ARC’s GPT-4 evaluation is cited in the FT article, in case that was ambiguous.
Hauke Hillebrandt
Karma: 202
Hauke Hillebrandt’s Shortform
Agreed, the initial announcement read like AI safety washing and more political action is needed, hence the call to action to improve this.
But read the taskforce leader’s op-ed:
He signed the pause AI petition.
He cites ARC’s GPT-4 evaluation and Lesswrong in his AI report which has a large section on safety.
“[Anthropic] has invested substantially in alignment, with 42 per cent of its team working on that area in 2021. But ultimately it is locked in the same race. For that reason, I would support significant regulation by governments and a practical plan to transform these companies into a Cern-like organisation. We are not powerless to slow down this race. If you work in government, hold hearings and ask AI leaders, under oath, about their timelines for developing God-like AGI. Ask for a complete record of the security issues they have discovered when testing current models. Ask for evidence that they understand how these systems work and their confidence in achieving alignment. Invite independent experts to the hearings to cross-examine these labs. [...] Until now, humans have remained a necessary part of the learning process that characterises progress in AI. At some point, someone will figure out how to cut us out of the loop, creating a God-like AI capable of infinite self-improvement. By then, it may be too late.”
Also the PM just tweeted about AI safety.
Generally, this development seems more robustly good and the path to a big policy win for AI safety seems clearer here than past efforts trying to control US AGI firms optimizing for profit. Timing also seems much better as things looks way more ‘on’ now. And again, even if the EV sign of the taskforce flips, then $125M is .5% of the $21B invested in AGI firms this year.
Are you saying that, as a rule, ~EAs should stay clear of policy for fear of tacit endorsement, which has caused harm and made damage control much harder and we suffer from cluelessness/clumsiness? Yes, ~EA involvement has in the past sometimes been bad, accelerated AI, and people got involved to get power for later leverage or damage control (cf. OpenAI), with uncertain outcomes (though not sure it’s all robustly bad—e.g. some say that RLHF was pretty overdetermined).
I agree though that ~EA policy pushing for mild accelerationism vs. harmful actors is less robust (cf. the CHIPs Act, which I heard a wonk call the most aggressive US foreign policy in 20 years), so would love to hear your more fleshed out push back on this—I remember reading somewhere recently that you’ve also had a major rethink recently vis-a-vis unintended consequences from EA work?
Ian Hogarth is leading the task force who’s on record saying that AGI could lead to “obsolescence or destruction of the human race” if there’s no regulation on the technology’s progress.
Matt Clifford is also advising the task force—on record having said the same thing and knows a lot about AI safety. He had Jess Whittlestone & Jack Clark on his podcast.
If mainstream AI safety is useful and doesn’t increase capabilities, then the taskforce and the $125M seem valuable.
If it improves capabilities, then it’s a drop in the bucket in terms of overall investment going into AI.
UK PM: $125M for AI safety
a large part of those ‘leaks’ are fake
Can you give concrete examples?
[Years of life lost due to C19]
A recent meta-analysis looks at C-19-related mortality by age groups in Europe and finds the following age distribution:
< 40: 0.1%
40-69: 12.8%
≥ 70: 84.8%
In this spreadsheet model I combine this data with Metaculus predictions to get at the years of life lost (YLLs) due to C19.
I find C19 might cause 6m − 87m YYLs (highly dependending on # of deaths). For comparison, substance abuse causes 13m, diarrhea causes 85m YLLs.
Countries often spend 1-3x GDP per capita to avert a DALY, and so the world might want to spend $2-8trn to avert C19 YYLs (could also be a rough proxy for the cost of C19).
One of the many simplifying assumptions of this model is that excludes disability caused by C19 - which might be severe.
Very good analysis.
I also thought your recent blog was excellent and think you should make it a top level post:
https://entersingularity.wordpress.com/2020/03/23/covid-19-vs-influenza/
Cheers—have taken this point out.
Cruise Ship passenger are a non random sample with perhaps higher co-morbidities.
The cruise ships analysed are non-random sample: “at least 25 other cruise ships have confirmed COVID-19 cases”
Being on a cruise ship might increase your risk because of dose response https://twitter.com/robinhanson/status/1242655704663691264
Onboard IFR. as 1.2% (0.38-2.7%) https://www.medrxiv.org/content/10.1101/2020.03.05.20031773v2
Ioannidis: “A whole country is not a ship.”
Thanks Pablo for your comment and helping to clarify this point. I’m sorry if I was being unclear.
I understand what you’re saying. However:
I realize that the Oxford study did not collect any new empirical data that in itself should cause us to update our views.
The authors make the assumption that the IFR is low and the virus is widespread and find that it fits the present data just as well as high IFR and low spread. But it does not mean that the model is merely theoretical: the authors do fit the data on the current epidemic.
This is not different from what the Imperial study does: the Imperial authors do not know the true IFR but just assuming a high one and see whether it fits the present data well.
