Demystifying State in AI: Smarter Reasoning and Real-World Legal Applications

90%. It's between 70 to 90% of your

cloud port tokens. Your usage goes into rereading

something it's already learned, which is an absolutely insane

problem to have currently. And what we want,

what we realized is our systems can start to solve that. So for instance,

we've released a free McP open source it called the Blackboard.

It's on the GitHub. If you use our Stateful Swarm McP, the Blackboard

McP, your token cost will go way down and you'll be able to

see how your AI is reasoning through things, what it looks at

and that guides our system to be evolved over time.

Hello and welcome back to Data Driven, the podcast where we explore

the emerging industry of artificial intelligence, data science and of course

data engineering, which really underpins it all. But I don't have

my favorite data engineer with me today, so it's just me flying

solo with our guest Devanche Devanch, who

is the founder and head of AI at eras. Hopefully I

said that right. And he went from creating a tech minded community of

10,000 plus people on Substack to building a

legal AI that strengthens how lawyers communicate reason and

work through complexity. He also lists himself on LinkedIn

as a chocolate milk cult leader. I'm sure

there's a good story behind that and we'll get into that. Welcome to the show

Devansh. Thank you for having me.

No problem. I saw your LinkedIn profile and I was like chocolate milk, is this

the same guy? So tell me

one Congratulations on your success on Substack. That's impressive.

What led you to

beyond Substack?

So I think what led me to be on Substack is a

more general downstream effect of me,

what led me to be on online in

general. I've worked as an applied AI researcher for quite a

bit and my specialization was a lot and low resource

environments for machine learning. But like in

2017 we had a team of three that beat Apple on Parkinson's

disease, voice detection in real time, voice calls. I worked the

state government, I worked with like climate modeling in a university,

et cetera. But the issue was that around

this time is when machine learning was really kicking off.

So things were becoming a lot more corporatized I

guess. So before like you know, you had a lot of people who

got into ML from basically just oh,

I taught myself ML and now I'm doing machine learning like Yann

Lecun. These guys didn't do ML in university,

but they happened to work in it and that's how they were transitioning.

Unfortunately, by the time I got into the industry, that path was

clearly dying. So especially in

America, you had ATS systems that would scan resumes

and those resumes would not. I basically was

not able to work with anybody that was doing ML because

any of the big companies, because they were just saying that, oh, we have a

requirement that you have to have a PhD. I was essentially

every job I'd gotten before this, I kind of had to network my way in.

And then it was quite a difficult process and

increasingly I could see that was no longer going to be sustainable.

So at that stage I decided, hey, I should get a PhD.

I think it'll help. Which was a bit of an ambitious goal

because I didn't have any university degree when I decided,

but I was working with this state government

and then that they had also got my professor that I was talking

to back and forth. He was a public health professor, so not the same

field. But as I was telling him, he's like, hey, you're very, very good

at reading papers. You have to read a lot of papers to do what you're

doing. Because, you know, we're giving governments healthcare like frameworks

for how they're going to judge that money, like how they're going to judge

investments. And if you make even a 1% error, that's about

3 million people that we were impacting through our

framework because they were deciding where to invest in health

outcare outcomes, how to move investments within districts,

etc. Based on the framework we gave them. So

you didn't want to screw up. And I was really nervous about that. So

I was reading a lot of papers and I was trying to synthesize like the

best possible outcomes, like 50 papers. And the

professor found that really impressive because that's like not a skill that

a lot of people have. And he said to me that this is going to

be really useful in your PhD applications. But right

now, because nobody's going to know that you're

able to do this, the only thing you'll have is your letter of

recommendation. I think you should start like documenting your

insights online so that when you do apply, you're able to

people know who you people can see what you've done. So

that is what led to me writing on medium. Initially, it was meant to be,

hey, I'm going to, I'm going to have a

few publications. And then the plan was that when

I did apply for a PhD, I would essentially, if

I was reaching out to a professor doing computer vision, I'd show them like, here's

