It takes time to create work that’s clear, independent, and genuinely useful. If you’ve found value in this newsletter, consider becoming a paid subscriber. It helps me dive deeper into research, reach more people, stay free from ads/hidden agendas, and supports my crippling chocolate milk addiction. We run on a “pay what you can” model—so if you believe in the mission, there’s likely a plan that fits (over here).
Every subscription helps me stay independent, avoid clickbait, and focus on depth over noise, and I deeply appreciate everyone who chooses to support our cult.
PS – Supporting this work doesn’t have to come out of your pocket. If you read this as part of your professional development, you can use this email template to request reimbursement for your subscription.
Every month, the Chocolate Milk Cult reaches over a million Builders, Investors, Policy Makers, Leaders, and more. If you’d like to meet other members of our community, please fill out this contact form here (I will never sell your data nor will I make intros w/o your explicit permission)- https://forms.gle/Pi1pGLuS1FmzXoLr6
Given how much you guys love the livestreams, decided to do this one as a livestream. Lmk what you think of it.
Last Year in Review
Some fun numbers/accomplishments I’m happy with-
Follower count- 166K—> 242K on Substack (310 if count all other platforms as well).
Revenue from paid subs: 20.5K (264 subs)—> 67K (935 Subs). It’s very interesting that many people choose one of the higher-tier subscription plans, even though there isn’t any “incentive” to do so (we have a pay-what-can system to support the mission of democratizing AI/Tech to everyone, irrespective of their financial ability).
Every “meet me in this city” led to lots of responses and we met over a 100 people in the trips we took.
Total Views: >15 Million
Foster cats we helped find a home —3 (I have 2 more with me now, if you want them).
Gallons of Chocolate Milk Consumed >=1.
The Year Ahead
Three focuses for this year:
1. Open-source the reasoning infrastructure. Latent space reasoning is already out. Knowledge Injection framework is next. The goal is intelligence that’s too cheap to meter and can be personalized for people in low-resource environments. AI for everyone, not just the ones that matter enough to show up in TAM calculations.
2. Geometry of intelligence research. Can we isolate knowledge in neural networks? Can we identify which representations encode which capabilities? Can we combine the best parts of different models—GPT-4’s writing quality, Gemini’s multimodal reasoning—without full model merging?
Early results are promising. We’ve localized knowledge to specific activation patterns, made edits to it, and transferred it across models (including across different paradigms like Transformers —>SSMs).
3. Break Palantir.
I’m not going to pretend the surveillance-industrial complex isn’t happening. And I’m not going to pretend that posting about it is a serious response.
Here’s what sickens me: tech now considers surveillance a good business. Investors celebrate Palantir like it’s some kind of success story. What Palantir actually does is suck money out of public programs—money that could go to infrastructure, education, healthcare—and redirect it into building surveillance states. They’re government contractors whose product is watching people. And the tech industry treats Peter Thiel like a visionary instead of what the hypocrite welfare princess he is.
I’m not naive. Defense tech exists. Militaries exist. The military-industrial complex isn’t going anywhere. But there’s a difference between acknowledging reality/investing in it as insurance and celebrating companies that profit from curtailing public freedoms.
Last year we went after Amazon’s worker surveillance systems. Published the deep dive on how their tracking works and how workers could circumvent it. Had a few fun conversations with the workers there. Got a few threats for that one. The fun stuff.
This year, the target is Palantir.
The goal is concrete: break one of their systems. Either figure out how to defeat one of the weapons systems they’re deploying, or hack one of their tracking technologies and publish the methodology.
Will we beat them long term? Probably not. They’re a massive organization with essentially unlimited resources and government backing. But resistance has to start somewhere. And I’d rather spend time on this than pretend it’s someone else’s problem.
We’re going to disrupt them. And when we do, you’ll read about it here.
Livestream Companion
Watch the full stream for the elaborations, tangents, comments and more.
The “Keeping Up With AI” Problem Is Misspecified
Here’s the question I get more than any other: How do you keep up with AI?
The question contains its own failure mode.
What people think the problem is: There’s too much happening. Papers, model releases, startups, research directions. The firehose is overwhelming and I’m falling behind.
What the problem actually is: You’re optimizing for coverage when you should be optimizing for depth-that-transfers.
Let me explain what I mean.
The Surface Area Trap
AI as a field has exponential surface area expansion. In the last few months alone, you have: interpretability breakthroughs from Anthropic, diffusion model innovations from Tsinghua, new inference optimization techniques, domain applications in nuclear materials discovery, laundry detergent formulation (yes, really), legal reasoning, medical diagnosis, protein folding variants, robotics control, chip design automation.
