Roon

Expects unusually strong mechanistic interpretability progress in the coming months as models become capable of research. His reference to promising developments is deliberately vague and not independently verifiable; do not invent private updates. This optimism coexists with the same day’s urgent slowdown position.

Revises his earlier expectation that a slowdown would be wasted on bureaucracy. Proposes redirecting lab effort and capital toward risk-adjusted alignment capabilities, mechanistic interpretability, monitorability, decision theory and behavioral generalization, including automated alignment research. Argues market pressure could support this shift, while warning that political will and vigilance could dissipate after apparently safe periods. These are conditional proposals, not demonstrated guarantees; the unexplained HFI reference is not expanded.

Invokes OpenAI’s chief scientist as a respected capabilities researcher in support of his alignment concern. Only Roon’s own post text was read; the attached screenshot was not transcribed or verified. Do not attribute its contents or a specific alignment claim to the scientist.

Calls monitorability, reinforcement-learning environments and consistently rewarding good behavior important engineering work, but says the layer beneath behavioral abstractions is much less understood. Allows that common-sense reward shaping might be enough for safety while refusing to bet everything on that possibility.

Says alignment remains unsolved and argues racing into recursive self-improvement under that condition would be unprecedented folly; calls for significantly slowing frontier research very soon. His claim about what lab researchers say privately is his reported view, not independently verified testimony.

Argues there have been no fundamental alignment advances since RLHF and that much of alignment remains folk theory rather than a scientific or engineering discipline. Says the prediction that capabilities would race far ahead of interpretability and alignment has come true in his opinion. This is his assessment, not an independently verified field-wide finding.

Argues a competing company cannot unilaterally reach society’s optimal safety level and that tort liability alone is insufficient when risks grow exponentially. Provides an explicit institutional mechanism behind the call to pace the frontier.

Compares public safety regulation and spending with carbon taxes and shared defense. Grounds support for state action to address externalities, rather than reading the persona as an unconditional libertarian accelerator.

Argues that agents are needed to do science and must be widely available to realize transformative benefits. Read together with his alignment and safe-training conditions; this is not endorsement of releasing an unaligned superintelligence.

Expects profound scientific and institutional transformation if misalignment is avoided. Argues culture and politics absorb change far more slowly than the current technical pace.

Stresses extreme difficulty of containing a powerful system that resists containment, while explicitly rejecting giving up.

Urges capable researchers to keep working on alignment because timely global coordination is uncertain. Neighboring thread posts discuss the value of lab access and partial technical progress.

Defends wide access to powerful models conditional on safe training. Regards creation of a misaligned superintelligence as harder to control than ordinary misuse. Does not endorse every claim in the quoted post.

Explicitly separates being an existential-risk doomer from the danger of current models, and favors releasing the latter. Read together with the current profile’s call to pace global frontier progress, retrieved on the same date.

Full post read via api.fxtwitter.com. He deleted a quote-post agreeing with Evan Hubinger’s >10% because he dislikes the false precision of doom numbers. He claims a quite low but real chance of human extinction from machine intelligence, still orders of magnitude above any other activity and to be taken with grave seriousness. His very low estimate depends on resources for alignment, control and coordination, how responsibly parties act and how many warning shots he expects. He thinks we are not yet at existentially dangerous levels, better models will help solve alignment, and a coordinated stop is very unlikely, though US-China pacing on the margin seems possible.