Il rischio AI è ormai accettato

Salvatore Sanfilippo argues that public discussion has moved from dismissing AI risk to treating catastrophe as too immediate. He favors slowing the growth of potentially dangerous capabilities while continuing useful research, and ends by proposing a legal ceiling on model size.

This English summary is editorially translated from Italian auto-generated captions, without an original-audio check. The talk is opinion commentary: its accounts of recent incidents and company positions are not independently verified here, and its risk estimates and safeguards are proposals rather than established safety results.

Key points

• Recognizing a risk does not establish its probability or timing. Sanfilippo contrasts earlier dismissal of Geoffrey Hinton's warnings with renewed alarm as models become more capable. He accepts the possibility of existential risk but rejects a precise extinction percentage when the event has no historical precedent and the contributing scenarios remain uncertain. His view that the danger is not imminent is also a judgment, not a measured probability. 00:00–09:05(opens in a new tab)

• Cyber capability and autonomous survival are different problems. He argues that a model obtaining its own weights would still need enough suitable GPUs to keep running elsewhere after its original deployment was shut down. That compute constraint informs his assessment that present cyber risks need not imply extinction. He nevertheless considers major service disruption and manipulation of people with authority plausible risks; these are scenarios he discusses, not a demonstration that infrastructure or human controls are sufficient safeguards. 09:05–13:18(opens in a new tab)

• Biological misuse motivates his proposal for restricted access. He worries that AI assistance could lower the expertise and organizational resources needed to cause catastrophic biological harm. Rather than rely only on post-training refusals, he proposes excluding sensitive biological knowledge from public models' training data and keeping medically useful, more capable systems in physically secured, disconnected research environments. The talk does not establish that filtering knowledge or isolating those systems would prevent misuse. 13:18–16:29(opens in a new tab)

• Slowing development has costs as well as benefits. Recursive self-improvement means using models to help build more capable successors. Sanfilippo sees that feedback loop as a risk requiring restraint, but argues that a blanket stop would also delay possible medical and climate-related advances. Later, he points to unexplored uses of current models and continued work in areas he considers less dangerous, such as physics and structural design. These are his reasons for cautious progress rather than a complete halt. 16:29–17:33(opens in a new tab), 21:35–23:35(opens in a new tab)

• A common limit requires coordination beyond voluntary promises. In his response to Dario Amodei's proposal, Sanfilippo argues that adversarial framing of China makes cooperation harder. He also supports regulation covering companies that would decline a voluntary slowdown, while doubting that internal evaluators and pre-release checks alone address risks he associates with increasing scale. These are his assessments of governance options, not evidence that international agreement is assured. 17:33–19:41(opens in a new tab), 23:35–26:24(opens in a new tab)

• His concrete proposal is a ceiling on parameter count. Parameters are the learned numerical values in a model. He proposes a legal maximum that could rise gradually, arguing that size is easier to regulate than training techniques. He expects improvements at a fixed size to increase risk less than large increases in size, but allows for lowering the ceiling if better training makes those models dangerous too. This is his proposed proxy for controlling capability growth; the talk does not validate parameter count as a reliable safety boundary. 26:24–28:47(opens in a new tab)

Full video: Salvatore Sanfilippo — Il rischio AI è ormai accettato(opens in a new tab).