LLMs are good enough at physics that we’ll soon be able to use them to red-team Drexlerian nanotech and actually see if it’s feasible, settling the question once and for all (and long before we actually attempt to build it).
Some years ago, LW user Muireall found that GHz mechanical nano-computers probably don’t work [1], but it’s not clear if this objection transfers to other nanotech. I think that GPT-5.6 could, with a sufficient token budget, make a lower quality but still acceptable version of this analysis for the rest of Nanosystems, and within 6 months make something of the same quality but going into far more detail with minimal human effort.
[1] I believe this for a variety of reasons, happy to share.
Note that Merkle et. al (2025) have proposed atomically precise FETs, claiming that their design would allow for computers with 10^25 SOPs[1]/cm^3 (using only 10 watts[2], though at 10MHz). Transistor physics is a much more settled science than nanomechanics, and I did not see any glaring issues in their technical notice (though my background in physics is limited). However, when I did ask GPT-5.6 (high reasoning, not Pro) to check the notice, it stated that ‘several claims are physically incorrect or substantially more definite than the analysis permits,’ the modeling of the 10W computer ‘silently changes from ordinary switching to ideal adiabatic switching,’ and also that it may rely on some unresolved physics, though the proposed atomic layout is physically plausible.
I unfortunately do not have the physical knowledge to check GPT’s claims, though I may continue working on stuff similar to your proposed research direction[3].
switching operations per second; we can upper-bound the amount of FLOPs thereby from just figuring out how many transistor ‘switching operations’ are used in a single standard floating point calculation. Claude Sonnet 5 says this is generally in the order of magnitude of 10^2 − 10^3 SO for FP16 (with it being dependent on the floating-point precision and arch. obviously). So our sugar-cube computer can have at most ~10^22 − 10^23 FLOPs/cm^3 (using FP16), though overhead will almost certainly reduce this number, and likely by a lot.
(note that this requires reversible logic, since 10^-24 J/op is past the Landauer limit at room temperature (10^-21 J/op); see page 19 for more details)
(I was thinking about trying to see if current frontier LLMs could reduce the known smallest sizes of general-purpose self-replicating machines in several famous cellular automata rulesets, chiefly among them von Neumann’s cellular automata. For semi-obvious reasons we probably do want to have lower-bounds on the capability for (even current) LLMs to design general-purpose self-replicating machines, even in very simplified environments)
LLMs are good enough at physics that we’ll soon be able to use them to red-team Drexlerian nanotech and actually see if it’s feasible, settling the question once and for all (and long before we actually attempt to build it).
Some years ago, LW user Muireall found that GHz mechanical nano-computers probably don’t work [1], but it’s not clear if this objection transfers to other nanotech. I think that GPT-5.6 could, with a sufficient token budget, make a lower quality but still acceptable version of this analysis for the rest of Nanosystems, and within 6 months make something of the same quality but going into far more detail with minimal human effort.
[1] I believe this for a variety of reasons, happy to share.
Note that Merkle et. al (2025) have proposed atomically precise FETs, claiming that their design would allow for computers with 10^25 SOPs[1]/cm^3 (using only 10 watts[2], though at 10MHz). Transistor physics is a much more settled science than nanomechanics, and I did not see any glaring issues in their technical notice (though my background in physics is limited).
However, when I did ask GPT-5.6 (high reasoning, not Pro) to check the notice, it stated that ‘several claims are physically incorrect or substantially more definite than the analysis permits,’ the modeling of the 10W computer ‘silently changes from ordinary switching to ideal adiabatic switching,’ and also that it may rely on some unresolved physics, though the proposed atomic layout is physically plausible.
I unfortunately do not have the physical knowledge to check GPT’s claims, though I may continue working on stuff similar to your proposed research direction[3].
switching operations per second; we can upper-bound the amount of FLOPs thereby from just figuring out how many transistor ‘switching operations’ are used in a single standard floating point calculation. Claude Sonnet 5 says this is generally in the order of magnitude of 10^2 − 10^3 SO for FP16 (with it being dependent on the floating-point precision and arch. obviously). So our sugar-cube computer can have at most ~10^22 − 10^23 FLOPs/cm^3 (using FP16), though overhead will almost certainly reduce this number, and likely by a lot.
(note that this requires reversible logic, since 10^-24 J/op is past the Landauer limit at room temperature (10^-21 J/op); see page 19 for more details)
(I was thinking about trying to see if current frontier LLMs could reduce the known smallest sizes of general-purpose self-replicating machines in several famous cellular automata rulesets, chiefly among them von Neumann’s cellular automata. For semi-obvious reasons we probably do want to have lower-bounds on the capability for (even current) LLMs to design general-purpose self-replicating machines, even in very simplified environments)