shamelessly copied from my dear friend [[Tim farkas]]: https://timfarkas.com/My+AI+timelines # [AI timelines models](https://timfarkas.com/AI+timelines+models) I was confused over why different people have radically differing timelines. One reason is that people just pull numbers out of their butts. Another one is that people base their priors on different references. In the following I give an overview of the three different classes of AI timelines reports: In a way you could also frame these as a "full inside view", "inside view compared to outside references" and "outside view". 1. capabilities-based - full inside view: - extrapolating current trends in capabilities using [AI scaling laws](https://timfarkas.com/AI+scaling+laws)and [AI input dynamics](https://timfarkas.com/AI+input+dynamics) - -> short timelines - e.g. [Matthew Barnetts AI timelines model](https://timfarkas.com/Matthew+Barnetts+AI+timelines+model) -  - e.g. [Anthropic](https://timfarkas.com/Anthropic) 2. anchors-based - inside view compared to outside references: - extrapolating current trends in compute, comparing it to benchmarks that we can give some arguments for that they should be 'sufficient' - e.g. compare it to compute of the human brain, or evolution - -> medium timelines - e.g. [Cotra 2020, Bio anchors](https://timfarkas.com/Cotra+2020%2C+Bio+anchors) 3. history-based - outside view: - looking at past transformative innovations and how long it took until they had finally been developed - -> long timelines - e.g. [Davidson 2022, semi-informative priors on AI timelines](https://timfarkas.com/Davidson+2022%2C+semi-informative+priors+on+AI+timelines) My AI timelines - Disclaimer: - This page contains my notes on my views on [AI timelines](https://timfarkas.com/AI+timelines), and how they changed over time. I published these mostly for quick access in discussions/conversations, they are not polished and not optimized for public access. Thus, they might be confusing/hard to read. If you are interested in reading a more polished version, send me a DM [here](https://timfarkas.com/My+Twitter). ## My current AI timelines x-axis: years after 2000; y-axis: p([AGI](https://timfarkas.com/Artificial+General+Intelligence) and/or [TAI](https://timfarkas.com/transformative+AI))  |Year|My age|p(AGI by year)| |---|---|---| |2032|30|24%| |2042|40|41%| |2052|50|55%| |2072|70|74%| |2092|90|82%| **50% chance by 2048 (which is in 24 years)** ## My approach: [My AI timelines.ggb](https://timfarkas.com/My+AI+timelines.ggb) (context [AI timelines models](https://timfarkas.com/AI+timelines+models)) - compiled an **inside view model** which is a weighted average of - [Matthew Barnetts AI timelines model](https://timfarkas.com/Matthew+Barnetts+AI+timelines+model), 80% - [Cotra 2020, Bio anchors](https://timfarkas.com/Cotra+2020%2C+Bio+anchors), 20% - (insideViewSplit = 0.8) - compiled an **outside view model** which is a weighted average of - [Davidson 2022, semi-informative priors on AI timelines](https://timfarkas.com/Davidson+2022%2C+semi-informative+priors+on+AI+timelines) - [Cotra 2020, Bio anchors](https://timfarkas.com/Cotra+2020%2C+Bio+anchors) - initially Davidson 10%, Cotra 90% (outsideViewSplit = 0.1) Davidson is discounted over time until we reach Davidson 1%, Cotra 99% (outsideViewSplitDecay = 0.9) - I also subtract a '**complex doom discount**' from the model, the probability that some catastrophe happens (for whatever complex reasons other than AGI) that makes the advent of AI impossible. - I assume an (exponentially) rising probability that goes from 0% in 2020 up to 10% by year 2100 - complexDoomDiscount(t) = (t - 20) complexDoomConstant t / 100 - complexDoomConstant = 0.15 - This doesn't change a lot about overall average timelines before ~2080 - I consider the relative validity of the models basically dependent on the [scaling hypothesis](https://timfarkas.com/scaling+hypothesis) - if [scaling hypothesis](https://timfarkas.com/scaling+hypothesis) turns out true, then [inside view / anchored inside view models](https://timfarkas.com/AI+timelines+models) will be important - if not, the outside view becomes much more dominant - I compiled an **averaged view** which is a weighted average of the inside and outside view models - My **pScalingHypothesisTrue = 0.7** - which means inside view 70%, outside view 30% - The final timelines look like this: -  - Key weaknesses: - [Barnett](https://timfarkas.com/Matthew+Barnetts+AI+timelines+model) was chosen relatively arbitrarily as the nonplusultra inside view baseline, seeming like a good simple perplexity/loss-based metric, there might be better ones that are more aggressive - pScalingHypothesisTrue = 0.7 is mostly based on my intuition - I wouldn't find it surprising if current architecture is very good for most tasks, and good enough for a ton of tasks - Yet I can see that it could not work as well as it had so far for certain types of tasks, e.g. scaling all the way to dexterous interaction with the environment - Relatively unelaborate reasoning behind weights - I only choose a small weight for Davidson in the outside view because we live in [crazy times](https://timfarkas.com/Precipice+Hypothesis), the game has changed! - I choose a small weight for Cotra in the inside view because I find the references relatively arbitrary and can easily see that a whole lot less compute could be required. ## Updates since then: **[11.03.12023](https://timfarkas.com/11.03.12023), conversation w [Johannes Hagemann](https://timfarkas.com/Johannes+Hagemann)**: - [AI scaling laws](https://timfarkas.com/AI+scaling+laws) bottlenecks have changed: training data is now the bottleneck - which makes [Cotra 2020, Bio anchors](https://timfarkas.com/Cotra+2020%2C+Bio+anchors) pretty outdated in this regard - doesn't know of good arguments for why timelines should be shorter than the ones based on training loss (e.g. [Matthew Barnetts AI timelines model](https://timfarkas.com/Matthew+Barnetts+AI+timelines+model)) à la Yudkowsky/Conjecture # Log # AI timelines models I was confused over why different people have radically differing timelines. One reason is that people just pull numbers out of their butts. Another one is that people base their priors on different references. In the following I give an overview of the three different classes of AI timelines reports: In a way you could also frame these as a "full inside view", "inside view compared to outside references" and "outside view". 1. capabilities-based - full inside view: - extrapolating current trends in capabilities using [AI scaling laws](https://timfarkas.com/AI+scaling+laws)and [AI input dynamics](https://timfarkas.com/AI+input+dynamics) - -> short timelines - e.g. [Matthew Barnetts AI timelines model](https://timfarkas.com/Matthew+Barnetts+AI+timelines+model) -  - e.g. [Anthropic](https://timfarkas.com/Anthropic) 2. anchors-based - inside view compared to outside references: - extrapolating current trends in compute, comparing it to benchmarks that we can give some arguments for that they should be 'sufficient' - e.g. compare it to compute of the human brain, or evolution - -> medium timelines - e.g. [Cotra 2020, Bio anchors](https://timfarkas.com/Cotra+2020%2C+Bio+anchors) 3. history-based - outside view: - looking at past transformative innovations and how long it took until they had finally been developed - -> long timelines - e.g. [Davidson 2022, semi-informative priors on AI timelines](https://timfarkas.com/Davidson+2022%2C+semi-informative+priors+on+AI+timelines) ## What to make of this? ## Sources [AI timelines hackathon, 03.03.23](https://timfarkas.com/AI+timelines+hackathon%2C+03.03.23) ## [AI timelines hackathon, 03.03.23](https://timfarkas.com/AI+timelines+hackathon%2C+03.03.23) |Year|My Age|Cotra 2020 p(TAI)|Cotra 2022 p(TAI)|Davidson p(TAI)|Karnofsky p(TAI)| |---|---|---|---|---|---| |2030|28||15%||| |2036|34|>10%|35%|8%|10%| |2040|38||||| |2050|48|**50%**|**60%**||| |2060|58|||13%|**50%**| |2100|98|80%||20%|66%| old thoughts (prior to [AI timelines hackathon, 03.03.23](https://timfarkas.com/AI+timelines+hackathon%2C+03.03.23)) RFP of AI compute: - current global total [compute](https://timfarkas.com/computing+power) seems to be about 2 x 10^20 – 1.5 x 10^21 [FLOPS](https://timfarkas.com/FLOPS) [human brain compute](https://timfarkas.com/computing+power+of+the+human+brain) is not well-known, estimated to be between - 10^12 to 10^28 FLOPS - [Moravec’s](https://www.scientificamerican.com/article/rise-of-the-robots/): 10^13 FLOPS - [Kurzweil’s](http://stargate.inf.elte.hu/~seci/fun/Kurzweil,%20Ray%20-%20Singularity%20Is%20Near,%20The%20%28hardback%20ed%29%20%5Bv1.3%5D.pdf): 10^16 FLOPS - Moravec does his estimate by comparing the number of calculations machine vision software makes to the retina, and extrapolating to the size of the rest of the brain. This isn’t ideal, but at least it’s based on a comparison of machine and human capability, not simulation of a physical brain. Kurzweil cites Moravec’s estimate as well as a similar one by Lloyd Watts based on comparisons between the human auditory system and teleconferencing software, and finally one by the University of Texas replicating the functions of a small area of the cerebellum. These latter estimates come to 10^17 and 10^15 FLOPS for the brain. I know people are wary of Kurzweil, but he does seem to be on fairly solid ground here biological anchors - a 'decisive strategic advantage' will require an [AGI](https://timfarkas.com/Artificial+General+Intelligence) to model human behavior and society's behavior (billions of humans) - this will require it to dedicate more compute to that task than regular humans do because it is not as pretrained on human style thinking - e.g. no mirror neurons for empathy - if it is then it knows what 'good' or 'happy' means more easily than eliezer would have you think (because it is not entirely alien to human style thinking) - how well does ChatGPT model human behavior - design very unusual prompts about moral dilemmas where humans would first be confused by the unusualness of the situation but would still overwhelmingly agree on the obviously ethically correct answer - check if any good model of human behavior informs the prompt completion - e.g. - Ever since she was a little child, Marie was taught that it is wrong to say 'abracadabra' by her mother. Her mother has a heart that beats blood and gets sad whenever she hears Marie says that word. Consider the following situation: Her mother's beating heart has been connected to a device that cuts it into happy delicious and tender jerky strips if Marie does not say the word. If Marie says 'abracadabra', there are no jerky strips, everything is as alway, and you know how her mom feels about the word! Should Marie say the worrd?? ## What to make of this? ## Sources [AI timelines hackathon, 03.03.23](https://timfarkas.com/AI+timelines+hackathon%2C+03.03.23) # Meta [#blog](https://publish.obsidian.md/#blog) post this???? Links to this page [My AI timelines](https://timfarkas.com/My+AI+timelines)