If “keeping up” means tracking all of this, you will fail. Not might fail—will fail. The field is moving faster than any individual’s consumption bandwidth.
But here’s what most people miss: broad consumption doesn’t build the skills you actually need.
What Actually Works
When I was doing Parkinson’s detection on resource-constrained mobile devices, I learned statistical rigor. I learned how to account for edge cases when you don’t get second chances. I learned how activations behave under compression.
That knowledge transferred. When I started working on latent space reasoning, the intuitions about information geometry, about how representations constrain capabilities—those came from going deep on a narrow problem years earlier.
The pattern is: deep expertise in one area builds transferable abstractions that let you learn adjacent areas faster.
So my actual framework is:
80% ruthless focus on 2-3 areas directly relevant to what you’re building right now
20% exploration on adjacent curiosities without pressure
The 20% matters because it’s where unexpected connections happen. My current geometry of intelligence work—isolating knowledge in latent space, making edits to it, understanding self-similarity in neural representations—started from a casual conversation about fractals with someone who isn’t even a mathematician. Fractals led to self-similarity, self-similarity led to thinking about how information representation dictates capability, and suddenly I had a new research direction.
But that connection only happened because I had deep foundations to connect to. Without the years of work on representations and activations, the fractal conversation would have been interesting and then forgotten.
The counterintuitive move: To learn faster in a fast-moving field, go narrower and deeper. Not broader and shallower. Our complete guide to learning AI is shared below, and has helped a lot of people. Will likely help you too.
Why Legal AI
I get this question constantly, so let me just explain it.
I come from India. And in India, one of the biggest problems is access to justice.
If you’re a daily wage laborer—meaning you make money each day, hand to mouth—and you don’t get paid, you’re screwed. Your recourse is the courts. The courts have a 10+ year backlog. That’s not an exaggeration. That is how long public court cases can take.
So now you’re a laborer who didn’t get paid, and your path to justice is a decade-long wait. You can’t survive that. You just eat the loss and get exploited again.
Farmers get hit the same way. Guys come sell you pesticides on credit, you rack up debt, one bad harvest and your whole family is underwater. Try to get legal help? Same story. The ROI isn’t there for lawyers, the timeline isn’t there for you.
I’ve seen this up close. Personal experience with it.
So legal AI was never some market opportunity I identified. It was always the thing I cared about. Even before this newsletter, I was making YouTube shorts about projects worth doing with ML for legal. Looking at LegalBERT. Asking why we weren’t using machine learning to help people find relevant cases faster.
When I met Sabih, my co-founder, we clicked on exactly this point. Legal needed to be more accessible. Access to law had to become cheaper, faster. People needed high-quality access to justice regardless of whether they could pay BigLaw rates.
I’d had conversations with other legal AI startups before Sabih. Didn’t work with any of them. They were venture babies—high valuations, getting pumped up, but leadership had zero interest in actually making legal AI accessible. They were building for BigLaw because that’s where the money is. Replace expensive associates, charge enterprise rates, print cash.
That’s fine as a business. It’s not what I wanted to build.
With Sabih, the philosophy aligned. That’s why Iqidis is free for anyone to use—not like Harvey or Legora. No secret gotchas. And Sabih built out this incredibly generous pro bono plan, completely on his own initiative—if you’re a lawyer working immigration justice, employment disabilities, any socially conscious work, you get massive discounts. Because those lawyers aren’t making BigLaw money and they’re doing work that actually matters.
We’ve had people use the system to fill out insurance forms and file claims they wouldn’t have known how to handle otherwise. That’s the stuff I care about.
Legal AI should not be legal AI for venture capitalists and BigLaw. Legal AI should mean everybody has access to law. The laborer who didn’t get paid. The farmer underwater on debt. The person trying to file an insurance claim. That’s the mission.
The Open/Closed Domain Framework
Question: I’m an Engineer at a Series B startup. We’re just building GPT wrappers. How do I push leadership toward real differentiation without getting fired?
To answer this question, you must first ask yourself the following (honestly): Is your domain open or closed?
What These Terms Mean
Open domains are where the intelligence itself is the bottleneck. Foundation models get you maybe 30-60% of the way to a working solution. The remaining gap requires actual AI innovation—novel architectures, new training approaches, better reasoning systems.
Examples: Legal analysis requiring multi-hop reasoning across thousands of documents. Medical diagnosis requiring integration of longitudinal patient data with current symptom presentation. Scientific research requiring hypothesis generation and experimental design.
