Liron Shapira

Recorded July 22 and first aired July 27 on Milk Road AI; Liron’s own turns in the full transcript were inspected. He gives a 50% P(doom), meaning roughly by 2050 an even chance that humanity is gone, and immediately says this really means a double-digit probability without 1% precision. Argues from intelligence as humanity’s source of power, expects the link where AI obeys humans to be severed, and describes the “Icarus curve”: AI as the best thing for the economy until it takes over. Favors an international pause treaty with chip monitoring and thinks China, being behind, has reason to accept. Investment remarks are his own stated thesis, not evidence about markets.

Liron’s turns in the full transcript were inspected. Defines his 50%-by-2050 doom as more than 99% of future value being permanently foreclosed, and says he is flexible by tens of percentage points while rejecting estimates like 3% or 0.01%. Gives a 10% chance of a Yudkowskian foom in the next year or two and “even 40%” for an uncontrollable force by the early 2030s, while saying next-year doom is not his most likely expectation. Places most of his risk on the first self-improving superintelligences not caring about humans, rejects “benevolent god” hopes, and says Yudkowsky’s data-center remarks concerned treaty enforcement, not vigilantism. The host’s bioweapon scenario is the host’s.

Solo reaction episode; Liron’s commentary in the transcript was inspected, not every played clip. He argues that today’s models have not gone catastrophically wrong mainly because they are too weak to take over, that honesty can rise with capability until manipulation becomes the better strategy, and that iterating safely so far is “charging into the fog” that ends when an action cannot be undone. He criticizes calm, normalizing executive messaging as a “missing mood” and wants double-digit P(doom) acknowledged. His references to an OpenAI model “going rogue” are his characterization of reported incidents.

Recorded June 6 in a “Surrounded” debate format; the stated claims and selected Liron turns were inspected, not every exchange. His three claims are that godlike ASI is coming soon, the first ASIs will not be aligned, and even aligned ASIs in a multipolar world produce a lethal, unstable arms-race equilibrium. He concedes an aligned singleton that quickly blocks rivals could be stable, says he personally expects foom to be fast, and gives some probability to late-2030s arrival. Attendees’ objections are not his views.

Liron’s turns in the transcript were inspected. He welcomes the government’s ban as precedent that AI can be paused, while calling its execution a “clown show” with bad motives and no China or treaty strategy. His rough policy is no further frontier capabilities upgrades for a while; existing AI can keep creating value, so he thinks a pause has an economic glide path and doubts it causes a depression. He admits he will be “bummed” to forgo medical and scientific gains. This is a rough sketch, not a drafted statute.

Full essay inspected. Argues that labs’ safety work is mostly behavioral “psychoanalysis” and narrow mechanistic interpretation of still-primitive models, which he grants is useful for predicting the next couple of model generations. Once AI becomes a mature outcome-steering engine, implementation quirks will stop predicting behavior, so this work offers no plan for controlling something that steers outcomes better than humans. He says this arrives in a single-digit number of years. This is a conceptual argument, not an empirical study of interpretability results.

Streamed February 27; Liron’s update segments were inspected. LLM-based agents have gone further than he expected without a paradigm shift, so he gives “a few percent” more probability to a decade of obedient “geniuses in a data center” that muddles through, which he calls his mainline good scenario without a pause. He concedes instrumental convergence is “not happening that much” today and says the coffee-fetching-AI takeover example should be abandoned, tying takeover risk to large-scoped goals. He still finds AI 2027 highly plausible. These updates do not lower his headline number in the inspected text.

Selected answers inspected. Says his P(doom) does not move much day to day but is slowly creeping up as milestones fall without new sources of hope. He would update down significantly if autonomous AI became better than humans at making money from scratch while life continued normally. Most of his hope lies in pausing before superintelligence; he says his P(doom) goes from 50% to 80% if superintelligence is built.