https://pacing.tech/ Arb Research's agenda for what "pacing the frontier" might actually mean.
summary of what the options are and what they would do, not advocacy of a single plan
pacing is already happening: delaying releases, rollbacks, training pauses, export controls. "pacing" here encompasses all proposals that slow or delay model development and diffusion, whether done by companies, governments, or other institutions.
tradeoffs, pros and cons, and open research questions considered throughout.
he thinks the argument is circular. the J-space is defined by the contents of the residual stream that predict what words will be output next, averaged across contexts. it is necessarily going to shift verbal output when altered!
"So they show that flexibly-reused intermediate variables are stored in a canonical format aligned with the unembedding. “Deliberate” computation reads these variables and routine computation bypasses them. This is weaker than a workspace, which under global workspace theory should additionally display competition-gated entry, capacity limits enforced by that competition, and broadcast to encapsulated modules."
"Companies do not always know why the EPA rejects their applications for engineered microbes as the EPA has never published a checklist of requirements. Companies submit applications based on their best guesses about what the EPA will want to know."
https://blog.spec.tech/p/bubbling-cauldrons-and-transient Ben Reinhardt has a model of healthy R&D ecosystems -- "bubbling cauldrons" of creative researcher-led work where nobody is bossing them around, augmented by "transient strike teams" where groups of researchers, with authoritative leaders, work in a coordinated fashion on an important project.