But indeed, on a meta-level the Oxford study (not the modelling itself) is evidence in favor of low IFR. When experts believe something to be plausible then this too is evidence of a theory to be more likely to be true and we should update. An infinite number of models can explain any dataset and the authors only find these two plausible.
By coming out and suggesting that this is a plausible theory, especially by going to the media, the authors have gotten a lot of flag for this (“Irresponsible”—see twitter etc.). So they have indeed put their reputation on the line. This is despite the fact that the authors are prudent and saying that high IFR is also plausible and also fits the data.
Cheers- corrected.
It looks more like you listed all the evidence you could find for the theory and didn’t do anything else.
That was precisely my ambition here—as highlighted in the title (“The case for c19 being widespread”). I did not claim that this was an even-handed take. I wanted to consider the evidence for a theory that only very few smart people believe. I think such an exercise can often be useful.
I don’t think this is actually how selection effects work.
The professor acknowledges that there are problems with self-selection, but given that there are very specific symptoms (thousands of people with loss of smell), I don’t think that selection effects can describe all the the data. Then he just argues for the Central Limit Theorem.
That the asymptomatic rate isn’t all that high, and in at least one population where everybody could get a test, you don’t see a big fraction of the population testing positive.
There’s no random population wide testing antibody testing as of yet.
I do not think that can be used as decisive evidence to falsify wide-spread.
This is a non-random village in Italy, so of course, some villages in Italy will show very high mortality just by chance.
That region of Italy has high smoking rates, very bad air pollution, and the highest age structure outside of Japan.
By the end of its odyssey, a total of 712 of them tested positive, about a fifth.
Perhaps other on the ship had already cleared the virus and were asymptomatic. PCR only works for a week. Also there might have been false negatives. I disagree that the age and comorbidity structure can only lead to skewed results by a factor of two or three, because this assumes that there are few asymptomatic infections (I’m arguing here that the age tables are wrong).
In my post, I’ve argued why the data out of China might be wrong.
Iceland’s data might be wrong because it is based on PCR not serology, which means that many people might have already cleared the infection, and it is also not random.
That’s true and that’s what they were criticized for.
They argued that the current data we observe can be also be explained by low IFR and widespread infection. They called for widespread serological testing to see which hypothesis is correct.
If in the next few weeks we see high percentage of people with antibodies then it’s true.
In the meantime, I thought it might be interesting to see what other evidence there is for infection being widespread, which would suggest that IFR is low.
No. My ambition here was a bit simpler. I have presented a rough qualitative argument here that infection is already widespread and only a toy model. There are some issues with this and I haven’t done formal modelling. For instance, this would be what would be called the “crude IFR” I think , but the time lag adjusted IFR (~30 days from infection to death) might increase the death toll.
Currently, also every death in Italy where coronavirus is detected is recorded as a C19 death.
FWIW, if UK death toll will surpass 10,000, then this wouldn’t fit very well with this hypothesis here.
The point remains: given that some people have such a different theory, it’s unclear how many supporting pieces of evidence your should expect to see, and it’s important to compare the evidence against the theory to the evidence for it.
Yes, that’s what I’m trying to do here. I feel this is a neglected take and on the margin more people should think about whether this theory is true, given the stakes.
Presumably some of these people are hypochondriacs or have the flu? Also, I bet people with symptoms are more likely to use the app.
With all due respect it’s not that hard to get data that you yourself find convincing, even if you’re a professor.
“”Our first analysis showed we’re picking up roughly that one in 10 have the classical symptoms,” he said. “So of the 650,000, we would expect to see 65,000 cases.
“Although you can have problems of self-selection and bias, when you’ve got big data like this you tend to trust it more. What we’re seeing is a lot of mild symptoms, so I think having this data should help people relax a bit more and stop seeing it as an all or nothing Black Death situation.
“Other symptoms are cropping up. Thousands of people are coming forward to say they have loss of taste, and we may start to see clusters of symptoms.”″
They do meet more different populations of people though. So if a small number of cities have relatively widespread infection, people who visit many cities are unusually likely to get infected.
You’d expect to see people to many severe cases amongst people who travelled for business a lot in January and February.
Not likely. About 1% of Icelanders without symptoms test positive, and all the stats on which tested people are asymptomatic that I’ve seen (Iceland, Diamond Princess) give about 1⁄2 asymptomatic at time of testing (presumably many later get sick).
I don’t quite understand what you’re saying here.
I’m not impressed by the comment about this paper here on LW or the twitter link in it.
This paper was written by an international team of highly cited disease modellers who know about the Diamond Princess and have put their reputation on the line to make the case that this the hypothesis of high infections rate and low infection fatality might be true.
I think it is a realistic range that this many people are already infected and are asymptomatic. Above I’ve tried to summarize and review the relevant evidence that fits with this hypothesis.
