Raed salah

@potencytoact Thats a heavy and much appreciated breakdown. I’m liking each as a bookmark, hope the spammed notification doesnt bother you 😂

On June 5, OpenAI merged a change to Codex that encrypts the messages its orchestrator agent sends to sub-agents. Until that PR, those messages were plaintext in your local session data. You could open the logs and read exactly what task each sub-agent had been given. Now, for the new GPT-5.6 models, that content is ciphertext generated on OpenAI's backend, and only OpenAI can decrypt it. The agents run on your machine but their instructions are no longer readable by you. Why would OpenAI do this, and what does it cost users? The answers are scattered across a GitHub issue, a long Hacker News discussion, and the follow-up PRs in the Codex repo. This thread distills the discussion. Github: github.com/openai/codex/i… Hacker News: news.ycombinator.com/item?id=489050…

Code is part of the problem you’re solving. Choosing the right stack, designing the database, defining relationships and constraints, writing queries, structuring services, applying patterns, managing memory, understanding why code is written a certain way, etc., are all part of software engineering. The code is not separate from the problem. So when people say, “I love solving problems, I just don’t like coding,” I don’t really buy it.

i like solving problems, the code was never something that truely interested me. im sure this is the case with a lot of other software engineers too.

@theo @ceo_on_bus “My employees”

@ceo_on_bus "slop" accounts? My employees are quality shitposters thank you very much Separately, it's a requirement that all of my employees make fun of me regularly. My haters suck at it so someone's gotta do it right.

does @theo has multiple slop accounts ? i won't dare to speak against my boss

@potencytoact His tweet radiates inexperience

This would only work if building, deploying, and maintaining software at scale was fully autonomous. Budget some of that for wet brains in the loop.

Claude’s tools feel fragmented. Projects, chat, code, designer, dispatch all work in silos. They need to simplify and make project and chat the core interface, everything else is a capability that works when needed and results can be accessed through project artifacts. I hate it when a project chat can’t speak to a design project and chats in a project hallucinates between different topics and streams. It doesn’t feel very seamless. I want it to feel like magic. A regular chat for basic questions and a project should feel more like a dashboard or a control panel. I hope for this and it would be even amazing if i can help working on this too.

i keep wondering why in the claude app there’s this sidebar to choose a tool/harness that requires clicks/taps but at the same time i could realistically see myself prompting to get to the same things prev: tap on artifacts next: simpler, unified ui; “hey show me my artifacts”

Depending on the product, having options across different price points is great but not for every case, i’ll use professional tools as an example. As a consumer, I prefer industry-standard tools because they’re directly relevant to my profession and career. Market demand and industry requirements largely determine which tools I invest time in learning and using. Learning a niche tool that has little industry adoption offers limited value if it doesn’t support my work or career growth.

imo we are already seeing pretty wild market saturation in software because of ai. it feels like there are sooooo many more products, new ones coming every day faster than ever before. probably great news for cloudflare/netlify/vercel/infra but i wonder how consumers feel

Depending on AI is all fun and games until you get laid off and realize you’ve become too reliant on the tools to confidently solve live coding challenges and get rehired. If you don’t have strong job security or a financial safety net, keep reading code, writing code by hand, and using AI to accelerate learning and upskilling. The competition is weakening. Don’t let your skills weaken with it. The majority of developers aren’t in San Francisco. From personal experience, tech jobs outside North America and Europe often lack basic benefits and job security. Salaries are frequently not high enough to build a meaningful financial safety net, and AI subscriptions can take up a significant portion of an engineer’s income. It’s common for many engineers not to have one. The trap is that many of these engineers don’t upskill. They’re still spending most of their time grinding through basic tasks as if it’s 2010. I don’t see a particularly optimistic future for software engineers. From my experience, as hardware has improved, code quality and performance have become less of a priority at many companies. Optimization is a skill I rarely see valued during hiring or day-to-day work anymore. I still remember pointing out performance regressions across several modules and hearing, “This isn’t Apple. No one cares.” To me, that attitude started long before AI tools became mainstream. Passion for craftsmanship and attention to detail had already been declining in many teams. My advice: focus on upskilling and becoming adaptable. Don’t spend all your time obsessing over micro-optimizations.

Software is slowly becoming a pay to win profession because of AI. While $20 a month isn’t much in many Western countries, engineers in countries with weaker currencies have a different reality. Software development and the internet were a gateway to a better life. They gave people the opportunity to build careers, start businesses, and compete on a global scale with nothing more than a basic Windows computer in a cybercafe. Now, tech is advancing so quickly that keeping up with the industry requires an AI subscription. For many developers, those subscriptions have become the new productivity power ups, creating another barrier for talent in economically challenging countries/states.

@potencytoact @neetcode1 you’re right

@r2dsl7 @neetcode1 You’re straw manning him. He didn’t say it’s a burden he said it’s a waste.

It was such a waste of cognitive capacity. People so busy debating AI that we forget it’s possible for one person to solve much bigger and complex problems now

I don’t disagree about the time factor. On the cognitive perspective, this is a very simple task. If tasks like this are labeled as a cognitive burden to taking too much cognitive capacity (assuming he meant thinking capacity), that points to a larger issue rather than the task itself.

@r2dsl7 @neetcode1 It is finite and limited by the container called Time.

Is @diabrowser a thing yet? Or is everyone still using Arc browser?

@thdxr @jarredsumner You are getting really good with this Twitter thing.

@jarredsumner what's everyone gonna do with their 8mb?

In the next version of Bun Bun uses 8 MB less memory on Linux at start

@TejasKumar_ Oh it’s a bitch to carry around. My commute is at least 45min and carrying this thing around has been a workout.

getting a 16” MacBook Pro is a mistake I will never repeat the smaller ones are always so much better

how do you guys gain followers from scratch on twitter without shit posting or rage baiting?

@_can1357 API versioning is typically used to signal breaking changes. But some teams iterate more frequently and only bump the version if the core has changed. Others just dont care anymore, or there’s no real decision tree for how versioning happens at all.

Idk why we prefix APIs with /v1, at this point all inference APIs are on their 5th iteration, I've yet to see a /v2

@rasbt I wonder Sebastian, if an AI includes your book in its training data, how do you think authors can protect it going forward. Would people then be able to learn the same ideas through AI instead of buying the book?

It's available from Manning (already shipping): mng.bz/Nwr7 and Amazon (for preorder, shipping in a few weeks): amzn.to/4aAKiFY

After 18 months of writing, coding, and experimenting, Build a Reasoning Model (From Scratch) is finally out! My first copies just arrived! 📚 440 full-color pages. Inference scaling, reinforcement learning, and distillation from scratch.

@AnthropicAI @claudeai Can you give us extra usage credit for just this one week 👀

We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow, and will share an update soon. We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.

It’s an open source native app that does the job. If you have time to point out what’s wrong with it, you have time to open a PR and improve it. Hating on it is just wasted energy.

@steipete Design first then automate. Try Claude Design and you will see the problem when you want to tweak things manually because the agent misses your vision Manual tools with automation are the right approach. Token limits make this even more important

This would have been amazing a few years ago, but in the age of AI, what’s the point?