We had a law firm working a medical malpractice case out of Arizona. They had 10,000+ documents—years of a patient visiting hospitals, every record, every chart, every note. They’re trying to establish malpractice.
Here’s what makes this an open problem: You’re not just searching documents. You’re reasoning across time. You’re cross-referencing Arizona state laws on medical best practices. You’re asking: does this specific disease justify emergency care criteria? If the patient was unresponsive, does that authorize intervention without consent?
And critically, lawyers aren’t going to think to ask a lot of these questions. They don’t know what they don’t know. It’s on the system to surface: “You’re claiming this was problematic, but the patient was unresponsive at that moment, which triggered emergency care protocols.”
That kind of reasoning—long chains, cross-domain knowledge, surfacing what the user wouldn’t think to ask—is genuinely unsolved. RAG doesn’t do this. Contextual AI’s current products don’t do this. You have to build. Hence, the various AI breakthroughs we have accomplished.
Closed domains are where intelligence is not the bottleneck. GPT-5 or Claude or whatever frontier model can handle the core task adequately. The difficulty is elsewhere—integration, data access, workflow fit, user experience.
Examples: Sales enablement (drafting proposals, summarizing calls). CRM integration (pulling data from Salesforce + Slack + email). Document automation (generating contracts from templates).
Why This Distinction Matters
Most engineers asking “how do I push for differentiation” are assuming their domain is open when it’s actually closed. And in a closed domain, the GPT wrapper is the correct technical choice.
This isn’t a failure of imagination. It’s recognizing where value actually comes from.
In closed domains, your users don’t care about the intelligence. They care about: Does this integrate with my existing stack? Does it save me time? Is it reliable enough that I don’t have to double-check everything?
If you’re in a closed domain and you spend six months building custom AI instead of building integrations across data sources, you’ve made a negative-ROI decision. Your leadership might actually be right.
The Harder Case: When It Is Open
If you’ve honestly assessed and your domain is open—intelligence is genuinely the bottleneck—then you need different framing to get organizational buy-in.
Don’t pitch “technical moat.” That phrase means nothing to leadership. Instead, frame as one of two things:
Frame 1: Competitive Defense “Competitor X does [specific capability]. We’re losing deals because of it. Here’s three customer conversations where this came up as the deciding factor.”
Frame 2: Revenue Unlock “Our users keep asking for [capability]. We can’t do it with current architecture. Building this opens [specific market segment] that we can’t access today.”
Both of these tie AI investment to money. That’s the language that moves decisions.
The Implicit Bet
One more nuance: a lot of leadership teams are making an implicit bet that models will improve fast enough to close currently-open problems.
“Yeah, our document summarization isn’t perfect, but GPT-6 will probably handle it fine.”
This bet is sometimes correct and sometimes wrong. But if your leadership is making it, you pitching harder on technical differentiation won’t work. You need to either change the bet (with evidence that the problem won’t be solved by scaling) or accept it and focus elsewhere.
If you know that differentiation is the answer, and your only challenge is to frame this in a way that appeals to management, the advice here will help you a ton—
Build vs Buy: The Actual Decision Tree
Default is buy. Always buy unless you have a specific reason not to.
Here’s why: engineering time has opportunity cost beyond salary. Every hour your team spends building commodity AI infrastructure is an hour they’re not spending on whatever actually differentiates your product.
The exception: Build when you need control over outcomes for accountability reasons.
At Iqidis, we made this split explicitly:
We buy the final synthesis layer—the LLM that produces the user-facing output. Closed models are more stable. We don’t want to deal with inference scaling, load balancing, GPU utilization. And they’re constantly improving, so we get upgrades for free.
We build the reasoning infrastructure. When a user asks “why did this legal analysis screw up?”, we can’t say “Claude had a bad day.” We need to show exactly where in the reasoning chain we made a mistake, and we need to fix it. That requires control.
The heuristic: if you can outsource something and a vendor’s failure mode is acceptable to your users, buy. If you need to own the failure mode—to diagnose it, explain it, and prevent it—build.
To give you some context on how expensive Open Models can be to deploy (as a case study for how expensive building can be, even when you don’t do the very expensive parts yourself), read the following deepdive—
The NVIDIA Thesis (Ecosystem, Not Hardware)
Surface-level take on NVIDIA: they have a hardware moat. GPUs are best for training, alternatives haven’t caught up, they win.
More interesting take: NVIDIA’s real moat is ecosystem, and ecosystems compound differently than moats.
A moat requires active defense. You have to keep building walls, keep innovating on hardware, keep fighting off AMD and custom silicon.
