Ex-Huawei researcher: AI is great at small-step optimization, but taste and system design remain out of reach
yangyi · x · 2026-09-16
Former Microsoft/Huawei researcher Bojie Li posted a long thread on how technology diffuses far slower than experts expect, extending it to AI's real limits:
- Azure was already using RDMA + SmartNIC-accelerated storage in 2016-2018, yet today many foundation-model companies still take minutes to load multi-TB models without RDMA;
- Polyhedral compilation from his 2019 Ascend operator work AKG was never widely adopted by compiler stacks — LLMs, not mathematical optimization, ended up replacing handwritten operators;
- He expected AI to do infra optimization after 2023, but operator optimization, scheduling and inference-system tuning remain deep problems: AI excels at goal-directed incremental optimization, while writing, taste and system design stay poor — he still has to constantly guide GPT-6 Astra to get a decent system design;
- Adoption of good AI tools remains low; many users can't even exhaust a $20 Codex plan.
His conclusion: talk of AI "replacing" engineers is premature until taste can be turned into a verifiable reward.
More from AGI Musings
- Agent monitoring is a precondition for safe harnesses: kill CLI, perfect sandbox, or trace surveillance — scottleibrand · 2026-09-16
- Faked alignment plus recursive self-improvement: firms can't tell real alignment from theater — birchlse · 2026-09-16
- Altman warns of open-source cyberattack wave; Byrnes backs selling frontier defense — beffjezos · 2026-09-16
- Veteran virus engineer rebuts claims that AI lets amateurs build bioweapons — aiamblichus · 2026-09-16
- WhatsApp head: AI agents threaten to disrupt India's BPO sector and bank lending — rohanpaul_ai · 2026-09-16
- ApprenticeBench claims AI agents surpass humans on real jobs; Liang Wenfeng calls it 'prompt engineering bullshit' — ysu_nlp · 2026-09-16