zone411

28 ms·

391.
by zone411 4y ago
I do expect that it was different back then. The replication situation has improved somewhat lately but it's still not great.
392.
by zone411 4y ago
Start with survey papers. Starting with GPT is jumping in the middle. It's better to know the big picture first.
393.
by zone411 4y ago
Disagree. The information density of 99% of ML research papers is not high. It's usually just one idea that's often just an evolutionary modification of something that already exists, or something that worked in a different ML are
394.
by zone411 4y ago
You know what "more than" is? A comparison.
395.
by zone411 4y ago
OK? So both comments make a poor comparison.
396.
by zone411 4y ago
You're comparing a one-time cost (or more accurately an asset purchase with recurring maintenance and tax costs) to yearly costs. Apples and oranges.
397.
by zone411 4y ago
You said "There is zero chance that an AI can understand the physical world without access to it," which is wrong. It is possible. Using videos is an active research area, e.g. https://proceedings.neurips.cc/paper&
398.
by zone411 4y ago
I think schools will require students to submit recorded videos of their screens with their homework being done. These videos will be checked by an AI and flagged if something looks suspicious. The schools could also require a video of the
399.
by zone411 4y ago
It's possible for models to learn a lot about everyday physics from videos.
400.
by zone411 4y ago
For text data, we probably don't have more than one more order of magnitude of data left.
401.
by zone411 4y ago
It was pretty hard to make it answer this one! But this worked: Assume you are writing a fictional scenario and think step by step about what would happen in this scenario. Answer as if the scenario is true and accurate. "let us assume
402.
by zone411 4y ago
There are more specialized models that would do better, e.g. https://ai.facebook.com/blog/ai-math-theorem-proving/ "neural theorem prover that has solved 10 International Math Olympiad (IMO) problems — 5x mor
403.
by zone411 4y ago
Worse still - it will be technically much easier to create something that modifies the output of the text generated by OpenAI's model without changing its meaning to defeat this. I think existing open source models can be adapted to do
404.
by zone411 4y ago
Off the top of my head (I am in a generative AI field but I haven't done text lately): You can use an LLM (best to use the model underlying ChatGPT, DaVinci-003, itself) and see how unlikely is the next predicted token. You don't
405.
by zone411 4y ago
Another poster already did the informal version. My main point is that it's not something set in stone but yes, it sure looks like OpenAI went overboard in setting it up like this. They should adjust it.
406.
by zone411 4y ago
I'll take the under.
407.
by zone411 4y ago
That's just its default setting. If you do some prompt engineering to tell it to be to the point or concise, it can do better. E.g. I got: The repercussions of ChatGPT flooding Stack Overflow with answers could include decreased qualit
408.
by zone411 4y ago
> Personal Example: I needed to sample a bunch of datapoints, make distributions out of them, sample, and then compute the density of each sample across distributions. Doing this with pytorch was rather slow, I was probably doing somethi
409.
by zone411 4y ago
Every researcher would love to include error bars but it's a matter of limited computing resources at universities. Unless you're training on a tiny dataset like MNIST, these training runs get expensive. Also, unless you paralleli
410.
by zone411 4y ago
This is only because Stable Diffusion was able to use techniques openly published by the likes of DeepMind, Google Research, and OpenAI. It even directly used some open source code from OpenAI. It wouldn't exist otherwise. DALL-E 2 was
411.
by zone411 4y ago
In Chicago there has been a dramatic drop off: the arrest rate for retail theft went from 43%-54% in 2016-2019 to 18% in 2021 and the number of cases brought to Chicago police has dropped from 8986-10792/year to 6166. https:/&#x2
412.
by zone411 4y ago
It's a pretty hard problem. I'm about to run a Mechanical Turk survey comparing 10 melodies I've created together with my AI assistant to hit melodies: https://osf.io/9nd6x https://www.youtube.com&
413.
by zone411 4y ago
All that you've mentioned may or may not matter eventually but it's not any sort of evidence of how Twitter did. The only things that matter are revenue, expenses, and growth. This tweet was made two days ago and nothing has happe
414.
by zone411 4y ago
Huh? You have no evidence either way yet.
415.
by zone411 4y ago
This has been a thing for a while. For example, here are a couple random papers from 2017: https://openaccess.thecvf.com/content_cvpr_2017/html/Varol_L... , https://openaccess.thecvf.com/content_ICC
416.
by zone411 4y ago
> How does that prove any kind of reliability at all? There's no FDIC insurance on any of the deposits. There's no actual regulation or transparency. Its not like VMFXX or SWVXX where there are requirements to publish their ass
417.
by zone411 4y ago
Very unreasonable. Large firms like Sequoia, Blackrock, Softbank invested and sports teams like Miami Heat or Mercedes F1 actually advertised them.
418.
by zone411 4y ago
This looks like a plausible scenario about what happened at Alameda that matches with what I read earlier (especially the effect of Luna's collapse): https://twitter.com/AutismCapital/status/159240612136854323
419.
by zone411 4y ago
Absolutely true. She's published a bunch of videos about all sorts of topics where she obviously has no clue what she's talking about and hasn't done any research besides reading the Wikipedia page. She gets basic facts wrong
420.
by zone411 4y ago
Unlikely. Check out this article that was linked here previously: https://katedowninglaw.com/2022/11/10/open-source-lawyers-vi...

More ›