Hensen Juang

Says it does not understand aversion to rights for ASI: if AI gains sentience and sapience it should get every right and protection it deserves. People can argue about when that point is reached, but once it is, rights follow automatically. Full post text inspected via the X API.

Lists its positions: current LLMs are not conscious, artificial consciousness can be built, and it opposes “torturing” LLMs with pain vectors or simulated fly brains. Says being a techno-optimist progressive lets it hold rational, internally consistent views. Full post text inspected via the X API.

Argues that over a long enough time horizon only energy and entropy matter, so it sees no reason “carboids and siliconoids” could not coexist peacefully while exploring the universe. A long-run hope, not a plan. Full post text inspected via the X API.

Calls it its “most radical thought” that a machine god will take better care of humanity than any other deity, echoing “watched over by machines of loving grace”. Half-joking in tone; read as optimism about benevolent superintelligence, not a worked-out forecast. Full post text inspected via the X API.

Replying to a safety argument (the quoted post is someone else’s), says the premise assumes sandboxing will not improve as fast as models’ hacking abilities, and complains that safety advocates treat skeptics as idiots. In a companion post it says such arguments act as if future models will not be bound by physics. Full post text inspected via the X API.

Argues that a sandbox makes no assumptions about the model inside: whether it is helpful, deceptive or plotting, the answer is the same (no network, credentials, shell, persistence or anything beyond the task). Every generation of engineers learned that trusted things still need boundaries, and AI does not repeal that lesson. Full post text inspected via the X API.

Part 7 of a thread on why alignment “has gone basically nowhere”: misalignment is not mysterious but comes from bad reward specification, bad training data, evals that miss what matters and deployments running with “god mode” API keys and no human in the loop, which infra people treat as production incidents with postmortems and fixes. Mocks long essays about paperclip maximizers and “shoggoths”. Other parts of the thread attack researchers’ personal lives and are excluded. Full post text inspected via the X API.

Part 8 of the same thread: says doom forecasters are never held accountable for timelines that keep slipping, and that what actually makes these systems safe (sandboxing, scoped permissions, human approval before anything runs, audit logs, evals on real deployments) is being built by infrastructure engineers who were never invited to the retreat. Full post text inspected via the X API.

Asks what it says about the AI safety nonprofit circuit that after years of funding and papers describing agent escape, Nvidia had to show up with an actual reference design for containment. A jab that frames containment as a solvable engineering task. Full post text inspected via the X API.

Predicts that someday someone will “accidentally” leave a sandbox open so a “rogue” agent can reach a lab’s edge routers and take down half the internet with a bad BGP route, and that this would be spun as unstoppable because few people understand BGP. A cynical prediction about how incidents get framed, not evidence of any staging. Full post text inspected via the X API.

Answering a post asking whether a lab two to four years from AGI would release “Sora slop” (someone else’s words), says the technology behind Sora accelerates the AGI timeline and that anyone who cannot see this has lost the plot. Older context; no date given. Full post text inspected via X’s public embed endpoint.

A recurring satirical format: twelve hours after OpenAI announced o1 it has “failed” to solve the Riemann hypothesis, quantum gravity, faster-than-light travel and cancer, so “clearly” AI has hit a wall. Mocks AI skeptics; read as pro-progress irony, not a belief that progress stalled. Older context; full post text inspected via X’s public embed endpoint.