"one of the authors recently acknowledged that “AI propaganda hasn’t been as effective as feared”."
That was about my previous scenario, written in 2021, not about AI 2027. I don't think there's been less AI propaganda so far than depicted in AI 2027, because AI 2027 didn't really depict AI propaganda happening by mid 2026. (If you are interested in the previous scenario, it's called What 2026 Looks Like. https://www.lesswrong.com/posts/6Xgy6CAf2jqHhynHL/what-2026-looks-like)
"By their estimate, uplift is only 17% of what they’d predicted at this point. The 65% figure primarily reflects metrics that are not directly related to real-world AI capabilities, such as benchmark scores and data center construction. I should note however that frontier lab revenue is at 80% of prediction (in their scoring, which may in fact be conservative here)."
After we published that, frontier lab revenue has continued to increase and now is higher than AI 2027 predicted. This matters because frontier lab revenue is pretty directly related to real-world AI capabilities!
As for uplift: We actually think we're being pretty harsh on ourselves here. The reason it's only 17% is because at the time we wrote AI 2027 we thought uplift was higher than it actually was. In our writeup we say: "AI software R&D uplift is behind pace. This is primarily because we have updated our estimate of uplift in early 2025 downward, and thus our uplift estimates for the end of 2025 are similar to our original estimates for the start of AI 2027." So basically, the slope of the line is similar to what we expected, but the whole line is shifted down relative to what we expected. Very different from progress only being 17% as fast as we expected, which is a possible interpretation someone might have from just seeing that figure.
Thanks for the clarifications! In hindsight I should have reached out before publishing this... I didn't frame the piece around AI 2027 until very late in the writing process and by then I wasn't thinking in terms of collecting feedback, mental note not to do that again.
I was a bit confused about the uplift figure (which should have been another signal to reach out!), your and Eli's comments here are very helpful. I'll make a few edits to the post.
> It can be difficult for businesses to integrate AI into their work.
I feel like these sorts of anecdotes should be given little weight relative to the insane revenue trend, which you discuss elsewhere in the piece, and suggests that AI is being integrated into many businesses quickly. Of course there are some difficulties, not denying that; but taken literally, the claim that it's sometimes difficult to integrate AI into business's work is a very weak one.
> I mentioned that the authors of AI 2027 recently reported that things have been progressing about 65% as fast as they’d predicted. But I would argue that the most important component of their forecast is “AI software R&D uplift” (the extent to which AI is accelerating its own development). By their estimate, uplift is only 17% of what they’d predicted at this point.
(edit: Daniel also covered this ground in his concurrently written comment, I said the same thing in more words and with a bit more detail.)
Reasonable to bring up. But this metric is unusual for a few reasons.
1. Unlike the other metrics, we can't observe it directly, so it's more of an educated guess.
2. Since we published AI 2027, we've changed our mind about the uplift value in Apr 2025 (when AI 2027 started). So the reason that 17% number is so low is mostly because of a change in mind about the absolute value of uplift, rather than the rate. Which is still a relevant change of mind, but of a different shape. I wasn't sure the best way to handle this in our calculations; sorry if this wasn't clear.
Also, it's not that uplift is 17% of what we predicted, it's that the pace of progress is 17% slower than we predicted (calculated using our best guess for Feb 2026 uplift, and comparing against our unrevised Apr 2025 estimate).
> I should note however that frontier lab revenue is at 80% of prediction (in their scoring, which may in fact be conservative here).
The 80% you're citing is the average of grading valuation and revenue predictions. Revenue predictions are ahead of AI 2027's predictions, and further so than when we did the grading. It looks like valuation is now about on trend, as opposed to our estimate of going at 44% speed when we graded a few months ago.
> Will the rapid progress in AI coding ability lead to full automation of AI R&D, or are “softer” skills also needed?
