Bayesian Investor Blog

Scott Alexander graded his predictions from 2018 and made new predictions for 2028.

I’m trying to compete with him. I’m grading myself as having done a bit worse than Scott.

Here’s a list of how I did (skipping a few where I agreed with Scott), followed by some predictions for 2028.

If AI translation becomes flawless outstanding, we will hear how language is just a formal system that can be brute-forced without understanding.

I predict that, regardless of whether AI translation is much better or slightly worse than human translation, people who want to deny AI is approaching human-level intelligence will focus on evidence that the AI sometimes makes dumb mistakes. Only a tiny fraction will talk about whether translation is a good test of intelligence/understanding.

Technological unemployment will continue to be a topic of academic debate that might show up if you crunch the numbers just right, but there will be no obvious sign that it is happening on a large scale. Everyone will tell me I am wrong about this, but I will be right, and they will just be interpreting other things … as obvious visible signs of technological unemployment, the same as people do now.

I’ll be more specific, and predict a 70% chance of a recession that causes reported US unemployment to exceed 6%. Roughly half of the people who comment on the causes will attribute that to technological causes. I will mostly agree with Scott Sumner that if the Fed had been more willing to increase the money supply quickly, unemployment could have been kept under 5.5% (Sumner will probably believe it could have been kept below 5%). Yet I also predict an 80% chance that I’ll believe technological unemployment is accelerating.

I was very wrong about the causes of unemployment in the 2020 recession.

1. Average person can hail a self-driving car in at least one US city: 80%
2. …in at least five of ten largest US cities: 30%
3. At least 5% of US truck drivers have been replaced by self-driving trucks: 10%
4. Average person can buy a self-driving car for less than $100,000: 30%
5. AI beats a top human player at Starcraft: 70%
6. MIRI still exists in 2023: 80%
7. AI risk as a field subjectively feels more/same/less widely accepted than today: 50%/40%/10%

1. 80%
2. 40%
3. 60%
4. 60%
5. 70%
6. 90%
7. 60%/30%/10%

I did much worse than Scott on 3 and 4, with little difference on the rest.

1. UK leaves EU (or still on track to do so): 95%
2. No “far-right” party in power (executive or legislative) in any of France, Germany, UK, Italy, Netherlands, Sweden, at any time: 50%
3. No other country currently in EU votes to leave: 50%

1. 90%
2. 70%
3. 60%

I did slightly worse than Scott here.

1. Trump wins 2020: 20%
2. Republicans win Presidency in 2020: 40%

1. 18%
2. 52%

I expect the timing of the next recession will be the main determinant of which party wins in 2020. I’m guessing that the recession will start right around the time of the 2020 election. If it’s earlier than I expect, the Democrats should win, otherwise the Republicans should win. Predicting recessions that far away is the kind of thing where I’m less than 50% confident of getting it right to the nearest year, so who knows?

1. Sanders wins 2020: 10%
2. Democrats win Presidency in 2020: 60%

1. 4%
2. 48%

I did a bit worse than Scott here. I overestimated the chances of Trump choosing not to run (due to not liking the job or due to health problems). But my reasoning about a recession still looks pretty good.

It will become more and more apparent that there are three separate groups: progressives, conservatives, and neoliberals.

I expect something other than this to happen. Most likely it won’t become more apparent that there are three political groups.

I beat Scott here.

1. At least one US state has approved single-payer health-care by 2023: 70%
2. At least one US state has de facto decriminalized hallucinogens: 20%
3. At least one US state has seceded (de jure or de facto): 1%
4. At least 10 members of 2022 Congress from neither Dems or GOP: 1%
5. US in at least new one major war (death toll of 1000+ US soldiers): 40%
6. Roe v. Wade substantially overturned: 1%
7. At least one major (Obamacare-level) federal health care reform bill passed: 20%
8. At least one major (Brady Act level) federal gun control bill passed: 20%
9. Marijuana legal on the federal level (states can still ban): 40%
10. Neoliberals will be mostly Democrat/evenly split/Republican in 2023: 60%/20%/20%
11. Political polarization will be worse/the same/better in 2023: 50%/30%/20%

1. 60% (I’m tempted to give a lower number, but Scott probably knows more about this than I do).
2. 25% (for at least one of MDMA, psylocibin, or LSD)
3. 0.2%
4. 1%
5. 30%
6. 1%
7. 15%
8. 10%
9. 60%
10. 60%/25%/15%
11. 30%/40%/30%

I think I did slightly better than Scott’s crappy performance here.

