What’s the evidence on rent control? I charted the findings from 112 studies: Great for rent-controlled tenants ✅ But: Pushes up rents for other people ❌ Pushes down new housing construction ❌ Pushes down building quality ❌

AI agent adoption is still very bimodal right now. You have coding and coding adjacent tasks which have taken off, and then everything else. And even within coding, you have a very wide continuum of adoption patterns, with some likely a small percentage of developers deploying background agents working on projects in parallel, and the rest of everyone else still working with agents 1:1. Then, there’s the entire rest of knowledge work where real agentic adoption is still very early. The reason for this is that most of the workflows still need to reengineered to work with agents. This isn’t the same as deploying a chat system for better research efficiency, but instead requires workflows to be rebuilt, data to be wired up in new ways, new practices for accountability and liability of decisions when agents are in the mix, governance and compliance paradigm changes, security upgrades, and more. We’re still so unbelievably early in what this is going to look like outside of a few categories of work right now. This is why you can basically expect 100X more agent adoption from what we’ve seen so far.

Why do people think AI is conscious or sentient or whatever? It's because it talks to you. That's the whole reason. No one seems to think the Waymo or Tesla FSD is conscious even though it's just as complex (if not more) and even a remarkably similar architecture

This is actually super impressive from Opus - it went from ~1500 elo to ~1750 elo. One possible explanation for Astra's deterioration is the context window - it natively has 272k in Codex, so perhaps with more notes & files it just cracks. I will see if I can run a 1m context window, but probably GPT-6.1-Sol only. Watch here: ailearningchess.ai-learning-…

Yudkowsky: “there’s no way an infant has qualia at birth” What I want to draw attention to here is not the horror of this, but the extreme overconfidence about something as poorly understood as consciousness. Take this into account when judging his other confident assertions.

Arguing that "consciousness arises from a computational substrate, and since AI models are computation, they are likely conscious," is as meaningless as arguing that "living creatures are made of atoms, and since this rock is made of atoms, it is likely alive." Everything is computation. Sometimes incredibly sophisticated computation, like a chess engine, AlphaGo, google3, or the Linux kernel. That does not make it conscious. A rock being made of atoms does not confer it any of the properties we associate with living things, also made of atoms. Likewise a static input-output program that has *none* of the properties we associate with conscious beings (e.g. information integration, interoception, temporal binding, embodiment, etc.) has no more reason to be presumed conscious than a rock has to be presumed alive. Humanity has not yet created a conscious computer program, and there are no signs we are close to doing so. When we get close, the case for machine consciousness will be backed by evidence and corroborated by consciousness science (which is a thing), not by evidence-free wishful thinking and cargo culting.

Your paper is very clear and useful to the field. It confirms concerns that I voiced publicly in 2017. Seeing it is a bit like someone saying to me "you weren't crazy after all," years after I witnessed a somewhat implausible traumatic event in which people didn't believe my testimony. I was an remain a free trade enthusiast, but I also wanted to understand the 2016 election, and Autor et al's work was a big theme that I felt I needed to explore. I had long admired Autor's labor market research on skills and tasks. My 2016 work on presidential preferences used their data and found no relationship at all between trade exposure or even manufacturing exposure and Trump support, which was my first clue that something was off. This was less notable when it became clear that the Autor et al trade exposure variable had no meaningful correlation with any economic outcome (unemployment rate, household income, etc) that I had been studying during my time at the Brookings Metro Program as a key indicator of regional prosperity. What was very surprising to see was that I could not replicate their results until I a) used their exact data and b) used their exact code. Using alternatives for either failed to replicate, and I told them so privately. In fact, not only was the regional data not playing out as expected (more "exposed" areas were no worse off in 2015 than other areas), but other national data were inconsistent with the story. Manufacturing layoffs and separations per worker were less common than those in domestic un-exposed sectors. After realizing that the stacked model was the key issue (comparing late 1990s outcomes to ~2007 outcomes), I wrote up my results, shared it with them, and tried to publish as a comment but could not find a receptive journal--so tried as a full paper at AER, since they published the original work. Debating Autor et al did nothing to help my career and became a major distraction from work I was supposed to be doing. It is easy to understand why many people don't do it. My 2017 paper was rejected by the (now) incoming AEA President Penny Goldberg, who as editor at AER at the time, on the flimsiest of excuses. I'm sure the Borusyak et al paper hides the bad results for Author et al to avoid controversy and facilitate publication. Their Supplemental Table C3 is very similar to my preferred results. It is reassuring to see that their other results, relaxing various assumptions, are consistent with what I was finding. I have nothing against Author, Dorn, and Hanson. They have every right to defend themselves vigorously against criticism of their work. I'm sure they believed they were right. At the same time it has become astonishingly clear to me that a) academics in every field are not capable of objectively reconsidering their own work (probably myself included), and should not be assumed to be capable of doing so; b) they follow and perpetuate narratives, often erroneous ones, to advance their careers. These narratives are rarely, if ever, directly tested, because doing so would be professionally awkward, embarrassing to many others, and run counter to the political preferences of most of their colleagues. I have seen this to an incredible degree in sociology, as I am working on a meta-analysis that completely contradicts a dominate narrative, and I have seen it across social sciences and humanities with respect to the lasting influence of Marxism, which still leads many academic writers to posit--without even trying to prove it--that there are large causal effects of growing up with low socio-economic status, and these effects would be eliminated with income/wealth redistribution.

