Stuff we figured out about AI in 2023
Large language models in 2023 are prompting people to rethink both artificial and human intelligence: how far text-only systems can go, what they reveal about multilayered human cognition, and whether digital machines can ever fully approximate analog brains. Commenters debate technical and philosophical questions such as gullibility in models, the limits of transformer architectures, the meaning of “AI” versus “AGI,” and whether current systems truly “understand” or are just powerful tools. Others focus on practical impacts, from code generation and creative work to how LLMs expose cultural biases and may reshape social sciences and our sense of humanity’s uniqueness.
Human vs. Machine Intelligence
- Several commenters argue LLMs reveal how much human cognition depends on language, but also how much more there is (visual thinking, motor skills, emotion, multi-modality).
- Some describe brains as multilayered: something LLM-like plus world-physics models, logic, and symbolic reasoning.
- Personal stories (brain injury, ADHD, disabilities) are used to contrast human flexibility and compensation with LLMs’ narrow text-based abilities.
- Others note that much everyday human discourse already resembles “LLM-like continuation” of well-worn scripts.
Analog vs Digital, Computation Limits
- One side claims human intelligence is fundamentally analog and beyond perfect digital emulation, invoking Cantor’s diagonal argument and a “cardinality barrier.”
- Opponents respond that digital systems can approximate physical (computable real) processes arbitrarily well; the issue is efficiency, not possibility.
- There is disagreement on whether brain-like intelligence is practically simulable or only in-principle.
Gullibility and Truth
- Some think gullibility can be mitigated by better training (counterfactual data, RAG, fact-checking behavior).
- Others argue transformers don’t truly understand meaning, pointing to inconsistent behavior across languages and sensitivity to frequency of contradictory training data.
- There’s broad agreement that humans are also gullible, making “ideal robustness” a very hard target.
Complexity, “Black Box,” and Intelligence
- Core LLM architectures are noted to be surprisingly small (hundreds of lines) once you assume existing stacks, though critics say this obscures the huge surrounding ecosystem.
- Debate over whether LLMs are “black boxes”: mathematics and code are understood, but emergent intelligence-like behavior is not fully explainable.
- Some insist LLMs are just powerful tools, not “smart,” drawing analogies to chess engines that play inhuman, “garbage” styles despite being unbeatable.
AI vs AGI, Terminology and Social Impact
- Ongoing disagreement over whether “AI” should cover ML/LLMs or be reserved for human-level/AGI; many note AI has historically been broad.
- LLMs are seen as mirrors of online culture: raw training data vs. post-hoc guardrails highlight societal conflicts and censorship desires.
- Discussion touches on Turing tests, philosophical zombies, and whether current systems truly challenge human exceptionalism or mainly expose limits of past concepts of intelligence.
LLMs for Coding and Work
- Some warn that “just run the code to see if it’s correct” is misleading; mental models and careful review remain essential.
- Others report LLMs significantly aid explanation, refactoring, and test-writing, but emphasize quality over raw speed.