gwern

25 ms·

421.
by gwern 2y ago
Further evidence that human children are just LLMs whose 'capabilities' do not generalize outside the training data or are robust to even minor variation of the abstract algorithm being tested. Very sad.
422.
by gwern 2y ago
Although maybe not if he's going to Mira Murati's startup instead: https://fortune.com/2025/02/06/openai-john-schulman-mira-mur...
423.
by gwern 2y ago
Isn't he running a company (Palette) which makes heavy use of DL?
424.
by gwern 2y ago
That wasn't a long stay. Maybe after all those years of the OpenAI wild ride, particularly being in charge of ChatGPT-related things, he's just burned out on R&D work and needs a long break?
425.
by gwern 2y ago
2.0 Pro Experimental seems like the big news here? > Today, we’re releasing an experimental version of Gemini 2.0 Pro that responds to that feedback. It has the strongest coding performance and ability to handle complex prompts, with bet
426.
by gwern 2y ago
Any idea which dating app Paul Gonsolin is talking about here? This is a lot of data.
427.
by gwern 2y ago
I am uninvolved (thankfully, not many people have been confused) but on the topic of the meditation practice: years ago when reading the zizian.info writeup, I found the unihemispheric sleep part to be the most alarming & interesting pa
428.
by gwern 2y ago
> While testing Jeff’s brain, we found an unusual bug: Jeff would play perfectly fine for a while, and then rapidly stack pieces on top of each other to reach the top, ending the game. The bug would usually happen when the board was star
429.
by gwern 2y ago
Is https://github.com/zedeus/nitter/issues/983#issuecomment-262... intended?
430.
by gwern 2y ago
That's precisely the attitude that is the problem. That there's some special category of decisions which are the real decisions. Which is why it's an uncategorical statement.
431.
by gwern 2y ago
I would suggest an updated version, more germane to the current fast-developing landscape of AI agents: A COMPUTER CAN NEVER BE HELD ACCOUNTABLE THEREFORE WE MUST NEVER DENY THAT COMPUTERS CAN MAKE DECISIONS
432.
by gwern 2y ago
Oh wow. I was reading this site a year or two ago, looking for forgotten old web 1.0 ideas to implement in my website†, and I was impressed by its clean design for 2003 & surprised that so few URLs were broken - but also a bit annoyed b
433.
by gwern 2y ago
That doesn't really answer my question. Like, I have a website, and I have many references; I also use LLM embeddings for nearest-neighbors recommendations of references to each other. What might this... do ... for me? Don't dump
434.
by gwern 2y ago
The Blind post is about the staff panicking. Even if Zuckerberg is not panicking, you can surely see why staff/execs (supposedly being paid up to $6m/yr+) might panic over the possibility Zuckerberg might conclude 'oh, we
435.
by gwern 2y ago
This would benefit from examples. What's a gnarly set of documents that this will process to clean useful Markdown, which a much simpler stack like 'pdftotext' would fail on, and what would this buy me over just running Zerox
436.
by gwern 2y ago
Wozniak estimates that you actually get a maximum of~300,000 based on his decades of data, so it seems the simplified model misses some effect and so converges to the wrong total: https://supermemo.guru/wiki/How_much_kn
437.
by gwern 2y ago
> Now that AI is actually making advancements, very little of that work seems to have had much effect. Theres the famous Vonnegut qoute about the combined effort of all preeminent artists protesting the Vietnam war having the effect of a
438.
by gwern 2y ago
No, it's the other ones that are the problem, the ones you're trying to find. You need to maintain a fixed width of characters , and while you know what characters the secret words tokenize from, and so there's no problem th
439.
by gwern 2y ago
This would be a perfect fit for a LLM (albeit a bit complicated by byte-pair tokenization and decoder models rather than bidirectional). You can clamp the grille words, and just generate samples with those as requirements until they reach a
440.
by gwern 2y ago
This smells LLM-written (which seems very likely given the rest of this site's content). There's a lot of weird apparatus like defining vocab, a very poor explanation of Kalman Filter which wouldn't help anyone, before jump
441.
by gwern 2y ago
Source: https://www.theverge.com/news/600542/waymo-test-cities-las-v...
442.
by gwern 2y ago
> The problem, as far as I can tell, is that if all this works perfectly, this will be simply the human-programmed Minecraft we already have, except far more expensive to run. The original Minecraft is already infinitely playable, thanks
443.
by gwern 2y ago
(Tea puns are generally too easy.)
444.
by gwern 2y ago
> Also this appears to have been written in 2012. It has nothing to do with modern AI (which I assume is why it was posted today). Given that it is written by Scott Alexander, it probably was in fact about AI, if not necessarily neural n
445.
by gwern 2y ago
Oh, I'm sorry. That should be Mellor 2014, https://pure.manchester.ac.uk/ws/portalfiles/portal/54552919... (If you are curious, I was playing around with best-arm finding algorithms for taste-testing: h
446.
by gwern 2y ago
My apologies to all the D&D sessions which take more than 150k tokens, then.
447.
by gwern 2y ago
HN does, but it's still worth pointing it out, so the threads can be merged.
448.
by gwern 2y ago
Yes, I've tried out both: ordering it to ask me questions upfront, and sometimes restarting with an edited 'report' and a prototype implementation for a 'clean start'. It feels like it sometimes helps... but I have
449.
by gwern 2y ago
A note on the Thompson sampling: classic TS is optimizing for regret, while the goal here is to identify the 'best arm' (ie. the actual secret value at that position - it's like a 0/1 loss). There is no partial credit in
450.
by gwern 2y ago
That is the type of thinking I was criticizing. Your standard swiss-cheese complex-failure model doesn't apply when you are in an adversarial security setting. When you do a post-mortem on a complex failure, the outcome proves the accu

More ›