I guess maybe, but then do those documents lose value as technical documents? Not necessarily at all, so I don’t see the point. How are you supposed to describe a useful technical thing to users?

If MS ever decided to discontinue VS Code or relicense it, there would be blood in the water. I guarantee you there would be multiple compelling competitors in under a year and probably a new open source winner with consolidation in 5.

So to answer your question: they would be forking Atom (which I think would’ve won otherwise).

There’s no “gap that becomes truly zero” at which point special consequences happen. By the time we achieve AGI, the lesser forms of AI will likely have replaced a lot of human knowledge labor through the exact “brute-force” methods Chollet is trying to factor out (which is why many people are saying that doing so is unproductive).

AGI is like an event horizon: It does mean something, it is a point in space, but you don’t notice yourself going through it, the curvature smoothly increases through it.

Whoa whoa whoa hold your horses, code has a pretty important property that ordinary prose doesn’t have: it can make real things happen even if no one reads it (it’s executable).

I don’t want to read something that someone didn’t take the time to write. But I’ll gladly use a tool someone had an AI write, as long as it works (which these things increasingly do). Really elegant code is cool to read, but many tools I use daily are closed source, so I have no idea if their code is elegant or not. I only care if it works.

There are lots of ways they could be doing this. And remember again, if they get you, they don’t have to tell you how they got you (so you might not be able to even glean information in return for the $200 you’d be losing).

Sure the internet has hundreds of thousands of super smart coders, but the subset who are willing to throw money and credit cards down the drain in order to maintain a circumvention strategy for something like this is pretty low. I’m sure a few people will figure it out, but they won’t want to tell anyone lest Anthropic nerf their workaround, so I doubt that exploits of this will become widespread.

And if you’re Anthropic, that’s probably good enough.

Many ways, and they’re under no obligation to play fair and tell you which way they’re using at any given time. They’ve said what the rules are, they’ve said they’ll ban you if they catch you.

So let’s say they enforce it by adding an extra nonstandard challenge-response handshake at the beginning of the exchange, which generates a token which they’ll expect on all requests going forward. You decompile the minified JS code, figure out the protocol, try it from your own code but accidentally mess up a small detail (you didn’t realize the nonce has a special suffix). Detected. Banned.

You’ll need a new credit card to open a new account and try again. Better get the protocol right on the first try this time, because debugging is going to get expensive.

Let’s say you get frustrated and post on Twitter about what you know so far. If you share info, they’ll probably see it eventually and change their method. They’ll probably change it once a month anyway and see who they catch that way (and presumably add a minimum Claude Code version needed to reach their servers).

They’ve got hundreds of super smart coders and one of the most powerful AI models, they can do this all day.

To the extent that I’ve heard people propose solutions, many of them have pretty big flaws:

- Retraining - AI will likely swoop in quickly and automate many of the brand new jobs it creates. Also retraining has a bit of a messy history, it was pretty ineffective at stopping the bleeding when large numbers of manufacturing jobs were offshored/automated in the past.

- “Make work” programs - I think these are pretty silly on the face of it, although something like this might be mecessary in the really short term if there’s very sudden massive job loss and we haven’t figured out a solution.

- Universal Basic Income - Probably the best system I’ve heard anyone propose. However there are 3 huge issues: 1 - politically this is a huge no-go at the moment (after watching the massive Covid stimulus happen in 2020 I have a sliver of hope, but not much). 2 - Even a pretty good UBI probably wouldn’t be enough to cushion the landing for people who make a lot right now and have made financial decisions (number of kids, purchasing a house, etc) on the basis of their current salary. 3 - Even if this happens in America (presumably redistributing the wealth accruing to American AI companies) it would leave non-Americans out in the cold, and we currently have no globally powerful institution with the trust and capability to manage a worldwide UBI.

In the past, AI coding agents could usually reason about the code well enough that they had a good chance of success, but I’d have to manually test since they were bad at “seeing” the output and characterizing it in a way that allowed them to debug if things went wrong, and they would never ever check visual outputs unless I forced them to (probably because it didn’t work well during RL training).

Opus 4.6 correctly reasoned (on its own, I didn’t even think to prompt this) that it could “test” the output by grabbing the first, middle and last frame, and observing that the first frame should be empty, the middle frame half full of details, and the final frame resembling the input image. That alone wouldn’t have impressed me that much, but it actually found and fixed a bug based on visual observation of a blurry final frame (we hadn’t run the NeRF training for enough iterations).

In a sense this is an incremental improvement in the model’s capabilities. But in terms of what I can now use this model for, it’s huge. Previous models struggled at tokenizing/interpreting images beyond describing the contents in semantic terms, so they couldn’t iterate based on visual feedback when the contents were abstract or broken in an unusual way. The fact that they can do this now means I can set them on tasks like this unaided and have a reasonable probability that they’ll be able to troubleshoot their own issues.

I understand your exhaustion at all the breathless enthusiasm, but every new models radically changes the game for another subset of users/tasks. You’re going to keep hearing that counterargument for a long time, and the worst part is, it’s going to be true even if it’s annoying.

It is clearly plain to anyone who is a musician or hangs out with a lot of musicians that the independent music world is livid about this stuff. Everyone I’ve talked to, from acoustic songwriters to metal singers to circuit-bending pedalheads are united in their absolute hatred of this technology.

(Yes, follow-up commenter, I’ve seen the Timbaland interview)

The tradeoffs of the different models are complicated and difficult to wrap your head around, and if you have the resources to try a bunch and form a conclusion, next week new models will come out and change the equation in small but difficult-to-understand ways again.

The solution is to ask your engineers which models they like, get them access to as many of those as you can, and expect their preferences to change and price that in.

“But I don’t have the budget to buy subscriptions to all the models my engineers want! And there are compliance issues with some of them!”

Note that I didn’t say you have to get access to all of them: as many as you can. And try to push the envelope as much as possible. Get creative. Perhaps give engineers a $200/mo AI coding budget and let them pick from a selection of subscriptions. You’re going to have different engineers using different AI coding tools, and if you refuse to let them, your competitors will.

Maybe in the future “standardizing on one coding agent” will be a thing that makes sense. But that time is not now.

At the moment I have a personal Claude Max subscription and ChatGPT Enterprise for Codex at work. Using both, I feel pretty definitively that gpt-5.2-codex is strictly superior to Opus 4.5. When I use Opus 4.5 I’m still constantly dealing with it cutting corners, misinterpreting my intentions and stopping when it isn’t actually done. When I switched to Codex for work a few months ago all of those problems went away.

I got the personal subscription this month to try out Gas Town and see how Opus 4.5 does on various tasks, and there are definitely features of CC that I miss with Codex CLI (I can’t believe they still don’t have hooks), but I’ve cancelled the subscription and won’t renew it at the end of this month unless they drop a model that really brings them up to where gpt-5.2-codex is at.