Hacker News new | past | comments | ask | show | jobs | submit | pietmichal's comments login

Pardon my ignorance but couldn't this also be an act of anthropomorphisation on human part?

If an LLM generates tokens after "What do you call someone who studies the stars?" doesn't it mean that those existing tokens in the prompt already adjusted the probabilities of the next token to be "an" because it is very close to earlier tokens due to training data? The token "an" skews the probability of the next token further to be "astronomer". Rinse and repeat.


I think the question is: by what mechanism does it adjust up the probability of the token "an"? Of course, the reason it has learned to do this is that it saw this in training data. But it needs to learn circuits which actually perform that adjustment.

In principle, you could imagine trying to memorize a massive number of cases. But that becomes very hard! (And it makes predictions, for example, would it fail to predict "an" if I asked about astronomer in a more indirect way?)

But the good news is we no longer need to speculate about things like this. We can just look at the mechanisms! We didn't publish an attribution graph for this astronomer example, but I've looked at it, and there is an astronomer feature that drives "an".

We did publish a more sophisticated "poetry planning" example in our paper, along with pretty rigorous intervention experiments validating it. The poetry planning is actually much more impressive planning than this! I'd encourage you to read the example (and even interact with the graphs to verify what we say!). https://transformer-circuits.pub/2025/attribution-graphs/bio...

One question you might ask is why does the model learn this "planning" strategy, rather than just trying to memorize lots of cases? I think the answer is that, at some point, a circuit anticipating the next word, or the word at the end of the next line, actually becomes simpler and easier to learn than memorizing tens of thousands of disparate cases.


I came to similar conclusion after pouring 95% of brain power for 8 years as an employee at a startup.

Make yourself visible in professional and social life.


Quick feedback: The "Polypane Portal" page hangs after opening it in Chrome on M1 Pro Macbook.


Can't seem to replicate that here (m4 macbook), but I'll keep it open for a bit and see what happens. Thanks for letting me know!


Scrolling is unusable on that page for me:

Computer: Apple M3 Max

macOS: 14.7.3 (23H417)

Google Chrome 134.0.6998.118

I also tried in my "clean" chrome profile (to rule out extensions) and it's still got really bad scroll lag. This happens as soon as I open the page.

Here is a video though I understand it's hard to convey since you can't see when/how much I'm scrolling. I can tell you I scrolled slowly down and back up consistently through this video.

https://cs.joshstrange.com/m71YtZdk

Even worse, I just found that having that tab open (and visible) makes Chrome (no other app) laggy everywhere. Something is definitely wrong with that page. Also that page was open in a different chrome profile and it still made my main chrome profile lag when just trying to click around the text area for this comment on HN.

Edit: Some extra details for my setup, I have external monitors (4) and the Macbook Pro is closed in clamshell mode. Not sure why either of those things would matter but I figure both those cases are more common for people on HN (external monitors/closed laptop) than the general public so I wanted to mention it.

Edit2: And here is a video of me doing the same scrolling on the homepage just to show the difference: https://cs.joshstrange.com/GWLr0Qwl


Can confirm, M1 Air 16 GB, Brave and macOS are up-to-date


M1 Pro here too, the page doesn't hang, but it's extremely slow to scroll on my end.


M1 Max MacBook Pro and no issues, fwiw. Just to help provide data.


M1 Mac mini here and instant loading in Brave.


It should be better now, as I made tweaks to the starfield canvas at the top of the page.


Have any extensions installed that may be blocking requests?


same


This sentiment reminds me of my teachers who complained that Wikipedia is not a reliable source to learn from.


Wikipedia and LLMs are completely different entities that work in completely different ways.


And yet, much of the same criticism applies to both.


Wikipedia is generally right and you can check it's sources. Also everything there was written by a human.

LLMs are often wrong and you cannot check their sources. Also since it is generated you can trick it into spitting out falsehoods intentionally.

They are not even close.


You can also check the sources of LLMs, just ask them for it, and then check that. An LLM is simply more flexible and more powerful than Wikipedia, and thus you have to be more cautious with regards to its results.

"Generally right" is not the same as "reliably right", and therefore if you really need to rely on a fact for something important, I would trust neither Wikipedia nor LLMs.


> LLMs are often wrong and you cannot check their sources.

Depends on which LLM you are using, perplexity and copilot both cite their sources.


Wikipedia is a marvel of the commons and did wonders for education. hostility to it from academia was surprising


Don't worry, WebKit devs will update Quirks.cpp soon!


Give Performance-Aware Programming series by Casey Muratori a try!

https://www.computerenhance.com/p/table-of-contents


I honestly don't understand how delusional you have to be to not think OpenAI wanted this to happen.


In London, it is normal to pay more for the coffee when you are staying in.


This was such a nice primer that inspired me to give Karpathy's series another try. Loved the explanation!


Very interesting! Apparently a lot of information on the Internet about the Jai language is outdated which makes me curious about the follow up post!

I don't know if OP is the blog owner. If yes, I'd love to have an option to subscribe to an email list. I've installed an RSS feed but email is still my preference. Thanks!


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: