>>/202642/
Claude thinks there are major difficulties with it:
"""
The analog optical computer is an impressive proof of concept, but it is a long way from practical use. The working hardware has only 256 weights, about five to seven orders of magnitude short of real workloads. Getting there depends on miniaturized 3D optical modules and integrated analog electronics that haven't been built yet. Its precision is limited to roughly 8–9 bits, signed weights introduced enough error to halve its capacity, and real latency is dominated by microsecond-scale settling and readout rather than the 20 ns loop. Its most striking results, such as beating Gurobi and reconstructing a 200,000-variable brain scan, came from a digital simulation of its algorithm running on a GPU, not from the optical hardware. The 500 TOPS/W efficiency figure is a projection for a system that doesn't yet exist.
Large language models face an additional, more basic mismatch. The AOC works by keeping fixed weights in a slow-to-reprogram light modulator and looping on them until a fixed point is reached. Transformers instead pass through each layer once per token. Attention, their core operation, multiplies activations by other activations that change with every token and grow with context length, so it can't be stored as fixed weights. At best the optics could handle the weight-heavy projection and feed-forward layers, while digital hardware does attention, softmax and normalization. That brings back the analog-digital conversions that are the source of the claimed efficiency. Frontier LLMs are also far larger than the 0.1–2 billion weights the roadmap envisions. The AOC becomes relevant to LLMs only if architectures shift toward recurrent-depth or other iterative designs that reuse the same weights many times.
"""