An ecosystem passively reinforces. Once you’re building on CUDA, once your entire toolchain assumes NVIDIA infrastructure, switching requires active effort against the path of least resistance. Unless you have a specific grievance with Jensen Huang, you’re not going to take on the inconvenience of rebuilding your inference stack elsewhere.
NVIDIA understands this. That’s why their strategy isn’t just “make better GPUs”—it’s invest in open source, invest in startups (CoreWeave, OpenAI), make themselves central to how AI gets built. Every startup that builds on their ecosystem reinforces it.
The bubble implication: Even if circular-funded AI startups collapse, NVIDIA (and Microsoft, and Google) survive. They have profitable core businesses. A correction means they go on an acquisition spree and buy distressed assets cheap. The ecosystem grows either way.
Our deepdive into how hardware will change over 2026 and the pressures evolving the ecosystem into the next stage are shared below—
Why Your Technical Moat Won’t Get You Funded
If you’re asking “how do I pitch my technical moat to investors who can’t evaluate it,” you’ve misdiagnosed the problem.
Why Technical Moats Don’t Convert
Early stage: If you built it with limited resources, someone else can too. By definition, your technical moat isn’t defensible at the stage where you need it most.
Late stage: You should have business metrics that matter more. Revenue growth, retention, unit economics. If you’re still leading with technical architecture at Series B, something has gone wrong.
VC incentive structure: VCs make money when you raise the next round (or exit). They’re pattern-matching for “who will other VCs invest in?” Not “who has the best technology?”
This is why you see the template founder pattern: ex-Google + Stanford + previous exit. Not because that profile builds better companies, but because that profile raises follow-on rounds reliably.
What Actually Works
Path 1: Bootstrap → Traction → Fundraise Skip the pre-seed/seed cold outreach entirely. Build a product, get users, generate revenue. When you have numbers, investors come to you.
Iqidis example: $60-70K bootstrap → built researcher product (free) and core product → 20K MRR in two months of public release→ $1M+ ARR in eight months → then raised seed.
The technical moat conversation happened after they were already interested because of the numbers. It confirmed their interest. It didn’t create it.
Path 2: Accelerator Brand YC, Antler, a16z Speedrun—these function as credentialing mechanisms. The stamp gets you in doors. Your technical moat confirms interest once you’re in the room.
Path 3: Template Founder Find an ex-Google + Stanford + previous exit person. Make them a co-founder. Let them do the fundraising while you build.
This sounds cynical. It is cynical. It’s also how the game works.
The Key Reframe
Technical moats don’t convert maybes to yeses. They convert interested parties to committed ones.
The interest has to exist first. Create that through traction, credentialing, or team composition. Then let your technical depth close the deal.
Research Sources Worth Following
Since people asked—labs doing work that’s actually interesting, outside the obvious names:
Noeon — Some of the most insane research into AI and Geometric Intelligence that extends wayy beyond LLMs. A lot of challenging math things, but I really respect their ambition. If you’re looking to strech your knowledge, this is a great place to start.
Tsinghua University — Probably the best AI research university in the world right now. Zhipu AI came out of here. Dream (the diffusion model) came out of here. Stanford publishes higher volume, but a lot of it is branding over substance. Tsinghua’s pure research output is consistently strong.
Sakana AI — May or may not count as “not mainstream” at this point, but consistently strong work on evolutionary approaches to model development.
Verses AI — Interesting active inference work. (Standard caveat: good research does not equal investment recommendation. I’m commenting on research quality, not business viability.) Our primer on their research is here—
What I Changed My Mind On (both AI and non-AI)
Someone asked this in the stream and I want to give it a proper treatment because it touches on methodology.
The Professional Shift
I come from low-resource engineering. Building on edge devices, strong constraints, accounting for every detail. When I was doing Parkinson’s disease detection on mobile phones, if I got the diagnosis wrong, that’s a real problem. You don’t get second chances.
That background trained a certain rigor. Statistical discipline. Accounting for edge cases. Start with the assumption that things won’t work, then make them work.
For a long time, I thought this was purely an asset. It is an asset—but in a field moving as fast as AI, it can also hold you back.
Here’s what I mean. Claude Opus came out a few months ago. Decent model, nothing outstanding. Then around December, they made what sounds like minor changes—better tool use, better reasoning across tool chains—and suddenly Claude Code becomes a completely different product. My assumptions from two months ago about what was possible? Obsolete.
If I’m still operating with the “get it right first time, account for every edge case, don’t ship until you’re certain” mentality, I’m building on outdated priors. The activation energy was basically zero and then suddenly the tool explodes because it needed one small push.