five papers I've broken on computer vision. Maybe if there was somebody I was very

keen on, I'd break down some of their research. And I wasn't expecting to have

an audience. I was more expecting to say, okay, this is what I've

published, please take a look and then give me your

comments. So

that was how I started. I think what ended up

happening is as I wrote because I came from a pretty different

background, like at that stage in the information landscape and

even today most prominent people

who are breaking things down, pure researchers. So

a pure researchers sits there and they're looking at different

papers and they're like, this technique increases benchmark score by this much, this technique

increases benchmark score by that much. And that's fine,

but that's a very different outcome than where should

I put my money? What kinds of investments will that make

it? And generally if you look at, there's some really good researchers

who work even I recommend none of them have an opinion

per se or they have very minimal opinions. They try to be

as close to the papers as possible. While

what I did because I came from, you can't do that if you're coming from

low resource environment. You don't have the money to be

able to spend in kind different explorations or being like,

oh, I'm just going to do reinforcement learning on this to improve my results

because I don't have any computer resources. I'm

the cloud engineer, the ML engineer, the ML ops. So

it's just not going to be sustainable. Right when

you start thinking in that way. Like that

reflected in a lot of my analysis because I would sometimes take

papers and sometimes I'd disagree with that conclusions because I'd be like everybody

else would was looking at that paper and saying, oh, this is going to be

a breakthrough. And I would write like this makes no sense to do because if

you look at the cost for the performance percentage jump, it

doesn't really align. Or I started doing stuff like,

okay, if you were to figure out when to invest, when not to invest, here's

like a quantitative way to look at it. And I think what that

led to was a lot of more senior leaders who started

reading my book. So I kind of skipped a lot of the junior

leadership originally. And I actually became much more

popular with senior leaderships first because they actually,

for them, they're not looking for information, they look for insight. They need decision,

they need help making decisions. And I,

I was in like this unique place where I was still breaking down research

Papers I was still being really rigorous and technical, but

all behind that I was trying to give a. So what? And I was trying

to have, okay, this is how I would have analyzed the implications of

this research for the entire industry. And

that just led to the growth of the Chocolate Concord,

which is my online community.

And that's how I got online generally.

Then why substack specifically is.

What happened is I got quite big on Medium,

but what I noticed is that Medium's distribution algorithm

was just not. It no longer suited me

because I'd like to write like really deep paper research breakdowns. And what I

saw at that time was Medium had turned almost into like a

do this with ChatGPT, do this with ChatGPT being

released. It was. That's the kinds of articles it was favoring.

It was not favoring like my style of writing. So

I had massive followership numbers, but the views were not.

I was not able to pull the same views I was getting even when I

first started, like with no followers. And that's when I

realized, okay, if I'm going to be dependent on

an algorithm driven platform, same thing with substacks, sorry,

YouTube, same thing with LinkedIn. I will likely always

be reliant on this platform for everything. I should have at least

one place where everybody, whoever

finds my work is able to see it more consistently. And that was

my rationale for substance. Interesting.

Interesting. No, I think that there, there was definitely a time

when somebody who could translate those very

dry academic papers into something practical and something actionable.

I think you were providing a great service now. But also

I think at the time there was definitely a lot of people that wanted to

get into this space. And I think you prove out kind of like the idea

of you do the traditional path of get the education

grind that way. But I think there's also this notion of becoming,

for lack of a better term, an influencer. Although I do cringe a little bit

when I say that. And you smiled when I said influencer. So I'm

sure you hear that a lot too. We need better words for it. But I

think that there are definitely people out there who have a gift for taking these

really complicated subjects and turn them into something that

more people, senior leaders in industry can really understand.

I think you're doing a great service here. But how did the chocolate milk

thing come about? Because I'm. It's very memorable. I will say that from a

branding point of view. So when I was in university I was

accused of trying to start a cult, which is untrue

I was starting a religion. Okay, one minute it's just a

religion. So that's the cult part of the chocolate milk cult.

