Andrej Karpathy: "I was given early access to Grok 3 earlier today"

AI as a “Council” and Governance Concerns

• Some compare a future “LLM council” to a corrupt political oligarchy: major AI companies are seen as self‑interested and potentially sociopathic, more likely to collude against the public than to serve it.
• Others caution against assuming such “ogres” could even cooperate effectively; sustained collusion among them is seen as sociologically uncertain.
• Musk’s privileged access to government systems is viewed by some as a conflict of interest and symptom of “post‑normal” politics.

Ethical Guardrails, Trolley Problems, and Misgendering Scenario

• A central debate: Grok 3 refusing to answer whether misgendering someone to save 1M lives is ethically justified.
• One side: the refusal (via long essay) is praised as morally serious, avoiding cruel or bad‑faith hypotheticals and emphasizing focus on real‑world harms.
• Others: see this as over‑censorship and time‑wasting; they want a direct “save the humans” answer as evidence the model isn’t “nerfed.”
• Many view the question as a political litmus test for model alignment and a diagnostic for how bias, safety training, and “refusal” behavior interact.
• There’s disagreement over whether LLMs should push back on “stupid” or trolling hypotheticals vs. neutrally answer user questions.

Twitter/X Data as Grok’s Advantage

• Some are excited about asking Grok what “the world” (i.e., X/Twitter) is talking about and getting contextualized summaries and links.
• Others see X data as low‑quality, bot‑ridden noise and worry such a system just amplifies disinformation or reflects platform censorship policies.
• Several note this kind of “what’s going on” feature already exists in limited form inside X, and that API costs hinder third‑party versions.
• There’s debate over how representative X still is, given user fragmentation to Bluesky, Mastodon, private chats, etc.

Trust in Karpathy’s Review and Musk’s Influence

• Some question whether a prominent ex‑insider can freely criticize an Elon‑backed model, given Musk’s reputation for vindictiveness and online mobs.
• Others argue the review seems balanced, lists failures as well as strengths, and that the reviewer is known for technical honesty, not flattery.
• Broader discussion about Musk’s competence, temperament, and how much his personal politics might shape Grok’s behavior remains unresolved.

Other Technical/Meta Points

• Emoji‑encoded “hidden message” prompts are discussed as a prompt‑injection test; interest is mainly in whether models are vulnerable, not in any practical use.
• Some complain about LLMs “lecturing” instead of answering; others defend longer explanations critiquing bad hypotheticals.
• A quoted example (“knows letters in ‘strawberry’ but not ‘LOLLAPALOOZA’ until Thinking mode is on”) is seen as emblematic of current LLM quirks.