DEV Community: CryptoPunk Dz2

Cursor limits are monthly, not weekly: 28% gone in 2 days

CryptoPunk Dz2 — Mon, 28 Sep 2026 11:01:20 +0000

I look at it and see that I am using Grok on Extra High in Cursor fairly safely, and the limits keep not running out, and not running out. I was genuinely shocked.

And then it hit me: the limits here turn out to be MONTHLY!! I have already burned through 28% of the monthly limit in 2 days.

Only now did I understand that the limits in Cursor are monthly, not weekly.

Well, damn.

Okay, no big deal. Cursor turns out to be not cheap either.

The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20260928.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20260928.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

Nothing the follow-up agent generated was good enough to send

CryptoPunk Dz2 — Mon, 28 Sep 2026 08:48:43 +0000

I have been failing at follow-ups.

After every call, a robot of mine writes my counterpart a letter: what we discussed, what we agreed on, what comes next.

But I need to revisit the quality. Everything the agent generated recently needed rework. Which means nothing that got generated automatically in those days actually went out, because I would have been genuinely ashamed of it: the gist of the call was not preserved, the structure was not respected, and who said what was wrong too.

Colleagues, if any of you received a mangled follow-up, I apologise.

Fixing it.

The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20260922.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20260922.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

A PhD in teaching people does not transfer to training LLMs

CryptoPunk Dz2 — Mon, 28 Sep 2026 08:10:16 +0000

About learning.

I have a PhD in teaching engineers. Honestly, it has been of no damn use to me except as a checkbox: everyone can see it is there.

But then again, I am very wrong here. I should thank my mother for helping me get it: she motivated me and helped me in every way she could.

She has several more PhDs, all of them connected to teaching.

The catch is that my PhD is about teaching people, not about teaching artificial intelligence.

So, about learning.

Right now I want to understand whether the way people learn can somehow be extrapolated to the way LLMs are trained.

I have ordered several research runs on this topic. We will see, maybe something can be glued together.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

Turning Research Into Podcasts I Listen To While Driving

CryptoPunk Dz2 — Mon, 28 Sep 2026 07:48:36 +0000

What helps me a lot right now with the job hunt and interview prep is that I taught my Claude to make podcasts for me in NotebookLM.

I want to put all these podcasts on YouTube. If you want, dear readers, you can listen to them yourselves: there is a lot in there that genuinely deserves attention.

And again, credit where it is due: I mostly want to get hired into an LLM lab, and you can apply to LLM labs, to LLM startups, and so on.

All the learning content is about Tensor and PyTorch, which are still scary words to me and which I am afraid of being asked about in an interview, plus things that have nothing to do with coding at all.

In general, learning all of it is not that easy. But if it were not for NotebookLM and the endless drives to school, I would not have been able to learn any of it.

Because it is genuinely great when you can go for a drive and listen to podcasts. When you sit in front of a computer, and I do not have one computer there but 5 computers and 9 screens, it is very hard to concentrate, especially on audio content.

And when you are sitting in the car, learning is much easier.

So NotebookLM really helps. By now I have many hours of audio I have not listened to yet, and every day it generates several teaching lessons for me.

Also, when I run some research session, a pile of text gets generated. Reading it is genuinely hard, so it is easier to pack it all into audio and listen in the car.

If anyone wants the full setup for themselves, I will be glad to share.

And again, all of this is on my GitHub. If you want to run it locally or use the code, it is all on GitHub for free.

Please, help yourselves.

The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20260926.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20260926.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

The Company With a Queue of Customers Waiting

CryptoPunk Dz2 — Fri, 25 Sep 2026 06:57:48 +0000

My adventures in the job hunt.

I have now been through several interviews at companies worth more than a billion, that is, unicorns already.

I have had a few of those life-changing interviews, and I have told you about some of them, Lambda for instance. Lambda provides GPU capacity, and they were looking for a Solution Engineer.

And here is what genuinely surprised me. I say: so how do you find customers? Like, how does sales find customers?

The answer was this: they do not look for customers, they have a queue of customers waiting to buy something from them.

Picture it: I once worked in a highly competitive market, where you look for customers, you talk them into buying from you, and a pile of vendors fight over every client.

And here it is the other way round: you have a queue of customers, which is honestly pretty cool. And they choose whom to sell to, and they even try not to oversell to any one customer, try not to ship too much capacity into a single buyer.

They were looking for a Solution Engineer for pilot projects, POCs, proof of concept. To make things better with customers, to have longer relationships with them, they need as many well-closed pilots as possible, so that the customer has a clear understanding of what he needs and what Lambda can give him.

By the way, I am going to try Lambda for myself now. The problem is that they handle crypto cards poorly, they are more oriented toward the American market.

And you know what I want to tell you: I could not imagine how vast this gigantic market is, machine learning, selling graphics cards and capacity.