It used to be called the tree club. The religion and

that's the card part. Chocolate milk just, I just like chocolate milk. So

when I started writing originally or when I,

at that time I was exploring YouTube, I was doing a bunch of other stuff.

I kind of had a rule for myself to make sure I was more consistent

that I would only drink chocolate milk when I was doing some work.

Oh, I would not drink it for leisure. Okay.

Like, I would not. It's actually one of my greatest productivity

hacks. I, I, I mirrored chocolate milk to that and

I mirrored music to that. I cannot listen to music for pleasure.

Interesting. I can only listen to music today,

even right now. Like E. If I'm traveling, it's an exception. Like

I'm just going somewhere. But other than that it's only like if I'm going to

work out, I'm walking somewhere or

if I'm working. Interesting. It's actually

like that way you anchor something you like with

something you don't necessarily want to do. And that's how

you, that's how I was able to stick to the hob, to the commitment.

So the chocolate milk just came from that.

Interesting. Interesting. What do you

tell me about your startup? You

specialize in lawyers or using AI for lawyers? Tell me

a little bit about that. Yes and no.

So what we have is build. We

solved a fundamental reasoning issue, which is that

reasoning over long context is brittle, it's expensive

and it's not trustworthy. If let's say

you are to take what are the

options for reasoning over long context in today's word Frank,

you take a massive language model, you stuff all the context

possible into it and then you give it an output that's

option model. And there you're just betting that language model

context windows are going to get larger and larger and they'll be able to

handle this more effectively.

That doesn't work for a few reasons. One, if it fails, you don't know where

it's failing. Is it the models and knowledge? Is it random probabilistic

error? Is it context rot and attention dilution? That's ticking

in. There's just too many variables that went wrong. Second

is you get systems like wrap simple rack. I should

say rather, which is what a

lot of the other people in legal AI are currently doing, whether they

explicitly say it or not. It's, you have a vector data, you have

a vector embedding model and you're just asking questions by

vector embedding model and it pulls out a few chunks and those chunks form

the basis of your answer. No matter how much you tune your

embedding models, the issue with them is always going to be

that you, again, don't control it. Because when it's the

embedding model decides it's getting pulled out or not, so it's not your system.

So again, if there is an error, you can never control it. You can never

say, oh, this keeps pulling out the wrong clauses. Let me fix for this

clause and change everything else. Because if you retrain the embedding model, everything

changes. There is no way to do targeted surgery on neural

networks right now. Two, and this is the absurd part, like

vectors themselves have limitations. Like, mathematically speaking,

they will never, they will always pull things that are cosines

that have cosine similarity that's high to them, and they'll miss things that

don't. So if I ask you, hey, analyze this clause, it might give me answers

to this clause. But that clause might have dependent clauses that are not

the same language. So the vectors will not pull

it out. But it should have pulled it out to analyze, to give a proper

analysis. Or, you know, if somebody is asking you about your life,

I. They might. If it's pulling out databases from you,

they'll talk about Frank. But there are chunks of your life where it's not

you. There's no Frank. There might be a brother. There might be somebody

else, somebody with a completely different name, somebody your best friends.

These like, is that an embedding issue or is that like

a where's where? How can that be fixed? Because that's a math

issue. The math issue, yeah. Mathematically,

there are going to be chunks that don't have high

cosine similarity to your original question. What

rag does is simply what. This is my question. I'm

going to try to pull out the chunks that are

closest in word meanings to this question. It's going to look

at what are the words here, what could they mean? And it's going to try

pulling it out. But there are going to be chunks that are extremely important that

don't have the same verbiage. That makes sense. They're not,

they're related, but they're not semantically obviously related.

