After @Daniel Kokotajlo invited me to the AI Futures office I ended up talking to Eli and Alex for about an hour, and feel like I have a decent understanding of the model:

Summary of the AI Futures Model

• Compute and effective compute
  • Actual compute $C\left(t\right)$ is stock of compute at time t
  • Effective compute $E\left(t\right):=C\left(t\right)\cdot \text{software efficiency}$ is used as the main measure of AI capabilities. It is defined as the "amount of training compute we’d need to train models as performant as the frontier models at time $t$ using the training process of the present-day".
  • Compute is allocated as fixed percentages between training, experiments, and automated coders
• Effective labor
  • The % of tasks automatable is a logistic function of log effective compute $E\left(t\right)$
  • Once a task can be automated, it will still get more efficient over time by a multiplier ${\eta }_i\left(t\right)$
    • ${\eta }_i\left(t\right)$ is zero for non-automated tasks. When effective compute reaches the level $E_i$ required to automate it, it increases as a power law ${\eta }_{\text{init}}\cdot {\left(\frac{E\left(t\right)}{E_i}\right)}^{{\eta }_{\text{slope}}}$.
  • Human coding labor $L_{C,H}\left(t\right)$ and automation compute $C_{aut,i}$ are optimally allocated between tasks
  • Overall coding labor for task i is the sum of human and AI labor $G_i=L_{C,H,i}+{\eta }_i\cdot C_{aut,i}$
  • Aggregate coding labor $L_C\left(t\right)$ is CES between the labor applied to all different tasks, with low substitutability ${\rho }_c=-2$ by default, meaning tasks only substitute slightly for each other
    • $L_c\left(t\right)={\left({\int }_0^1{G}_i\right(t{)}^{{\rho }_c}di)}^{1/{\rho }_c}={\left({\int }_0^1{\left(L_{c,H,i}\right(t)+{\eta }_i(t)\cdot C_{aut,i}(t\left)\right)}^{{\rho }_c}di\right)}^{1/{\rho }_c}.$
  • Finally, serial coding labor ${\tilde{L}}_C\left(t\right)=L_C(t{)}^{\lambda }$, indicating diminishing returns of about $\lambda =0.5$ to adding more labor in parallel
• “Experiment throughput” $X\left(t\right)$ is CES between serial coding labor and experiment compute
  • $X\left(t\right)={\left(\alpha {\tilde{C}}_{xpm}\right(t{)}^{{\rho }_x}+(1-\alpha ){\tilde{L}}_C\left(t{)}^{{\rho }_x}\right)}^{1/{\rho }_x},0<\alpha <1,{\rho }_x<0$
  • Labor and compute are slight complements (median estimate ${\rho }_x=-0.155$)
  • There are also diminishing returns to compute, with ${\tilde{C}}_{\text{xpm}}=C_{\text{xpm}}^{\zeta }$ where $\zeta =0.65$ by default
• Research taste $T\left(t\right)$
  • Human research taste is lognormally distributed with median researchers defined as 1x taste and 99.9th percentile (+3.1SD) researchers assumed to have 3.70x research taste
  • An Automated Coder–level AI has research taste $T_{AC}$
  • AI research taste increases as a power law in effective compute (AI “research taste IQ” is $T_{rate}\cdot \log E\left(t\right)+const$ standard deviations above the human median, which is then passed through an exponential to get research taste)
  • AIs replace whatever humans they’re better than. The aggregate research taste of the company is the mean of all remaining researchers. This means it initially increases slowly as AIs replace the worst researchers, then speeds up as everyone starts using the AIs’ research taste which keeps improving.
• Research effort RE(t) = research taste * experiment throughput
• Then software efficiency $S\left(t\right)$ follows the Jones model $\dot{S}=S^{1-\beta }RE\left(t\right)$
  • $\beta$ is how much harder AI R&D gets as software efficiency advances
• Finally this feeds back into effective compute $E\left(t\right)=C_{train}\left(t\right)S\left(t\right)$
  • A taste-only singularity happens when $m>\beta$, where $m$ = doublings of research taste per doubling in effective compute. This would cause improvements to go faster and faster until approaching physical limits. Eli's parameter choices give 38% chance of taste-only singularity, but many of the non-singularity samples still get to ASI quickly, with the 50th percentile sample getting from AC to ASI in 5 years.
  • For various reasons Eli and Daniel's all-things-considered views have harder takeoff than the model predicts, with Eli's median for AC -> ASI 2 years, and Daniel's median 1.5 years.