To be more precise, I think everyone agrees non-coding skills are needed to fully automate AI R&D, but some people think it's likely that either automating coding is enough to bootstrap you to automate the others, or that automating coding will generalize well enough (but not perfectly well) to other aspects of AI R&D, such that full automation is close by.
And I think missing from your list of questions is how fast AI will progress after full AI R&D automation to different levels of superhuman AI R&D capabilities, and how high the ceiling is above the human range.
As I just responded to Daniel – thank you for the clarifications and feedback! I've inserted two corrections into the post (search for "EDIT"), please let me know if further corrections are warranted.
Regarding revenue trends vs. integration challenges: I agree that we should pay a lot of attention to frontier lab revenue figures. I guess my vibe here is as follows: Yes, current revenue figures (on the order of $60B combined ARR for OpenAI+Anthropic alone, plus Gemini, plus open-weight models, plus I believe Microsoft's revenue for direct sales of OpenAI models, plus all the other players) are crazy high for a technology this young. They are also crazy *low* compared to the seeming potential of the models, agents, and applications that are already available – especially if we view the models through the lens of benchmark scores (e.g. METR horizon lengths, GDPval). AI as powerful as current benchmark scores imply should be having a super super super super super big impact, but (going by revenue figures) it is only having a super super super big impact, which seems consistent with the existence of integration challenges, jagged capabilities not well reflected in benchmark scores, and other limitations and frictions. But still super super super big impact!
> To be more precise, I think everyone agrees non-coding skills are needed to fully automate AI R&D, but some people think it's likely that either automating coding is enough to bootstrap you to automate the others, or that automating coding will generalize well enough (but not perfectly well) to other aspects of AI R&D, such that full automation is close by.
Thanks, I'll add a footnote regarding this. It seems related to question 3, but worth calling out.
> And I think missing from your list of questions is how fast AI will progress after full AI R&D automation to different levels of superhuman AI R&D capabilities, and how high the ceiling is above the human range.
“> I should note however that frontier lab revenue is at 80% of prediction (in their scoring, which may in fact be conservative here).”
“The 80% you're citing is the average of grading valuation and revenue predictions. Revenue predictions are ahead of AI 2027's predictions, and further so than when we did the grading. It looks like valuation is now about on trend, as opposed to our estimate of going at 44% speed when we graded a few months ago.”
Maybe I am looking at different numbers but it appears that Anthropic has an estimated valuation of 387.97B while OpenAI has an estimated valuation of 852B. It seems to me that when you say “about on trend” that means one of the leading AI companies will have a valuation of that is less than 2.5 trillion dollars. I am predicting a valuation somewhere in the range of 1 trillion to 2 trillion dollars. Do you agree with this assessment?
I'm not sure I understand your question. When I said about on trend I was referring to current valuations being close to the $1T predicted by AI 2027 for Apr 2026.
In terms of valuation, I have been going off of official numbers based on what companies have said. I am skeptical of secondary markets because they are not official. This is why I believe that 1 trillion dollars won’t be reached until the end of June and that by the end of December the leading AI company will have valuation 2 trillion dollars or so. Sorry for any confusion! Please let me know if this makes sense!
That makes sense. I trust the private markets more than you. there's also been news reports of Anthrpic raisng at $900B valuation. But 2 trillion valuation end of year seems like in the range of reasonable predictions, nonetheless. Looks like AI 2027 depicted $3T for Dec 2026
Thanks for the sharing your perspective. I looked back at AI 2027 but also at the post from February: https://substack.com/@elifland/note/p-187696603?r=6lp84s&utm_medium=ios&utm_source=notes-share-action. In the February post both you and Daniel say: “AGI company revenues and valuations. In AI 2027, we depicted the leading company reaching $55B in annualized revenue and a valuation of $2.5T by 2026, making it one of the most valuable companies in the world. We think these are decent indicators of the real-world value that AI is providing.” However AI 2027 as you mentioned the leading AI company has a much higher valuation. Maybe I misinterpreted the post from the beginning of the year but I was under the impression that both you and Daniel were predicting an end of year valuation of 2.5 trillion dollar. Is this correct or did I misread?
