That’s all fair, and given what has been said, despite my initial impression, I don’t think this was “obviously wrong”—but I do have a hope that in this community, especially in acknowledged edge cases, people wait and check with others rather than going ahead.
Davidmanheim
Karma: 2,762
On the topic of growth rate of computing power, it’s worth noting that we expect the model which experts have to be somewhat more complex that what we represented as “Moore’s law through year ”—but as with the simplification regarding CPU/GPU/ASIC compute, I’m unsure how much this is really a crux for anyone about the timing for AGI.
I would be very interested to hear from anyone who said, for example, “I would expect AGI by 2035 if Moore’s law continues, but I expect it to end before 2030, and it will therefore likely take until 2050 to reach HLMI/AGI.”
I don’t have a very good (or even halfway decent) memory for phrases, so I have no idea, and since no-one else heard it, I assume it wasn’t said. Still, it seemed clear to me that the request was intended for the talk to be off the record, in the journalistic sense.
The phrase “no recording and no transcript,” which you seem to agree was said explicitly, seems to indicate that he didn’t want there to be a record of what he said. At that point, maybe you didn’t technically do anything he requested you not to do, but it seems like the responsible and decent thing would be to have asked Sam if he minded.
I don’t recall what phrase was used, but I thought that it was clear enough. If someone said that they agree to do a talk on the condition that there be no recording and no transcript, unlike every other talk in the series, it seems to take a really weird model of the situation to claim that you had no idea that they would not want people publicly posting notes. At the very least, it merits checking.
Sam has said he thought this was “off-the-record-ish” and it was clearly known that it not being recorded was a precondition for giving the talk. I don’t recall what terms were used, but I thought it was pretty obvious—and Sam’s later responses seem to agree—that he expected notes like this not to be made public.
Edit to add: I thought at the time that it was clear that this was off the record, despite that phrase likely not being used. If not, I would not have asked the question which I asked during the meetup.
The request for this to be off the record was explicit during the introduction to the talk, so I’m not sure why it’s ambiguous. And “off the record” has a pretty clear meaning—I certainly had the clear expectation that my question, and his answer, weren’t going to be published.
Edit: I do not recall the phrasing, but as I said below, I was under a distinct impression that the request for no recording and no transcript was at least indicative, and that asking him if you could share notes publicly would have been the right thing to do.
My simple question is whether he checked with Sam Altman and/or Joshua Fox about whether posting notes of this explicitly off the record talk, as mentioned during the talk, was OK before posting—and as far as I can tell, he did not.
It depends on the antibody test—some will not register anything for vaccinated people, others will register the vaccine reaction (I’m not sure which is which.) Either way, these results are not comparable to the reaction you have from getting COVID-19, so any comparison you make is ad-hoc, and you’re not going to be able to infer how good or bad your post-vaccine immunity is this way, especially with sample size n=1.
I mostly agree, but we get into the details of how we expect improvements can occur much more in the upcoming posts on paths to HLMI and takeoff speeds.
Modelling Transformative AI Risks (MTAIR) Project: Introduction
Ahh. I was aware that there was an issue, but I definitely didn’t understand clearly enough what the issue was. I assumed that the larger number of tokens would allow better splitting, but the way you explain it, this wouldn’t help. (And given that, I wonder what good tokenizer development looks like—because I presume it’s hard to optimize alongside the model itself.)
The key advance here seems to be the tokenizer, with larger vocabulary, which has been identified by others as a potentially critical limitation for GPT-3. I’d be very interested in seeing its performance on multi-digit addition tasks, for example.
I would be very interested in seeing how well this model works as a drop-in replacement for GPT-3 in various applications, both because it would undermine the market value of building AI systems which can be duplicated by others, and because it would say something about how flexible the architectures built around AI systems are to improvements.
This is unfortunately defunct, replaced by another site on a different topic.
It’s not true that the only protection lost would be from asymptomatic people, though that would still be a big deal if a quarter of cases are asymptomatic and R is above 4, which it likely is in a population taking no other precautions. And even without masks, people who actively feel very sick often seek treatment and are diagnosed, and when not, mostly aren’t going out in public much. But there are two other groups that matter;
Presymptomatic spread is a big deal for COVID, and accounts for much of why it spreads quickly. That’s why we saw such short serial transmission intervals. And if you don’t eliminate the rapid spread, you’re not getting much benefit from masks.
Paucisymptomatic people, who have a slight runny nose or temperature and nothing else, are fairly common, might not notice, or will assume it’s not COVID, since it’s mild, and spread the virus. (And this category partly overlaps with the previous one—people often start manifesting minor symptoms before they notice all of them.)
Agreement reached!
Mechanistic explanations are good for priors, but don’t replace, much less refute, empirical evidence. If we see that there is ~0% impact in RCTs, the fact that we know better because it “must” decrease transmission isn’t relevant. And the mechanistic model could be wrong in many ways. For example, maybe people wear masks too poorly to matter (they do!) Maybe masks only help if people never take them off to blow their nose, or scratch it, or similar (they do all those things, too.)
And we see that even according to the paper, the impact is pretty small, so it mostly refutes the claim made by the mechanistic model you propose—the impact just isn’t that big, for various reasons. Which would imply that there is no way to know if it’s materially above 0%.
In fact, I think it’s likely that the impact is non-trivial in reducing transmission, but the OP is right that we don’t have strong evidence.
Based on this approach, optimal allocation for equities for younger folks is probably well over 100% - and this isn’t particularly complicated to do, contra the statements in the article. Long dated out-of-the-money stock index options are a viable retirement investment. I’d tell people to seriously consider buying out of the money calls. As an illustrative example, a 120% of future price once a year for 2-3 years away with, say, 5% of your portfolio.
BUT—warning to readers: If you don’t know / understand the argument I’m making, please don’t just go buy stock options. Certainly don’t spend more than a small portion of your long-term savings on them!
I think we agree—I’m certainly in favor of massive investments in surveillance and in PPE. The key question was whether I was missing something in the push for vaccines and antivirals, as if both were similarly promising.
I guess I could have cited more data on the claim that antivirals work poorly—but I wasn’t trying to write an academic paper, and I don’t think you cited anything that refutes my point.
You seem unconvinced about how much this generalizes, so in addition to the obvious relative lack of efficacy for HIV, noted earlier, it might be somewhat useful to note that, AFAIK, the entire set of viral diseases we have antivirals for is HIV, HPV, Flu, Hep-B and C, and various herpesviruses (HCMV, HSV, VZV,) and that most of these (HIV, Hepatitis, and the herpesviruses,) seem to be used mainly to treat chronic disease by reducing viral load, rather than cure the disease, and the the remainder aren’t particularly effective as cures.
Some, in fact, only seem to work in studies funded by their manufacturers. You, and others, claim that Neuraminidase inhibitors like tamiflu seem to work. Some people, like the people who wrote the Cochrane review, disagree. That’s fine—evidently you know lots about this, and I only looked into it briefly, though the evidence seems at best shaky to me. And I’m not going to try to convince you, or write a paper on this. But I was asking for feedback and corrections, so thanks.
I agree that in general there is a tradeoff, and that there will always be edge cases. But in this case, I think judgement should be tilted strongly in favor of discretion. That’s because a high trust environment is characterized by people being more cautious around public disclosure and openness. Similarly, low trust environments have higher costs of communication internal to the community, due to lack of willingness to interact or share information. Given the domain discussed, and the importance of collaboration between key actors in AI safety, I’m by default in favor of putting value more on higher trust and less disclosure than on higher transparency and more sharing.