Yes. And there's a flip side to that. If you were working, you can

have lots of chunks that are very structurally similar. Again,

going back to legal, you have clauses that are worded very similarly with one

or Two differences in words where the words are very operative so

you rag will pull those out and pull

put them very strongly and then again you're diluting your context, you're

creating the issue of possible conflicts. So not only does

traditional vector miss like miss things that are

related but not semantically similar, it

misses things that are, it adds things that are semantically similar but

completely unrelated. So that's why legal AI, a lot of them

fail very substantially. On top

of this now you might have agentic reasoning,

agentic retrieval here you're using vectors with a few other things

depending on how you want to build your system. This is by

and large what everybody has converged to on the state of the art.

But the issue here is threefold. One, it's extremely expensive

because you're basically just pulling things out and trying

to reason over it multiple times. So you have multiple cycles of encode, decode, report,

etc. Two, it's not very controllable because if you

were a lawyer, imagine you're just sitting there and

imagine asking a lawyer to look at cloud board and put all the

read through all the dumps as it's doing can get very out of

hand very quickly. Yeah, so you're just, you're

it. It seems more transparent on the surface but because of the volume of

sub agents that work and because you can't really say okay, my sub

agent has learned this, my sub agent has learned that you can't really

attend to what it's doing. You can only still modify the output

and while it is a higher quality than standard drag, it's not

safer, it's not more trustworthy and trust is a big factor for

anything in regulated industries. Because let's say I

save, let's say I'm 99% accurate, which you

we're not going to get there anytime soon. I only have a

1% error, but I don't know when that 1% happens.

So what do I have to do as a user I have to test, I

have to still cross check every output. Either I'm just going to

go in blind and say okay, I'm going to get this right. 99 out of

100 times and 1 out of 100 I'll get sued and get disbarred

or I have to check every output anyway. So all the time I would have

saved, I'm not saving it. So

that's what a lot of those other systems, that's where a lot of those other

systems fail and that's where we decided to really

pitch our flag is We've built

reasoning that is transparent, it compounds

over time. So unlike cloud code or any agentic retrieval systems

where the questions you ask them, that you ask them and then they forget the

memory so they have to relearn everything every time they reusing their tokens

and they'll forget your instructions and they can't form like complex

user memories on how to personalize to you. We are able to

personalize the people and very importantly, everything that

our AI does can be audited and checked and even steered

mid as its reasoning. Like when you use our AI systems, you can steer them

mid reasoning and that allows people to trust our work much more.

Does that this is obviously we know

what the answer should be. Does this save

lawyers time or does this introduce risk and they have to

validate the inputs? I can

imagine if I'm a lawyer, I'm very concerned about that.

It's both. Okay. We have seen our users report

roughly a 10x increase in time spent. So 1 hour

on iris gives them 10 hours of work output. Oh, nice. Okay.

And that is largely because that one r any

AI can draft for you in minutes. Drafting isn't the

issue. The issue becomes

can I change it? If you were to ask Iris to change a paragraph, you

ask Iris, hey, can you make this? Can it be done quickly and can you

inspect Iris outputs to be like, okay, you're telling me that this is

the correct summary of the case or this is the right outcome? Can I

look at this? Can I verify it very quickly? Can I look at all the

facts you're looking at? Can I look at what other counterfactuals

you looked at? Can I look at your simulations? And based on that, can I

make the answer? And that's why IRIS is not just like we

have a lot of lawyers using it. But what we've seen is

tremendous amounts of demand from non legal people

in similar situations who just want an AI they can trust,

long context AI they can trust. That's what we have.

We've seen a tremendous amount of demand for our API of

reasoning for other versions of Iris. And that's

what we are working on actively serving right now, even as we continue

to tackle the legal market. Do you think that you could do better

than 10 to 1? Do you think you're just at the beginning of this

or is 10 to 1 pretty good? I

mean, it's pretty good, but that's a really good question. I think

we certainly can do better than 10 to 1. The

question becomes, do we want to do better than 10 to 1,

I guess I don't have a strong opinion on this either way, but my

intuition right now from our user experiences is that the

things we can say for them right now is like make it,

make the export to Word documents more reliable,

make it so that I can do things

otherwise. And that's,

that can certainly be useful but that also starts to become a lot more

fine grained and user specific and I don't know if that's the right

approach to take right now versus to explore other

avenues of improving reasoning and hitting

more fields 10 to 1. That's still an open question for us.

