rowbin

I do miss programming instead of reading code. But I enjoy designing and picking languages and tools. For me AI has been great (so far): I revived old projects and came up with new ones, including branching to hardware products. This would have never been achievable without AI. Not because a lack of skill. I'm sure I could have picked any of the projects and see it through. But I wouldn't have, because the time and effort wouldn't have justified the result. And I would never have been able to do so many of them. I would have needed to pick one, pour many, many hours of work into it and all the other things would need to wait. So I could have done a fraction, but in reality I would have hardly done any. So for me it's enabling.

It's not that I couldn't have. I wouldn't have.

Probably more than a year. Maybe even two. I used to be on there all the time. Kinda nostalgic for it.

And the write up: https://blog.leonbecker.de/i-built-a-shell-that-hallucinates...

I've built something very similar but with off the shelf models: https://github.com/beleon/vibeSH

EDIT (comparison)

Very impressive, and a lot more work than it looks from the outside (the DAgger loop especially). I built something similar in June, vibeSH, a shell that hallucinates the whole computer, but from the other direction: no training, just a frontier model and a long prompt, so maybe the comparison is useful.

The granularity is different though. You model keypresses and predict which rows change, I batch keystrokes and have the model repaint the whole screen each round trip. Non interactive cli programs work very good, but interactive cli programs like vim, less and top do work, but it feels like SSH over a 1996 modem, you type a burst, wait, and the screen catches up. Yours is going to be a lot more responsive at that.

The big model buys consistency, for as long as the session fits in the context window. Files come back the way you wrote them, python output is usually right instead of guessed, nested paths stay where they are. The price is an API call per round trip and a cost that grows with the session (and it's not 208 MB and doesnt run offline).

Mine offers an @ai prefix that talks to the director instead of the machine, so mid-session you can say "@ai make this a haunted Solaris machine from 1996" and the filesystem, the motd and the error messages all change to match, and the directives stick.

Your hardest problem and mine are the same one I think. For you it's a 360M model losing track of state, for me it's the context window filling up, after which the box gets amnesia and old files have quietly rewritten themselves. I decided early not to fight that completely and made it canon (the box has bad memory), your DAgger setup is an actual fix for it, which I don't have.

rhci runs standalone on on x86 and ARM on Linux, Windows, macOS and the BSDs via a cosmopolitan compiled mruby. Or anywhere CRuby runs via a single .rb file.

GHCi is super useful for quick checks (is `round 3.5` 3 or 4, what is `(-9) div 2`, what's the type of `(.) . (.)`, ...). But its slow to start and a huge install. rhci is a small version of GHCi that exactly matches GHCi semantics.

While GUIs have some legitimate use cases, I prefer TUIs if I can get away with it. Its just text, ssh friendly, constraint in a beautiful way that still allows endless possibilities.

Can recommend https://codefloe.com

What do you mean with "became bad"? It started out with a 140 characters limit in 2006 and kept that for more than 10 years. So its the other way around: it became worse, when the character limit increased. Though I don't think its due to the character limit either way.

I always did this on all my projects. Until I switched my approach and forced myself to do it the other way around: do the thing first and improve the workflow late. Better to do something annoying one time too often before improving the workflow, than to only improve the workflow and not do the thing.

I'm still a little worried that I'll stop posting on my blog soon, but it's actually been going better than expected. Rooting for Ben to keep posting.

Anti-AI mentality claims another victim

Maybe have a seasoned mathematician check if the counterexample is really bogus before saying people fell for ai psychosis.

as can be seen in this comment section

Maybe, maybe not. The "proofs" may not have helped at all with finding a counterexample. Either way, it doesn't matter. A counterexample was found, no one found one before even though clearly a lot of people have tried who also had access to the prior "proofs".

Absolutely!

Works

Hopefully fable 5 will be back soon. I had so much more fun on vibeSH with it. Its remarkable how much better it was at hallucinating a coherent shell than Opus 4.8. It was especially good with the ansi escape codes to get the colors right.

I get "JSON.parse: unexpected character at line 1 column 1 of the JSON data"

Author here. Happy to answer questions.

One clarification: vibeSH does not run commands or inspect your machine. The model only sees the terminal transcript, and the transcript is the fake machine's state.

Clickable link: https://stelae.eu

Thanks, ditto!

Oh, I released something with the same name recently: https://codeberg.org/beleon/mdview

I don't think anybody's worried about that. There's much more to it than just having a good idea and letting ai vibe code it.

Anyway, soldering as a service is nothing to worry about so you're good either way.

Still working on stelae.eu (private WP editor -> static deploy: more secure, faster, cheaper). Its pretty solid already, only working on minor things. The main issue is that I think that I have a real cool product (maybe a bit boring, but in a good way) with good values (anti lock-in, privacy respecting, EU centric, fair pricing, no VC money -> sustainable business approach) but I can't reach the people that would love to use it. So thats what I'm really working on: trying to be more visible.

I think this conflates setup cost with operating cost. The painful part of self-hosting was always to get it working: writing configs, reading docs, SSH ceremony. That's exactly the part that agents can help with a lot, so the perceived gap to Vercel shrinks a lot. But the problems that come later are still there, you just discover it later. DDoS absorption, zero-downtime rollbacks, cert rotations failing at 3am, someone else getting paged when they do. An agent can write my firewall rules, but it doesn't carry a pager.

So I'd reframe the thesis a bit: AI didn't destroy the moat, it moved it. Saving the customer setup time is no longer worth paying for, taking operational liability for the customer still is. SaaS that only offer the former are indeed in trouble, and honestly founders should verify that their product doesn't have this issue.

Many have theorized this. But i wouldn't bet money on it, the timing is just too unpredictable IMO.

He doesn't really use the reliance argument. Whether dropping the pause was the right call is a separate question from whether people were allowed to rely on it. They were: the v1.0 language alluded to a much higher bar for changing commitments than they ended up applying, and Anthropic employees described the RSP as binding many times (Habryka says he heard it on more than a dozen occasions). Also, Hubinger's post announcing v3 says Anthropic is responsible for that impression. If you made vendor decisions based on that like the author did, the complaint is still valid, even if the change itself was the right decision. Both can be true at the same time, they were right to change it and you were right to feel misled.

I agree, his takes should not be dismissed lightly. I'm not sure about "demand is fixed" though. I feel like software demand has been declared saturated at least a few times.

I saw what changed syntactically. I meant I don't really understand what changed semantically. And whether there is any context to why the change was necessary.