AI is likely to control quantum computers first, then use them for narrow science tasks
imjustnewatai · x · 2026-07-28
A three-stage AI × quantum roadmap to 2040
The post argues the most likely path is a relay: AI first makes quantum computers reliable, then QPs are used for a small set of hard scientific calculations, and finally AI learns from those results to design better technology.
Core claims
- Quantum computers do not brute-force every answer; their speedups only apply to problems with the right structure.
- The next frontier LLMs will still train on GPUs/TPUs, because shuttling internet-scale data and trillion-parameter models into qubits would erase much of the advantage.
- The real crossover is already happening in the other direction: AI is learning to operate quantum hardware.
Timeline proposed
- 2026–2030: AI becomes standard in quantum control, calibration, error decoding, circuit optimization and fabrication; post-quantum cryptography migration is the first broad impact, but QPUs do not yet lower frontier LLM costs.
- 2030–2035: early fault-tolerant systems become a scientific instrument for narrow questions about molecules, catalysts and materials; AI uses them as a high-cost oracle inside a classical search loop.
- 2035–2040: the loop becomes AI proposing materials, QPUs evaluating the rare hard cases, robots building and testing candidates, and the results feeding the next round of search.
Concrete numbers and examples cited
- Google’s Willow work is described as RL controlling 1,000+ parameters, cutting logical errors by 20% and making the system 3.5× more stable against drift.
- The post cites a gap between Google’s logical-memory error rate of about 7.72 × 10⁻⁴ per cycle and the much lower error rates needed for long computations.
- It also references NVIDIA’s GPU-QPU link with under 4 microseconds of round-trip latency.
The overall thesis: a QPU will likely become a specialized scientific instrument inside an AI supercomputer, not a replacement for GPU-based frontier model training.
More from AGI Musings
- Blog post argues inference is synthesized learning about people, not token generation — Scobleizer · 2026-07-28
- AGI Will Replace All Human Jobs by Year-End — davidpattersonx · 2026-07-28
- LLM authorship credit really bundles tracing, attribution and ownership — voooooogel · 2026-07-28
- Why patients turn to chatbots first: access and trust, not just misinformation — EricTopol · 2026-07-28
- Trask says the open-vs-closed debate deserves nuance, but calls out Dario’s double-speak — iamtrask · 2026-07-28
- Joseph Jacks says open-weight models may now be only 1–3 months behind frontier AI — JosephJacks_ · 2026-07-28