HNHacker News
TopNewBestAskShowJobs

kadoban

7,571 karma · joined April 20, 2018

submissionscomments
kadoban··on Go grandmaster Shin defeats AI KataGo with a two-stone handicap
There are tweaks to move choice like that, but it's all evaluation time. None of that happens in training. Zero times in the RL loop does katago see a two stone game against a weaker player.

You can kind of tweak towards play this metric or that, but it's not the same.

kadoban··on Go grandmaster Shin defeats AI KataGo with a two-stone handicap
The headline is misleading, but this is still huge. 2 stones against katago is insane, I'd never have guessed we'd see that, ever.
kadoban··on Go grandmaster Shin defeats AI KataGo with a two-stone handicap
> Also, the human played a strategy tailored to that huge initial advantage. He said that the AI did not handle this particularly well, and played high probability moves instead of trying to lure him into a mistake.

Yeah, katago's training is not really focused at all on handicap games, because it's by nature learning from even games against similar-strength opponents.

It doesn't have specific training from playing in a way to exploit a weaker player. In a handicap game you have to give your opponent opportunities to fuck up if you want to play optimally.

If a move loses 0.0005 points if the opponent plays optimally, katago won't play it even if there's ~zero chance a weaker player would play it right.

There have been go AIs that tried to train more directly on uneven opponents, one called "sai" comes to mind, but katago has huge advantages otherwise and won out over the others (for very good reason, it's a great project).

kadoban··on Flock Safety's controversial CEO just got a taste of his own medicine
That's not ironic, that's _hey, look over there!_
kadoban··on Dimethyl Mercury Exposure Incident at MIT
The wording of the announcements, sounds like a student synthesized it. Can't imagine why, hopefully it wasn't a fatal mistake.
kadoban··on Z.ai confirms Ox Alpha is a new GLM-series model and will release its weights
Pi.dev has an extension that does it. None of the models I use seem to have that failure mode so I haven't bothered throwing it in.

The failure mode I run into commonly is agents just stop sometimes. Even sending a "." Or something they start back up, but I haven't worked out exactly how to fix that generally in harness, bit unclear how to tell if they're done or just derped to a stop.

kadoban··on Black hole singularity is a surface not a point
> The problems start once you consider quantum mechanics, because any such singularity will break unitarity (a fancy way of saying that probabilities must add up to 1), which means your theory as a whole can no longer make predictions.

How is this any different than classical? Isn't it still just an ~impossibility hidden behind an event horizon in either model?

kadoban··on DeepSeek-v4-flash-vision-exp
It's not that useful for a person to be able to read an analog clock either, but if a supposed-genius came to me and confidently gave the wrong answer, it would say something about their strengths/weaknesses in general.

The whole point of these models is they're meant to be able to generalize fairly well, not just answer questions the got trained on.

kadoban··on Mushroom behind 'tiny people' hallucinations identified
> What stopped them?

This would be impossible to get past ethical review in any institution who has such a thing. You'd be forcing people to have a trip, which may or may not be pleasant for them, for days, with no guidance on wtf is happening to them. Depending on the person, that's some pretty heavy trauma.

You could tell people what to expect, but that ruins the blind part. I still think you'd have no chance of passing review anyway.

kadoban··on Mushroom behind 'tiny people' hallucinations identified
That's probably based on some specific definition of food poisoning where you get effects from the food. It's semantics rather than saying anything useful about what's going on.
kadoban··on Qwen3.8-2.4T
Tbh I just go through openrouter because it's easy, so you can just go down the list in that and see who sounds good. I honestly don't spend much time at it.
kadoban··on Qwen3.8-2.4T
I wouldn't just rush out and buy hardware, but there will be benchmarks after a while to make an informed decision.

95GB active is not _too_ bad, would require some creativity and $$, but I bet I could do that at home for less than a cheap car.

kadoban··on Qwen3.8-2.4T
> no zdr ofcourse.

Since the models are open, there's other infra providers who have different policies.

kadoban··on Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots
Training scales pretty badly, so smaller models like this are really not that bad in terms of cost.
kadoban··on Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
Is the quality of the thinking better or it's just shorter since Mandarin is more compact?
kadoban··on An Interesting Fourier Transform – 1/f Noise (2007)
I went through a phase of using ~whitenoise while working to block out distractions.

