Evalöf argues AI stochasticity prevents full coding automation
Alex Evalöf contends that large language models cannot replace human accountability in high-stakes software due to their stochastic nature. He advocates for a risk-based approach to AI adoption, noting that current tools lack deterministic reasoning.
Alex Evalöf, author of 'Reliability Engineering Mindset' and former Senior Staff Engineer at Volvo, Klarna, Discovery, and Sheepstead, argued that coding is not fully solved because large language models are fundamentally stochastic engines. He explained that these models predict the next token via autoregressive architecture, lacking the deterministic reasoning required for high-stakes applications. Evalöf highlighted a key distinction: human engineers face accountability through performance reviews, legal penalties, or imprisonment, whereas AI has no feelings and cannot be held accountable.
Evalöf maintained that while AI can handle low-risk tasks such as personal software or proof-of-concepts, high-stakes software requires human oversight. He noted that current improvements in AI are incremental, involving better data, larger contexts, faster GPUs, and better harnesses, but the models still do not truly reason or learn in real-time. In his own workflow, Evalöf uses DeepSeek V4 Flash for coding, citing a cost of $10 per week and speeds over 200 tokens per second. He also uses Obsidian to manage context to reduce hallucinations.
The engineer criticized the degradation of software quality in major products like GitHub and Google, attributing it to excessive AI usage without proper safeguards. He warned that treating AI-generated code as a black box is risky for critical systems. Evalöf advocated for a pragmatic, risk-based approach to AI adoption, emphasizing that humans must remain accountable for the final output.