HNHacker News
TopNewBestAskShowJobs

ChaitanyaSai

1,227 karma · joined January 27, 2008

We run a microschool in Mumbai. https://playbook.comini.in/ Along the way,we ended up building a microschoolOS to help run everything. We are, essentially, an AI-native microschool where the AI is invisible.

We wrote a book explaining Consciousness - Journey of the Mind: How Thinking Emerged from Chaos https://www.amazon.com/Journey-Mind-Thinking-Emerged-Chaos/dp/1324006579/

I write here https://saigaddam.medium.com/ and here https://blog.comini.in/

sai email comini dot in

submissionscomments
ChaitanyaSai··on I miss thinking hard
I miss the thrill of running through the semi-parched grasslands and the heady mix of terror triumph and trepidation as we close in on our meal for the week.
ChaitanyaSai··on 'World Models,' an old idea in AI, mount a comeback
Interesting! Documenting this anywhere?
ChaitanyaSai··on What brain surgery taught me about the fragile gift of consciousness
Yup, it's the short sentence cadence.

"And yet, hours before surgery, with death still in the room, I didn’t feel fear. I felt something quieter. Stranger. I felt connected. To her eyes. To my breath. To the weight of my feet against the floor. To the wind brushing the window."

The usage of these short sentences (which, people do use, but sparingly) is a good marker. My hunch is this is because of how they call attention to themselves and are rewarded by human RLHF participants. I don't know if incentives including spending time on essays like this but if they don't and the rater is trying to do a speed-read, these stand out.

Have written about other markers here: https://saigaddam.medium.com/it-isnt-just-x-it-s-y-54cb403d6...

One along those lines: "Not just that we think. But that we feel. That we can marvel. That we can sit in silence across from someone we love and feel time slow down and become something."

ChaitanyaSai··on What If We Had Bigger Brains? Imagining Minds Beyond Ours
As brains get bigger, you get more compute, but you have to solve the "commute" problem. Messages have to be passed from one corner to the other, and fast. And there are so many input signals coming in (for us, likely from thirty trillion cells, or at least a significant fraction of those). Not all are worth transporting to other corners. Imagine a little tickle on your toe. Should that be passed on? Usually no, unless you are in an area with creepy crawlies, and other such situation. So decisions have to made. But who will make these decisions for us? (Fascinating inevitably recursive question we'll come back to)

This commute is pretty much ignored when making artificial brains which can guzzle energy, but matters criticallyfor biological brains. It needs to be (metabolically) cheap, and fast. What we perceive as a consciousness is very likely a consensus mechanism that helps a 100 billion neurons collectively decide, at a very biologically cheap price, what data is worth transporting to all corners for it to become meaningful information. And it has to be recursive, because these very same 100 billion neurons are collectively making up meaning along the way. This face matters to me, that does not, and so on. Replace face with anything and everything we encounter. So to solve the commute problem resulting from a vast amount of compute, we have a consensus mechanism that gives rise to a collective. That is the I, and the consensus mechanism is consciousness

We explore this (but not in these words) in our book Journey of the Mind.

You'll find that no other consciousness model talks about the "commute" problem because these are simply not biologically constrained models. They just assume that some information processing, message passing will be done in some black box. Trying to get all this done with the same type of compute (cortical columns, for instance) is a devilishly hard challenge (please see the last link for more about this). You sweep that under the rug, consciousness becomes this miraculous and seemingly unnecessary thing that somehow sits on top of information processing. So you then have theorists worry about philosophical zombies and whatnot. Because the hard engineering problem of commute was entirely ignored.

https://www.goodreads.com/en/book/show/60500189-journey-of-t...

https://saigaddam.medium.com/consciousness-is-a-consensus-me...

https://saigaddam.medium.com/conscious-is-simple-and-ai-can-...

https://saigaddam.medium.com/the-greatest-neuroscientist-you...

ChaitanyaSai··on How to build Intrinsic Motivation: a review of the science
Great article! SDT has fascinating parallels in consciousness science that no one to my knowledge has actually explored. This is because in consciousness research, the experiencing self is a given, it just happens to be taken for granted that there's an "I" experiencing, and the wonder and magic is focused on the experience itself. What about the self that is experiencing? On the other hand, SDT operates at a level where a biological and even experiencing conscious self is taken for granted, and the focus is on how the cognitive self operates (in many ways). And this is also where the criticism comes from. This is all in the domain of the self and motivation and whatnot articulated in language. To go deeper, we need a bridge between these two that can explain how the self is constructed. And we do have a beautiful theory/framework for that

Consciousness is a consensus mechanism by which the self is constructed. It is a recursive loop where the self emerges, experiences, and folds in the next experience to create an evolving, expanding self. With language we have the ability to freeze many of these ideas and we are able to go much further. "I can think, feel, experience and reflect on this"

And why a consensus mechanism? Because "you" are actually a constellation of cells and experiences that needs to be sufficiently decentralized but also be able to act and plan in the very short and long term. How do you get 87 billion cells (in our case) to decide as one? That is actually a pretty difficult engineering problem where you have to think about both compute (all the different data streams coming from different sources need to be digested and acted on) and commute (one cell group in the prefrontal cortex needs to immediately broadcast a danger message to other corners of the brain, and we dont have direct wiring)

Now the natural question to ask is, what do synthetic beings need to develop both? If you are interested you might want to read our book Journey of the Mind

Here is a short read on the idea of consciousness as a consensus mechanism https://saigaddam.medium.com/consciousness-is-a-consensus-me...