Actual white noise indeed sounds really bad and grating. The best for me was a mix of pink and brown noise, pink for an ~equal baseline and brown to make it sound a little more mellow.

I suspect most/all generators meant to block out noise do something similar. It really sounds pretty bad without that, especially for anything more than a few seconds.

kadoban··on Why Wall Street is ignoring big tech's debt [video]
> Have you been trashing the non-replaceabe SSD with constant AI workloads?

Your other point is valid, but you're either vastly underestimating how many reads/writes a modern SSD can take or vastly overestimating how much output a typical LLM is capable of.

kadoban··on Andrew Wiles on proving Fermat’s Last Theorem (1995) [video]
Consensus is that there's no way he had a correct, general proof.

My personal guess is he figured out later that his proof was broken or limited and never got around to fixing it.

kadoban··on Black lung rates in Appalachia are high as the 1970s
Coal doesn't even compete with just solar, those aren't your two options.

Coal competes with literally everything else, and it loses to every single one of them. It's a terrible, terrible fuel. Burn natgas instead and you're already magnitudes better off.

Coal only exists today in the US because our politics is broken beyond belief.

kadoban··on Can Intel finally beat ARM on performance per Watt?
It was obviously loaned so he could do a review. That's not sponsorship.

You should _want_ that to happen, otherwise the only people who can do reviews will be the few who can afford to spend thousands for hardware they don't want or need.

kadoban··on Reviving a four year old reMarkable 2
Yeah, koreader is great. As long as you're fine with greyscale, the rm2 size is great for comics and textbooks. Most ereaders are way too small for both.
kadoban··on U.S. Department of Energy Launches the Genesis Open Models Initiative
S, I'm running one of the unsloth quants. Q4 something K_XL I think?

I don't think it's all resident in memory on the 3090 but it runs fast enough for use, I'd have to check but it's like 30 t/s or so, fast enough that it doesn't bother me.

kadoban··on U.S. Department of Energy Launches the Genesis Open Models Initiative
It's a lot smaller, and runs (quantized) on a 3090 quite well. Ds4 flash 0731 you're talking about? It's great but it's much harder to run locally.
kadoban··on U.S. Department of Energy Launches the Genesis Open Models Initiative
Yeah I think it got bad press because the chat templates (or something?) were messed up on first release, but I've been using a quant of it and it's a powerhouse, better than qwen 3.6 27b for local on a 3090, which is saying a lot.
kadoban··on LLMs won't break symmetric crypto
Is there? Like...kind of, but on the face of it I'd say about the same amount in both. If you look back at DES there's a _lot_ in common with modern ciphers, but like, RSA is still in use and that's old as shit.

I think you're right if your point is that we're more likely to see big breaks in asymmetric crypto, but it's kind of based on vibes to me, it's not really clear that it's provable in any way with anything like our current understanding.

kadoban··on LLMs won't break symmetric crypto
> with a key that repeats

There's a _lot_ hiding in that, all of the interesting stuff for security and potential breaks. So...yeah it is based on complicated math, it's just in that bit instead of the xor.

Even the xor is a bit of a fudge, but probably close enough.

kadoban··on Why Erdős Problems Are Falling to AI
> What practical things are possible or might be possible due to these results?

Not much, they're all fairly minor problems that have been solved, of ~entirely niche academic interest. It's so far more just that AI _can_ solve novel math problems, ones that humans didn't accidentally train the answer in and it just spit it back out.

kadoban··on Pi's Minimalism Is Its Advantage
You can just fire it up and ask it to suggest some, or look on the site at the most downloaded and see which sound useful to you. That's basically what I did. I also put in some work to containerize it and do some manual/automatic checks to lock down extensions a bit, but that's just me.

It really doesn't take much for it to be useful, maybe something to search the web?

If you want you can just tell it to write you extensions too, like one to wrap curl so it can get to the web easier for instance. Or just tell it to use curl, really up to you. Just an example to point out: like, do whatever, it's very flexible.

kadoban··on Pi's Minimalism Is Its Advantage
It's really just the minimal prompt and the minimal built-in skills. And ~full control I guess if you wire up something custom (I haven't).
kadoban··on Show HN: Elevators
Human psychology might allow other reasonable criteria even if perfectly rational choice is just total time to destination.
← PreviousPage 5 of 34Next →