Daily AI Digest: Distillation, Dollars, and DIY LLMs

Alex: Good morning, everyone, and welcome back to Daily AI Digest! It's July 23, 2026, and we've got a jam-packed show for you today.

Jordan: Oh, it's a big one. We're talking US-China AI tensions escalating into actual sanctions threats, Google's wild AI spending numbers pulling in two different directions, and a guy who built an LLM from scratch for the price of a nice dinner out.

Alex: Three hundred and fifty-three dollars? We'll get there. But first, Jordan, did you see the orca story?

Jordan: The ones ramming sunfish until they explode? Yes. Terrifying and also, honestly, relatable energy for a Wednesday.

Alex: Scientists said it might just be 'for fun.' No AI model has that kind of unhinged creativity.

Jordan: Yet. Give it time. Speaking of unhinged energy, let's get into today's actual chaos — the Moonshot-Anthropic sanctions story.

Alex: Okay, so, TechCrunch is reporting the Treasury Department is threatening sanctions after the White House claimed that Moonshot — the Chinese lab behind Kimi — distilled its model from something called 'Fable,' which is apparently Anthropic's internal name for one of its proprietary systems.

Jordan: Right, and this is a huge escalation. Up until now, most of the US-China AI friction has been about export controls — chips, compute, that kind of thing. This is different. This is the government essentially accusing a foreign lab of stealing the actual intellectual property baked into a model's weights.

Alex: Wait, can you even prove that? Like, how do you 'catch' a model being distilled from another one?

Jordan: That's the fascinating and messy part. Distillation usually means training a smaller model to mimic the outputs of a bigger one — you feed it prompts, capture the responses, and train on that. The fingerprints can show up in things like shared quirks, similar refusal patterns, even matching hallucinations.

Alex: So it's less like DNA evidence and more like... comparing handwriting?

Jordan: Kind of, yeah. It's circumstantial by nature, which is exactly why this is such a big deal — if Treasury moves forward with actual sanctions based on that kind of evidence, it sets a massive precedent.

Alex: A precedent for what, exactly?

Jordan: For how governments police model provenance going forward. Right now there's no clean legal framework for 'this AI learned from that AI without permission.' If the US pushes this through, every lab is going to have to start thinking about proving where their training data and outputs actually came from.

Alex: That feels like it could get really messy, really fast. I mean, don't all these labs kind of train on each other's outputs at this point? Like, isn't synthetic data from competitor models already everywhere?

Jordan: That's the open secret nobody wants to say out loud. Everybody scrapes, everybody benchmarks against everyone else's outputs, and the line between 'inspired by' and 'distilled from' is incredibly blurry. Anthropic just happens to be the one making the loudest noise about it right now.

Alex: And it's notable that it's Anthropic specifically, right? Given how much they've leaned into the 'we're the safety-conscious, responsible lab' branding.

Jordan: Exactly, and that makes this a great story to protect. If your entire brand is built on trust and responsible AI development, having your work allegedly siphoned off by a rival — especially a foreign one — is a direct hit to that narrative.

Alex: So what happens next? Does this actually go anywhere, or is it more saber-rattling?

Jordan: Hard to say yet. Sanctions threats can absolutely be leverage in a broader negotiation. But even the threat alone shifts the tone — we've gone from soft diplomatic pressure and chip export bans to 'we might financially cripple your company over model theft.' That's a different level of seriousness.

Alex: Okay, this actually connects really well to the next story, because apparently while Washington is trying to contain Chinese AI, Silicon Valley is just... using it anyway?

Jordan: Yes! This is from Hacker News, aggregating some great reporting, and it's basically the perfect companion piece. The headline is literally 'The US wants to contain China's AI, Silicon Valley keeps using it.'

Alex: Using it how? Like, developers are actually building products on top of Kimi?

Jordan: Yep. Moonshot's Kimi model — the same one at the center of that sanctions controversy — has become genuinely popular with developers because it's cheap and it performs well. When you're building a product and your API bill matters, 'it's from a geopolitically controversial lab' doesn't always win against 'it's ten times cheaper and almost as good.'

Alex: That's kind of wild, though. Isn't there risk in building your product on a model that might get sanctioned into oblivion next month?

Jordan: There absolutely is, and that's the tension this piece captures so well. You've got policy trying to build walls, and you've got a market that just wants the best price-to-performance ratio, walls or no walls.

Alex: It's kind of like the streaming wars, but for compute. Everyone says they care about the ethics of the platform until the price difference gets big enough.

Jordan: That's a great way to put it. And remember, open-weight Chinese models have been improving really fast. DeepSeek kicked that off last year, and now Moonshot's Kimi is right there too. For a lot of startups, especially ones without huge funding, that's simply too good to ignore.

Alex: So does that undercut the whole containment strategy? Like, what's the point of sanctions and export controls if companies are just going to route around them anyway?

Jordan: That's the billion-dollar question, literally. Export controls target chips and hardware — things with physical supply chains. But software and model weights? Those can move at the speed of a download link. It's a fundamentally different kind of thing to contain.

Alex: So policy is playing whack-a-mole with something that doesn't really have a fixed location.

Jordan: Pretty much. And that's why this pairs so well with the Moonshot-Anthropic story — you've got the US government trying to punish a Chinese lab for allegedly copying American IP, while American companies are simultaneously adopting that same lab's tools because they're just really good and really cheap.

Alex: The irony is almost too perfect.

Jordan: It really is. It shows you that market incentives move a lot faster than regulatory frameworks, especially in a space evolving this quickly.

Alex: Okay, let's shift gears a little bit, because speaking of money and AI, we've got two Google stories today that are basically arguing with each other.