> This doesn’t mean that things will continue to progress on the AI 2027 timeline. It’s under-appreciated that the explosive and it’s-complicated views do not diverge in their predictions until the advent of “strong AGI”
Taken literally, this would imply that we will get zero evidence to distinguish between these views until strong AGI. However, in much of the rest of the piece you seem to attempt to decipher evidence pointing one way or the other. This seems contradictory?
My take is that there will be a bunch of evidence that point one way or another between these views before strong AGI, though I think it will take some care to interpret the evidence well and to pay attention, do informative experiments, etc. to generate better evidence. So I worry that in practice people won't update, because I agree that there may not be obvious smoking guns.
Agreed, "do not diverge" is an overstatement. I just edited it to add the word "much", which might still be a bit inadequate?
And I think it's very important to look for such evidence as will be possible to collect, and draw as much attention to it as possible – an effort I'm hoping to contribute to.
Recent AI company revenue growth has nothing to do with the accuracy of the 2027 forecasts. LLMs are inherently limited and those limitations are increasingly apparent. For continued major improvements -- including superhuman intelligence -- we will need different approaches.
Great scoping of breadth for the post and then execution. The one thing that didn’t track was why you think we’re almost certainly going to see a singularity in a few decades. Maybe it’s easy to say with such a long time window but I found it a little jarring relative to the rest.
Thanks, that's a good flag – I really didn't lay the groundwork for that here. If you're interested, I sketched my reasons for this belief in a December post: https://secondthoughts.ai/p/the-unrecognizable-age.
My main issue with checking the predictions from AI 2027 is that it's really two stories: things continue walking along normally until early 2027, and then a MASSIVE discontinuity happens as AI leaps from merely doing well on arbitrary benchmarks to swallowing most white collar work in a short amount of time. But the former does not imply the latter: there's tons of schlep in peoples' jobs that AI is not showing signs of being able to overcome. It can help with individual tasks, but is still very far from serving as a drop-in replacement. The AI 2027 story doesn't do enough to show how AI gets from benchmarks to real-world impacts or how it solves for the jagged frontier, it just kinda motions at "Recursive Self Improvement" as a binary yes/no toggle that somehow solves everything like magic.
It also massively underestimates the amount of civil unrest all that disruption is going to cause. It straight up says in the "good" scenario "everyone loses their jobs, but no one is mad because they get welfare." It's the trillionth example of Silicon Valley having no idea how people actually are. This is already a big deal that the government is under increasing pressure to rein in. You don't think that's also going to grow exponentially? At a certain point, "shut down the AI labs" will become the only thing any voter in either party cares about.
Like, the Verizon CEO saying that the unemployment rate will be 20-30% in two years. If he is correct, we will have anarchy and civil war, and it's hard to see AI progress still chugging along at that point.
I feel like you may have adopted a rather loose definition of what would count as the "AI as a normal tech" view winning. Even if all three are true, I still wouldnt say AI is a normal tech:
"• AI’s impressive showing on benchmarks and coding will struggle to generalize to broader real-world settings.
• Even as AI develops genuinely useful capabilities, there will be (or should be) delays and limits on how it is adopted in practice.
• AIs will not develop radically superhuman abilities, because the real world is often too complex for abstract genius to eliminate the need for trial and error."
Thanks for this review! Some comments:
We co-authored an article with the "AI as Normal Technology" authors you might be interested in: https://asteriskmag.substack.com/p/common-ground-between-ai-2027-and
"one of the authors recently acknowledged that “AI propaganda hasn’t been as effective as feared”."
