BlackHC Rebuts the “AI Is the Worst Technology” Thesis

On July 16, BlackHC published a thread systematically rebutting an article that labeled AI “the worst technology.” He argued the article’s case rests on outdated assumptions about inference cost, model scaling, and system bottlenecks, making its conclusions about AI economics and scalability unreliable.

Core arguments

BlackHC said inference costs for GPT-3.5-level performance fell from $20 per million tokens in November 2022 to $0.07 in October 2024, a roughly 286x drop. He also argued that frontier progress is not coming purely from ever-larger models: recent models are becoming smaller while also getting stronger and cheaper, which he presented as evidence that efficiency gains are materially improving the economics.

He further cited a claim that inference revenue is growing at roughly 3x per year even as models themselves become smaller and cheaper. In his view, the original article cited that source but drew the opposite conclusion from it.

Rebuttal on technical bottlenecks

Against the claim that attention is fundamentally bad, BlackHC said the quadratic complexity of vanilla attention scales with context length, not with the number of users. For an 8k-token chat workload, he estimated attention is only about 14% of the FLOPs in a 70B model. He also pointed to optimizations such as FlashAttention as evidence that this bottleneck has already been significantly improved in practice.

Dispute over chart interpretation

BlackHC also said the article misread an Epoch AI chart. According to his explanation, the x-axis measures tokens per second for a single request, not overall system throughput. He argued that making one request faster often requires smaller batching on the server, which raises unit cost; that trade-off cannot be directly used to infer how many requests a system can serve at the same speed.

2026-07-16 ~ 2026-07-16 · 5 related posts