Jack Clark used the May 11, 2026 edition of Import AI to endorse a governance framework called 'radical optionality,' drafted by researchers at the Institute for Law & AI, that asks democratic governments to build the legal and institutional machinery for an AI crisis before one arrives. The pitch: avoid heavy-handed rules today, but spend aggressively to give regulators real tools for tomorrow. Clark, who writes Import AI alongside his policy work, called the cost 'modest' relative to the downside of inaction. Issue 456 also covered a Meta and KAIST paper proposing neural computers in the 10T-1000T parameter range and an economics paper modeling recursive self-improvement as a growth shock.
The Institute for Law & AI authors frame the goal plainly. 'At its core, radical optionality is about preserving democratic governments' ability to make good decisions about how to govern transformative AI systems as circumstances evolve,' they write. 'In the short term, this means avoiding overregulation while rapidly building the institutions, information channels and legal authorities needed to respond competently to a broad range of scenarios.'
The specifics are concrete. Governments should impose transparency and reporting requirements on frontier developers, set up third-party auditing to verify those disclosures, protect whistleblowers inside frontier labs, and harden information-sharing channels across allied governments. They want flexible 'if-then' regulatory triggers rather than fixed thresholds, plus government and third-party capacity to evaluate model capabilities and safety properties. They also push for stronger physical and cyber protection of model weights and algorithmic secrets, primarily through voluntary standards.
Key facts
- 01Import AI 456, published May 11, 2026, endorses the Institute for Law & AI's 'radical optionality' framework for governing transformative AI.
- 02The proposal calls for increased funding for AISI (UK) and CAISI (US) while avoiding broad use of the Defense Production Act.
- 03A Meta and KAIST paper led by Juergen Schmidhuber introduces the Neural Computer, a learned runtime unifying compute, memory and I/O.
- 04The authors estimate a mature neural computer would need 10T-1000T parameters, sparser and more addressable than current models.
- 05Prototypes built on the Wan 2.1 generative video model render basic CLI and GUI workflows but show limited symbolic stability.
Funding follows from the framing. The authors argue 'governments should be willing to spend an extraordinary amount of money, effort, and political capital on preserving optionality' and that policymakers should be 'wary of counterproductive interventions, but not much concerned with the actual pecuniary cost of any realistic measure that seems likely to have net-positive results.' A meta-recommendation runs through the paper: hire more technical staff, and increase budgets for AISI in the UK, CAISI in the US, and their counterparts elsewhere.
“Schmidhuber and co-authors argue a mature neural computer points toward a 10T-1000T parameter substrate that is sparser, more addressable, and a little more circuit-like than today's models.”— Jaeden Schafer
The authors also pre-empt the obvious pushback. On the risk that governments concentrate power over AI, they say they are 'basically convinced' the risk is real, which is why they reject sweeping moves like broadly expanding the Defense Production Act. They float using only 'law-following AI systems' inside government as a partial mitigation. On private governance, they say independent verifiers can play a role but still require a competent technical core inside government to function.
The second piece in Import AI 456 is more speculative. Researchers at Meta and KAIST, including Juergen Schmidhuber, ask 'can a neural network act as a traditional computer?' Their proposed answer is the Neural Computer, which they describe as 'a neural system that unifies computation, memory, and I/O in a learned runtime state.' One co-author put the framing more bluntly: 'a new machine form is starting to emerge.'
The prototypes are early. Using the Wan 2.1 generative video model and curated training data, the team built neural-computer demos for both command-line and graphical interfaces. 'The NC learns to render and execute basic command-line workflows,' the paper says, capturing behaviors like scrollback, prompt wrapping and window resizing, 'though symbolic stability remains limited.' On the GUI side, the team tested world-model designs across cursor supervision, action injection and encoding choices, with mixed but suggestive results.
The endgame is bigger than the demos. The authors describe the target as a Completely Neural Computer, 'a fully learned computer whose compute, memory, and interfaces are unified in a single learned runtime substrate rather than engineered as separate modules.' Schmidhuber's own guess is that the mature version 'points toward a different substrate: something more like a 10T-1000T machine that is sparser, more addressable, and a little more circuit-like' than current dense transformers. The paper concedes that 'progress toward CNCs will therefore depend not only on stronger models, but also on whether reuse, consistency, and governance become sustained and testable.'
Import AI 456 closes with a third strand: a paper from researchers at Forethought, Columbia University and the University of Virginia arguing that recursive self-improvement, or even very deep automation of the economy, could trigger a compounding feedback loop and an unprecedented growth boom. Clark has covered RSI dynamics in earlier issues, including #455, and the new paper formalizes the macro side of that argument with explicit economic models.
The counterweight to radical optionality is the one Clark flags himself. Transparency requirements, auditing regimes and reporting authorities are not load-bearing on day one, but a future government with different priorities could turn them into something far more aggressive than the drafters intended. The neural-computer line of research carries its own caveat: the prototypes are barely functional, comparable in maturity to early powered flight, and the leap to a 10T-1000T parameter substrate would require step-changes in compute, data and training methods that nobody has yet demonstrated.
Taken together, the three strands sketch the dominant policy and research bet of the moment: AI capability gains will keep compounding, the substrate underneath those capabilities is still up for grabs, and the governments that want a seat at the table later need to start writing checks for institutional capacity now. For frontier labs, the practical implication is that the reporting, auditing and weights-security regime being modeled today is the one they will operate under for the next decade, regardless of which party holds power in Washington or London. The cheaper path is to help shape it now rather than litigate it later.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
