thieme

@unclebobmartin Maybe Rust next? I imagine that it will have the longest compilation times, as it will check much more. But then again will it offer the optimal runtime conditions. It might be okay to prototype an application in Typescript, which is the fastest development for AI and then Rust.

I re-ran the java experiment and had it avoid maven for the test runs. That improved things mightily. The new time for Java is ~2:52. This compares to GO at 3:29 and Clojure at 2:24. CPU overhead was significantly reduce on this run. So, yeah, it was maven that slowed things down so much.

@unclebobmartin You want to jump to some specific screen/scenario inside a graphical UI to save click time. That will require a high quality software underneath. ADD => Algebra Driven Design.

Running UI QA tests will likely require special application states so that the tests can be driven through the desired pathways. These might be engaged with command line arguments or special QA test commands.

@unclebobmartin They automatically suggest a commit message after having done work. And when using my acceptance-tdd skill they do text book tdd and commit after each green test (automatically).

one of the great advantages of working with agents is that they actually understand git.

@MePeterNicholls @raphaelmansuy @GitHubCopilot If you are a company and want EU-residency and EU-inference.. then what provider will offer it cheaper?

@thieme @raphaelmansuy @GitHubCopilot $50 A DAY?!? lol. No. I’m using other providers instead.

Formal Complaint: GitHub Copilot Token-Based Billing Model @GitHubCopilot Subject: Critical Issue with New Token-Based Billing — Product Has Become Unusable Summary of the Issue I am writing to formally complain about the recent shift to token-based billing for GitHub Copilot, which was rolled out this morning. This change has fundamentally broken the value proposition of the product and is rendering it unusable for paying subscribers, including myself. Specific Problems Observed Within just a few hours of the new billing model going live, the developer community is already reporting alarming consumption patterns: Pro+ subscribers paying $39/month are reporting that 60% of their monthly credits were depleted in only 2 hours of normal usage. One user reported losing 20% of their entire monthly allowance from a single file review — no code generation, just a review. At this rate, a paying customer will exhaust their plan in less than a single working day, despite paying a premium subscription fee. This is not "normal usage at scale" — this is a broken pricing model that punishes the very developers who rely on Copilot daily for their work. Why This Makes the Product Unusable The core promise of Copilot was a predictable, always-available AI coding assistant integrated into the developer workflow. Token-based billing destroys that promise because: Developers cannot predict costs. Every keystroke, every file review, every refactor becomes a financial calculation rather than a productivity boost. The tool actively discourages use. Users will hesitate before invoking Copilot, defeating the entire point of an AI assistant. The $39/month Pro+ tier is misleading. Customers signed up expecting reliable access, not a pre-paid metered service that runs out mid-morning. Heavy users — your most loyal customers — are penalized the most. The Competitive Reality While GitHub Copilot is moving toward a restrictive metered model, competitors are moving in the opposite direction: Cursor offers Composer 2.5 with unlimited usage once token limits are reached on their plans, ensuring developers can keep working without interruption. Other tools (Windsurf, Cody, Continue) offer flat-rate or far more generous usage tiers. Developers will not stay on a platform that runs out of credits before lunch when alternatives offer uninterrupted productivity at the same or lower price point. My Demand If GitHub does not revise this licensing model, the product is effectively dead. I am requesting: Reinstatement of a flat-rate unlimited (or effectively unlimited) tier for Pro and Pro+ subscribers. Transparent, upfront communication of what each interaction actually costs in tokens. A grace period or credit refund for users who burned through their allowance under the new model without warning. A long-term commitment that core IDE-integrated features will not be metered into uselessness. Without these changes, I — along with a growing number of developers — will be canceling our subscriptions and migrating to Cursor or competing alternatives. The decision to monetize aggressively at the expense of usability will not be remembered as a successful pivot; it will be remembered as the moment GitHub Copilot lost its market. Please escalate this to the product and pricing teams immediately.

@MePeterNicholls @raphaelmansuy @GitHubCopilot In what way is it unusable? Or do you mean you can’t afford it? Not sure. But you could try to give yourself a budget of $50 per day and see how far this goes… why not?

@raphaelmansuy @GitHubCopilot I’m cancelling because it’s in unusable.

@raphaelmansuy @GitHubCopilot *Somebody* has to pay for it. What if they bring back the old pricing but you have to buy your own $800k system to run a model, and pay electricity? Try to sue them or just move on or stop using AI or whatever.

@unclebobmartin You can create a screenshot skill. Put a tool/script into the folder where your SKILL.md lives that can take screenshots. Then your agent *can* see the screen.

The agents cannot see the screen, and cannot tell the difference between things that work in the tests, and things that appear on the screen. Therefore, QA testing through the UI is necessary. This can be automated by the agents. UI testing is very fragile because the UI is the most volatile part of the system. It changes for lots of reasons. This used to put pressure on us to avoid testing through the UI since all those tests would become as volatile as the UI. Agents are workhorses. They don't care about rewriting the UI tests. So that pressure has shrunk appreciably. It is now much more feasible to create, and maintain, a good UI test suite. I'm cautious about this, but I also see no alternative.