1. 1 Bitcoin costs above $1K: 80%
2. …above $10K: 50%
3. …above $100K: 5%
4. Bitcoin is still the highest market cap cryptocurrency: 40%
5. Someone figures out Satoshi’s true identity to my satisfaction: 30%
6. Browser-crypto-mining becomes a big deal and replaces ads on 10%+ of websites: 5%

1. 75%
2. 40%
3. 5%
4. 30%
5. 20%
6. 20%

I did slightly worse than Scott here.

1. At least one person is known to have had a “designer baby” genetically edited for something other than preventing specific high-risk disease: 10%
2. At least a thousand people have had such babies, and it’s well known where people can go to do it: 5%
3. At least one cloned human baby, survives beyond one day after birth: 10%
4. Average person can check their polygenic IQ score for reasonable fee (doesn’t have to be very good) in 2023: 80%
5. At least one directly glutamatergic antidepressant approved by FDA: 20%
6. At least one directly neurotrophic antidepressant approved by FDA: 20%
7. At least one genuinely novel antipsychotic approved by FDA: 30%
8. MDMA approved for therapeutic use by FDA: 50%
9. Psilocybin approved for general therapeutic use in at least one country: 30%
10. Paleo diet is generally considered and recommended by doctors as best weight-loss diet for average person: 30%

1. 25%
2. 6%
3. 5%
4. 90%
5. 15%
6. 15%
7. 20%
8. 60%
9. 40%
10. 10% (I expect I have a lower opinion of doctors than Scott does)

I did slightly worse than Scott here.

Predictions on other topics:

1. alz.org will say Alzheimers is preventable/maybe sometimes preventable/not preventable/none of the above: 1%/29%/25%/45% (as of 2018-02-19, it says it “cannot be prevented, cured or even slowed.”)

The result is either maybe or none of the above. I didn’t find anything by browsing from their home page. A Google search for “prevent” yields “While Alzheimer’s prevention has no definitive answers at this time, research has shown that we can take action to reduce our risk of developing it.”

For the following, I’m giving 80% confidence intervals. I will score them using Scott Garrabrant’s scoring rule.

1. Crude oil $/barrel: 30-70
2. BTC: 500-50000
3. ETH: 100-10000
4. XRP: 0.1-100
5. SP500: 2200-3800
6. Treasury bond yield: 3.1-4.8
7. inflation expectation: 1.7-2.4
8. Nvidia 2022 revenues: $14-$24 billion
9. S&P 500 Financials weighting: 13-16%
10. S&P 500 Energy weighting: 4-7.5%
11. S&P 500 Technology weighting: 18-27%
12. S&P 500 Healthcare weighting: 12-17%

1. Berkeley LW meetup attendance (the number of people who show up by 8pm at the most recent meetup I’ve attended before President’s day 2023: 5-14
2. Number of pills I take on a typical day (excluding caffeine / coffee / tea pills): 8-16

Results (I don’t see a good way to aggregate these scores):

1	-6.16 – 20*40	-806.16
2	0	0
3	0	0
4	0	0
5	-197 – 20*1600	-32197
6	0	0
7	-0.7 – 20*0.18	-4.3
8	-2.91 – 20*10	-202.91
9	0	0
10	0	0
11	0	0
12	0	0
	[I didn’t find enough data to resolve the GiveWill + OpenPhil donation forecasts]
1	0? It’s been months since I attended a LW meetup, so my memory is hazy. Maybe 10 people?	0
2	-24 – 20*8	-184

I expected that inflation would be less than 10% over the five years. It turned out to be 20%. That single mistake seems sufficient to explain the 4 financial prediction failures.