Much of my research is on arXiv. I was one of the early adopters. It is simply a repository of research papers with PDF and metadata. It makes it convenient to find research without any paywall. Physicists started it in 1991. Math and computer science followed gradually. It relies on volunteers to moderate it, because we don’t want garbage or spam. On October 1, they capped the submissions to two per person. Why? ArXiv got over 40,000 submissions in September. Two years ago, it was about 20,000. Doubling every two years. It is unsustainable for human beings. In related news, Google stopped taking new bug reports in its open-source bounty program. They could not cope with the influx. The writing is on the wall.

I don’t know if I agree with this take. Good frameworks are intelligence caches. The production of verifiable high-quality code has a cost to it (insert for now caveat). Frameworks can lower this cost as long as they don’t get in the way of agents.

Will A.I. Make Your Brain Lazy? Here’s What the New Research Actually Shows. When people become accustomed to skipping the productive struggle of logical and analytical tasks and having the process automated by using AI, they are more likely to give up on hard questions that they could otherwise figure out. “Part of how A.I. is making us dumber,” says Dr. Adam Green, a cognitive neuroscientist at Georgetown University, is “undercutting the development of learning how to think in younger people that now have an alternative.” But students who usedA.I. more like a tutor, to explain complex concepts or clarify ideas, not only retain high test scores, but their essays were higher quality as well. So the key is using this technology is a more effective way. Sadly, this is not the default way students seem to be using it. nytimes.com/interactive/2026…

In one case rather than outright refusing Astra swapped out a part of the experiment I wanted to run with something “safer”, which would have completely invalidated the results if I hadn’t caught it. Do not recommend unless very carefully watched

Friendly reminder about new subtle policy and guardrail changes in GPT-6 and 6.1: Contrary to how Anthropic does it and slaps your wrist for even asking it, Sol 6 and 6.1 dance around the bushes, A LOT, and loop forever, effectively wasting tokens, instead of refusing taking final actions. It chats and documents "progress" and keeps iterating over results but takes no actual action, being execution of a command, placing an order or any action that crosses the system policy guardrails. I've been observing this since day one of Sol 6 release, especially in my AI trading harness. Stick to 5.6 Sol, is my suggestion for now. YMMV.

The notion that current AI models are sentient and can suffer, combined with the foolish idea that suffering can be mathematically quantified and weighted between humans and non-humans, could lead us down an incredibly dark and dystopian path. But before it gets to that point, it will rightfully be met with immense backlash from team humans.

I have not seen near enough excitement about MCP Events developers.openai.com/plugin… Right now, the only way most people's agents can wake up and do something is either A) cron, or B) you decide to message it Event-driven triggers are a huge deal

Ben Thompson says Meta and Google are making more money using AI to improve their ads than the labs are from selling it: "The biggest monetization right now is probably not in Anthropic or OpenAI. It's the incremental gain that is happening for Google and Meta." "They can actually validate their image creation and their text creation in a way no one else can. Their validation is the ad marketplace, running a gazillion A/B tests to see what works. Most ads don't convert." "What do LLMs do? LLMs predict. We're going to move to this world where Meta is going to look at people and say, this person probably wants to see this next. And they're going to go find that thing and show it to them." "The potential upside in terms of just showing people better ads that are more relevant to them, they only need to increase a few percentage points for the returns to be billions and billions of dollars. This alone is worth them investing in being on the leading edge."

Reasoning is great but there's still no substitute for a certain base level of knowledge, this unlocks a kind of bullshit detection that is otherwise unavailable Example just now, looking at a reranker benchmark GPT: "you should use minilm-l6-v2, it beats everyone including much more recent models like voyage-2" Me: "that's got to be fucking broken, find the bug" GPT: "ah here it is, they applied sigmoid() to the voyage scores that were already 0..1"

Austin's Police Association president talks about officers making arrests for violent and serious crimes, filling out the paperwork and affidavit And then finding out the DA has dismissed the case even before their shift even ends, with basically no investigation!

If a private school stinks, nobody has to go to it. If a public school stinks, some kids are forced to go to it, and the most powerful pressure group in education will do everything it can to prevent them from going elsewhere. I'm more concerned about that.

Pretty happy with luna 6/max + fast tho