All of Simon Berens's Comments + Replies
As a counterpoint, Sydney showed aligning these models on the first go, and even discovering unsafe behavior is non-trivial.
The flip side of this is that it helps to speak in a way that’s almost repetitive to make sure someone who zoned out can catch back up.
For example, frequently once the subject/object of a sentence is known, people refer to the subject/object with pronouns instead of repeating the subject/object. If someone misses the initial definition, they are immediately lost.
(This also applies somewhat to writing — above, I’m repeating “subject/object” instead of “it”. People also zone out during reading, and this repetition saves them from having to scroll back up and figure out what a pronoun is referencing.)
Did you feel a subjective increase in your intelligence? E.g. feeling like you’re thinking faster, more clearly, having a better memory?
Cool project!
I’m curious—what does the long tail of websites look like for you? For me, it’s the small number of sites that i repeatedly go to (twitter, youtube, hackernews, etc…) that take up the vast majority of my wasted time.
(Btw, I also built my own website blocker: https://chrome.google.com/webstore/detail/webblock/jeahkphmdfbddenabgndnooheiciocka)
I think the main beneficiaries of being able to sideload apps will be incumbents, not startups. Big companies like Spotify, Netflix, and Tinder will offer users discounts if they sideload because it will spare them the 30% Apple tax.
I am confused how to square your claim of requesting extra time for incontrovertible proof, with Ben’s claim that he had a 3 hour call with you and sent the summary to Emerson, who then replied “good summary!”
Was Emerson’s full reply something like, “Good summary! We have incontrovertible proof disproving the claims made against us, please allow us one week to provide it?”
1Yes, Ben took Emerson’s full email out of context, implying that Emerson was fully satisfied when in actuality, Emerson was saying, no, there is more to discuss - so much that we’d need a week to organize it.
He got multiple extremely key things wrong in that summary and was also missing key points we discussed on the call, but we figured there would be no reason he wouldn’t give us a week to clear everything up. Especially since he had been working on it for months.
3Bloomberg reported 2 weeks ago that Twitter resumed paying Google Cloud: https://www.bloomberg.com/news/articles/2023-06-21/twitter-resumes-paying-google-cloud-patching-up-relationship
I agree AI should be regulated, but lumping social media in with AI only weakens your argument.
Also, is it just me or does this writing style feel like a NYT op-ed?
You might want to clarify that, because in the post you explicitly say things like “if your goal is to predict the logits layer, then you should probably learn about Shakespearean dramas, Early Modern English, and the politics of the Late Roman Republic.”
This is probably obvious, but maybe still worth mentioning:
It’s important to take into account the ROI per unit time. In the amount of time it would take for me to grok transformers (let’s say 100 hours), I could read ~1 million tokens, which is ~0.0002% of the training set of GPT3.
The curves aren’t clear to me, but i would bet grokking transformers would be more effective than a 0.0002% increase in training set knowledge.
This might change if you only want to predict GPT’s output in certain scenarios.
I think you would get diminishing returns but reading a few hundred thousand tokens would teach you quite a lot, and I think likely more than knowing Transformers would. I'm not convinced that Transformers are all that important (architectures seem to converge at scale, you can remove attention entirely without much damage, not much of the FLOPS is self-attention at this point, etc), but you learn a lot about why GPT-3 is the way it is if you pay attention to the data. For example, BPEs/poetry/puns: you will struggle in vain to explain the patterns of GPT ...
I don't think researchers should learn world-facts in order to understand GPT-4.
I think that (1) researchers should use the world-facts they already know (but are actively suppressing due to learned vibe-obliviousness) to predict/explain/control GPT-4, and (2) researchers should consult a domain expert if they want to predict/explain/control GPT-4's output on a particular prompt.
I agree that recursive self-improvement can be very very bad; in this post I meant to show that we can get less-bad-but-still-bad behavior from only (LLM, REPL) combinations.
Yeah, this isn't something I have an ugh field around, but having portable versions of travel stuff like shampoo, skincare, and chargers ready to go is nice.
Awesome! What GitHub integration are you talking about?
I think the hackernews comment section, though still somewhat emotionally charged, is of substantially better quality.
Also, I responded to some comments/questions there.
This is awesome! I highly encourage you to write up your experience; I think this should be more normalized!
I didn’t feel a difference; I guess because it was my iron reserves that were low, not my actual iron levels.
But yeah, if it does continue to be a problem I will do something like that.
Sure, but my point was I don't know what dose per unit time to take.
Whoops, I meant year. Edited.
A relevant experience of mine: after being vegan for ~1 year, I got my blood tested to see if there was anything I should watch out for. It turned out that my iron reserves were low, so I was told to take iron supplements. One year later I got another blood test where my iron levels came back as too high. My doctor was confused until I told him I still sporadically took iron supplements; he told me to stop immediately, and that as a male I didn't need to supplement iron much.
Now I feel like I'm in a damned-if-I-do-damned-if-I-don't limbo; it would be great to have some sort of framework of when/if to take iron supplements.
I think a wrench and 20 minutes of time is (1) a higher barrier than you think (not to say it justifies the lack of bidet) and (2) is an underestimate of how inconvenient installing it is. I installed a bidet a few years ago and it probably took 2+ hours and ended up leaking. When I moved, I opted not to install one again and instead follow #2 with a shower.
However, I agree bidets are great modulo the installation, and I will try to buy a toilet with one pre-installed.
I actually had that as an example in a draft, but my friends told me it didn't quite fit the mold so I deleted it :)
Thanks! Here is the extension I made: https://chrome.google.com/webstore/detail/webblock/jeahkphmdfbddenabgndnooheiciocka
Admittedly, I copped out on the UX in a few places (motivation section, link parsing, etc...) but better to publish than wait for perfection!
Edit:
Right after I published my extension, I found these similar tools:
...Aha upon further searching I came across this post: https://www.lesswrong.com/posts/HaESgyu2BqDMTpujS/a-better-web-is-coming
It's not the post I remember reading but it's definitely about the same same topic so ¯\_(ツ)_/¯
I shouldn't have used a company for my example because as you illustrated that's not a very interesting case; if it wasn't a company but the government (e.g. the TSA), does economic theory say those jobs will come back in a more efficient form?
How did you meet your roommates? I would like to surround myself with similar people (perhaps less extreme).
Re twitter’s profitability, Musk about doubled EBITDA despite revenue halving, i.e. he more than tripled EBITDA margin
https://www.teslarati.com/elon-musk-x-doubled-ebitda-since-2022-takeover-report/amp