@JonKBateman
Pinned post
www.youtube.com AI Is Coming for Your Job. Now What? Even as AI dominates the headlines, the gulf between what insiders ... AI is all over the headlines. But the gap between what outsiders believe about the technology and what insiders know has never been wider.View quoted post on X
@GarrisonLovely has interviewed hundreds of these insiders — and he joins @JonKBateman on The World Unpacked to explain: youtube.com/watch?v=J_mzJj…
NEWS: Ds review 181-page dossier from ex-Warren/Sanders staffers ID’ing hundreds of executives & corporations to targetView quoted post on X
Envisions “resignations of major CEOs, hauling people in, material restitution”
Tech giants, crypto, sports gambling, gas & more
washingtonsun.com/congress/democ…
🧵 Astra covertly underperforms on safety research tasks.
We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this.
To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.View quoted post on X
I have a lot of thoughts on what to expect in AI policy in 2027.View quoted post on X
Here are 5 things I’m buying, and selling, after the election.
Buy
1. A legislative reset and more ambitious proposals.
Current bills are largely stuck in 2025, having been adapted from proposals predating this summer’s spate of cyber incidents. With a shuffle in leadership + voting coalitions, more offices will write from scratch, and that’s a good thing.
2. Hearings and oversight.
The filibuster still exists, so this is the easiest thing to get off the ground. And boy, will Congress have questions for the AI companies. I expect this to be largely popular and bipartisan, and my only hope is that Congress acts on what it learns.
3. Codifying the current White House oversight regime.
I’m hopeful here. The WH has already constructed some form of pre-vetting and incident reporting. Those processes belong in legislation, and there’s some version that is sufficiently aligned with the admin and minority leadership to get the votes. The vote margin in the Senate really helps here, as every vote counts in getting past a filibuster. Either party securing a big vote advantage would have helped here.
4. Auditing, incident reporting, monitoring, and internal use visibility.
The first two are the bread and butter of current bills and state law, but those bills have tended to be exceptionally narrow (eg not covering any current models or incidents). That will change.
The latter two are the next frontier for company-centered laws. Governments and the public want to know what is going on inside these companies and to avoid surprise. This isn’t a partisan thing, but I expect the election reset and the increased salience of AI to shift this in a useful direction.
5. Government capacity.
This is a natural cross-party synergy. How can we enable the exec branch to do its job. And this can be done in various forms, starting with funding CAISSI, but also bringing in talent via partnerships. I don’t expect this to be easy, but this has a route to 60 votes.
Sell
1. A lame duck deal.
If there had been split control, we might have gotten last-minute negotiations and a greater willingness to deal from both parties. But with a full switch, I think we’ll have to wait for real action on AI.
2. Entirely voluntary standards.
We’ve known for a long time that this isn’t enough. Companies will not willfully self-constrain. The only reason this is still the standard for most bills is inertia — in 2025, when political will was low (and most policymakers thought autonomous cyber attacks were a distant possibility), voluntary rules seemed like a good starting point. And without a change in leadership, those bills stuck around. Now, with a shuffle in leadership and control, it’ll be easier to move on.
3. Broad preemption.
This has aged so poorly. The new Dem majority won’t support it, and many R’s have likely changed their minds in the last year. The election puts the nail in the coffin.
4. Oversight of executive branch use of AI.
This one pains me, and I hope I’m wrong. If we see widespread AI use in the executive branch, Congress will be at a considerable disadvantage and will lose even more relative power. We want oversight both to maintain checks and balances and to avoid concentration of power. But I think it’ll be considerably more controversial for a D Congress to restrict this admin than a split or R Congress.
5. Rulemaking being a dirty word.
Up until now, it’s been hard to get rulemaking in AI leg. The problem is, AI is uniquely hard to predict and it adapts rapidly. So static leg text is really ill-equipped. If you want minimum standards for auditors or adaptive thresholds for which companies or models are covered, you need rulemaking. This is one place where partisan preferences make a difference — D’s are less allergic to rulemaking. They’ll need to figure out ways to bound those powers to get past a filibuster, but I’m hopefully that will be a productive process
re: agentic shopping
I have NEVER in my life heard so much hype and discussion about a tech/product experience that basically does not work at all for me.
I feel like I’m using theranosView quoted post on X
