Attention Bottlenecks Were Optimized Long Ago
BlackHC · x · 2026-07-16
The author refutes the claim that "the attention mechanism is terrible," pointing out:
- The quadratic complexity of vanilla attention scales with context length, not user count
- In an 8k token chat scenario, attention accounts for only about 14% of a 70B model's FLOPs
- FlashAttention optimized memory complexity to linear back in 2022
- This is already the default in almost every serving stack
The point is that these engineering challenges were solved long ago, and it's inaccurate to treat outdated assumptions as current reality.
Related event: BlackHC Rebuts the “AI Is the Worst Technology” Thesis(5 posts)→
More from Infra
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- Gary Marcus: With Q3 data added, Anthropic's compute-efficiency growth shows deceleration — GaryMarcus · 2026-10-27
- SpaceX locks in $40B GPU financing deal with Apollo, the largest ever committed — TechCrunch AI · 2026-10-27
- Inference is just Kubernetes with a multi-trillion-dollar market, and the communities overlap — matt_slotnick · 2026-10-27
- vLLM eyes Rust migration as inference engines race to rewrite the stack — AccBalanced · 2026-10-27