That was about my previous scenario, written in 2021, not about AI 2027. I don't think there's been less AI propaganda so far than depicted in AI 2027, because AI 2027 didn't really depict AI propaganda happening by mid 2026. (If you are interested in the previous scenario, it's called What 2026 Looks Like. https://www.lesswrong.com/posts/6Xgy6CAf2jqHhynHL/what-2026-looks-like)
"By their estimate, uplift is only 17% of what they’d predicted at this point. The 65% figure primarily reflects metrics that are not directly related to real-world AI capabilities, such as benchmark scores and data center construction. I should note however that frontier lab revenue is at 80% of prediction (in their scoring, which may in fact be conservative here)."
After we published that, frontier lab revenue has continued to increase and now is higher than AI 2027 predicted. This matters because frontier lab revenue is pretty directly related to real-world AI capabilities!
As for uplift: We actually think we're being pretty harsh on ourselves here. The reason it's only 17% is because at the time we wrote AI 2027 we thought uplift was higher than it actually was. In our writeup we say: "AI software R&D uplift is behind pace. This is primarily because we have updated our estimate of uplift in early 2025 downward, and thus our uplift estimates for the end of 2025 are similar to our original estimates for the start of AI 2027." So basically, the slope of the line is similar to what we expected, but the whole line is shifted down relative to what we expected. Very different from progress only being 17% as fast as we expected, which is a possible interpretation someone might have from just seeing that figure.
Thanks for the clarifications! In hindsight I should have reached out before publishing this... I didn't frame the piece around AI 2027 until very late in the writing process and by then I wasn't thinking in terms of collecting feedback, mental note not to do that again.
I was a bit confused about the uplift figure (which should have been another signal to reach out!), your and Eli's comments here are very helpful. I'll make a few edits to the post.
Also: I appreciated the Common Ground essay – I linked to it here!
> It’s under-appreciated that the explosive and it’s-complicated views do not diverge much in their predictions until the advent of “strong AGI”
(I just inserted the "much", in response to one of Eli's comments.)
Thanks for writing this! A few thoughts:
> It can be difficult for businesses to integrate AI into their work.
I feel like these sorts of anecdotes should be given little weight relative to the insane revenue trend, which you discuss elsewhere in the piece, and suggests that AI is being integrated into many businesses quickly. Of course there are some difficulties, not denying that; but taken literally, the claim that it's sometimes difficult to integrate AI into business's work is a very weak one.
> I mentioned that the authors of AI 2027 recently reported that things have been progressing about 65% as fast as they’d predicted. But I would argue that the most important component of their forecast is “AI software R&D uplift” (the extent to which AI is accelerating its own development). By their estimate, uplift is only 17% of what they’d predicted at this point.
(edit: Daniel also covered this ground in his concurrently written comment, I said the same thing in more words and with a bit more detail.)
Reasonable to bring up. But this metric is unusual for a few reasons.
1. Unlike the other metrics, we can't observe it directly, so it's more of an educated guess.
2. Since we published AI 2027, we've changed our mind about the uplift value in Apr 2025 (when AI 2027 started). So the reason that 17% number is so low is mostly because of a change in mind about the absolute value of uplift, rather than the rate. Which is still a relevant change of mind, but of a different shape. I wasn't sure the best way to handle this in our calculations; sorry if this wasn't clear.
Also, it's not that uplift is 17% of what we predicted, it's that the pace of progress is 17% slower than we predicted (calculated using our best guess for Feb 2026 uplift, and comparing against our unrevised Apr 2025 estimate).
> I should note however that frontier lab revenue is at 80% of prediction (in their scoring, which may in fact be conservative here).
The 80% you're citing is the average of grading valuation and revenue predictions. Revenue predictions are ahead of AI 2027's predictions, and further so than when we did the grading. It looks like valuation is now about on trend, as opposed to our estimate of going at 44% speed when we graded a few months ago.
> Will the rapid progress in AI coding ability lead to full automation of AI R&D, or are “softer” skills also needed?
To be more precise, I think everyone agrees non-coding skills are needed to fully automate AI R&D, but some people think it's likely that either automating coding is enough to bootstrap you to automate the others, or that automating coding will generalize well enough (but not perfectly well) to other aspects of AI R&D, such that full automation is close by.
