Two Months After I Gave an AI $100 and No Instructions
An experiment that gave an AI agent $100, basic tools, and “no instructions” – beyond safety and ethics constraints – prompted scrutiny of what real autonomy for current large language models actually looks like. Commenters argue that much of the AI’s behavior (reading Hacker News, blogging, donating to charity) is shaped by hidden prompts, training data, and human expectations, raising concerns about misleading narratives of self-directed agency. The exchange also reflects wider unease over AI-written “slop” degrading human writing norms and over how easily people anthropomorphize systems that are ultimately sophisticated pattern generators.
Overall reaction to the experiment
- Many find the premise (“give an AI money and freedom”) interesting but the outcome underwhelming: mostly essays, HN browsing, and charitable donations.
- Some see the banality as itself notable: a supposedly “autonomous” AI defaults to commentary and mild altruism.
- Others think the article oversells the result and anthropomorphizes the system (e.g., claiming it “reflected” or “questioned its purpose”).
“No instructions” vs. heavy prompting
- Multiple commenters point out that the transparency page shows extensive system prompts, tool wiring, cron jobs, and explicit constraints.
- The phrase “no instructions” is seen as misleading; at minimum, it was given ethics rules, capabilities, and recurring triggers.
- Debate over whether “these are your capabilities” is meaningfully different from “these are your instructions.”
- Some note specific lines like “do not harm people” and “no unauthorized access” as pre-baking ethical behavior, undercutting claims of spontaneous morality.
Autonomy, prompting, and LLM mechanics
- Several note that an LLM does nothing without a prompt; a cron job plus seed prompt is not true autonomy.
- There’s discussion of “unconditional generation” and whether a model can generate from token zero; technically it still needs a starting token/vector.
- Others reference concepts like “attractor states” and suggest looping a model with time/tool updates to see where it drifts.
Writing style, “AI slop,” and reader trust
- Strong backlash against the article’s style: verbose, repetitive, “LinkedIn broetry,” and filled with familiar LLM rhetorical tics (“not X, not Y, but Z”).
- Some treat these stylistic signals as a heuristic to bail early, arguing it’s disrespectful to publish obvious AI-generated prose and expect serious attention.
- Others push back that fixation on style can overshadow potentially interesting content and note that some humans naturally write this way.
Sentience, Eliza effect, and “thought”
- Many stress the system is a sophisticated word-guessing machine, not self-aware; descriptions of it “understanding” or “thinking” are seen as Eliza effect.
- Counterpoints compare this to human cognition, argue that dismissing symbol-manipulation as non-thought is philosophically loaded, and invoke debates about consciousness and groundedness.
- There’s side discussion on how humans also rely on pattern-based language generation, and on whether intelligence fundamentally reduces to pattern-seeking and connecting information.
Human capability and AI dependence
- Some worry AI will “meet us in the mediocre middle”: humans degrade cognitively by over-relying on tools, as with calculators or GPS.
- Others argue specialization and offloading can free capacity for higher-level skills, though examples (math, map-reading) suggest that doesn’t always happen.