3,655 karma · joined August 8, 2021
I feel this is actually somewhat straightforward. I assume deep water on roadways is not commonly in the training set, because frankly it isn't common in real life, and when it is common people do not drive and do not gather that training data. As a result the proper response has not adequately been beaten into the models. There are probably also challenges of world-sensing, since water can act as a mirror, and maybe other complications. So waymos are bad at handling deep water on roadways. However, deep water on roadways is also not common in the areas where waymos are deployed. As a result, waymo's have a great safety record, and at the same time they make mistakes that are obvious to a human.
A common criticism of AI discourse is that people act as if LLM's "think". I don't want to be a vocabulary purist, but I suspect that's related to the astonishment here -- the Waymo doesn't know what flooding is, it doesn't fear drowning, it doesn't think. So unless it's been repeatedly trained, or a special case has been hard coded by manual effort, it doesn't know that flooded roadways are dangerous.
I have made a lot of assumptions here, and I don't truthfully know what the training data looks like. Feel free to push back if you think my assumptions are wrong. I'd especially be interested if somebody can show that water on roadways _is_ in the training data
Assume that n% of foundational studies are demonstrably incorrect if replication were attempted. Then if you replicate X foundational studies, there's a 1 - (1-n/100)^X chance that you'll show one foundational study to be incorrect. Then I think you just need to find an advisor who believes the incorrectness rate n, and the replication attempts X is high enough for you to revert a foundational study over the course of your PhD.
Lol, I appreciate the sprinkling of well timed humor in an article that is mostly straightforward facts and analysis.
[0] https://www.youtube.com/watch?v=x3uoB8zvLdI
edit: just watched the youtube video in full, and it isn't an admission. It's just discussing the shooting in general, and he puts blame on others.
> Nearly a year of "full time" work at 40hrs/week went into this
Sorry, I think you may be mistaken? From the section you quote, it looks like it's about 370 hours, or 9 weeks of full time effort. That's still a lot, but certainly not nearly a year.
They were just putting it off.
Do I smell a cursor-analogue from openAI?
[0] https://www.klfy.com/louisiana/interactive-map-shows-which-p...
tl;dw -- louisiana has a ton of natural resources and wonderful harbors, but has allowed corporations to get away with almost anything without paying their fair share of taxes, cleaning up their waste, or generally being held accountable. And likely related, there is a long history of rampant corruption in the area. With sea levels rising and cancerous pollution filling densely populated areas, life in Louisiana has generally become very difficult :(
I think this is a good candidate to hang on the wall, or perhaps for a tattoo.
This is such a ridiculous, preposterously self-confident claim. Claims like this give HN (and engineers in general) a bad reputation.
Edit: relevant xkcd https://xkcd.com/793/
I appreciate the backwards compatibility they highlight in the article, and more generally the ownership oss gives you over your files. My small business uses gimp, canva, and Inkscape, and they all have their purpose, but one of the complaints we have with canva is that it’s really hard to get the original files backed up somewhere, aside from canva itself. You have to download each file individually , which is quite a pain (please correct me if there’s an easier way!). Gimps openness is an obvious path for oss, but really is a great feature. If the gimp team stops, we’ll still be able to edit our gimp files. If the canva team stops, we mostly lose all the canva stuff.
> A lot of new and existing contributors have submitted improvements to GIMP’s user interface and its user experience. We wanted to highlight their efforts, and encourage you all to continue sharing your feedback on our design issue tracker.
I see no active resistance. If you have feedback, maybe share it with them?
They claim that spacex has competitive advantage. They take no stance on condoning spacex’s behavior. Personally I strongly dislike musks behavior, but I appreciate discussion of competitive advantages/disadvantages, independent of moral views. I liked the thread, it adds new info to the conversation.
1. Human reward-seeking becomes longer-term, with less of an emphasis on short term rewards. I'd bet this trend has probably already been ongoing for a long time. But I wonder if there are some negative effects of this -- for example perhaps it takes us longer to learn things that truly do have very quick, direct feedback loops. Maybe it takes us longer to learn how to walk, swim, etc as a result.
2. Screen-avoidant communities start to dominate the human population. Eg, Amish communities, "traditionalist" communities like that, your friends that religiously don't use social media, etc. I'd bet this is already happening too! Would be really interesting to look at the proportion of new births, perhaps controlling for usage of birth control. I wonder if this effect could result in humanity generally becoming more religious (as opposed to the recent trend of becoming less religious). Admittedly, this is more of a cultural change than a change in our dna, but I feel it fits the category -- it's a widespread change across humanity in response to the pressure of distracting apps.
3. Humans grow a stronger distaste for blue light. I suspect this wouldn't be the "end" state, because if it happens we'll probably just start buliding devices that put off less blue light.
This is just like animals that didn’t run from humans — they didn’t know to be scared of us and now they’re dead. Even massive things that could crush us can be killed if we slip through their defenses. The ones that survive are the ones that learn to fear the dangerous thing.
This is also an absolutely great way to ensure the participants really understand what they’re saying, even aside from ai-related concerns. You could always pull definitions straight from textbooks, and write something which mostly makes sense but that you don’t really understand. But you really can’t teach something unless you do understand it, imo. The best command I ever got on a topic was due to TA’ing