Predictions for 2028

Scott’s predictions in blockquotes, mine underneath:

Some big macroeconomic indicator (eg GDP, unemployment, inflation) shows a visible bump or dip as a direct effect of AI (“direct effect” excludes eg an AI-designed pandemic killing people) : 15%

12% or 20%, depending on which indicators count and how strictly Scott interprets “direct”. 12% if we only use the 3 indicators he mentions and a strict meaning of direct effect. 20% if we include interest rate futures and it’s sufficient that I see more than a 50% chance that those futures are forecasting a rise due to AI.

The leading big tech company (eg Google/Apple/Meta) is (clearly ahead of/approximately caught up to/clearly still behind) the leading AI-only company (DeepMind/OpenAI/Anthropic) in the quality of their AI products: (25%/50%/25%)

20% / 50% / 30%.

I predict that DeepMind will be more likely than any other single entity to have the best technology, but I’m unclear on whether that will have been translated into products or whether it will be separate from Google products.

Gary Marcus can still figure out at least three semi-normal (ie not SolidGoldMagikarp style) situations where the most advanced language AIs make ridiculous errors that a human teenager wouldn’t make, more than half the time they’re asked the questions: 30%

25%

AI can play arbitrary computer games at human level. I will count this as successful if an off-the-shelf AI, given a random computer game and some kind of API that lets it to against itself however many times it wants, can reach the performance of a mediocre human. The human programmers can fiddle with it to make it compatible with that particular game’s API, but this is expected to take a few days of work and not involve redesigning the AI from scratch: 25%

35% true, 25% false, 40% too hard to resolve.

As above, but the AI can’t play against itself as many times as it wants. Using knowledge it’s gained from other computer games or modalities, it has to play the new computer game about as well as a first-time human player, and improve over time at about the same rate as a first-time human player (I don’t care if it’s one order of magnitude slower, just not millions of times slower): 10%

30% true, 25% false, 45% too hard to resolve.

Some product like “AI plus an internal scratchpad” or “AI with stable memory” fulfills the promise of that model, and is useful enough that it gets released for some application: 50%

“fulfills the promise” seem too vague to bet on.

AI can make a movie to your specifications: 40% short cartoon clip that kind of resembles what you want, 2% equal in quality to existing big-budget movies.

75% and 6%.

AI can make deepfake porn to your specifications (eg “so-and-so dressed in a cheerleading costume having sex on a four-poster bed with such-and-such”), 70% technically possible, 30% chance actually available to average person.

85% and 25%.

AI does philosophy: 65% chance writes a paper good enough to get accepted to a philosophy journal (doesn’t have to actually be accepted if everyone agrees this is true)

70%

AI can write poetry which I’m unable to distinguish from that of my favorite poets (Byron / Pope / Tennyson ): 70%

80%

There is (or seems about to be) a notable increase in new drug applications to the FDA because of AI doing a really great job designing drugs: 20%

15% (but a 30% chance that AI causes a notable increase in drug applications – it’s more likely that AI will make it much easier to deal with the FDA).

Something else in scientific research at least that exciting: 30%

30%

At least 350,000 people in the US are regularly (at least monthly) talking to an AI advertised as a therapist or coach. I will judge this as true if some company involved reports numbers, or if I hear about it as a cultural phenomenon an amount that seems proportionate with this number: 5%

6%.
I see almost no chance that an AI will be advertised as a therapist. I expect millions to be using AIs in coach- and therapy-like ways. I expect this claim to resolve as false mainly because companies will be cautious about how they advertise.

At least 350,000 people in the US are regularly (at least weekly) talking to an AI which they consider a kind of romantic companion. I will judge this as true if some company involved reports numbers, or if I hear about it as a cultural phenomenon an amount that seems proportionate with this number: 33%

40%

AI not-say-bad-words-ists treat AI not-kill-everyone-ists as (clear allies/clear enemies/it’s complicated): 25% / 35% / 40%

20% / 30% / 50%

AI is a (bigger/equal/smaller) political issue than abortion: 20% / 20% / 60%

55% / 25% / 20%

Artificial biocatastrophe (worse than COVID): 5%

2%

Ukraine war cease-fire: 80%

80%