Gaps

The developers of one pipeline report experimental success rates of 10 to 100% depending on the target. An earlier study found the overall design success rate low and raised it nearly 10-fold with deep-learning filtering. An independent group testing another method on six targets, five designs each, reported that most targets gave no working binder.

Held open by: AI for science

Small batches leave the arithmetic units idle while the memory system is saturated. The measured token energy falls from 7.46 to 0.72 J/token as output length grows from 10 to 512 tokens at batch 16, because fixed costs are amortized; batching gains shrink as context grows (6.31x at 512 tokens of context against 1.17x at 4K for 10 output tokens).

Held open by: Low-energy data movement

Production energy per prompt includes idle machine capacity and data-center overhead, not only active accelerator power. Google reports a 33x reduction in energy per median Gemini Apps text prompt over one year from software efficiency and clean-energy procurement (the abstract states the combined effect on energy and does not separate the two). Further gains depend on batching, routing and model choice.

Single pairs of superconducting qubits reach two-qubit gate fidelity of 99.94% (error 6e-4) with stability over 24 days, but a million-qubit machine needs every pair near that level at once, with leakage and crosstalk held down. The Willow memory ran at 0.143% logical error per cycle at distance 7, short of the target by about nine orders of magnitude. Closing it means both lower physical error and larger code distance on many more qubits.

Approaches

• Fluxonium qubits with direct two-qubit gates (TRL 3) lin2025days

Muons and gamma rays create quasiparticle bursts that cause correlated errors across a chip, which error correction cannot absorb. A measurement on a 63-qubit processor separated the contributions of muons and gamma rays. Back-side phonon downconversion cut correlated poisoning by two orders of magnitude on three-qubit chips; it has not been shown on a million-qubit array.

Approaches

• Phonon downconversion with back-side normal-metal reservoirs (TRL 4) iaia2022phonon

The Willow decoder averaged 63 microseconds at distance 5, six times the 10 microsecond reaction time assumed in the RSA-2048 estimate, and at a cycle time of 1.1 microseconds, far from 1 microsecond at thousands of logical qubits. FPGA decoders report tens to hundreds of nanoseconds per measurement round in modeled-noise studies, which is not the same quantity as end-to-end latency in a running experiment.

Held open by: Real-time quantum error-correction decoder

Approaches

• Distributed Union-Find decoder on FPGA (TRL 3) liyanage2024fpga

Each wired qubit adds passive heat load from cables and active load from signal dissipation. The strongest dilution refrigerator in the evidence delivers 2 mW at about 100 mK with a base temperature of 6.6 mK; no source in the atlas states the load of a million-qubit machine. Resource models of modular machines predict power and thermal load, and point to splitting the machine across cryostats.

Held open by: Dilution refrigeration

Approaches

• Modular machine split across several cryostats (TRL 2) saadatmand2026superconducting

A multinode machine needs entanglement between cryostats. Internode gates may be two to three orders of magnitude noisier and slower than local operations, and a systems analysis finds link performance must improve by 10 to 100 times.

Held open by: Quantum interconnect

Approaches

• Optical interconnects with entanglement distillation (TRL 2) ang2024arquin

The 2023 demonstration decoded 62 words per minute, which the authors describe as beginning to approach the speed of natural conversation, 160 words per minute. The result comes from one participant, so how it generalizes across people is not established.

Held open by: Long-lived neural electrodes

Reusable first stages reduced launch cost, and providers on several continents plan reusable systems. The cost per kilogram falls with cumulative payload launched, following a learning curve, so further reductions depend on reuse at higher rate and with more of the vehicle recovered.

Reversible reverse osmosis and electrodialysis consume the Gibbs free energy of separation, and the practical energy of both approaches that minimum only as the number of stages grows. Reviews state that most desalination technologies already work near their limit, so the remaining reduction is bounded and each step costs capital. The numeric value of the minimum is in the body of the cited papers and is not recorded here yet.

Reverse osmosis membranes trade water permeability against salt rejection. Analysis of biomimetic membranes indicates what a more permeable, defect-free membrane could offer for seawater desalination, but such membranes are not yet a commercial product.

Held open by: High-selectivity membranes

The NHS-Galleri randomized trial of 142,250 participants reported that its primary endpoint, a reduction in stage III/IV diagnoses in the test arm, was not met. Detecting cancers is not the same as improving outcomes; longer follow-up and other trials are needed.

Legumes host nitrogen-fixing rhizobia in root nodules. The objective of engineering nodulation in non-leguminous crops has not been achieved; the open questions are the signalling, infection and nodule-organogenesis programs.

Held open by: Plant genome engineering