I do not share your utopian optimism on AI, for a pragmatic reason. As far as policy is concerned, it makes sense to adjust downward any projected growth, to be prepared for difficulty. I will be happy to be wrong, but I would rather strap myself in for another decade of <1% YoY economic growth than abandon the political struggle with the assumption that everything will be dissolved into AGI. Having said that, my blog isn't an effectual way to create change so it's a bit of a cargo cult or LARP, so there is a very small marginal cost to your attitude. But if I did have power I think my attitude would be justified. Your attitude is healthier for individual growth and prosperity in either case.
I know nothing about technology. The expert consensus on AI seems to be "it will improve exponentially and rapidly." However, that was also the expert consensus on internet streaming in 2000, and it was about 7 years too early. I think there is little cost to being a "delayer", assuming that AI predictions will stall out before reaching the critical mass necessary for singularity. I'm not going to be mad if it happens, just not banking on it. I like to be pessimistic as a risk-avoidance strategy. I wouldn't know how to invest in AI if it's real.
AI capabilities have improved consistently and rapidly in recent years. Check out https://epoch.ai/trends for an overview. At current scaling rates compute grows 4.7x/year, algorithmic efficiency improves 3x/year, and cost for a given performance drops 10x/year. It's estimated progress can continue for at least another 4 years at this rate before bottlenecks are hit.
Arguing for a plateau is, imo, more exceptional than expecting continued growth.
Few benchmarks remain where AI underperforms humans: desktop navigation (video-gaming/agency tasks) and ARC-AGI. Yet these gaps are closing fast:
o3-mini (Jan 31): 22.3%
o4-mini (Apr 16): 41.8% (a 20-point jump in three months)
AI is now tackling superhuman challenges like the “Humanity's Last Exam” and “Frontier Mathematics”-and historically it doesn’t take long before benchmarks are saturated (https://tinyurl.com/yus6wnu3).
You might already know this, but AIs have their own 'g'. That is performance on one benchmark predicts performance on all other benchmarks - even ones the AI couldn't have trained on. This implies AIs are also getting better at AI research. It's pretty universally agreed that the AI labs are explicitly using AI to automate their own progress now. This isn't speculation, they're literally just saying that they're doing this. This is important for obvious reasons, it creates an intelligence enhancement feedback loop.
Singularity isn’t imminent this year or next, but within 3-5 years, white-collar work will likely be disrupted or revolutionized. You'll start seeing AIs make their own startups, education will be a mess with mass AI cheating, etc. Things will start to get weird and unpredictable. I think AIs will be smarter than all humans in most meaningful ways in the next decade.
If you want to invest in this, Google, Microsoft and NVIDIA are good places.
Also: TrackingAI frontier LLM performance on IQ test that doesn't exist online (so not in training data).
This time last year: ~90
Today: 115 (o3 and Gemini 2.5)
This matches the competence levels I have seen from them. If this trend continues another year, we will have 140 IQ frontier models by mid-2026. If we also posit that they are strongly agentic by that point you have an arbitrarily expandable swarm of Ivy League STEM PhD level researchers who will proceed to solve the software space in the late 2020s.
Good idea. I get plenty of Anatoly in text. To little in audio. I like
I do not share your utopian optimism on AI, for a pragmatic reason. As far as policy is concerned, it makes sense to adjust downward any projected growth, to be prepared for difficulty. I will be happy to be wrong, but I would rather strap myself in for another decade of <1% YoY economic growth than abandon the political struggle with the assumption that everything will be dissolved into AGI. Having said that, my blog isn't an effectual way to create change so it's a bit of a cargo cult or LARP, so there is a very small marginal cost to your attitude. But if I did have power I think my attitude would be justified. Your attitude is healthier for individual growth and prosperity in either case.
Beautiful background.
What makes you think "AGI"/singularity isn't coming?
I know nothing about technology. The expert consensus on AI seems to be "it will improve exponentially and rapidly." However, that was also the expert consensus on internet streaming in 2000, and it was about 7 years too early. I think there is little cost to being a "delayer", assuming that AI predictions will stall out before reaching the critical mass necessary for singularity. I'm not going to be mad if it happens, just not banking on it. I like to be pessimistic as a risk-avoidance strategy. I wouldn't know how to invest in AI if it's real.
AI capabilities have improved consistently and rapidly in recent years. Check out https://epoch.ai/trends for an overview. At current scaling rates compute grows 4.7x/year, algorithmic efficiency improves 3x/year, and cost for a given performance drops 10x/year. It's estimated progress can continue for at least another 4 years at this rate before bottlenecks are hit.
Arguing for a plateau is, imo, more exceptional than expecting continued growth.
Few benchmarks remain where AI underperforms humans: desktop navigation (video-gaming/agency tasks) and ARC-AGI. Yet these gaps are closing fast:
o3-mini (Jan 31): 22.3%
o4-mini (Apr 16): 41.8% (a 20-point jump in three months)
OS-World (AI using computers): 5% → 42% (Apr 2024 → Apr 2025)
See:
https://arcprize.org/leaderboard
https://os-world.github.io/
AI is now tackling superhuman challenges like the “Humanity's Last Exam” and “Frontier Mathematics”-and historically it doesn’t take long before benchmarks are saturated (https://tinyurl.com/yus6wnu3).
You might already know this, but AIs have their own 'g'. That is performance on one benchmark predicts performance on all other benchmarks - even ones the AI couldn't have trained on. This implies AIs are also getting better at AI research. It's pretty universally agreed that the AI labs are explicitly using AI to automate their own progress now. This isn't speculation, they're literally just saying that they're doing this. This is important for obvious reasons, it creates an intelligence enhancement feedback loop.
Singularity isn’t imminent this year or next, but within 3-5 years, white-collar work will likely be disrupted or revolutionized. You'll start seeing AIs make their own startups, education will be a mess with mass AI cheating, etc. Things will start to get weird and unpredictable. I think AIs will be smarter than all humans in most meaningful ways in the next decade.
If you want to invest in this, Google, Microsoft and NVIDIA are good places.
Also: TrackingAI frontier LLM performance on IQ test that doesn't exist online (so not in training data).
This time last year: ~90
Today: 115 (o3 and Gemini 2.5)
This matches the competence levels I have seen from them. If this trend continues another year, we will have 140 IQ frontier models by mid-2026. If we also posit that they are strongly agentic by that point you have an arbitrarily expandable swarm of Ivy League STEM PhD level researchers who will proceed to solve the software space in the late 2020s.
This is pretty cool. I thought I missed out on NVIDIA.
Do you still think Ethereum is the long term winner?
Yes - I think if crypto has a future beyond "store of value"/"digital gold" (Bitcoin) then Ethereum will play the central role in it.
Crypto question - What do you think will reverse the grizzly ETHBTC chart trend in the near term?