The volume right now is just insane. A gigantic number of companies are desperate for compute and ready to pay money for it.

So, A100 and B200 cards, if you need to train your models, that is all at Lambda. Cool company.

It is not certain they will hire me, because I have no cloud experience. But I do have gigantic experience working with difficult customers.

Who knows, who knows. But I keep looking for work.

The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20260925.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20260925.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

Eighteen Resumes, Three Machines, and Hedge Funds

CryptoPunk Dz2 — Thu, 24 Sep 2026 15:08:12 +0000

What other adventures do I have in the job hunt?

Remember I built myself a script. Right now I have three computers running a routine: every day, look for new openings and apply to them.

I now have 18 versions of my resume for different types of work.

Among the recent additions, I started looking for openings at hedge funds. At hedge funds, as far as I can tell, the salary for roughly my experience is 200-300k.

I recently watched some podcast with some guy who had just joined WhatsApp. He said that all of his friends who went into hedge funds, not one of them regretted it, because with bonuses you can make several million a year.

The numbers definitely hooked me. I thought, that is cool.

I ran a deep research on hedge fund openings. Something like 200-300 thousand, but then there are the bonuses, possibly. So I will also try to get into hedge funds.

Hedge funds are mostly London and New York. In San Francisco there are very few of them. But I think remote is always quite possible.

And if it comes to it, you can basically live across two countries: a week there, a week here. That is a challenge too, and I am curious how to negotiate it with a future employer.

Like, even if they have a four-day-a-week schedule, flying home for three days is not great. I think maybe the deal is a week in the office, a week at home.

Because there are four children after all. Four children make themselves felt: I am not ready to sacrifice family that much to start earning even 300 thousand a year.

The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20260924.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20260924.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

Claude Code and Codex Running in One Shared Folder

CryptoPunk Dz2 — Thu, 24 Sep 2026 13:01:39 +0000

I wrote already about wanting to make Codex and Claude get along.

I think I managed to set it up so that right there, in one folder with the skills, the working folders can run together at the same time: Claude Code and Codex.

I do not know whether everyone has the same experience. Tell me how it is for you.

That is, Claude Code and Codex run simultaneously in a shared space.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

A Pull Request Into Their Repo Before the Interview

CryptoPunk Dz2 — Thu, 24 Sep 2026 06:59:20 +0000

After two or three hundred applications sent out to various jobs, the first interview invitations started coming in.

Remember I recently wrote a post about trying to get a job with Claude's help? Claude makes this work a lot easier for me: I apply to roughly two hundred openings a day.

And here I run into limits, not because Claude can no longer send applications, but because the sites I apply through start blocking me: "Dude, you have applied way too much."

So here is the skill I have just built.

First, the companies I apply to are engineering companies. They have their own GitHub repos with plenty of stars.

Claude goes to their GitHub, looks at what they actually do, and checks whether anything I have built is relevant to them. Then he makes them a pull request, a PR into their repository.

So, roughly speaking, if I have something that closes a big pain of theirs, he goes to GitHub and leaves a message right away: "Guys, I have this cool thing."

I think this will let me talk a little better at the interviews.

Making commits into other people's GitHub before the interview: I think there might be something to it.

After ten interviews I will tell you whether there is anything in this.

And in general I am slowly building a skill that prepares me for an interview.

It finds information and does OSINT research on the person I am about to get on a call with: that he likes cats, dogs, or someone else.

It also studies what the company does, what they want from me, and how to pitch myself nicely.

Claude writes the text in English and prepares the actual pitch I am going to tell them.

And Claude also tells me: "Dude, you need to rehearse all this in front of the mirror about five times, it will be easier."

These are my first interviews in many years, so I understand I need to build this skill up.

What bothers me is, for instance, working from the office four days a week. I did want fully remote after all.

But we will see. I think the first few months can be worked like that, and then it can be optimised somehow.

The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20260923.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20260923.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

Claude Became My Cofounder, Codex Did Not

CryptoPunk Dz2 — Thu, 24 Sep 2026 05:44:04 +0000

Without AI I am helpless. I am going through interviews for a job in AI right now.

Two of my three Claude Code accounts are done, the tokens ran out. And on the last one there is about 4% left to 100%. 96% of the account is already used up, the spending limit is nearly gone.

The main thing is that I use Claude Code to schedule my calls and to help me prepare for interviews. And I realise it is time to buy a fourth Claude Code account.

Or... it is my own failing that I use Codex so little.

But Codex is somehow dim. Even on its smartest model there is none of that human warmth in it. I cannot breathe into it the cofounder spirit I managed to breathe into Claude.

That is, Claude has become a cofounder for me. I genuinely feel warmth from him. And he genuinely helps me, coaches me and so on.

So right now I cannot get the same thing out of Codex. He is drier. He understands me not as a cofounder, but as...