ChaitanyaSai··on OpenAI releases image generation in the API
Almost every image has a yellow tint. Any discussion of why and when that's being fixed?
ChaitanyaSai··on AI Horseless Carriages
Loved those! How are those created?
ChaitanyaSai··on Where are all the self-directed learners?
Absolutely. Wonderful book that talks about this in a different context https://www.amazon.com/Why-Greatness-Cannot-Planned-Objectiv...

Also wrote about it here https://blog.comini.in/p/what-should-replace-grades

ChaitanyaSai··on Where are all the self-directed learners?
Interesting analogy!
ChaitanyaSai··on Where are all the self-directed learners?
OP here. The Indian context is important. We have a massive talent pool (many outside the US just don't know how many engineers graduate yearly (1.5 million!) in India ) and we are not competing with FAANG / well-funded startups. There is a lot of raw talent out there. We know because we have been hiring for a long time. Working with us has been the springboard to well paying jobs and senior roles. What we cannot compete on at the moment (salaries) we make up for with growth opportunities and learning (and this isn't us selling this; we have the evidence to back it up). What we might have is a problem where the raw talent doesn't even know how to make a case for themselves. What really surprises me is how the numbers havent increased in this day and age where one can realize that the median college (in India) is awful in educating them and they can find what they need online.
ChaitanyaSai··on Where are all the self-directed learners?
OP here. Yes, that is a jump, which I bridged now.

OK, technically the absence of genuinely interested applicants is NOT an indication of drive. However, please note the context: It’s a brutal job market at the entry level in India at the moment. That’s why we are flooded with applications. Candidates either know this, or quickly figure this out. And yet, that does not translate into many giving their application serious and creative thought?

It is a problem they have never been asked to solve for: How to stand apart and demonstrate ability without relying on the usual credentials. It needs creativity and independent thinking and some agency. It is the lack of these that baffles me and makes me search for answers.

Learning is not just the know-how and knowledge you acquire for doing a job. It is everything else you need to navigate life and also, in this narrow sense, the skills you need to apply for a job, make your case that you are the right candidate, or at least that demonstrate that your candidature deserves attention. It is a very learnable skill. Especially now. And given the market, a skill anyone with initiative should try to acquire.

ChaitanyaSai··on Where are all the self-directed learners?
That's one of the questions in the form.
ChaitanyaSai··on Where are all the self-directed learners?
OP here. What we hope the independent thinks will realize is that when the cost of applying drops down to a single automatable click, their cursory click application is lost in a sea of spam. This allows them to stand apart. Recruiting and dating have a tone of parallels. They are time-intensive for both parties. And that time spent does not always result in a positive outcome. And yet that high time-cost (for both sides) is what makes the relationship meaningful.

Here's a question. In a difficult job market with a deluge of candidates, how should a candidate try and stand apart? And we have to answer that for the new AI age where output for many tests can generated very cheaply.

ChaitanyaSai··on Where are all the self-directed learners?
Here's the problem we are trying to solve for. Filter from 2500+ applicants. What would you suggest. And we do not want to rely on credentialing and leetcode filters. What non-garbage tests would you suggest? Please note that we are talking about an initial test that's generous enough in error, but not so much so that we end up with 1000 people to seriously interview.
ChaitanyaSai··on Where are all the self-directed learners?
OP here. Added this conxtext: OK, technically the absence of genuinely interested applicants is NOT an indication of drive. However, please note the context: It’s a brutal job market at the entry level in India at the moment. That’s why we are flooded with applications. Candidates either know this, or quickly figure this out. And yet, that does not translate into many giving their application serious and creative thought? It is that lack of agency that baffles me and makes me search for answers.
ChaitanyaSai··on Where are all the self-directed learners?
OP here. We are fully aware of how this might seem like a lot of work for job seekers. But the practical reality is the Indian job market has a massive supply of talent. As with any, it is of variable quality. We are very clear we do not want to use the usual credentialing and leetcode interviews as filters. They suck. We want to work with people who are genuinely interested in the domain we work in (education) and this has worked reasonably well for us.

I would love to find a better way to filter through 2500 applications. What markers do we rely on. AI has made the usual filters even more irrelevant.

ChaitanyaSai··on Where are all the self-directed learners?
OP here. It's a question to easily detect for people who use chatgpt to fill the form. Anyone who enters alpha-gal syndrome has probably used chatgpt. That isn't an automatic reject though (very generic template answers are). Guess we'll have to find another question soon now this is out here for LLMs to slurp up!