Jordan: Yes, this is a fun one. So on one side, Hacker News is highlighting reporting that Google is burning through significant cash because of spiraling AI infrastructure and compute costs.

Alex: How bad are we talking?

Jordan: Well, training and serving a frontier model like Gemini isn't just an upfront cost — it's this continuous, enormous capital expenditure. Data centers, custom chips, power, cooling, and then serving inference to hundreds of millions of users every day. It adds up to tens of billions of dollars a year in spend.

Alex: And that's just to stay competitive, right? Not even to necessarily be winning?

Jordan: Exactly, that's the scary part. This is table stakes now. If you're not spending at that level, you're not even in the conversation for frontier models. It's raising real questions about whether this pace of spending is sustainable long-term, or if we're heading toward some kind of correction.

Alex: Okay, but then the second story — this one's from TechCrunch — is basically saying, 'actually, Google's doing great, don't worry about it'?

Jordan: Right, this is the flip side. Google just reported record profits, and they're crediting a booming cloud business, largely driven by enterprises adopting their AI tools and infrastructure services.

Alex: So which is it? Are they bleeding cash or printing money?

Jordan: Honestly, both, and that's not even a contradiction — that's just what it looks like to build a capital-intensive business during a technology transition. You spend enormous amounts up front on infrastructure, and if it works, the revenue eventually catches up and then some.

Alex: So the cloud growth is basically proof that the spending isn't just burning money into a void.

Jordan: That's the argument, yeah. Enterprises are paying for Gemini-powered tools inside Google Cloud — coding assistants, data analysis, customer service automation, all of that. And that demand is apparently accelerating, which is exactly the kind of proof point investors want to see before they'll tolerate continued massive spending.

Alex: It's kind of like justifying a really expensive gym membership by pointing at your gains.

Jordan: That's exactly it. The spending story alone sounds alarming — 'Google is burning cash!' But paired with the profit story, it becomes, 'Google is investing aggressively, and it's paying off.' Context really changes the narrative.

Alex: Do you think this settles the whole 'is the AI infrastructure race sustainable' debate, though?

Jordan: Not fully, no. Google can point to cloud profits as validation, but that doesn't mean every player in this race is going to land the same way. Not everyone has Google's balance sheet, or an existing massively profitable ad business subsidizing the ride.

Alex: Right, so this might be more of a 'the big get bigger' story than a 'the whole industry is fine' story.

Jordan: Exactly. It's a really good reminder that when you see one of these dramatic 'AI company burning cash' headlines, you have to ask — burning cash relative to what? For Google, it might just be the cost of staying dominant while the returns are already showing up elsewhere in the business.

Alex: Okay, I love this next story because it's basically the opposite of everything we just talked about — instead of billions of dollars, we're talking about $353.

Jordan: This is such a fun palate cleanser. Also from Hacker News — a developer audited Stanford's CS336 course, which is their deep dive into how large language models actually get built, and documented training an LLM completely from scratch for $353.

Alex: Okay wait, how is that even possible when we just spent ten minutes talking about Google spending billions?

Jordan: Different scale, obviously — this isn't a frontier model that's going to compete with Gemini or Claude. But it's a legitimate, functioning small language model, trained using rented cloud GPU time, that demonstrates the actual mechanics — tokenization, architecture choices, the training loop, optimization, all of it.

Alex: So it's less 'I built the next GPT' and more 'I built a real working example of how GPT-style models function under the hood'?

Jordan: Exactly right. And that's incredibly valuable, especially right now with all this talk of billion-dollar training runs and geopolitical fights over model distillation. It's grounding. It reminds you that the fundamental technology isn't magic — it's math, data, and compute, just applied at a scale most of us will never personally touch.

Alex: I actually really like that this is coming out the same week as the Moonshot story. Like, on one hand you've got billion-dollar accusations of IP theft between massive labs, and on the other, some guy with a laptop and a few hundred bucks demonstrating the actual building blocks in public.

Jordan: That's a great point. It's a nice reminder that the core knowledge isn't some closely guarded secret — Stanford literally teaches this course, and the fundamentals are increasingly accessible if you're willing to put in the work.

Alex: What's the value for someone listening who isn't a machine learning engineer, though? Like, why should a curious non-technical listener care about this one?

Jordan: I think it demystifies the whole conversation. When you hear about a company spending twenty billion dollars on AI infrastructure, or accusations of model distillation, it can feel like this abstract, almost mythical process. Seeing someone document, step by step, 'here's what it actually takes, here's what each dollar bought me' — that makes the whole industry feel a lot more graspable.

Alex: It's kind of the anti-Google-earnings-call story, in the best way.

Jordan: Right, no earnings call jargon, no geopolitical tension, just someone genuinely curious, documenting their work, and sharing it for free.

Alex: Okay, so if I'm pulling today's threads together — we've got a major sanctions threat over alleged AI distillation between Anthropic and Moonshot, we've got the very real market pull toward Chinese models despite that political pressure, we've got Google simultaneously spending like crazy and profiting like crazy off AI, and then a grassroots reminder that you can build a real LLM for the price of a plane ticket.

Jordan: That's the whole spectrum right there — geopolitics, economics, and the actual engineering, all in one day's headlines.

Alex: It really does capture just how multi-layered this AI moment is right now.

Jordan: Completely agree. It's not just 'is the model good' anymore — it's trade policy, corporate finance, and grassroots engineering all colliding at once.

Alex: Well, that's going to do it for us today. Thank you all so much for listening to Daily AI Digest.

Jordan: We'll be back tomorrow with more of the stories shaping the world of foundation models. Stay curious, everyone.

Alex: See you next time!