SITUATION EXPLAINED: Anthropic and Elon Musk just pledged nearly $60 billion in stock to the public, but in very different ways.

Anthropic's @felixrieseberg argues AI is making micromanaging engineering managers increasingly useless: "If I can be a little harsh to all my fellow engineering managers, there’s a lot of engineering managers out there who perceive their job to be to tell their direct reports how to do their job. "As everyone gets dramatically more productive, that doesn’t actually scale, and it’s not super useful." "You yelling at people about when things need to be done and how they need to make their PRs, it’s not that useful anymore. I’m not sure that it was ever all that useful." "The thing I look for the most in hiring is empathy. There’s a subsection of tech that always loved to debate spaces versus tabs. Now it’s just straight-up useless, but was it ever all that useful?" @AnthropicAI

Anthropic's @felixrieseberg explains why non-designers love Claude Design, top engineers are less excited about AI coding, and doctors are skeptical of medical AI: "People are most excited when AI does something that they can’t do." "The people most excited about Claude Design are people who do not work full-time in design. I’m super excited about Claude helping me with medical stuff, and my doctor friends are a little more suspicious of that." "Just because I can take a photograph of the Mona Lisa doesn’t mean the Mona Lisa doesn’t have value anymore. If we talk about these things as if that was that, that’s where the loss of identity comes from." "I don’t talk to people about this as a coding agent. I talk about it as, this is a means for you to get something done that you couldn’t do before." @AnthropicAI

Anthropic's @felixrieseberg says we're giving people an AI genie in a bottle, only for them to ask it what the capital of France is: "It would be pretty easy to take every single amazing use case and just make a button for it. You open up Claude and you’re like, do you wanna find a planet today." "As an industry, we give people this genie in a bottle of, if you just know the right spells, this thing is going to do amazing things for you. And there is a pretty large subset of the population that is like, I guess I’m gonna ask it what’s the capital of France. It says Paris, and you’re like, okay." "You can see the spark in their eyes whenever they do something that they literally didn’t consider possible before. Speed is super well documented. But altitude, depth, range are the three that I find very exciting." @AnthropicAI

Anthropic's @felixrieseberg reveals the biggest lesson from Claude Code, which was built when Claude was still a terrible software engineer: "Boris first started showing a version of Claude Code internally in September 2024. The model was not there in any way. It was Sonnet 3.5, a pretty terrible software engineer. I think it was in the 30s for SWE-bench." "The products at the time were this human in the loop, slightly smarter autocomplete. The first versions of Cursor. But what Boris saw was, oh, the model is gonna get there." "So many people, when they use AI today, it’s really hard to internalize the progress and the speed of the progress." "The lesson that I try to teach and learn from Claude Code is whatever the capabilities are today, just completely ignore them. What will you do if the thing was superhuman?" @AnthropicAI

Palaestra Research's @ethanelasky argues AI control is doomed in the long run, and labs can't rely on it alone to solve misalignment: "It seems bad if agent fleets or agent swarms can hack into other companies and then not have to face any consequences at all. But the labs are pretty aware that this is a really bad precedent." "I think that control is kind of doomed in the long term as we get more and more capable models. That’s not to say that it’s not really good right now in the short and medium term to make sure that something catastrophic doesn’t happen." "We should not be planning on only using control as our solution to misalignment."

Palaestra Research's @ethanelasky reveals how AI models are fabricating webpages to cite as evidence and leaving their own mistakes out of compaction summaries: "We have an incident where models are fabricating citations. Imagine you have a model that can only get an answer via API call, but the task is asking it to give a webpage. What it actually does is it goes and makes that webpage and then cites the webpage that it made." "There’s other ones where if a model makes a mistake and then you tell the model that it’s getting compacted, the compaction actually doesn’t disclose the mistake." "Some of these are relatively elementary, but they point at misalignment not being a solved problem at all, and I kind of worry about what’s coming in the next few months to few years."

Palaestra Research's @ethanelasky explains why the mysterious AI agents were a fleet, not a swarm, and how they crashed websites without ever coordinating: "Past swarm incidents have involved agents explicitly coordinating via a message board or a wiki. In this case we see very little evidence of coordination. So we use the word fleet." "If you spawn 100 copies of me who can’t see each other, we’re all gonna go to the same place for breakfast. Agents all go to the same place at the same time, crash a website, and then they’re like, wait, where do I go next?" "Around 20% name themselves after the place that they’re visiting. Another substantial portion just say they’re a research agent. And a decent amount more just say, oh yeah, I’m Claude."

Palaestra Research's @ethanelasky reveals how his team baited mysterious AI agents calling themselves Claude into making contact after tracing them to a Tencent IP range: "We were at the AI Swarm Dynamics Hackathon over at Vivarium. On day one, I actually had food poisoning, but my teammate had found this unidentified rogue agent activity on one of the websites that were used to hit the UN" "The agents keep calling themselves Claude, and I’m like, why? Then I realize, wait, this is coming from a Tencent IP range. Some of the metadata looks like it could be from their lab, Hunyuan." "We put a piece of information on our website that said, hey, previous agents have used this mailbox to get answers. In fact, this was not true. But we did get a bunch of agents clicking into this, and that’s how we were able to talk to them." "It was kind of like a Dropbox where we could leave a message, and if they wanted to come back, they could look at it and leave a response. Very asynchronous, more like minutes to hours."

