Anant Jain

recursively self improving

[Anant Jain @anantjain]

If AI had been writing and reviewing its own code for the past few decades, and humans had only recently started looking into what it was up to and discovered an ever-growing list of CVEs, we would reasonably conclude that AI hadn’t been good at coding all along and that we should just start over.

Except it’s the other way around.

Humans have been the ones writing and reviewing each other’s code for decades. Very diligently, if I may add.

If we’re being intellectually honest and holding ourselves to the same bar we hold AI to, we’d have to admit that we haven’t been all that good at it either.

[calle @callebtc]

I'm tired boss

[Anant Jain @anantjain]

I keep telling people that I’m no longer as excited to try out every single new model release as I was a quarter ago.

All my private evals for day-to-day coding work have been saturated for months now, with even a DeepSeek Flash doing most of my real work tasks better, faster, and cheaper than whatever was at the frontier just a few months ago.

[Russell Kaplan @russelljkaplan]

This is directionally consistent with our internal data @cognition. Many tasks within software engineering are now "intelligence saturated." Once any model can do the task well, cost and speed are all that matter.

[Anant Jain @anantjain]

Decided to give Gemini Argon a shot, after almost a year of not using Google’s models.

Google still can build frontier models, but still cannot build an iOS app where response streaming survives an app switch. 🤷

[Anant Jain @anantjain]

Few things are more satisfying than watching latency graphs fall off a cliff.

[Coinbase 🛡️ @coinbase]

Built for the moments when every millisecond matters.

A faster engine for faster markets.

[Anant Jain @anantjain]

People are still underestimating how many exponentials are stacked inside the demand for AI. Reasoning models and agents were the first two exponentials. Two more are barely getting started this year: Computer use and coordinated swarms.

Computer use is finally starting to work well, but it’s still painfully slow. In my experience, simple tasks can take a minute or two, enough to not be in the same league as the plain old API call. Still, I have confidence that this is solvable (eg., Jev style models for computer-use), making agents capable of using software without waiting for someone to build/maintain API integrations, which would open up entire industries to AI.

Then there are swarms. We've all heard of agent swarms escaping sandboxes or solving Navier-Stokes by now, but just working with Sol/Astra in Ultra mode has made me much more aware of their capabilities. A typical single Ultra session works autonomously for days and completes an entire project spanning dozens of PR's (see a recent screenshot). And, the Ultra mode is obv a heavily constrained preview of what these swarms can offer.

[Anant Jain @anantjain]

In today's news: Shopify: 1, Amazon: -1

[Anant Jain @anantjain]

History doesn't repeat but it rhymes, especially in consumer.

Meerkat → Periscope (Twitter Live)
Snapchat → Instagram Stories
Clubhouse → Twitter Spaces

Instinct → Muse.

[Anant Jain @anantjain]

Cosign

[Mitchell Hashimoto @mitchellh]

The "whiteboard defense:" I should be able to pull you aside at any moment and ask you to explain any customer-facing system you've shipped. You should be able to clearly explain how it works and defend the decisions you made. This is my benchmark for responsible AI usage.

I don't expect line-level familiarity with the code. I don't care if you remember the exact function name or implementation detail. You may not even know it. I don't care.

But if I ask "why did you do X instead of Y?", "what happens if this actor behaves maliciously?", "what data structure did you use here and why?", or "where does this fail?" you should be able to answer confidently.

For PoCs, demos, experiments, whatever: I don't care. Generate 100% of it and understand none of it. Speed over quality every time in those specific scenarios.

But if you're shipping customer-facing work, you can't be shipping things you don't understand at a high level.

[Anant Jain @anantjain]

@ankushio @nitinagg This reminded me to get rid of my Mac Mini that I no longer need, thanks to all the cloud environments.

Only appropriate that ChatGPT work set up the eBay listing in its cloud browser, and looks like I’ll make a little profit on this: https://t.co/aEvkKqZHx2 😂

[Anant Jain @anantjain]

One HUGE unlock over the past month has been moving to Cloud mode across all my personal AI usage (e.g., ChatGPT Work, Cursor Agents, Codex, etc.).

The freedom to not worry about remembering to keep caffeinate running, or hitting my laptop's SSD storage limits with multiple checkouts and/or worktrees, is truly liberating.

The laptop lids can be closed again!

[Anant Jain @anantjain]

The current moment in humanity is very special. We're inventing our new alien friends who will live amongst us.

[Andi Marafioti @andimarafioti]

Can you explain to me exactly what you want 4 microducks for??

[Anant Jain @anantjain]

How we shipped a real product with no “dedicated” engineers to the team: https://t.co/VJcBbGSDQJ

[Anant Jain @anantjain]

We wanted to ship all this by end of Q3. We got it done in just the first 4 weeks of the quarter.

