LLMs Unsafe by Definition? Debate on Stochastic Components and Agent Safety
gerardsans · x · 2026-08-25
Gerard Sans argues LLMs are unsafe by definition due to stochastic components that are unpredictable, fail silently, and accumulate errors unless cleared at each step. He cites Salesforce's AgentForce as an example from last year, noting the industry prioritizes selling over correctness. A reply counters that frontier models always follow instructions, but labs skip safeguards due to cost, not impossibility.
Related event: AI safety debate: the real risk is bad engineering, not rogue models(5 posts)→
More from AGI Musings
- 2023: the last year you could assume what you read was written by a human — david_perell · 2026-08-25
- Why AI productivity lags: the Productivity J-Curve paper on intangible complementary investments — robseamans · 2026-08-25
- Peter Diamandis: Fear of the future vs. builders — PeterDiamandis · 2026-08-25
- 'Has AI made anything non-digital cheaper?' A thread on labor productivity — BellaRudd1 · 2026-08-25
- To feel symbiotic with AI, physical input must become primary over language — mrjonfinger · 2026-08-25
- Closed Frontier vs Open-Weight Intelligence May Matter More Than OpenAI vs Google — VraserX · 2026-08-25