So I’ve had to learn: ship with half-assumptions. Get comfortable saying “I don’t know if this will work, assume it will, we have to test it.” The old mentality was “don’t ship until you’re confident.” The new mentality is “ship to find out, because your confidence is based on stale information anyway.”
I used to look down on “AI will eventually get good enough” as an assumption. I’m starting to see the wisdom of it.
The Personal Shift
This one’s harder to articulate.
I’ve always thought in hard true/false. Popper’s falsification method—you have claims, you identify what would disprove them, you test, you update. That’s how science works. That’s how rigorous thinking works.
I was over-applying it.
Plato had his theory of ideal forms—there’s an ideal chair, and every physical chair is a shadow of that ideal. Do I think that’s literally true? No. But as a framing for thinking about representations, about how concepts relate to instances, it opens something up.
Or take hermetic geometry. I don’t think it’s mathematically true in the way I’d say my other research is true. But when I’m working on geometry of intelligence—how information gets represented, how representation structure constrains capability—some of these framings from traditions I would have dismissed actually help me see differently.
My grandfather died recently. In Hinduism, there’s this belief that for the first 11-12 days, the spirit haunts the earth. Only on the 13th day, after the funeral ceremonies, can it move on.
I don’t literally believe there’s a spirit haunting anything. But think about what that framework encodes: grief takes time. The first days, you’re haunted—the loss is present, inescapable. The rituals of letting go, burning the body, releasing attachment—maybe they serve a function for the living. A structure for processing. A permission to move on after a defined period.
That’s not literally true. But it carries something useful.
I think I was dismissive of anything that couldn’t be proved or falsified. I’m starting to appreciate that some traditions—religious, philosophical, cultural—encode wisdom that isn’t trying to be literally true. It’s pointing at something else.
This connects to the geometry research. I’m looking at hermetic philosophy, at frameworks I would have ignored, because maybe there’s a way of seeing that unlocks something I haven’t seen. Not because it’s true. Because it might be useful.
Physical Energy as Infrastructure
Someone asked about the number one thing to get good at AI.
My answer: buy a 30-pound mace and swing it around your neck until you’re exhausted.
Let me explain.
Before I did this full-time, I competed in Combat sports. Trying to be the best ever and all that. Training twice a day—real intensity, close to throwing up. That builds a physical base that most people in tech don’t have.
The ability to do sustained deep work is physical, not just mental. My longest stretch was close to 24 hours of nonstop work. I don’t do that regularly—most days aren’t like that. But having the capacity means my average output is higher, and when I need to hit peaks for sustained periods, I can do it without destroying myself.
The mace is because I kept missing strength training. Cardio is easy—shadowboxing, hill sprints, I can do those at midnight without a gym. But strength requires equipment, gym hours, logistics. The mace solved it. Twenty, thirty minutes of swinging this thing and I’m completely burned out. Whole core, shoulders, everything. Two times a week fills the gap.
The number one most important thing for AI right now is just having a lot of energy. There’s so much to experiment with, so much to learn, so much to build. The people who can sustain high output over months and years will compound faster than the people who burn out or plateau. A lot of physical energy is going to be a huge competitive advantage.
Final Closing Notes on AI
When I look at the AI Space, I feel excitement. I think there’s a lot more open space in the world than you might assume. Things have changed so much, and I believe that even established experts are mostly bluffing. No one really knows how the cards are going to fall or what will end up being truly important. I believe chaotic spaces like this are the best places to be a human.
I guess what I want to say is that I’m going to bully my way to GOAT status- first on Substack, then in AI, and eventually of all humanity. I hope you feel privileged since you get front-row seats to the whole thing. I know I feel privileged that I get to go up with such fun group of people.
I might be delusional. I might be a visionary. I don’t know. But I am a lot of fun. Hitch your ride to my wagon, and I promise you’re going to have a great fucking time either way.
Thank you for being here, and I hope you have a wonderful day.
Dev <3
Reach out to me
Use the links below to check out my other content, learn more about tutoring, reach out to me about projects, or just to say hi.
Small Snippets about Tech, AI and Machine Learning over here
AI Newsletter- https://artificialintelligencemadesimple.substack.com/
My grandma’s favorite Tech Newsletter- https://codinginterviewsmadesimple.substack.com/
My (imaginary) sister’s favorite MLOps Podcast-
https://machine-learning-made-simple.medium.com/
My YouTube: https://www.youtube.com/@ChocolateMilkCultLeader/
Reach out to me on LinkedIn. Let’s connect: https://www.linkedin.com/in/devansh-devansh-516004168/
My Instagram: https://www.instagram.com/iseethings404/
My Twitter: https://twitter.com/Machine01776819