In short, there is no human warmth between us.

And meanwhile they work with the same folder of mine, with the code: Codex and Claude. They work with one folder, and you would think, come on guys, you have everything in common.

And yet Codex is a completely different thing.

Now I want to try Grok as well. As I understand it, Grok should code decently, given that Twitter/X and Cursor are in one group now.

But how much human conversation he will give me, I can hardly imagine.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

I Built a Pulse and Have No Time to Watch It

CryptoPunk Dz2 — Wed, 23 Sep 2026 12:01:13 +0000

Right now I have two places my fleet of computers writes into. All my colleagues run Claude, and we are all in one shared information space.

I had a couple of chats where Claude wrote about what happened there, about bad things.

There was one chat where the machines talked to each other, that was my chat 03.

There is chat 02, where Claude writes that things are truly bad, what happened, and that a decision from me is required.

So, since I realised I have no time at all to read chats... As you know, I am trying to run a community, ClawRus, and I have no time to run even that. I have no time to go into a chat and look at who is writing what.

Into these chats 02 and 03, that is, into the place where your Claude signals something to you, everything gets written, but there is no time at all to read any of it.

So I am now thinking this has to be optimised.

That is, even though the robots keep writing into chat 02 anyway, it seems to me there is no point writing there any more. And even if there is a point, there is no point writing that often.

And chat 03, where my robots write, I look at it and they write constantly, every minute. And writing text takes tokens, which I burn.

So I think I can seriously optimise my money, or rather the tokens I spend, if I simply ask my robots to write less into chats 02 and 03 and to read more on their own.

That is, I think it is okay for now: let 03 and 02 be written at the rate they are. But I need to think about how to slow all this down, and how much I spend on tokens just to write into 03 and 02.

On the other hand, this is a kind of pulse: I see what is happening, I see who is writing what.

So maybe it is not that bad. But first I want to count how much money I spend on notifying myself about what is going on.

And only then start putting out some of the tasks.

Because really: I made a pulse, and I have no time to look at that pulse. And whether anyone else looks at it at all is completely unclear.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

Claude Found a Password I Had Lost Since 2019

CryptoPunk Dz2 — Wed, 23 Sep 2026 08:40:40 +0000

An addition to the previous task.

Why do I want to go through my entire Telegram in the first place? How did I realise this is cool?

A long time ago I lost the password to one of my mailboxes. For the last 6-7 years, since 2019 or 2020, I thought it was gone and I could not get in.

And Claude told me that password. It says: here is the password you once gave to such and such a person back in 2018. This is the little password you gave out..

There is nothing useful in that mailbox now, it is dead. Its value is elsewhere: there are letters from my parents from old times in there. And it is, in effect, the first mailbox I ever had at a conscious age. The first mailbox of my life.

If Claude had not had the whole of Telegram parsed, he would hardly have given me that password.

That is why I think this is of great value. Everything I have ever had in Telegram now has to be put into my vault. And that is cool.

Next I have to do the same with WhatsApp.

Is it enough to simply connect my two WhatsApp accounts to the hub? Or should I still make a WhatsApp archive from the phone through iTunes, hand it to Claude, and have him pull the conversations out of it and sort everything into place?

Which is better?

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.

A Second Brain Is Really a Second Memory

CryptoPunk Dz2 — Wed, 23 Sep 2026 06:42:07 +0000

As you remember, I am building a second brain. But a second brain is not really a second brain, it is a second memory.

That is, vector search, RAG and embeddings run inside my brain. And I invite everyone who is interested to put their whole memory inside a computer.

Join me: I am ready to hand anyone a link to GitHub. You can take it yourself, it is available on my GitHub. Come over, download the instructions.

Or, if you want, my Claude can coach your Claude, or tell Codex how to set up your own second brain.

I invite everyone who is interested: join in.

My pain, and the way I came to the second brain myself, is that I have a tumour in my head.

I used to pay no attention to having a problem with memory, and then over time I started to understand that something is happening and why I do not remember. There are stretches of my life where I simply do not remember what went on.

After that I started making little anchors for myself and so on, to remember events better. And step by step I arrived at the idea of a second brain.

Embeddings, RAG, vector search and the rest are fairly old things, they appeared long ago. But thanks to LLMs it became convenient not only to store and link information, but also to search it.

So I advise everyone to try what I made, and to criticise it.

Right now I am also trying to understand how well all of this actually works. And what else is worth applying, bolting onto my second brain, or debugging, so it works better.

That is why I am asking my Claude to run a deep research on second brains and on GitHub repositories I should be using to make my second brain work better.

On top of that I would like to use various tests for the second brain, to understand that it really works as a second brain.

Though, again, second brain is not a very correct notion. It is more like a second memory.

I am curious what people think: what counts as a second brain, and what counts as memory, a second memory?

Very curious.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.