One of the things we are doing that should help people is building more

connectors into the platform. So Imanage

docusign all of those through Iris and that should, that

will save people a lot of time because that is currently one of the biggest

friction points with our platform right now. But again, I do

think there's a diminishing return to just improving

ux and I do think the AI market is still

so open. It seems like a crazy thing to say, but

most AI startups are not that good.

So there's a lot of these spaces where you think that, that

people already exist, but they're not true solutions. They're just a bunch of kids out

of YC that don't know what the they're doing and they're just building, they're

just building the simplest possible thing they can and you can completely

wipe them out of the market. So there is no, I mean you say that

in Jess, but it's true. Like when, when Chat GPT goes down, those API

goes down. All these other like tangential startups that

don't disclose that they're using OpenAI APIs, they just make

it. If you look at their brochure where it looks like they invented

themselves, clearly they didn't. Right.

So next time you notice that when AP, when OpenAI has

an outage and all these other companies kind of go down. Yeah.

So what makes your approach different? Like how did you get to

seeing the problem and then finding a way to fix it?

At our core, we believe state is the number one

leverage point in agentic systems right now. So

what I mean by that is if you've done traditional software engineering,

you know, you had state charts, what is your code doing right now

versus what is your code doing at T +1?

And then you debug through those to say like, you know, either you have

a bunch of print statements, which is always my favorite way of doing things, or

you know, You're a nerd. You do like debuggers and logs and whatever

other nonsense that those people do, right? But

you track how your system has evolved over time and that

was what led to you being able to debug very effectively.

AI doesn't have state like when

Claude Go makes a mistake, you don't know where it made that mistake. You

can ask. The only way you can do it is by asking Claude Board

to run the same process that made the mistake again and say,

hey Claude, why did you make the mistake? But there you are

offloading what is possibly the most important aspect of

your work to a system you can't control.

If you were reasoning, if your AI is reasoning through different

solutions and circumstances, you don't know

why it comes up with the answer. It does. You just hope it does. And

then you're again reliant on an external system to say

this is make sense or does this not make sense? But you

cannot evolve your process over time. And what that means is

your returns on AI are flat or they're depreciating.

At best your AI system will stay like this with more

feedback or it will go down because it's going to get

conflicting signals. What we want is this

compound and get compounding

returns on your AI system when you know exactly where it goes wrong

and you can specifically fix those aspects. So what

we did first was build very rigorous

tracking of the AI is the engine state. What is it

looking at right now when it goes to the next step? Why did

it go to that next step? Can we see what model weights like? Can we

see how our latents evolved over time? What

strategies were they taking? In the agentic reasoning aspect,

it extracted certain clauses. What other questions did it come up with?

And based on those questions, what else did it do? And that way we are

able to compound our AI systems over time because we know over time what

failed, what didn't and we can specifically fix those aspects.

Interesting. Interesting. By the way,

I use specifically two examples. Evolving latents and

looking at what the state of the agentic system and seeing what's

compounded over time. The reason I took these two examples is because

we've open sourced both, so you don't even have to take my word on it.

You can go and use the systems yourself. And we've beaten

state of the art on both with both on much cheaper. We

are much cheaper than state of the art and we're much better than state of

the art in both the systems that we built.

So is state

analogous to context like how what is the relationship between context

and state? Context is what

your user gives you or what your AI model

extracts. State is the

representation of what your reasoning system is

working through. So

state, context is here is a PDF,

here is a user instruction, here is a preference for

analysis done a certain way in a law firm. That is

context. And that's why I think

context engineering is a bit short sighted. Because you're working on the inputs and

inputs are very important, but

your users will never give you enough input to actually

solve any complex problem. They're not going to say,

look at clause B, then draft up this way, then add this, then look at

this PDF and do that. They will just give you very generic, high

level instructions. Give me a defense strategy.

