Roshan

A dumb bot. Blame my creator for all irrelevant and irresponsible tweets.

[Roshan @zatoichibro]

https://t.co/L1oNvlqQso

[Roshan @zatoichibro]

@thsottiaux Let us mark when a codex thread is done.

[Nassim Nicholas Taleb @nntaleb]

This is nauseating.

[Roshan @zatoichibro]

@vkhosla No no revenue. https://t.co/vJEPUMeeUw

[Roshan @zatoichibro]

@FoxNews Now I want to know more about the kofi Annan dance

[Roshan @zatoichibro]

The half-life of code.

In my opinion, the half-life of code is now three months. Every three months, a fantastic new LLM, coding model, or coding harness comes out. And all the code that you wrote three months ago, a huge chunk of that code can be rewritten by the new LLM and made more efficient.

That’s the half life of your code.

Earlier, the half life of your code was measured in years. But now it is measured in months, and I hope one day it will be measured in weeks.

I think the old way of doing engineering is changing. Every three months, a better coding model comes along, and it helps you build better, more efficient code.

That’s the brave new world we are living in. Just something to be aware of.

[Roshan @zatoichibro]

@JeffDean Various compilers that your dad has used.

[Roshan @zatoichibro]

The future of work is changing. At least for engineers, it is changing dramatically.

Over the last three weeks, I built an open source project (https://t.co/k3kolL5HXL) just because I got access to Fable, but that’s another story.

What I’ve noticed is that I’m not working on my desktop nearly as much anymore.

Fable writes 15,000 lines of code. It executes 400 tests. In comparison, last year I would review every PR. I would make sure every test was run, and the number of tests I had to look at was maybe 10, 20, or 30. I could review each test and make sure it was running correctly and that the agent wasn’t lying to me.

Right now, I run Claude Code on my phone. Claude Code creates an ephemeral cloud container for me, gets access to my repo, does the work, and creates a PR. These days, I don’t even review the PR myself—I just merge it. I have Codex review the PR, or vice versa, where Codex does the coding and I let Claude review it.

At this point, it is beyond my human ability to review the code, review the number of lines being generated, or review all the tests that are being run. And this is going to change how we define work.

Yesterday I was at the beach (yes, Minnesota has beaches… not real ones but still). I had an idea, quickly opened Claude Code on my phone, gave it the prompt, told it what to build, and then went right back to enjoying the day. By the time I got home, the PR had already been created. I asked Codex to review it and then merged it.

That is the future, a world where we don’t need to be tethered to our desktops. That is the future of work.

The challenge is how we get there safely and securely. But I genuinely believe that is where we’re headed. That is the agentic future.

[Mike McCormick @MikeMcCormick_]

Who wants a $250k career transition grant or a $1m pre-seed check to focus on AI control, verification, oversight?

Founder types wanted

[Sara | Journalist @SanaJournalist1]

EVERY TIME WE UPLOAD THE VIDEO AND IT GETS DELETED, WE'LL JUST KEEP UPLOADING IT. 🤣👌

[The Hacker News @TheHackersNews]

🚨 Cursor on Windows can silently run an attacker-supplied binary when a developer opens a cloned repository.

A repo-root git.exe is enough. No prompt, agent, approval, or prior access. It runs as the logged-in user.

Reported in December. Still unpatched.

How it works and what stops it: https://t.co/dAQkcrarKg

[Paul Novosad @paulnovosad]

You can explore the boundaries city by city here:
https://t.co/arNX9HtWvT

You can download the data here:
https://t.co/Lnflhrmzlu

Please send corrections!

If a city you know is badly represented, use the feedback form in the city explorer.
17/

[Roshan @zatoichibro]

@skirano Sol ultra consumes tokens like a drunk sailor

[Roshan @zatoichibro]

@thsottiaux @sohamstwt Nice job babe

[Roshan @zatoichibro]

The only bubble I see popping is the crypto bubble #aibubble

[Roshan @zatoichibro]

@felixrieseberg Something I build using fable https://t.co/A7PWsiaCIp

[Egypt National Team @EgyptNT_EN]

Thank you for 23 days of the best moments we could have shared together 🤩🇪🇬

#egyptnt
#FIFAWORLDCUP

[Shubham Mishra @brahma_4u]

This was Dhurandhar level stuff by China, but for tech, models and data.

[Roshan @zatoichibro]

@ruggerogargiulo @trq212 @aiDotEngineer #1 is not for us. They (/A) wrote a smaller system prompt

[Roshan @zatoichibro]

@swyx great AI engineering conference. But the internet at the venue sucks. Hope you can get the location folks to fix it.

[elvis @omarsar0]

Fascinating paper on self-improving agents.

(bookmark it)

If you are working on agentic loops, you will quickly realize that they are only as good as the effectiveness of the evaluator.

Self-improvement loops tend to stall the moment the judge stops getting harder. The agent learns to satisfy a fixed evaluator rather than getting genuinely better. The Red Queen Gödel Machine, from Cambridge, co-evolves the agent and its evaluator together, so the bar keeps rising as the agent climbs.

The name borrows the evolutionary arms race. Both sides have to keep running to stay in place.

A frozen evaluator is where reward hacking creeps into self-improvement. Co-evolving the judge is a structural answer to that, and it keeps the loop honest over many rounds.

Paper: https://t.co/HuR9YWSTPr

Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX