🎧 Language Models: The Week in Review

_Audio podcast — two AI hosts discuss fresh AI news. Full transcript below._

Host A (00:00): So Meta is laying off 1,000 employees amid record profits just to buy silicon chips. Meanwhile, the Vatican is preparing to release a historic tract on protecting the soul from algorithms.

Host B (00:13): Yes, it sounds like the beginning of some cyberpunk novel.

Host A (00:17): Exactly. And somewhere in a lab, a tiny neural network that can literally fit on an old smartphone is right now learning to control a real caterpillar tank.

Host B (00:26): And doing it successfully, mind you.

Host A (00:28): Yes. You know, if someone still thinks that artificial intelligence is just some smart chatbot sitting comfortably in its digital enclosure, helping write emails, then today's pile of materials from Wired, The Verge, TechCrunch, and a bunch of developer blogs proves the absolute opposite.

Host B (00:46): The enclosure is open.

Host A (00:47): Exactly, it's open, and our deep dive today is dedicated to how all these scattered news flashes add up to a large-scale shift in our entire reality—both physical and corporate.

Host B (01:03): And here, you know, it's critically important not just to skim the headlines, like, okay, someone got fired, someone got bought. You need to understand the very mechanics of what's happening, because behind each of these news items there's a fundamental change in how resources are distributed in the world, who controls infrastructure, and what's most interesting—how these systems start to interact directly with the physical world. Anyone trying to understand the trajectory of technology this decade needs to see the connections between these events.

Host A (01:37): Okay, let's break this down with probably the harshest and most obvious example of resource redistribution. I'm talking about the Meta corporation news.

Host B (01:47): Yeah, the numbers there are impressive.

Host A (01:49): The company announced its largest layoff since 2023. That is, 8,000 employees are being let go, plus another 6,000 already open positions are being canceled. And at the same time, the company has record profits.

Host B (02:04): Absolutely right.

Host A (02:05): And all this freed-up budget, all this money is being directed to some insane goal—they're investing $145 billion in infrastructure.

Host B (02:15): In hardware.

Host A (02:16): Yes, in servers, in chips, in their own Trainium and Inferentia processors. Listen, but I'm experiencing cognitive dissonance. If a company is building cutting-edge technologies, doesn't it need engineers to write code for them? It looks like a direct transfusion of blood from living developers to silicon servers.

Host B (02:34): Well, the cognitive dissonance arises from our habit of thinking in old categories. Like how technological growth used to work. More people equals more written code.

Host A (02:44): Yeah, that's logical.

Host B (02:45): Which in turn equals more features and products. But Meta's leadership right now is betting that this equation simply doesn't work anymore. $145 billion is, in essence, an acknowledgment of the fact that computational power—that computer itself—has become more important than human capital.

Host A (03:05): Wow! So people are no longer needed?

Host B (03:08): Well, let's say the models themselves now write code. They optimize processes, they create architecture. Hiring a huge army of junior and mid-level developers no longer makes economic sense. What makes sense is owning the machines on which this new intelligence is printed. And whoever has more servers with these Trainium and Inferentia chips will simply train the next iteration of the model faster, which in turn will replace even more line personnel.

Host A (03:37): A vicious cycle. And this battle for these machines and tools is taking quite aggressive forms. There's a news item that at first glance seems kind of niche. But if you read into it, it's a real corporate power play.

Host B (03:53): You mean the Anthropic deal?

Host A (03:55): Yes-yes-yes. TechCrunch writes that Anthropic, creators of the Claude model, is buying the New York startup Stainless for $300 million. The startup was founded in 2022 by former Stripe engineer Alex Rattei, and they make SDKs—toolsets for developers. And what's Anthropic doing? They're not just integrating the team, they're closing Stainless cloud services for everyone else.

Host B (04:24): Harsh move.

Host A (04:26): And among the clients, by the way, were OpenAI, Google, Cloudflare. Let's clarify this moment—why is the purchase of some developer tool worth $300 million and makes such giants nervous?

Host B (04:39): Listen, to understand the scale of this step, you need to look at how AI interacts with the outside world. So. A language model itself is, roughly speaking, just a brain in a jar.

Host A (04:51): Great comparison.

Host B (04:52): Yeah. And for that brain to be able to, say, book tickets, check the weather, or pull data from a corporate database, it needs API—application programming interfaces. And SDK, or Software Development Kit, is a set of adapters and ready-made instructions. It allows developers to easily and without errors connect this AI to any service. Alex Hartley once said that SDKs deserve the same care as the APIs they wrap, and Stainless automated the creation of these extremely complex adapters.

Host A (05:31): Aha. So if API is, say, the language that programs speak to each other, then SDK is like a ready-made dictionary with grammar rules, right? So nobody has to reinvent the wheel each time.

Host B (05:45): Excellent analogy, I completely agree. And now imagine Anthropic buying the best printing house in the world that prints these dictionaries, and just puts a huge lock on the door.

Host A (05:58): Wow.

Host B (06:00): In a world where autonomous AI agents executing tasks in 1,000 different applications are the future, the quality of this integration is literally a matter of life and death for a product. By bringing the Stainless technology inside, Anthropic guarantees that their Claude model will integrate with the world faster and more reliably than all the others.

Host A (06:21): And in doing so, they're kicking the chair out from under competitors?

Host B (06:24): Exactly. That they've taken this tool out of the hands of OpenAI and Google is ruthless, but, you know, absolutely brilliant business strategy. Development tools are now strategic weapons.

Host A (06:37): And here's where it gets really interesting. If billions of dollars are at stake, like Meta's, and there's this monopolistic control over tools like Anthropic's, it's logical to ask: who exactly is calling the shots?

Host B (06:51): Oh yeah, who are these people?

Host A (06:52): Yeah, who's making the decisions. And here we have materials from Wired and The Verge about a high-profile lawsuit between Elon Musk and Sam Altman from OpenAI.

Host B (07:03): Very telling story.

Host A (07:05): Musk accused the company of betraying its original non-profit mission for huge money. But wait. The jury took just 2 hours. 2 hours to deliver a verdict in favor of Altman and dismiss the case based on statute of limitations, recognizing commercialization as lawful.

Host B (07:23): Yeah, the speed is incredible.

Host A (07:25): Doesn't that mean Musk's position was, well, legally null and void from the start? Why did the press grab onto this case if it was all decided in a couple hours? Seriously?

Host B (07:36): Well, of course. From the letter of the law perspective, the suit fell apart quickly, right? But the trial itself lasted 3 weeks. And these 3 weeks became just a public vivisection of all of Silicon Valley.

Host A (07:48): Correspondence.

Host B (07:49): Yes-yes-yes. We saw these armies of lawyers that cost tens of millions of dollars. We read personal messages full of mutual accusations of greed, manipulation, intrigues. Both sides suffered reputational damage. And the press grabbed onto it because the trial exposed a very frightening reality.

Host B (08:10): Industry leaders—people who control technologies capable of actually changing the course of human history—they're behaving like ordinary capitalists from the Wild West era.

Host A (08:20): Who can't agree.

Host B (08:22): Yes, it seems mutual destruction is the only way.

Host A (08:24): So masks are off, and society just saw that behind all these beautiful words about safe AI for all of humanity hides a banal battle of egos and capital.

Host B (08:38): Exactly. And the consequence is a total crisis of trust. The legitimacy of the entire AI industry suffered a serious crack. Society no longer believes that tech giants can regulate themselves. A huge moral vacuum emerged.

Host A (08:57): And nature, as you know, abhors a vacuum. And precisely at this moment, a player enters the stage from whom, well, personally I least expected to hear news about algorithms. The Vatican.

Host B (09:09): Yeah, it was a surprise.

Host A (09:11): Guardian writes that Pope Leo XIV—by the way, the first American pope in history—is preparing a historic encyclical, and it's entirely devoted to protecting human dignity in the age of AI.

Host B (09:23): Sounds massive.

Host A (09:25): And invited to the presentation of this document at the Vatican is the co-founder of Anthropic. Wait, but why specifically Anthropic? Is it because they position their Claude model as the most, you know, ethical and safe?

Host B (09:37): Well, Anthropic's stated emphasis on ethics and this constitutional AI of theirs is, let's say, just the façade part of the answer.

Host A (09:45): And what's behind the façade?

Host B (09:46): The situation is much deeper; it has an obvious political subtext. Anthropic is currently in a fairly intense confrontation with the Trump administration over the question of state regulation of AI.

Host A (09:56): Oh, so politics got involved?

Host B (09:58): Yes. The administration might lean toward deregulation to speed up innovation and outpace competitors on the global stage, but Anthropic insists on strict safety protocols. And by inviting their co-founder, the Vatican is sending a very clear signal.

Host A (10:13): Which one?

Host B (10:13): Look, an encyclical is the highest form of papal document. By focusing on the problems of algorithmic bias, total data collection, and the threat of synthetic media and deepfakes, the Church is trying to build a moral framework—and mind you, outside of state or corporate interests.

Host A (10:32): So they're showing their independence?

Host B (10:36): Absolutely. The Vatican is demonstrating independence from current US politics and betting on those market players who are ready to discuss existential risks rather than just revenue growth charts. It's important to understand that we're not taking sides in these political debates; we're simply analyzing why the Vatican made this choice.

Host A (10:57): It's a striking contrast, really. While billionaires are suing each other and the Pope is trying to save our souls from deepfakes, the technologies themselves aren't standing still. They're rapidly penetrating places where mistakes cost far too much.

Host B (11:11): Oh yeah, into the real sector.

Host A (11:13): I'm looking at a fresh post from OpenAI's blog. So they partnered with the giant Dell to deploy their Codex model locally—meaning On Premise. This means banks, intelligence agencies, pharmaceutical corporations can now install AI directly on their own servers, inside their closed corporate perimeter. But why is this even being presented as a breakthrough? Aren't we already using neural networks everywhere, from any smartphone?

Host B (11:41): We use them as consumers through the cloud. But for large enterprise, the cloud is often simply an impenetrable zone.

Host A (11:49): As in because of security?

Host B (11:50): Of course. Imagine a large European bank or pharma company—they operate data protected by the strictest standards like GDPR in Europe or HIPAA in the US. They literally don't have the right to send even 1 line of their code or even 1 patient medical record to OpenAI or Google servers.

Host A (12:10): So they were just watching this party of life through glass?

Host B (12:13): Exactly. It was a colossal barrier. Corporations saw AI's effectiveness but couldn't let it near their main secrets. And this partnership with Dell changes everything.

Host B (12:25): Deploying the Codex model on local hardware means that source code and data never leave the company's closed loop. This translates AI from the category of some experimental cloud tool into the status of a basic, absolutely secure corporate infrastructure.

Host A (12:42): Listen, since we've talked about pharmaceuticals and those working with super-sensitive data, there's one more piece of news that sounds like science fiction that just became routine.

Host B (12:54): Are you talking about Sandbox AQ?

Host A (12:56): Yes. Sandbox AQ integrated its proprietary models for developing new drugs directly into Claude's interface. And now biologists can analyze complex molecules by just chatting with a chatbot. In regular human language. No programming required.

Host A (13:11): Fantastic. Do I understand correctly that competitors in this field, like Chai Discovery or Isomorphic Labs, are trying to make their models bigger and smarter, while Sandbox AQ simply made it so anyone could talk to theirs?

Host B (13:27): Yes, they chose the path of accessibility.

Host A (13:30): Wait, how does a text language model understand chemistry at all—they're completely different things?

