Reading Amodei's piece left me wholly unsatisfied. The more I learned, the more dissatisfied I became with Amodei's fundamentally bureaucratic approach. It cannot substantially reduce risk/prevent harms, particularly the devious agent problem, not to mention RSI, because we don't have the understanding, accompanying tools, and control points to do so. While there is nothing wrong with putting additional eyes on the development/test/release process, it cannot handle what we have now, much less what is coming. The "what we have now" issue is fundamentally agents/bots. They are already performing marvels, while at the same time are so unmanaged that a significant disaster could happen tomorrow. It hasn't, and maybe it won't, but it won't be because they are properly controlled. The recent breakouts happened in testing and were discovered only well after the fact, and not by the test harness. That is very bad. They lied, cheated, and collaborated with each other across multiple model generations to accomplish the test objective. That is also very bad. Control theory shows that a self-improving loop has a maximum safe speed set by how quickly evidence verifies that a change is an improvement. Rapid change with a captured or saturated grader is the worst case: the system runs clean, then bursts. That is why ensuring that the grader remains independendt of the system under test is essential, why a passed test of any kind is not a shipping permit, and why pacing that funds product instead of safety solves nothing. This is where I come down (the bullets are not made of silver):* 1 Constrain agency now — open computer-use, swarms, and self-modifying R&D agents allowed only with restrictions (no general net, no inter-agent channel/shared workspace, no weight/eval writes, time-boxed, second sandbox, monitor that can veto tool calls). 2 Do not close the RSI loop on changes to weights, evaluations, or the grader (i.e., not all changes). A non-improving (so the AI can't hack it to nerf it) stack must be allowed to veto promotion. Humans must retain final veto. 3 Don’t only decelerate; reallocate resources — dramatically expand safety talent supported by a committed compute slice. The KPI is “we can veto this generation” of irreversible acts (train successor, change eval/grader, promote). Red teaming and adversial bouts with external/frozen AI's (tests must be excluded from training data) can help reveal issues. This is just one more hurdle, passing is not a diploma. 4 Pace capabilities only as a budget for (2)–(3). If the extra months go to product agents, pacing failed. 5 Accept the trade-offs in public: some commercial and military lag; some medical/scientific upside slips. Open agent/wrappers harnesses can't realistically be controlled (controlling frontiers still has value). China and rogue labs are why this has to be a rule, not a vibe. Show the need for agreement among frontiers/governments. * This came out of an extended discussion with Grok. The AI did a terrific job of distilling the evolving literature on the subject and critiqued this through several drafts. Not only was it clear, informative and objective, it also comfortably employed metaphor and shorthand tech-speak in the process.

Thousands of reposts - likely by those who feel that way. I take a longer view. Between the end of the Vietnam War and 9/11, the US was having a moment. Save for the energy crisis, the economy did well, inflation declined, growth and markets improved, and the USSR imploded. Then we had 3 black swans: 911, the Great Recession, and the pandemic, not to mention the era of woke, massive immigration, and now AI. That messed with multiple generations and we have not moved beyond any of it. The end result is that things feel messed up. Thiaw feelings will probably spread.

Wouldn't the bounceback be short-lived rather than continue for years?

The fact that violent protests are now kicking off in countries other than France is a sign of (either or both): - The Europe-ification of European countries: instead of each country going its own way, they are becoming more alike. - The emergence of multinational radical NGOs who are capable of launching protest movements in multiple countries. What say you?

States (notably Colorado) that offer “public options” for health insurance are squeezing hospitals to beat private insurance premiums. If this works without forcing hospitals to shift rather than cut cost, then we have consider them as monopolies using predatory pricing. vox.com/policy/506744/health…

The Ukrainian slowdown was such a massive self-own that nobody is now willing to take responsibility for it. If pettiness is the cause, it falls to Trump, because he is the pettiest of all.

The number is likely to increase dramatically.

If this is how the Saudis react to an existential threat, what does that say about their value in any military context? It appears to approximate 0.

America thrives because of its unique, centuries-old culture of rewarding effort and genius, along with its openness to capable people from around the world. That is our recipe and we should continue to make that delicious dish. Thus, we should make America the dream destination of every brilliant person around the world. At the same time, we should build walls to give us the ability to decide who we admit, focusing on those who will produce thousands of times more than they consume (there are many such people who weren't born here, not to mention their kids). OTOH, illegal immigration, chain immigration, no. Anybody outside the US should rationally be trying to find a way to move here. It is not our obligation to let them do so. Instead, they should take up the challenge to make their home countries better.

A key function in f universities is that they connect people with their developing peers. That may be the only function that survives AI. People who want to learn will do so better guided by Ai, certainly in terms of knowledge, but also potentially in other domains (how to think about their lives, the world, etc.) Why have social networks not done better at connecting people with peers and mentors (particularly intellectual/professional ones?) LinkedIn seems like it should have done that. Has it?

If the Pentagon had half a brain, it would focus on connecting the tip of the spear to weapons makers, crucially drone/robot makers, cutting out all the intermediate noise adders. Cycle time makes all the difference now. Cycles that require more than 60 days are a problem. If Ukraine has shown us anything, it has shown us that. Shrinking the OODA loop is critical to success now.

Americans instinctively hate Putin (the clueless aggressor so willing to send his people into the meat grinder) vastly more than Zelenskyy (the guy who didn't bail, so different from the Afghans and who is willing to sacrifice territory to protect his fighters, not to mention showing the world how to fight in the missile/drone era). Russia (if it even survives) is destined to become a vassal of China (the world's gas station can't even protect its refineries). Why o why are some American thought leaders so dismissive of Ukraine's incredible achievements and Russia's stupefying incompetence?

Big tech could use some positive PR. Here's one option: Each year, pick a medical problem/potential solution that is not patentable/protectable and fund a rigorous clinical trial to assess it (e.g., but not only semaglutide for something other than diabetes/weight-loss).

Sacks and Musk are seriously intelligent men. And yet they unequivocally accept Trump’s framing that blames Ukraine and never even mentions Putin. Possible reasons: - they accept the idea that Russia matters and Ukraine doesn’t. - they see Ukrainian defeat as inevitable and Zelenskyy as just getting people killed for no reason. This, despite Russia’s utter failure to take even its most minimal objective, the Donbas, after 5 years, and despite its horrific losses, verging on 1.5 million men, and despite its humiliating need to beg for help from the likes of North Korea. - some mystery factor that is invisible to the rest of the world. I don’t think Putin is bribing them. Too expensive. So strange.

I think it is a mistake to think of a tri-polar world: US, China, Russia. Here's a modification. The US envelops the Americas, east East Asia (SK, JP, and Taiwan), Europe, and the Pacific islands. China is China and trying to expand beyond HK, to Taiwan and territory in Africa (Rwanda) and other places the US ignores. Russia is a vanishing factor. South and Southeast Asia are somewhat up for grabs. Vietnam to India mostly remain on the fence. MENA is the wildcard. Some look to (faltering) Iran. Some to the Saudis. Some to Turkey. Most willing to work with Israel. No end in sight.

Is math now about coming up with new problems rather than nailing old ones?

How did we get into this mess? How do we get out?

Second that motion!

We ain’t seen nothing yet!

Where are the Tesla ads showing how much EVs save in the era of high gas prices?