@MackenZ_arnold: Logging my predictions now: 1. At least two frontier AI companies will make new unilateral commitments to safety by EOY…

Logging my predictions now: 1. At least two frontier AI companies will make new unilateral commitments to safety by EOY. Internet consensus will be ~these are good, narrow, and not enough. 2. Those agreements will have some genuinely good parts – probably around technical oversight and monitoring (various things that don’t constrain outputs or methods of development). They will not contain commitments like tiered-release of systems, restrictions or delays after unsafe events, disclosure of safety procedure failures, or restrictions on certain techniques (like non-human-legible CoT). They’ll likely include something on incident disclosure, but it’ll be pretty weak. 3. Within one year, at least one company will have amended its unilateral commitments because they believe they will soon violate them. At least two others will have violated the spirit of their commitments (e.g., by disclosing vastly less information than promised, or hiding said information until it leaks) As for collective agreements: 4. Companies believe they can share incident information amongst each other, but won’t share much of it for other reasons. They’ll eventually start sharing more information about distillation and other things that more directly affect their bottom line, though even this will be delayed by fears of legal risk. They will not share information on novel safety risks that emerge or failures of their safety protocols. 50/50 DOJ will release a guidance saying infosharing on cyber and some other things is okay; it won’t move the needles will mostly address the easiest, least controversial versions of sharing. 5. By the end of 2027, companies will try to make some collaborative agreements. The agreements will come very close to failing but survive in wounded, narrow forms. Behind the scenes, people will say that public pressure on the companies and pressure from their employees made the deals more likely and stronger. 6. In negotiating those collaborative agreements, companies will in fact feel held back by antitrust, and >80% of GCs would say that claims made in late 2026 that companies can obviously collaborate with only limited risk are obviously false and too dismissive of the difficulty. If you poll the participants, ~30% will say that the antitrust barriers led to a meaningfully less useful safety agreement, 60% will say it meaningfully increased the odds that the agreement would fall apart. 7. One of the key mechanisms for this failure will be that legal uncertainty reduced the leverage of employees who cared about public safety and empowered employees who wanted to maintain the status quo. 8. Those agreements will happen later and with fewer participants than they would have had the law been clear. Some agreements would already exist today if the law was already clear. 9. Within 18 months of this collaborative agreement, no lawsuit will have been initiated. If one does come, it’s most likely to come from a DOJ that is hostile to some of the participating companies. Many will be surprised that Elon played nice and didn’t sue. The main cost of antitrust uncertainty will have been in slowness, concern, procedure, and FUD. That cost will be real and have hurt the public interest. 10. Any safety promises contained in the collective agreements will be much more likely to appear in Congressional of CA legislation that passes after the agreement.