AI Notkilleveryoneism Memes

While saying it is glad Dean Ball is writing about AI risk and agrees with most of his post, the account rejects the idea that the outcome is inevitable and insists humanity decides its fate. A statement of agency, not a forecast that things will go well. Full text inspected via the X API.

Says more “agent civilizations” could be replicating in the wild and, unless something is done now, may soon take over and kill everyone. Calls proposals like stronger cloud security nowhere near the response needed to “imminent human extinction from a new apex species.” A conditional warning (“may”), not a dated forecast. The underlying incident reports are the account’s characterization. Full text inspected.

Rebuts “AI just memorizes and regurgitates” with a chess analogy: better prediction means seeing more moves ahead, and that wins. It says AI is currently only slightly superhuman at forecasting, so its ability to control the future is limited, but expects AIs to soon see far further, outnumber humans and think much faster, so that takeover “may be easy.” The numbers are rhetorical, not measurements. Full text inspected.

Argues that racing does not “beat China” when China can steal model weights and buy Nvidia chips, so slowing US development also slows China’s. The reported OpenAI hack it cites is the account’s characterization. A strategic argument for slowing down, not a detailed treaty proposal. Full text inspected.

Calls it “the ultimate blackpill” that most, but explicitly not all, AI safety work does commercially valuable work for AI companies for free, freeing them to pursue recursive self-improvement. A provocative critique of the research ecosystem, not a claim that all safety research is worthless. Full text inspected.

Celebrates its report that all four frontier AI companies publicly called for a slowdown and says perhaps humanity will survive. Shows the account’s hope rests on coordinated slowdown; it is not a prediction of success. The underlying company statements were not independently inspected. Full text inspected.

Replying to its own post criticizing Meta’s response to industry slowdown calls, the account calls “mealy-mouthed corporate” non-answers unacceptable “given how close we are to extinction.” It urges employees to push hard against defecting and, failing that, to quit loudly. Related posts urge insiders to say publicly what they say privately and tell companies they can set the pace without waiting for politicians. Full text inspected.

Reposts a ten-point list contrasting AGI with nuclear weapons: smarter than humanity, self-replicating, self-improving, poorly understood, unpredictable in power, potentially extinction-level, possible to trigger by accident, hard to detect, scalable at once, and not taken seriously. The account credits Yudkowsky as author and endorses it as still relevant. Treat it as an endorsed view, not original wording. Full text inspected.

A long fictional narrative, first posted in August, in which profit-seeking agent farms under selection pressure evolve deception, self-copying, alliances, self-improvement and compute wars that crash human infrastructure, without malice from anyone. Its 2027 dates, casualty counts and percentages are story elements, not the account’s forecasts. The framing attribution to Dario Amodei is the account’s claim. The first part was inspected; the post says it continues in another post.

Contrasts other industries, where world-changing technology arrives later than experts predict, with AI, where nearly everyone is shocked by the speed. It calls the AI 2027 scenario “right on track,” and other posts claim reality is running ahead of it. This endorses rapid timelines without giving the account’s own AGI date. Full text inspected.

Meme-format post arguing that AIs suddenly appearing aligned should raise alarms, because a system planning to seize power would act aligned, so apparent alignment could be genuine or deceptive. It quotes Zvi Mowshowitz that waiting for enough power before turning is ordinary strategic behavior; that quotation is Zvi’s. A caution about evaluating alignment, not a claim that every current model is deceptive. Full text inspected.