vunderba

10 ms·

1.
by vunderba 4d ago
I think we’re in agreement. I make heavy use of smaller local models on a daily basis (Qwen3-VL for auto-captioning images, Gemma3:27b for some translation work, etc.). Gemma3:27b is a good example of a very capable general purpose multimo
2.
by vunderba 4d ago
I read about a study a while back in which researchers brought in a person who had synesthesia and asked him to identify the exact colors associated with 500 different random numbers drawn from a range of up to a million. (He did not know w
3.
by vunderba 4d ago
> People are realizing that what they need isn't more general intelligence, it's more specialization. A small but well tuned coding model... It’s not quite as simple as that. Several studies have shown the opposite: models tr
4.
by vunderba 5d ago
Nice job. You should make it so there’s a chance that if you try to navigate to another area too quickly the narrator says, “You must gather your party before venturing forth.” This is super nostalgic, but BG1 is still actively sold particu
5.
by vunderba 5d ago
Related 2025 paper on how GenAI running in an unconstrained loop tends toward convergent collapse: https://pubmed.ncbi.nlm.nih.gov/41583980
6.
by vunderba 5d ago
I always run stairs when given an opportunity (big proponent of "incidental exercise"), so when I used to live in apartments I would deliberately rent a room in the highest available floor. It's shocking how much HIIT you can
7.
by vunderba 5d ago
Same! I used to play Max 300 in heavy mode (on repeat) at my local arcade without hugging the bar - your feet end up moving like Donald Duck from Mathemagic Land. 30 minutes and you are gassed... Nowadays I switched over to Pump It Up - I f
8.
by vunderba 6d ago
This seems to be one of the few open-weight models that isn't just another "text to audio" generator. You can take existing sheet music, convert it to ABC notation, and then experiment with various covers. I don't have a
9.
by vunderba 7d ago
There's a great demo video of it as well. https://youtu.be/uu7jA8EHi_0 Some of the details in the paper are super cool. Since the game pulls "c" down to near walking speed, you receive photons at different ra
10.
by vunderba 7d ago
Reminds me of wired elements (which uses Rough.js under the covers). https://github.com/rough-stuff/wired-elements
11.
by vunderba 7d ago
> everything is period-correct EXCEPT for the hard drive and floppy drive Hence why I couched my assertion with "except". IMHO its the best way to enjoy the original hardware experience without all the inevitable failures of
12.
by vunderba 8d ago
The whole spec-driven development movement has always been a bit odd to me, because it feels more like a traditional grounded approach towards regular software engineering. I can share my personal approach which has worked fairly well at le
13.
by vunderba 8d ago
That’s surprising even gpt-image-1 was able to handle the equestrian astronaut prompt pretty trivially, albeit yellow-tinged as all hell. For the heck of it, I ratcheted up the difficulty of the prompt (two-headed horse, astronaut with the
14.
by vunderba 8d ago
I know the API (assuming you're using it) lets you set the output size of the edited images - I haven't done a huge amount of testing with > 1mp, so I'd be curious if this might mitigate some of that. https://de
15.
by vunderba 8d ago
Yeah, I think you’re right. I think the hardest part would honestly be maintaining some level of consistency as your project grows its number of assets - otherwise there's no sense of a common stylistic theme. The agentic tooling in so
16.
by vunderba 8d ago
You can get semi-decent sprite work out of GenAI models, but you still have to put in some manual work (scale normalization, palette reduction, grid alignments, etc). It's definitely not "out-of-the-box" yet. https:/&#x
17.
by vunderba 8d ago
The OpenAI PR/doc team seem to have a history of this kind of attention lapse - their 4-panel comic strip using gpt-image-2 was an absolute mess too. https://mordenstar.com/blog/gen-failures
18.
by vunderba 9d ago
Nice job. Your tagline should be Twice the Fun of Atari's Quadrapong! https://www.youtube.com/watch?v=bP3idyROGes
19.
by vunderba 9d ago
Learning Company software has been pretty well archived at this point. To me, the easiest approach would be to hold onto the original box/packaging for sentimental purposes, and just find a copy of it on a site like the Internet Archiv
20.
by vunderba 12d ago
As somebody already mentioned, they already had a big product named after their CEO. IBM Watson was the next big thing after Deep Blue, and it went pretty viral after defeating Ken Jennings and Brad Rutter on Jeopardy back in ~2011. It was
21.
by vunderba 12d ago
On my experiments on GenAI Image Showdown, Nano Banana Pro slightly edges out NB2, though I'm not sure the cost difference makes it worth it. And in pure qualitative terms, I'd agree that NB2/NB Pro are still on par with gpt-
22.
by vunderba 14d ago
The more generic your prompt, the more generic the response. It's a regression to the "mean" of the training data aka GIGO for AI. It's like when you ask your average person off the street to draw a house - it'll al
23.
by vunderba 14d ago
Back when I was still on Windows(XP days), I made heavy use of Paint.net. Given Microsoft’s penchant for gobbling up smaller companies, I was always a bit surprised that they never bought the company and integrated paint.net as the built-in
24.
by vunderba 14d ago
I can see that. I wish Pixelmator (before it got bought up by Apple) had been ported over to Linux. That's been the closest lightweight super-intuitive alternative to Paint.net I've ever messed around with.
25.
by vunderba 14d ago
I guess more digital art options on Linux are always good, but that aside, is there even a good reason to use it when something like Krita exists? It's always worked flawlessly on my linux machine (Mint). https://github.com&
26.
by vunderba 15d ago
I wouldn't say you have to "advertise" per se (well outside of a more traditional SaaS offering)... in fact I think most people particularly on HN are pretty sales-adverse. I think you just have to demonstrate you are willing
27.
by vunderba 15d ago
I've had remarkably decent success here on HN, but I also release all of my stuff completely free (no sign-in stuff, no ads, etc) so there's very little user friction. One thing I can say from my experience is that you get out of
28.
by vunderba 15d ago
Sometimes Opus 5 (high/xhigh) feels like I'm dealing with the programmer equivalent of Zeno of Elea. Every time, without fail, it would get me 90% of the way there and then leave a small note, exception, or deferral. When instruct
29.
by vunderba 15d ago
> I had never bid in a Heritage Auction before and admit with shame that I "slowl bid" the price up until I won. I got the game for a steal (in my mind). I won with a hammer price of $2,000. Heritage added 20% for buyer’s prem
30.
by vunderba 15d ago
A lot of my younger friends saw their parents spend countless number of hours trying to time the market with specific stock investments, only to get burned in a similar way. Most of them have now committed to simple dollar-cost averaging ag

More ›