The reason for this questionnaire is culture fit. It's also the very practical matter of weeding through 2500 applications when AI has made all straightforward code-based questions very pointless

ChaitanyaSai··on Children's arithmetic skills do not transfer between applied and academic math
Agree! And that's because there are some who just get this and are able to explore this in the space of ideas the same way we all have very different tastes in music and literature. Math has both practicality and poetry and many problems are introduced when the focus is one on. What one responds to is personal.
ChaitanyaSai··on Children's arithmetic skills do not transfer between applied and academic math
I run a microschool where I teach math (and I am a neuroscientist) and this is pretty obvious, and also something we see every day. And it's because we still suck at understanding learning.

Learning is not the acquisition of knowledge. Learning is all the things by which we learn to model the world. And we do that, our brains do that, because it matters.

Math in classrooms is pretty much designed to leach out all context. Meaningless symbols that need to be manipulated to arrive at some mysterious answers. Why bother, our minds scream out. The why is evident in street settings. Also, we are designed to pick up patterns in such meaningful settings and learn. You don't even need to teach it.

Fixing learning and education means being able to articulate answers to five questions: why, what, how, when, where we learn. And all our current answers are outdated and one-size-fits-all

There is also the matter of learning design. Math we use is highly compact because of its efficiency. But these can be hostile starting out. There are ways to explore mathematical concepts without using mathematical symbols.

Resonant learning requires the interplay of building competence (how) and comprehension (why) and this can be done well only in meaningful settings (where)

https://blog.comini.in/p/why-khanmigo-will-fail

https://blog.comini.in/p/what-happens-in-math-class

Also, the most profound revelation I've had about learning is from working on our book. Journey of the Mind. Spent twenty years at the intersection of neuroscience and AI but never really got to "understand" and "learn" how the mind works and learns until diving in and then trying to tell its story in an accessible manner. If you are interested in any of this lease do check it out

ChaitanyaSai··on What Is Entropy? A Measure of Just How Little We Know
Lovely article. The subjective nature of entropy and information immediately makes me think of the IIT theory (integrated information theory) of consciousness and its foundational futility. Information cannot be discussed without perspective. Someone has to define the states. A die have 6 states only to us humans. What about an ant that's likely to have the die land on it? Bringing the observer back into discussions of information is fascinating because it then begs the question: How is the observer put together? And how does a perspective, an I, emerge in a multi-trillion-cell entity? For those interested in this detour, you might like reading this and our book (mentioned in there)

https://saigaddam.medium.com/consciousness-is-a-consensus-me...

ChaitanyaSai··on Not yet panicking about AI? You should be.There's little time left to rein it in
There's a lot that's been written about AI and its potential impact. Found this particular piece interesting in its details and also because I just picked up a book by the author (Measuring The World) and loving it
ChaitanyaSai··on Education and Healthcare Suck for the Same Reasons
Point taken. Used a different image.
ChaitanyaSai··on Education and Healthcare Suck for the Same Reasons
Agreed, the scaling is because of the CMS. Changed the image
ChaitanyaSai··on How to Study Mathematics (2017)
>Step 3. Memorize the exact wording of the definition.

Huh. Any mathematicians who want share their own opinions and experiences about this?

This pretty much goes completely against my experience with other grad school level neuroscience/ML

You don't want to be so familiar with stuff as to make it second nature but NOT from memorization. That, at least an other areas, leads to surface level recognition

Does the author mean internalize and not memorize?

ChaitanyaSai··on Everyone is capable of, and can benefit from, mathematical thinking
Tiring because the answer wasn't revealed? That was the whole point :) It's the path, not the summit.
ChaitanyaSai··on Everyone is capable of, and can benefit from, mathematical thinking
"mathematics is a game of back-and-forth between intuition and logic" I teach/guide Math at our school (we run a small school and currently have kids under age 10) and this is so so true.

I just wrote about this. In fact, you can even see this at play in the video of the kids talking https://blog.comini.in/p/what-happens-in-math-class

ChaitanyaSai··on Dobble (The Mathematics Of) (2018)
Haven't tried Dobble XL.

> delusion-motivated homeschool.

The words that trigger us...who'd have guessed :) Well, it's more DIY schooling on a small scale to allow for meaningful personalization. There is some thinking and structure to it. Here's a little more about it https://blog.comini.in/p/how-can-we-personalize-learning

ChaitanyaSai··on Dobble (The Mathematics Of) (2018)
Ha! We decided to make a version of these for kids at our microschool (with their faces on the cards) for an end of year gift. Didn't realise the math was going to tricky. Found a page with code to generate combinations. Luckily, being a microschool meant we had the right number of kids plus educators for a small pack that covered all of them equally. Was fun diving into it and the kids loved the cards.
ChaitanyaSai··on Reports of the death of dental cavities are greatly exaggerated
Down to the island!
ChaitanyaSai··on We built the city of Colombo in Cities:Skylines
That sounds great. How would one go about this with a city like Mumbai?
← PreviousPage 3 of 11Next →