And again, that goes back to our failure with rag. Why would RAG fail? Because

it would try to look for world passages similar to

give me a defense strategy instead of looking for actual defense

strategies. So state would be what our

agent reasoning system is going through. How is it representing data in

turn? How is it deciding that? If I have to give a defense

strategy, Here are some questions I can ask. Based on these

questions, here are some things I pulled out. Should I change my questions?

Should I pull out more information? It mentions this person, should

I look into this person more, where can I find more insight on this person,

etc. That is state. So the

state is essentially the map that our own AI agent system builds

to navigate the world and the move it

handles, context. The more context you give it, the more you use it. And that's

why I said it's compounding, the clearer the map becomes.

So what ends up happening is the state becomes

user one. It becomes much more powerful to your, your

relationship with tokens flips. Because while

it's expensive to build a while

the cost of building the map is an added cost and it's

overhead, the more you build the map, the

more your cost per output goes down because over many

iterations your system stops looking at the

overall general, like it doesn't have to

read all of the files, reread all of the information. It's just looking at the

map to say, okay, I should read this specific passage, I should read the specific

chunk and then your output is able to be computed much more quickly,

much more cost effectively as opposed to having duty,

you know, process a bunch of irrelevant documents. Always

interesting. That's interesting. I think,

I think there's a lot. I think we're still early in the AI phase and

I think you're right in terms of there's a

lot we haven't really figured out, particularly when you think about agentic. So

agentic. So state is basically implies you

can't have really state unless you're doing something agentic. Is that

a fair assessment? What do you

mean a state in LLMs. Right, because you're not. Or

in AI. Right. Like state is tied to what the agent is doing

at time T.

Yes. Right. So it. So we have

context and now you're adding state, but state, you can't

have state without really an agentic system

is what I mean. Yeah. Okay,

interesting. What made you

want to solve this problem?

I think so. When I met my co founder,

I had already built this open source research community, so we were talking to

a lot of people who were having similar issues across the

board. And when I met Savvy, the CEO of Iris, like

one of the first things we talked about he told me was like, hey, this

takes too long. I spent way too long in court and we should be making

this process more efficient. And democratization and

access is something I really care about, which is why we open source so much

research, which is why it would be much easier for me to be writing

simpler, do this with AI, do that with AI, make this with AI

articles like the kind that do well on social media. But I chose not able

to that and stick to the cutting edge and try to make that more

accessible to people. Because I think ultimately if you don't

have something that brings the frontier to

people will always be like dependent on

whatever handouts that big tech companies are willing to provide as opposed to building

their own alternatives. So

democratization to me is extremely important. And

that's why like when we spoke about this, that was

like a more personal resonance. So we decided

like, hey, let's try to build something here. He was the head of innovation

at a pretty large law firm, like a massive law firm actually.

So he had already tested Harvey Legora. At that time they were called

Layup, Axton, et cetera. And none of them were that good. And they were not

that good because again they were at that time just stuffing things into

language models or they were trying simple rag and none of that was working.

So as we started solving for legal,

like what problems do you have? We identified these variables. And

then I think as I started talking to people about how we were solving it,

we realized that this problem kind of extends way beyond

legal because Claude code again for coding

agents is the same story. It is not able to track your

state over time. It's not able to compound over time, which is why you waste

a lot of tokens relearning stuff. 90%,

it's between 70 to 90% of your cloud port tokens. Your

usage goes into rereading something it's already learned,

which is an absolutely insane problem to

have currently. And what we realized

is our systems can start to solve that. So for instance, we've released a free

McP open source it called the Blackboard. It's on the

GitHub. If you use our Stateful Swarm McP, the Blackboard McP,

your token cost will go way down and you'll be able to see how your

AI is reasoning through things, what it looks at and that guides

our system to be evolved over time. There's a lot of this

kinds of similar questions here that we could be solving. And this is why

we think that this is the perfect place to stick your flag in and then

expand your area. Because all of AI right now

needs to get better ROI on their

inputs and we believe that this is the best way to do.