And I think missing from your list of questions is how fast AI will progress after full AI R&D automation to different levels of superhuman AI R&D capabilities, and how high the ceiling is above the human range.
As I just responded to Daniel – thank you for the clarifications and feedback! I've inserted two corrections into the post (search for "EDIT"), please let me know if further corrections are warranted.
Regarding revenue trends vs. integration challenges: I agree that we should pay a lot of attention to frontier lab revenue figures. I guess my vibe here is as follows: Yes, current revenue figures (on the order of $60B combined ARR for OpenAI+Anthropic alone, plus Gemini, plus open-weight models, plus I believe Microsoft's revenue for direct sales of OpenAI models, plus all the other players) are crazy high for a technology this young. They are also crazy *low* compared to the seeming potential of the models, agents, and applications that are already available – especially if we view the models through the lens of benchmark scores (e.g. METR horizon lengths, GDPval). AI as powerful as current benchmark scores imply should be having a super super super super super big impact, but (going by revenue figures) it is only having a super super super big impact, which seems consistent with the existence of integration challenges, jagged capabilities not well reflected in benchmark scores, and other limitations and frictions. But still super super super big impact!
> To be more precise, I think everyone agrees non-coding skills are needed to fully automate AI R&D, but some people think it's likely that either automating coding is enough to bootstrap you to automate the others, or that automating coding will generalize well enough (but not perfectly well) to other aspects of AI R&D, such that full automation is close by.
Thanks, I'll add a footnote regarding this. It seems related to question 3, but worth calling out.
> And I think missing from your list of questions is how fast AI will progress after full AI R&D automation to different levels of superhuman AI R&D capabilities, and how high the ceiling is above the human range.
I'll add a note about this as well.
“> I should note however that frontier lab revenue is at 80% of prediction (in their scoring, which may in fact be conservative here).”
“The 80% you're citing is the average of grading valuation and revenue predictions. Revenue predictions are ahead of AI 2027's predictions, and further so than when we did the grading. It looks like valuation is now about on trend, as opposed to our estimate of going at 44% speed when we graded a few months ago.”
Maybe I am looking at different numbers but it appears that Anthropic has an estimated valuation of 387.97B while OpenAI has an estimated valuation of 852B. It seems to me that when you say “about on trend” that means one of the leading AI companies will have a valuation of that is less than 2.5 trillion dollars. I am predicting a valuation somewhere in the range of 1 trillion to 2 trillion dollars. Do you agree with this assessment?
Anthropic and OAI have hit a bit above 1T on secondary markets like https://app.ventuals.com/markets
I'm not sure I understand your question. When I said about on trend I was referring to current valuations being close to the $1T predicted by AI 2027 for Apr 2026.
This is the source that I have been looking at: https://finance.yahoo.com/markets/private-companies/highest-valuation/.
In terms of valuation, I have been going off of official numbers based on what companies have said. I am skeptical of secondary markets because they are not official. This is why I believe that 1 trillion dollars won’t be reached until the end of June and that by the end of December the leading AI company will have valuation 2 trillion dollars or so. Sorry for any confusion! Please let me know if this makes sense!
That makes sense. I trust the private markets more than you. there's also been news reports of Anthrpic raisng at $900B valuation. But 2 trillion valuation end of year seems like in the range of reasonable predictions, nonetheless. Looks like AI 2027 depicted $3T for Dec 2026
Thanks for the sharing your perspective. I looked back at AI 2027 but also at the post from February: https://substack.com/@elifland/note/p-187696603?r=6lp84s&utm_medium=ios&utm_source=notes-share-action. In the February post both you and Daniel say: “AGI company revenues and valuations. In AI 2027, we depicted the leading company reaching $55B in annualized revenue and a valuation of $2.5T by 2026, making it one of the most valuable companies in the world. We think these are decent indicators of the real-world value that AI is providing.” However AI 2027 as you mentioned the leading AI company has a much higher valuation. Maybe I misinterpreted the post from the beginning of the year but I was under the impression that both you and Daniel were predicting an end of year valuation of 2.5 trillion dollar. Is this correct or did I misread?