But I’m not ruling out the more common theory (that we have maybe only 10x the 500k confirmed cases). I just find it less likely.
AI labs should escalate the frequency of tests for how capable their model is as they increase compute during training
Comments on the doc welcome.
Inspired by ideas from Lucius Bushnaq, David Manheim, Gavin Leech, but any errors are mine.
—
AI experts almost unanimously agree that AGI labs should pause the development process if sufficiently dangerous capabilities are detected. Compute, algorithms, and data, form the AI triad—the main inputs to produce better AI. AI models work by using compute to run algorithms that learn from data. AI progresses due to more compute, which doubles every 6 months; more data, which doubles every 15 months; and better algorithms, which half the need for compute every 9 months and data every 2 years.
And so, better AI algorithms and software are key to AI progress (they also increase the effective compute of all chips, whereas improving chip design only improves new chips.)
While so far, training the AI models like GPT-4 only costs ~$100M, most of the cost comes from running them as evidenced by OpenAI charging their millions of users $20/month with a cap on usage, which costs ~1 cent / 100 words.
And so, AI firms could train models with much more compute now and might develop dangerous capabilities.
We can more precisely measure and predict in advance how much compute we use to train a model in FLOPs. Compute is also more invariant vis-a-vis how much it will improve AI than are algorithms or data. We might be more surprised by how much effective compute we get from better / more data or better algorithms, software, RLHF, fine-tuning, or functionality (cf DeepLearning, transformers, etc.). AI firms increasingly guard their IP and by 2024, we will run out of public high-quality text data to improve AI. And so, AI firms like DeepMind will be at the frontier of developing the most capable AI.
To avoid discontinuous jumps in AI capabilities, they must never train AI with better algorithms, software, functionality, or data with a similar amount of compute than what we used previously; rather, they should use much less compute first, pause the training, and compare how much better the model got in terms of loss and capabilities compared to the previous frontier model.
Say we train a model using better data using much less compute than we used for the last training run. If the model is surprisingly better during a pause and evaluation at an earlier stage than the previous frontier model trained with a worse algorithm at an earlier stage, it means there will be discontinuous jumps in capabilities ahead, and we must stop the training. A software to this should be freely available to warn anyone training AI, as well as implemented server-side cryptographically so that researchers don’t have to worry about their IP, and policymakers should force everyone to implement it.
There are two kinds of performance/capabilities metrics:
Upstream info-theoretic: Perplexity / cross entropy / bits-per-character. Cheap.
Downstream noisy measures of actual capabilities: like MMLU, ARC, SuperGLUE, Big Bench. Costly.
AGI labs might already measure upstream capabilities as it is cheap to measure. But so far, no one is running downstream capability tests mid-training run, and we should subsidize and enforce such tests. Researchers should formalize and algorithmitize these tests and show how reliably they can be proxied with upstream measures. They should also develop a bootstrapping protocol analogous to ALBA, which has the current frontier LLM evaluate the downstream capabilities of a new model during training.
Of course, if you look at deep double descent (‘Where Bigger Models and More Data Hurt’), inverse scaling laws, etc., capabilities emerge far later in the training process. Looking at graphs of performance / loss over the training period, one might not know until halfway through (the eventually decided cutoff for training, which might itself be decided during the process,) that it’s doing much better than previous approaches- and it could look worse early on. Cross-entropy loss improves even for small models, while downstream metrics remain poor. This suggests that downstream metrics can mask improvements in log-likelihood. This analysis doesn’t explain why downstream metrics emerge or how to predict when they will occur. More research is needed to understand how scale unlocks emergent abilities and to predict. Moreover, some argue that emergent behavior is independent of how granular a downstreams evaluation metrics is (e.g. if it uses an exact string match instead of another evaluation metric that awards partial credit), these results were only tested every order of magnitude FLOPs.
And so, during training, as we increase the compute used, we must escalate the frequency of automated checks as the model approaches the performance of the previous frontier models (e.g. exponentially shorten the testing intervals after 10^22 FLOPs). We must automatically stop the training well before the model is predicted to reach the capabilities of the previous frontier model, so that we do not far surpass it. Alternatively, one could autostop training when it seems on track to reach the level of ability / accuracy of the previous models, to evaluate what the trajectory at that point looks like.
Figure from: ‘Adjacent plots for error rate and cross-entropy loss on three emergent generative tasks in BIG-Bench for LaMDA. We show error rate for both greedy decoding (T = 0) as well as random sampling (T = 1). Error rate is (1- exact match score) for modified arithmetic and word unscramble, and (1- BLEU score) for IPA transliterate.’
Figure from: ‘Adjacent plots for error rate, cross-entropy loss, and log probabilities of correct and incorrect responses on three classification tasks on BIG-Bench that we consider to demonstrate emergent abilities. Logical arguments only has 32 samples, which may contribute to noise. Error rate is (1- accuracy).’