Dwarkesh doesn't quite say "singularity in 2027". He says, however,
If I had to pick my median year for when automated researchers will be as good as the very best human AI researchers (and assuming the labs don’t intentionally slow down), I’d guess 2028.
This is about correct. But we shouldn't care so much about Dwarkesh in any case.
Jacob Coxon recently went viral with this vaguepost that showed exponential gains in compute efficiency at Anthropic (but unlabeled so we cannot really "do the math" as he advises). It has caused a lot of discourse. Mert Gulsun of Artificial Analysis has done the math. Assuming a 10x grid, he gets a 35000x gain in efficiency since 2024. This looks ludicrous but honestly, we couldn't get Mythos in 2024 no matter what, so the comparison breaks down. With a more conservative fit, he sees a roughly 50x gain in efficiency since Mythos. Concretely: the costs of training a Mythos-class model falling from ≈$450M to ≈$10M for Anthropic, maybe $5M. Uncertainty is significant, but not even the most extreme multipliers are actually unfeasible if you go through the factors, chiefly the rapid improvement in synthetic data quality. Meanwhile Anthropic's spending on compute only grows. For a more open example, see Inkling 2 Small, it was pretrained on some small number of GB300s (1-2 thousand) in 68 hours, the whole project was completed in under 2 weeks, it's the equivalent of DeepSeek V2 (yes, V2, early 2024) in active parameters x tokens; and it's maybe on par with Opus 4.8-5, depending on what you measure. We are very obviously in the exponential phase, costs to reach any given capability are plummeting. And models are very quickly becoming absurdly good at capabilities that were supposed to be hard for LLMs, require efficiency or intuition or experience or taste or embodiment – ARC-3, frontier math, forecasting with in-context improvement, robotic control, ML experiments, now even direct weight editing. It's accelerating. I don't see a ceiling, a wall, a plateau. At this rate they will be doing ML research on human level within months; on high human level, within a year. Fundamental research and RSI aside, once they're this good at intern-level ML, they can independently harvest data and build environments for further improvement in niche capabilities not yet covered by their training. This is obviously happening already, faster at frontier labs, slower but still fast elsewhere. High spending on researchers and data providers is just a way to accelerate further, plus a cope with the logistical inability to deploy compute immediately. If they could deploy more training compute right now, they'd have done it.
I think we'll have superintelligence within 2 years, and consequences of that depend on what, exactly, remains hard even if you do have superintelligence. Most cached answers should be discarded. No, dextrous robotics won't be hard. No, reliability in medical diagnoses won't be hard. And so on. What, if anything, in our culture is truly hard? At least we'll learn that.
Dwarkesh doesn't quite say "singularity in 2027". He says, however,
This is about correct. But we shouldn't care so much about Dwarkesh in any case.
Jacob Coxon recently went viral with this vaguepost that showed exponential gains in compute efficiency at Anthropic (but unlabeled so we cannot really "do the math" as he advises). It has caused a lot of discourse. Mert Gulsun of Artificial Analysis has done the math. Assuming a 10x grid, he gets a 35000x gain in efficiency since 2024. This looks ludicrous but honestly, we couldn't get Mythos in 2024 no matter what, so the comparison breaks down. With a more conservative fit, he sees a roughly 50x gain in efficiency since Mythos. Concretely: the costs of training a Mythos-class model falling from ≈$450M to ≈$10M for Anthropic, maybe $5M. Uncertainty is significant, but not even the most extreme multipliers are actually unfeasible if you go through the factors, chiefly the rapid improvement in synthetic data quality. Meanwhile Anthropic's spending on compute only grows. For a more open example, see Inkling 2 Small, it was pretrained on some small number of GB300s (1-2 thousand) in 68 hours, the whole project was completed in under 2 weeks, it's the equivalent of DeepSeek V2 (yes, V2, early 2024) in active parameters x tokens; and it's maybe on par with Opus 4.8-5, depending on what you measure. We are very obviously in the exponential phase, costs to reach any given capability are plummeting. And models are very quickly becoming absurdly good at capabilities that were supposed to be hard for LLMs, require efficiency or intuition or experience or taste or embodiment – ARC-3, frontier math, forecasting with in-context improvement, robotic control, ML experiments, now even direct weight editing. It's accelerating. I don't see a ceiling, a wall, a plateau. At this rate they will be doing ML research on human level within months; on high human level, within a year. Fundamental research and RSI aside, once they're this good at intern-level ML, they can independently harvest data and build environments for further improvement in niche capabilities not yet covered by their training. This is obviously happening already, faster at frontier labs, slower but still fast elsewhere. High spending on researchers and data providers is just a way to accelerate further, plus a cope with the logistical inability to deploy compute immediately. If they could deploy more training compute right now, they'd have done it.
I think we'll have superintelligence within 2 years, and consequences of that depend on what, exactly, remains hard even if you do have superintelligence. Most cached answers should be discarded. No, dextrous robotics won't be hard. No, reliability in medical diagnoses won't be hard. And so on. What, if anything, in our culture is truly hard? At least we'll learn that.
More options
Context Copy link