I think 2.5T was the value in our internal spreadsheet for end of 2026, but it was rounded to 3T in the side panel display
One more thought.
> This doesn’t mean that things will continue to progress on the AI 2027 timeline. It’s under-appreciated that the explosive and it’s-complicated views do not diverge in their predictions until the advent of “strong AGI”
Taken literally, this would imply that we will get zero evidence to distinguish between these views until strong AGI. However, in much of the rest of the piece you seem to attempt to decipher evidence pointing one way or the other. This seems contradictory?
My take is that there will be a bunch of evidence that point one way or another between these views before strong AGI, though I think it will take some care to interpret the evidence well and to pay attention, do informative experiments, etc. to generate better evidence. So I worry that in practice people won't update, because I agree that there may not be obvious smoking guns.
Agreed, "do not diverge" is an overstatement. I just edited it to add the word "much", which might still be a bit inadequate?
And I think it's very important to look for such evidence as will be possible to collect, and draw as much attention to it as possible – an effort I'm hoping to contribute to.
Recent AI company revenue growth has nothing to do with the accuracy of the 2027 forecasts. LLMs are inherently limited and those limitations are increasingly apparent. For continued major improvements -- including superhuman intelligence -- we will need different approaches.
Great scoping of breadth for the post and then execution. The one thing that didn’t track was why you think we’re almost certainly going to see a singularity in a few decades. Maybe it’s easy to say with such a long time window but I found it a little jarring relative to the rest.
Thanks, that's a good flag – I really didn't lay the groundwork for that here. If you're interested, I sketched my reasons for this belief in a December post: https://secondthoughts.ai/p/the-unrecognizable-age.
My main issue with checking the predictions from AI 2027 is that it's really two stories: things continue walking along normally until early 2027, and then a MASSIVE discontinuity happens as AI leaps from merely doing well on arbitrary benchmarks to swallowing most white collar work in a short amount of time. But the former does not imply the latter: there's tons of schlep in peoples' jobs that AI is not showing signs of being able to overcome. It can help with individual tasks, but is still very far from serving as a drop-in replacement. The AI 2027 story doesn't do enough to show how AI gets from benchmarks to real-world impacts or how it solves for the jagged frontier, it just kinda motions at "Recursive Self Improvement" as a binary yes/no toggle that somehow solves everything like magic.
This is similar to my view. I tried to get at this in #3 and #4 (and, to some extent, #1) in my list of questions at the end of the post.
It also massively underestimates the amount of civil unrest all that disruption is going to cause. It straight up says in the "good" scenario "everyone loses their jobs, but no one is mad because they get welfare." It's the trillionth example of Silicon Valley having no idea how people actually are. This is already a big deal that the government is under increasing pressure to rein in. You don't think that's also going to grow exponentially? At a certain point, "shut down the AI labs" will become the only thing any voter in either party cares about.
Like, the Verizon CEO saying that the unemployment rate will be 20-30% in two years. If he is correct, we will have anarchy and civil war, and it's hard to see AI progress still chugging along at that point.
"It frequently mistakes correlation with causation and it is not able to course-correct for different hypotheses."
Ah, so only human level intelligence.
I feel like you may have adopted a rather loose definition of what would count as the "AI as a normal tech" view winning. Even if all three are true, I still wouldnt say AI is a normal tech:
"• AI’s impressive showing on benchmarks and coding will struggle to generalize to broader real-world settings.
• Even as AI develops genuinely useful capabilities, there will be (or should be) delays and limits on how it is adopted in practice.
• AIs will not develop radically superhuman abilities, because the real world is often too complex for abstract genius to eliminate the need for trial and error."