What does Alan Kay think about LLMs?
Skepticism about large language models runs through this exchange, centering on their tendency to confidently generate plausible but unverifiable or incorrect output and the resulting lack of “trustability” for anything beyond low‑stakes assistance. Commenters link Alan Kay’s long‑standing ideas—message passing, late binding, systems that are inspectable and explainable—to critiques of opaque, correlation‑driven models whose inner workings and training data are neither transparent nor auditable. Others counter that, despite real dangers like scaled misinformation and “strip‑mining” society, LLMs already outperform traditional tools in areas such as code generation and intent parsing, and that their real potential lies in carefully engineered, hybrid systems rather than as standalone question‑answering or teaching agents.
Overall stance on LLMs and “trustability”
- Central concern: LLMs are not “trustable” for running commands or teaching, because they don’t reason, only correlate and generate plausible text.
- Trust is linked to auditability: people want verifiable chains of reasoning, explainable decision paths, and reproducible results.
- Current LLMs can’t show how they derived an answer beyond vague token attributions, so they’re seen as unsuitable for critical tasks.
Message passing, late binding, and system design
- Large subthread revisits classic ideas: true message passing, late binding, and live, image-based systems (e.g., Smalltalk-style environments).
- Debate over whether modern systems (HTTP, microservices, browsers) embody these ideas well or are “bastardizations”.
- Some argue message passing significantly improves security and scalability; others note microservice messiness and need for stronger typing.
- There’s disagreement on whether Alan’s guidance is too vague or actually quite concrete when you look at his systems and research programs.
LLMs in programming and education
- Power users report LLMs as very useful but frequently wrong, especially in less popular languages or niche libraries.
- “Obvious” errors (invented APIs) are easy for experts to catch; subtle ones (deprecated, insecure, or inefficient patterns) are dangerous for learners.
- Concern that students will “cheat” through CS curricula with LLMs, further exposing how dated some teaching already is.
- Some like that traditional programming is literal and debuggable; they fear opaque AI layers undermine this transparency.
Epistemology: correlation, superstition, and BS
- Strong theme: LLMs exemplify “reasoning by correlation,” likened to superstition and BS generation, especially when they rationalize wrong answers fluently.
- Counterpoint: correlation-based empiricism can still be testable and useful; superstition arises when people misread or overinterpret correlations.
- Several comments note humans are also unreliable, biased, and prone to BS, so comparisons must be against actual human experts, not an ideal.
Societal and economic concerns
- Worry that LLMs will be used to “strip mine society” more efficiently, increasing extraction, surveillance, and large-scale manipulation.
- Others argue this dynamic predates LLMs and applies to most major technologies; LLMs are just a new lever.
- Additional fear: a future internet flooded with semi-plausible nonsense, eroding trust in digital information and possibly even literacy norms.