally everyone, from an L1 to an L7 engineer here is doing the same thing. Talk to Claude. There is no sense of victory. Nobody is resolving bugs. In reality, nobody is thinking anymore. Everything is done by LLMs. It is so soul-sucking. I would not mind it, to be honest, if we were at least given the time to check out the code and see what is going where. But no, the goal is to just ship. No matter what happens.
great to see that, people are getting back to core Machine Learning by working on systems like classifers.
Decision making systems like Jev by @typesafeai is trained through a methodology called RLCD, but every without it there were architectures like using bidirectional BERT for the same task.
I think Laya does smthg same im not sure coz havnt read it yet, but Julia -1 works on similar principles of mmBERT. mmBERT demonstrates that low-resource languages can be effectively learned during the decay phase of training and Julia works on CPU so yup.. good to see good durable and cheap systems again.
Introducing Julia-1:
Our first classification model that runs on almost anything.
Learn more 👇
supersoniclabs.ia.br/julia-1…
Priyam retweeted
i have been getting alot of email reachouts and people on multiple channels asking to be a part of @metacognitionai
This is actually pretty interesting to see cause our journey as a team hasnt been very easily tbh.. constant backlashes..
-ppl saying noone funds a research lab..
-sell your research to get money to present at top conferences.
- research isnt a moat..
- team isnt a moat..
- ai research will take you no where etc..
but then i see a different spectrum of ppl who admire our work, our hardwork and goals.. inbound messages from top industry engineers, researchers admiring our work from top ai labs, engineers from well established startups asking to work with us..
At the very end we are just naive 20 year olds dreaming about putting a dent in the universe.
And for godsake we will do it.
Dont let anyone kill that potential inside you kiddo, you are meant to change the world. Remember who you are.
Ad astra.
everyone should actually read the Deepseek V2 paper tbh, personally i feel its way more important and bigger than any other architectural paper
me recently while running training set on @modal 🥀🥀🥀
Priyam retweeted
I remember the time when @PriGoistic and me started @metacognitionai with literally such naïvety, and discussing to solve this massive AI memory gap using Physics, we were clueless what it'll lead to.
And now, 9 months, the work feels surreal with the team pushing towards a positive unknown to solve the AI Memory issue. Every work is a breakthrough in itself NGL.
We've met people who question why we exist and how we are different and generally my response is - our differentiation is a cracked team who knows the importance of research and not some shitslop.
We believe that -
Once in your lifetime, you get opportunity to work on something that makes a significant dent in the industry, and then, you should step out of your shoes to address them with full pace and complete humility.
I was revisiting @RichardSSutton 's famous essay, "The Bitter Lesson " again after seeing all the fiasco in the AI space today, especially this fear mongering of some frontier labs towards "humanity threatening" outcomes.
Sutton's work has been exceptional, the godfather of RL, and this essay in 2026 makes way more sense tbh..
Human Knowledge < Scaling Computation : is the way to how we can attain better performance and something we call intelligence maybe.. This is what most frontier labs today believe tbh.
But again the industry today doesnt actually understand the core meaning of what Sutton actually meant.. even in a podcast with @dwarkesh_sp , he specifies the importance of "The Bitter Lesson "
Especially the last paragraph, is more like a reminder to us, just putting huge compute and data wont make intelligent systems that has the ability to reason like us.
Humans attain cognition through experiences, mistakes and meta-learning. We discover things accidentally.
We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.
Humans attained the state of intelligence through Millions of years of experiences and information that got passed through through our genome which triggered evolution. To adapt. Current AI archs or systems are no where close to this ability to learn and adapt continually.
Humans arent perfect ofc, but the level of our intelligence is the highest in the known universe. I think its really important for us to understand the way we think and make systems which dont just mimic frontier intelligence with plethora of compute and data but can attain a state of meta learning where with data scarcity and limited experiences we can learn fast.
The core idea of this essay was that, neuroscientists and computer scientists in the late 80s and 90s were fixated with building systems which can be traced through human knowledge only. We were basically forcing the vaguely abstract systems with complex human language and understanding -> which showed poor performance and slow outcomes.
Now the issue isnot way human knowledge.. it was with the architectures.. the way HARPY was designed or even earlier methodologies like SIFT for vision where designed.. we very naively attached human knowledge with poorly represented architectures.
it was an architectural problem always. not data.
there's nothing more thrilling than charging towards uncertainty...
We know nothing about anything yet its a human nature to keep feeding our curiosity..
Our paper on dynamics of memory is officially on @IEEEXplore which we presented recently at IEEE ICACI at Tianshui, China
cc : @venky1701 @sauhard_07
small win for us @metacognitionai