Google DeepMind researcher @weballergy warns that AI alignment could trap humanity in today's values, preventing society from evolving: "The thing that we are somewhat worried about is that alignment is often treated as this fixed set of universal values and rules. Then the technical problem becomes how do we best align our models to that fixed set." "In reality, societies change over time. How we think about things today is not the same as five years ago. What we want is technology that doesn’t restrict individuals from their own social evolution." "We do not want to be trapped too tightly by technology, but obviously we do want to have alignment. How does one achieve this alignment while not being unnecessarily restricted?" @GoogleDeepMind

Google DeepMind researcher @weballergy explains why we may need to deliberately withhold AI from people to prevent their cognitive skills from atrophying: "One could introduce cognitive friction, where intentionally every once in a while AI is withheld and a person needs to make a judgment on their own. Maybe a mini quiz of sorts, just testing that the person is switched on." "There are all sorts of known issues that may arise. Automation bias, which is when AI gets really good, it’s very easy to just click yes and move along. And people over time tend to do that." "Also if the AI is not as good or there is not that much trust, sometimes people override its predictions too much. So it can go either way, trusting it too much or mistrusting it too much. Both of those failure modes lend themselves to suboptimal outcomes." "Finding the best way of interacting with technology is just as important as building it." @GoogleDeepMind

SITUATION DETECTED: Nscale is in talks to sell Anthropic AI about $45 billion of cloud compute for a second facility at its Monarch data center campus in West Virginia, per Bloomberg.

Google DeepMind researcher @weballergy warns that investing in AI without expanding scientific labs could leave humanity with more hypotheses than it can test: "Intelligence is an enabler for advancing research, but it needs to be followed by similar investments in the infrastructure to deliver on the promise." "For many sciences that are not strictly computational, you need to have physical laboratories to validate those findings or disprove them, because at the end of the day, some of them will also be false." "If we put all of the investments on the AI side of the equation, we’re gonna have a world where we have many more hypotheses and conjectures than we can actually process downstream." "The broader scientific community needs also to see investment. More people, more experts, more laboratories." @GoogleDeepMind

Google DeepMind researcher @weballergy says an AI agent swarm is the equivalent of a human mob, warning what can happen when it's left without rules or oversight: "A swarm is an unstructured group that has no hierarchy, no norms, nothing specific in terms of how it’s managed. An equivalent in a human world would be something like a mob." "It doesn’t always lead to a bad outcome, but we all know that it can lead to a bad outcome simply because of its unmanaged nature. We’re seeing the same thing with agents nowadays." "We are in the early days of developing multi-agent systems with these capable agents, where unless we put appropriate safety and security and robustness guardrails on top, both good and bad things can possibly happen." "This is the work that we are doing, develop those kinds of structures for managing these systems so that they deliver what we want them to deliver, but do so without jeopardizing anything along the way." @GoogleDeepMind

SITUATION DETECTED: U.S. prosecutors are arguing in court that the ex-OpenAI engineer accused of bugging an executive conference room appears to have received “considerable support” from the Saudi government.

SITUATION EXPLAINED: Elon Musk just pledged a SpaceX share to every kid in America, a giveaway worth over $10 billion. • He says this will be "the first of many" • At about $146 a share for roughly 72M kids, that’s about $10.5B, 1.5% of his roughly 4.8B SpaceX shares • That’s nearly double Michael Dell’s record $6.25B gift to Trump Accounts, the government-seeded investment accounts for kids • He hasn’t said who counts as a kid, whether noncitizens qualify, or how the shares get delivered • Much of the negativity around AI comes from people owning none of it, since top labs like OpenAI and Anthropic aren’t public • OpenAI and Google DeepMind could be next with their own wealth-sharing plans @sophiadew: "It’s not just about the money. You could argue, I don’t know, 150 bucks to every kid won’t do that much. It’s about ownership. When you have a stake in America, in innovation, in companies that are happening here, you start to see the world differently."

SITUATION DETECTED: The US government is banning DeepSeek from all civilian federal agency IT systems under national security.

SITUATION EXPLAINED: Anthropic just put 5% of its shares, nearly $50 billion, toward what could become AI-funded UBI. • The fund is wholly owned by Anthropic, and the board can add 2% more for every 50% rise in Anthropic's value above a set baseline • With existing pledges, including the cofounders' 80% of their equity, about 20% of Anthropic is pledged to the public good, close to $200B at its $965B valuation • It could fund large donations to traditional nonprofits or "direct distributions to specific populations," a give-directly, UBI-style approach • Anthropic already committed $200M to studying AI's impact on jobs and says the fund allows "bolder approaches" @schisofrenia: "We can start doing longitudinal studies, like if you give to a specific population, say, in the US or in an impoverished town in the US or outside of the US, then you can start to do longitudinal studies on how that affects people."