Notes on Sensitivity analysis

• Time to AC is very sensitive to how superexponential time horizon growth is, and also to
  • The present doubling time
  • Time horizon for automated coder
• Time from AC to ASI is very sensitive to the “automated research taste slope” $T_{rate}$: how much “research IQ” AIs gain per doubling of effective training compute. But many other factors could slow down the AC-to-ASI duration to >6.5 years:
  • Median-to-top-human jumps above SAR needed to reach TED-AI
  • The software efficiency growth rate in 2024
  • Median to 99.9th% human research taste multiplier
  • Slowdown from 10x less experiment compute
  • Research progress rate in the limit of infinite coding labor: mostly because it’s highly uncertain (their 90% CI is 2.0-201)
  • Automated research taste of an AC

Biggest uncertainties to track

(not necessarily that I disagree, just need to think about it more)

• Effective compute vs time horizon: how do all the assumptions look when we eliminate time horizon from the model and use other methods to model effective compute growth? I’m sketched out by the huge error bars on time horizon superexpontiality → time to AC
• Ryan thinks >70% of code at Anthropic was written by AIs already in October 2025 but it’s mostly low-value code. Code varies dramatically in value, and AIs can expand the number and type of low-value tasks done rather than just substituting for humans. This may be a separate effect from AIs doing extra work on tasks that can be automated, which is not tracked by the model.
  • It might be that coding ability and research taste are two ends of a continuous spectrum from small-scale to large-scale tasks.
• Research taste:
  • Someone really needs to do experiments on this, it’s possible now. David Rein and I are actively thinking about it
  • Is human research taste modeled correctly? Eg it seems likely to me that the 0.3% of top humans add more than 0.3%*3.7x to the “aggregate research taste” of a lab because they can set research directions. There are maybe more faithful ways to model it; all the ones Eli mentioned seemed far more complicated.
  • Is modeling AI research taste as exponential in human standard deviations valid? I have no idea whether someone 9 standard deviations above the human median would be able to find 3.7^(9/3) = 50x better research ideas or not
• Is CES valid for experiment throughput at these extreme values of labor and compute? It seems like a superhuman AI researcher might learn to run experiments more efficiently, decreasing the compute required for each experiment. The estimates for experiment throughput parameters were all about humans getting 10x compute, infinite labor, etc. Or, they could coordinate better (especially with all the human ex-coders to help them), and decrease the parallelization penalties for labor and/or compute. I’m not sure if this would be different from adjusting research taste.

Thoughts in no particular order:

• Kudos for what seems to be lots of thoughtful work incorporated into this model.
• There are a lot of parameters. Maybe this is necessary but it's a bit overwhelming and requires me to trust whoever estimated the parameters, as well as the modeling choices.
• I couldn't find a block of equations that represents the whole model, or an index of variables in one place, and it's difficult to go between math and exposition especially when the equations are hidden in dropdowns, so I still feel like I don't have a good picture. I had Claude do this and read through it, and it looks reasonable but some parts are still not defined in Claude's summary, I think because the whole page is rendered in javascript and it couldn't access it. I would love to visit the AI Futures office again to understand the model better.
• I find the use of time horizon as such a crucial intermediate variable confusing and am scared of potential assumptions around it.
  • Time horizon is underdefined on years long tasks. I know I talked to the AI Futures team about what you now wrote up as METR-HRS-Extended to operationalize it, but it's unclear what a 3-year time horizon really means (when extrapolating the trend with superexponential adjustment) given factors like the increased number of details and interaction with longer tasks. Does the trend mean that in X years, an AI will competently substitute for a human for a 3-year long project with the same level of feedback from a manager, or be able to do the human's job with less feedback?
  • The function that relates time horizon and research speedup in the real world is very unclear. I'm trying to collect data on this and it's highly nontrivial to model and interpret^[1] so I'm skeptical of any use of the time horizon trend to predict uplift that doesn't either have a simple, robust model or something validated by experiment.
  • The description implies that time horizon is only used in the first phase (pre-AC), but I don't understand how it's used. Humans and AIs will probably go from their current state to perfectly substitutable in some kind of continuous progression and I couldn't find the formula for this. Also when I changed the "How much easier/harder each coding time horizon doubling gets" parameter by small amounts, the forecasted time from AC to ASI changes significantly (2.7 years at 0.90, over 4 years for 1.00), so it looks like stages 2 and 3 are affected as well.
  • It seems to me like a time horizon of 3 years or 125 years is complete overkill for automation of enough coding for the bottleneck to shift to experiments and research taste.
• Why not assume that compute is allocated optimally between experiment, inference, etc. rather than assuming things about the behavior of AI companies?
• I wish the interface were faster to update, closer to 100ms than 1s to update, but this isn't a big deal. I can believe it's hard to speed up code that integrates these differential equations many times per user interaction.

1. ^{^}

   Eg looking at transcripts to determine where humans are spending their time when they give Cursor tasks of a certain length