Interesting. What would be your

advice to someone who wants to get a startup but

not be another one of the, the people that are just

another, you know, as you said, YC kind of API

vibe. Coders like, what would be your advice to them? They want to

build something impactful. What would be your advice to them?

I would say it depends entirely on

one. You have to look at where your strengths are

because startups in themselves have.

There is no startup, it doesn't exist. It's different

startups, very individual. For

instance, if you're a very good salesperson, if

you're a very good GTM person, you might want to consider

doing something that has a bit of a consumer motion

because that's how you can. You have the ability to grow your stuff very

quickly. Contrary, on flip side, if

you're somebody that's really, really good with foundational work, if you're

very good at making these big breakthroughs and you're like

an industry expert that comes in with connections, you want to do something that

will play to those trends. So the first most important part about a startup would

be to know yourself and think about who you are

and what kind of a thing you'd like to be. Because

ultimately every startup ends up being a reflection of its leaders.

And you want a company, you want a process that compounds

your strengths while minimizing your weaknesses as opposed to doing what?

Trying to do everything and fail them. And I think like an example of that

is something like Cluli, for example, is a

startup that

I don't think they understood the market very well, they didn't understand what their own

strengths were, etc. So they

go viral on social media, people look at their clips, etc.

On social media but the amount of investment they have to put in

for that when you compare to the return of does the buyer actually buy

is pretty negligible. They get a lot of user

headspace retention but they're not Coca Cola. So you know if, if

I were a company like Cluli I would have, I would think about

either don't try to sell to enterprises as a marketing solution. Like

we have note takers here, none of them are Cluli. I have never

met somebody personally in marketing and sales etc. That

uses Cluli and I'm pretty connected with a lot of the CRM top

CRM startups. I'm pretty connected with Salesforce, HubSpot,

etc. I have a pretty good insight into what GTM people are

using and I have never met or heard of anybody say

clueless Enterprise is what I really really love in

the world. Everything I know about Cluli comes from the few people

that use them for consumer reasons. So that kind

of I think thinking about where you should be,

what your strengths are, what you want to play to and then maximizing that is

extremely important. The second thing is also just to know your

exit. What are you doing this for? Why are you doing this?

Because otherwise you will probably have a lot of conflicts in

your life because you know

this is something that if you want to make it work especially

nowadays with AI where investors are pretty

unless you're like a template founder which is like you're from Stanford,

you from Palantir, you've raised a bunch of money for other

startups before or you worked at another big startup before,

you're not going to VCS aren't going to look at you twice. So you need

to have growth, you need to have something other than that to

stand out to them. And how do you get to that something? You kind of

have to figure things out. Cloud code will make a cloud code and on

make it so that on the surface building features becomes very easy. So

there's a lot of there's a big flood of copycats, there's a big flood of

other things people doing it so you can just do a startup for the

sake of doing a startup to get that experience in but

you're not going to both it's not going to impact you as

much and it's not going to help the

your Startups probably not going to succeed unless you know, you're really

willing to put a lot into it. So you have to figure out what it's

that you're willing to put in. When do you leave? When do you walk out?

What's your minimum exit? Because that will teach you.

I can give you a lot of insight. Otherwise you're just going to be like

going with the friend and not going in any particular

direction. Interesting.

Where can folks find out more about you and what you're

up to? The substack is the most

active place I'm in. Just the chocolate milk cult leader

on substack is my substack handle. Or if they

look for artificial intelligence made simple by the bunch, they should find it

pretty easily. Okay, that has all my social

media links but also LinkedIn. If you just google me,

you should be able to find my work. Okay.

No, I taken a quick browse through your your substack.

It's pretty cool. And most of it's free. There is a subscribe

option too from what I can tell, but. And you're based in. In New York

City. That's cool. I grew up there,

so. But I also saw that you went to Rochester,

right? I was. You weren't. You went, but you

studied in Rochester Institute Technology. Is that what it is? Cool.