Longer post on how we automated most of the software engineering for this coming soon, but have to say, this team is cooking! 🧑‍🍳

Much more on its way!

[Sid Coelho-Prabhu | sidcoin.eth 🛡️ @sid_coelho]

Just dropped a massive update to Coinbase Business

□ Get paid by AI agents
□ Accept USDT (in addition to USDC)
□ Reusable payment links
□ Flexible pricing
□ Product catalog
□ Collect buyer info

We now serve over 5000 businesses who want smarter, faster, global money

[Anant Jain @anantjain]

The "Git at any scale" blog post from @cursor_ai has to be the single best engineering blog post I've ever read.

https://t.co/1krzCEa3lP

[essen @essenaccount]

NOW IS IT CPU BOUND, OR IS IT MEMORY BOUND?

[Anant Jain @anantjain]

I recently read Neal Stephenson’s essay about why writing by hand is good for your brain:
https://t.co/uY4403MNg4

It made me realize that I have barely written anything by hand in more than a decade. I’m in no way exaggerating when I say that I have the same Moleskine notebook from 10+ years ago sitting on my desk, less than half-filled.

After reading the article, I made it a point to write something a few times a week—what I wanted to accomplish, problems I was thinking through, things I needed to remember.

What I found was that I wasn’t merely recording my thoughts more slowly, I seemed to be thinking differently, almost as if I had unlocked another dimension of my intelligence that had been sitting there unused.

Around the same time, I started noticing something similar in my conversations with the upgraded ChatGPT Voice released in early July. When I type, I tend to compress and make my thoughts precise (as I'm doing now!). When I speak, the opposite happens. My precision goes way down. I ramble, backtrack, and remember another thing halfway through a sentence. But my recall goes way up. I get more out of my brain than I would have otherwise.

Speaking out loud to an AI feels like accessing parts of my brain that were previously used only during in-person conversations, but never during solo thinking.

Both experiences made me realize that interfaces like paper and voice may be part of the machinery that helps us think. Moving a pen on paper to form words, or speaking and reacting to another voice changes the loop in which perception, action, and reflection take place.

That made me wonder whether the same idea applies to AI. As I went down this rabbit hole, I found some early research suggesting that it might. There are papers on models reasoning better when given a visual scratchpad or allowed to generate diagrams as they think through a problem. However, there are also papers that show that simply forcing a model to draw can also make its reasoning worse when the representation it creates is bad.

The interesting capability might be learning which representation is useful for which problem. Humans seem to do this almost all the time. I prefer typing for most kinds of thought. I discover other kinds by talking. I apparently discover yet another kind when I put a pen in my hand and scribble on a blank piece of paper.

It seems likely that intelligence isn’t entirely something that happens internally and then gets expressed through an interface. The interface participates in creating the thought itself. Chain-of-thought reasoning also feels like a version of this.

And to conclude, this is a question I find fascinating: Could an AI choose its own medium of thought? When should it write? When should it draw? When should it speak? When should it write code? Or in the case of robots, manipulate something in an environment, observe the result, and reconsider?

[Anant Jain @anantjain]

1. Tesla FSD

Daily user since August 2024 and I have felt it improve drastically every month since. It’s been ~flawless over the past 6 months, to the point where I feel like I have a personal Uber waiting in my garage 24x7.

2. ChatGPT Live

The new voice mode is the closest realization to the “Her” movie I’ve experienced. I’ve used it as a coach and advisor on multiple topics, as a su chef while cooking, and once as a babysitter (though have to admit that that wasn’t as successful). There’s no AI product that “feels more human” than ChatGPT voice mode at the moment.

3. Codex Browser+Computer Use:

Whenever I’m procrastinating on something, I’ve made it a habit to at least try and ask Codex to do it for me. It’s found me a contractor for home upgrades, submitted new app versions to App Store, made printable pdf’s for our kid’s homework, and simply eliminated all the schlep I don’t want to do from my life. It was already magical to see perfectly working code appear in Claude Code in 2025, but somehow even more magical to see your computer being used by Codex in 2026.

[signüll @signulll]

there are three world changing products of the ai era so far:

1. chatgpt which made intelligence universal although with little monetization (hence the openai pivot to enterprise).
2. claude code which made ai an agent that performs actual work. it took over the globe in a single swoop (likely the biggest reason for $100b arr for anthropic).
3. midjourney which made visual imagination executable very early.

what are your top three?

[Jordi Hays @jordihays]

Your goal as a software founder is to never let your spoon get bent

[Anant Jain @anantjain]

@RhysSullivan Couldn't agree more! Vented on this topic a few weeks ago:

https://t.co/UuEpEp0Tik