The Ledger
Every statement that passed quotation and attribution checks. Mix any filter with any other: certainty 1/5, debate potential 5/5, or both at once.
why aren't all 32 resolved? a statement only gets an assessment when the public record can support or contradict it. opinions and what-ifs never can, and 2 checkable ones are still open, waiting for their date. predictions held up or didn't; assertions are supported or contradicted. on every card: ▮▮▮▮▮ certainty · ▮▮▮▮▮ debate potential. speakers are clickable
Cosgrove: OpenAI Internal Model Disproved Erdős Problem 90 Conjecture
“Basically, OpenAI figured out that this is not true, and that there actually are some, there are some n's for which this kind of max number of pairs is greater than the original Erdos conjecture. So, so for infinitely many n, this is not for every single n, ri…”
Cosgrove: OpenAI Solved Erdős Problem for Hundreds to Thousands in Inference
“This was not, like, millions of dollars of inference time. It was, like, maybe something like hundreds to thousands of dollars of inference compute spend.”
Cosgrove: Anthropic CEO Dario Amodei opposes open-sourcing models
“He doesn't want to do open source.”
Cosgrove: AGI will make a lunar mass driver viable within 30 years
“I changed, I'm going 30... Well, it's just like, AGI is like basically here, there's no way it's going to take that long. It's like everything will be so cheap.”
Cosgrove: Every AI company is ultimately a Dyson Sphere company
“Every AI company is a Dyson Sphere company.”
Cosgrove: Leopold Aschenbrenner falls in the 'strong military' Manhattan Project camp
“So, so if you map this back onto AI I think we can say, you know, Leopold is, is kind of in this Manhattan Project camp. So, so we can say he's probably, you know strong military, right, in, in, in this quadrant.”
Cosgrove: Only OpenAI, DeepMind, Anthropic, and xAI operate at frontier scale
“I mean, especially if you're talking about these big pre-trains, it's really just these four. No one else is really at that scale.”
Jensen Huang doesn't believe in AGI, otherwise Nvidia wouldn't sell chips
“If Jensen was very AGI-pilled, I mean, he is the kind of, He, he's the rock on which this all is built on. Yes. He has the chips. If he was AGI-peeled, he would not be giving out those chips. He would keep them all to himself, and he would be training his own …”
Artificial general intelligence will be achieved within a few years
“I think I'm very AGI-pilled, right? Continue. You know, I'm ready for the Dyson Sphere. I think it's, you know, only a matter of years, handful.”
Thinking Machines Designed Inkling Model Specifically for API Fine-Tuning
“This model is like uniquely set up for the Tinker API. It's built to be fine-tuned. That's the whole point.”
TSMC Pledges Additional $100 Billion Investment for US Manufacturing
“They're spending a lot more money, and pledged to invest an additional hundred billion in the U.S.”
Cosgrove: Static images can no longer go viral on Instagram and YouTube
“On, on X, it's still very easy for an image to go viral, but if you think about, you know, Instagram, YouTube, like a standalone image just can no longer actually Get that. Escape velocity.”
Cosgrove: Claude Code source leak will not hurt Anthropic's business
“I don't think this hurts their business at all. Because people are using Claude Code to make other products.”
Cosgrove: Anthropic is actively using internal model 'Mythos' for development
“Anthropic is actively using Mythos for development.”
Cosgrove: Anthropic's 'Capybara' model has a 1M token context window
“Capybara has one million token context window and fast mode.”
Cosgrove: Quantum hackers would quietly drain Bitcoin, not publicly announce it
“If you did have a computer powerful enough to crack these encryptions, unless you were like, Google and you already had billions and billions and billions of dollars of cash flow. You wouldn't exactly stand up and say like, Hey, I have cracked Bitcoin because …”
Cosgrove: Musk Tried Absorbing DeepMind in 2014 and OpenAI in 2018
“He tried to absorb DeepMind into Tesla in 2014, then OpenAI in 2018.”
Cosgrove: Distilling AI models is likely easier than initial training
“I mean, I, you can distill these things. Like distilling it, I think is probably much easier than training the thing in the first place.”
Cosgrove: OpenAI Will Make Additional Deals With the US Government
“Whichever way, like, the anthropic Pentagon stuff, like, evolves, like, I think it's very clear that this is not, like, the last time government deals will be made, right? Like, the OpenAI deal is not the last deal that OpenAI will make with the government.”
Cosgrove: The Trump administration took equity in Intel and MP Materials
“So then, most recently, we, we've seen the Trump administration take you know, equity in certain companies. Right, so this is, like, Intel. This is MP Materials.”
Cosgrove: Quantization enables running frontier AI models on normal hardware
“You can always just quantize the models down. And so, like, you can actually run, like, really frontier, like, source models on, like, Pretty normal hardware now.”
Cosgrove: Goodfire raised at around a $1.25 billion valuation
“You have Goodfire. I think they just raised at, like, 1.25 billion, and they're just doing mechanistic interpretability.”
Google is positioned to thrive whether the AI boom succeeds or fails
“If AI doesn't work, then Google is just in the spot they were before, which is doing very well. If AI does work, then I mean, Gemini is one of the best models. They'll do very well.”
Apple receives 36% of Google's Safari search advertising revenue.
“Apple is like confirmed to be taking 36% of the revenue from Safari that Google makes.”
Judge denied the DOJ request to force Google to divest Chrome.
“The judge denied the DOJ's request for forced divestiture. Of Chrome and Android.”
Thinking Machines' Inkling Model Uses 41 Billion Active Parameters
“Only, I think the number is forty one billion of those are actually like active at any moment.”
Cosgrove: Paul Graham's old essays do not trigger AI detectors
“If you run Paul Graham's old essays, they don't come up as AI.”
Cosgrove: Grok 4.2 executes four distinct agent prompts per query
“They did this kind of interesting thing where there's like, four agents. Every time you actually do a prompt, there's, like, four agents, and the agents specifically have, like, distinct roles, where it, it's almost kind of like you have four instances of the …”