@unclebobmartin With this I agree. However. By using the AI *you* are now an implicit OpenAI employee who isn’t get paid but who pays to work for them. You are creating the new training data. Current AIs have never seen this. Next year they will.

There is no way to get the human out of the loop. The Agents simply do not understand obvious user expectations.

@mattwridley In the last 300k years noone has ever moved faster than 100 km/h. By your logic it should not be possible today.

I disagree. Maximum human lifespan has hardly changed in 30 years. The oldest person alive today, Ethel Caterham, is 116, a long way short of setting a record (122). Only 4 people have ever gone past 117. Increasing average lifespan does not lead to increased maximum lifespan. It would take genetic engineering to change this.

@embossedly @jasperdevs How can something be a scam when they offer you more than you paid for? We signed a contract with them and understood there are no extra resets included. Even a one-minute reset should make us slightly thankful.

@jasperdevs it's the biggest scam they reset limits for 4.5+ million devs simultaneusly a week ago then reset them an hour before they would've reset anyway. their excuse last time was: "tons more devs have signed up and aren't on the same cycle!!!" doesn't work this time

Why are 90% of codex/claude resets conveniently like 30 minutes before the weekly limit already would of reset 😭

@koltregaskes I’d be totally fine with everything I want to buy or use being cheap or free. Failing that, Elon just wiring me the money works too😄 Seriously though — OpenAI would love you to pay 1000x more. So maybe the middle ground is: we just pay extra or use the service less.

@thieme Add tokens sounds like it costs me more? My suggestion is the same price but to move the tokens to when I need them in my timezone. :-)

These 5-hour token resets are brutal during weekends. That's when I actually have time to run my agents, but I'm constantly hitting limits. I'd happily sacrifice overnight tokens for more during the day and especially weekend days. But even if I could shift the allocation, it wouldn't solve the real problem. Agents still aren't autonomous. They can't loop through hundreds of queued tasks. They don't complete tasks satisfactorily without fixes. They don't wake up when the 5-hour reset happens and carry on. Most of the time I'm pushing them along, feeding them corrections, forcing them to follow instructions. Recent research confirms this - what they're calling "calibrated autonomy" is just a polite term for constant hand-holding. We're nowhere near autonomous agents. We've got expensive assistants that need micromanaging.

@HelpDefeatAging Do it!

We need to solve cancer ASAP.

@daily_romania That would be amazing! Microplastic is inside our bodies and brains. We MUST get rid of this stuff. Plastic should not touch food. Clothes should not be made out of plastic.

EU bans single-use plastic packets for ketchup and other condiments, customers will have to use shared bottles or refillable dispensers

@tomfgoodwin It’s a skill issue. AI wasn’t usable at all. Now it’s tricky. Now millions of people use it and this usage becomes the next training material. Our human intelligence/skill will get copied to models of 2027, and then outputs will have much higher quality.

errmmmmm, not to be miserable but has anyone noticed that agentic AI doesn't really work at all. Like the errors compound, fragile integrations ( any external change breaks it ) , observability is an issue, no verification, context loss, the whole thing seems VERY tricky Not sure this can ever be fixed.

@plainionist Those are the ones that are truly important but you are right… it is so much reading and thinking about details. AI removes the simple tasks. The ones that remain are hard.

When AI writes the code, reviewing acceptance specs becomes mission-critical ... But damn ... it’s boring 🤷‍♂️

@koltregaskes @DimitrisPapail What are you using Codex for? Weren’t you a non-programmer until recently?

@DimitrisPapail I did that last month. I think I'm actually going to go even further by dropping Claude to Pro and Codex at times 20

Codex is now too good to be on the Plus plan. I just downgraded my Max 20x Claude plan to 5x and upgraded Codex to Pro (5x). The (1/2,1/2) combination is Pareto Optimal.

@Star_Knight12 What does this even mean? You started with a lowercase letter? Why not write correctly? Is your `is` placed correctly? Is there an exclamation mark or questionmark missing? On my computer it’s called “Codex” with an uppercase C.

is codex better than Claude code

@asaio87 Think of evolution. It took billions of years and a quantum computer the size of Earth to output us. This is what I call a lot of good hardware.

hash truth: AI is more about powerful enough hardware than actual intelligence given by software innovation.

@nickprince I liked the video and I would hope that Twitter will allow us to make 3-5 posts a day with an AI bot, just for simple personal use. Just sayin’ that current terms make it unnecessarily complicated.

@thieme yeah to be clear I still posted this myself, along with uploading the video. figured it would be cool demo though. I’ve found computer use most high leverage when actually building things (recently a remote mcp) but I wanted to show something everyone can relate to

watching codex control my browser to do things it can't do in the harness is a holy shit experience