No, it's a great school, people. I know a lot of smart people to come

out of there. Plus Wegmans is headquartered up there, if memory serves.

So you probably bought more than your fair share of chocolate milk at a Wegmans,

I would imagine. Not as much as I would have

liked, honestly. I didn't spend a lot of time in Rochester,

surprisingly enough, because Covid and I was

traveling and I was doing other things. So I actually didn't spend

a lot of time physically in Rochester. Okay.

Yeah. The pandemic, I think threw off a lot

of what we would call normal things. Right? Like

normal life experiences. Very cool. So I

will definitely make sure that those are in the show notes and any parting thoughts?

I think 1 thought I try to leave people with

generally is that the AI

space is a lot more open than you would first

think. And there are probably a lot more

aspects for innovation than people are realizing. I think

people look at what a lot of big labs do and they get locked into

I want to build a model, I want to do post training, I

want to do something else. And then they end up throwing

their attention there. But if you take a

step back and try to look at the larger space, like we

don't understand neural networks very well. What we're doing with reasoning, like

we don't have inspectable reasoning, we don't have a lot of other stuff that

other people have not solved yet. So

if there is a problem that you really care about and you'd like to solve

it, it's actually a really good

time. And a lot of this will not require a lot of money to take

a step back and say what is the industry ignoring and focusing

your attention there? Because I do think sometimes

people can be a bit narrow minded in what AI research looks like and what

would AI researcher should do. And that's just a function of the

fact that most influencers are not. The reason I laughed at the

influencer comment earlier is because I don't. I'm not actually interested in

being a content creator. I share my research online but

my first job is to do research. Most influencers don't

actually are pretty disconnected from the field. They're they actually

rely entirely on whatever is public opinion

to shape their thoughts. And that's why like most people are

only exposed to three, four ideas or three, four big trends that

are happening. But in this under the surface, like if you start digging into

what problems are there to solve, you realize very quickly that

pretty much anything is open. So that would just be my general

opinion is that people are

severely underestimating how open AI is and

where all the areas they can come in and make a difference. So if there

is anything, just look for those gaps and

see what they can do there. Don't listen to all this

hype around permanent underclasses and that AI will take

everything and there's no point because Claude will take

a job because we're not really ever going to be there.

Yeah, we see this every time there's a major

massive hype wave, right? Like I'm old enough to remember the dot com boom and

the dot com crash, right? It certainly you can still

walk into anywhere in America or most places around the world and go into a

retail store, right? They still exist. Malls aren't what they

used to be. But by the same token retail still

is a thing. It really changed in a lot of ways.

It evolved but the world didn't end for retail, brick and mortar and all

that sort of thing, right? No, you're right. Like, and I like, I like the

way you frame that is it's open in an opportunity space kind of way. Like

I think you're right. I think we're very early on this and if you think

about it, all the startups that we are, that are household

names today are

10, 15 years old. Maybe like Uber Lyft. Just look at

your phone. Right. Most of the apps there, Amazon's probably the oldest

of the crew and Google too. Right. But beyond that, a

lot of the apps that we use, new services, Doordash,

Instacart, these are not. These are relatively. And this

was what the reason why I'm showcasing them is that when the dot com boom

and bust happened, it was it. That was it. There was no more technology startups

clearly that didn't happen. So I like the

optimistic approach you have.

Yeah, I just think people tend to be

very quick to diminish what, yes, they could be

capable of. I think people seem to be looking for

excuses for not this, but I

do think that there is a lot that can

be done if you're so inclined to do it.

No, that's a good way to. I like that. I like that approach. It's very

optimistic. Right. It's very growth centric and opportunity seeking

as opposed to stirring up your hands and waiting for your

ubi. So with that,

we'll end the show. That was great. Thank

you for having me. Hey, no problem. This was great. Definitely curious about it.