tluyben2
Joined 16 years ago
Karma
11,388
Submissions
5,899
Member since
Dec 18
Recent submissions
Comment · 3 days ago
Because the result is the same. We did split test and when something is wrong we retry with fable or opus to see what it would do. But I did say seem; I am not a researcher, just someone running a business and it seems faster, cheaper and with the same outcomes so far. Subscriptions last much longer now for no visible degradation yet.
Comment · 3 days ago
I now run Jev in front of claude code and the results so far have been very good. It picks Haiku far far more often than I would think to use it and it works very well.
Comment · 9 months ago
Nice work; I remember browsing it during covid.
Tried AI for a standalone html page snippet renderer[0].
I tried just feeding it snippets one by one and that worked, but the raymarching ones it could not get going until I gave it the dwitter github repos. Now most, but not all (simple to fix manually though), work.
Also interesting to see Claude is terrible at trying to write the art (the demos itself) and seeing what it tries to do; not surprising given the challenge.
[0] https://claude.ai/public/artifacts/d8b357df-5982-48c6-be58-7...
Comment · 1 year ago
I like it, and indeed neatly shows the power of Lisp. The JS variety (well the one that I could come up with) is far less elegant, but works [0] (well, mostly). It really shows how the different LLMs stack up; some really cannot get anything right, but something like openai/gpt-4o-mini seems to get it right mostly (8/10).
Comment · 1 year ago
Similar to others, the unevenness of prompt results makes pricing per message quite tricky; it generated a nice looking app for me in the first go, I asked for an enhancement and it spent the rest of my free messages on trying, and failing, to fix 1 TS error that came from that enhancement. As this is using one or more of the openai/anthropic/google models, I also know that it probably won't be able to actually fix that error without my (either in code or specifically telling it what to do, which is coding) explicit help and just loop until I burn through whatever plan I have that way.
I got a cheap 1 year account on replit and it suffers from this of course; first few prompts yield amazing results and then it gets stuck. This is fine for me as I can fix it myself and by now I have good feeling of which type of error the current llms will just loop forever on, but it is hard to justify the pricing model of per message because for people who cannot fix it themselves and need to vibe on it; spend 50 bucks/mo for looping over one error for a day and that's that; pay more or wait or eject and go to another; I see non devs (vibe-only devs?) doing the latter so you never retain clients; they pick the good deals every month and move from one to the other.
So far only cursor is reasonable (that I know off) because it continues after the paid messages run out with the slower requests; you can go on forever.
Comment · 1 year ago
That was what I tried on the train [0] a few weeks ago. I used Groq to get something very fast to see if it would work at least somewhat. It gives you a PDF in the end. Plugging in a better model gave much better results (still not really readable if you actually try to; at a glance it's convincing though), however, it was so slow that testing what kind of impossible. Cannot really have things done in parallel either because it does need to know what it pushed out before, at least the summary of it.
Comment · 1 year ago
I tried to vibe code (I don't like the term and normally would never) this as someone here suggested, so here [0]. It did not 'quite' take less than an hour and it's not usable yet for writing (just reading); it does have working dark mode though.
I only helped out when it was completely stuck (twice), other than that I didn't read or check the code, just talked (voice) ideas to it when it was done while doing other (not vibe) coding while it was looping over it. I did, I guess stupidly, dictate the tech; I wanted to try Go + htmx; that probably didn't help; probably Python and/or React would've lead to better results.
Comment · 1 year ago
Turbo c was slow, Turbo Pascal and later object pascal were (and are) very fast.
Comment · 1 year ago
Cline works well for me.
Comment · 1 year ago
I think this was the video?
Comment · 1 year ago
I for one partially wrote a few mobile games with LOAD81 ; I was travelling and only had my Pandora with me; I wrote all the logic tested them all out with load81. Of course I had to actually get them on a mobile platform device and create a better UI, however all the logic and most of the UX was all done in lua on load81.
Comment · 1 year ago
This is great; I have been thinking about this for a long time. I like reading about past and current implementations that try to better sql; from a programming and a data science and performance perspective. I am aware of the ones you linked and some others like 'Real' (shakti.com) sql and some enhancements from papers.
Anyway; nice one! Will try.
https://flexlists.com
Email: tycho at e-lab dot nl