Host B (13:35): Well, this is one of the most elegant tricks in modern machine learning.

Host A (13:39): So-so.

Host B (13:40): For a language model, there's no fundamental difference between the English language and the language of biology. Amino acids in proteins or chemical structures of molecules can be written as text strings—well, for example, through a format called SMILES.

Host A (13:55): Aha, so chemistry turns into text?

Host B (13:58): Exactly. And if a model can learn English grammar and predict the next word in a Shakespeare sentence, it can absolutely learn the grammar of chemistry and predict how a protein will fold or how a molecule will bind to a receptor.

Host A (14:16): Wow!

Host B (14:17): And as for integration into Claude, it's purely a matter of lowering the barrier to entry. Before, a biologist wanting to test some hypothesis had to basically go beg the Data Science department. Wait weeks while they write code and run the data.

Host A (14:33): And now?

Host B (14:34): Now a researcher just types "Show me analogs of this molecule with lower liver toxicity." And that's it. Sandbox AQ's built-in models do all the complex math, and Claude translates that back into an answer a human understands. This is true democratization of discovery.

Host A (14:53): Okay, we've discussed how AI is penetrating bank server rooms and biologist test tubes. But there's material from Habr.AI that made me read it twice. Interesting. An experiment that the author himself called "neuropunk." So, an enthusiast took Google's Gemini Nano language model—it has just 270 million parameters.

Host B (15:19): That's tiny.

Host A (15:20): Against the backdrop of trillion-parameter monsters from OpenAI, it looks like dust. But what did he do? He trained this tiny model to control a physical caterpillar robot with a manipulator arm. The training happened in a virtual simulation, and then this neural network was simply transferred to a real robot, and it drove.

Host B (15:40): Yeah, I saw that video.

Host A (15:41): Without internet, without connection to any cloud, completely autonomous, navigating space and grasping objects.

Host B (15:48): And it's far more complicated than it seems at first glance.

Host A (15:52): Listen, explain to me why it's so hard? Doesn't it work like, well, in the movie The Matrix, when they upload the kung-fu program into Neo, and he just opens his eyes and goes, "I know kung-fu"?

Host B (16:03): The Matrix analogy works great. But the devil is in the physics.

Host A (16:08): What do you mean?

Host B (16:09): Look, in simulation gravity is perfect. Friction, wheels, surface are absolutely predictable, and lighting never changes. But the real world is chaotic.

Host A (16:20): Oh, dust, dirt?

Host B (16:22): Yes. A wheel can slip on a dusty floor, a manipulator can encounter unexpected material resistance, light from a window can just blind the camera. And for a long time, models that brilliantly worked in simulation turned out to be absolutely helpless in reality.

Host A (16:39): And this experiment changed something?

Host B (16:41): The uniqueness of this experiment is in 2 things. First, the model learned to generalize its experience and cope with this chaos of the real world. And second, the size. 270 million parameters. We're used to thinking that intelligence requires giant servers, right?

Host A (16:59): Yeah, we just discussed Meta's $145 billion investment.

Host B (17:03): Right. But a physical robot that needs to, say, catch a falling glass in real-time, it doesn't have time to send a request to the cloud and wait for an answer. It needs millisecond latency. A compact model works locally on a cheap chip, consumes a pittance of energy, and responds instantly. This proves we can create smart, autonomous agents for factories, warehouses, or agriculture without tying them to data centers at all.

Host A (17:33): Listen, what does all this mean on the scale of the global race? I'm just struck by the speeds. Let's look at the final 2 cases in our pile today to just appreciate the pace.

Host B (17:45): Let's do it.

Host A (17:46): First case. A Russian team launches AIHub in the Max messenger. In 54 days. From idea to launch. They gathered more than 10 models in one chat, added 6 multimodal input formats, connected subscription through Yucassa, which launches in May 2026, and they already have 5,903 users and 300 daily active ones. 54 days.

Host B (18:18): The speed is insane.

Host A (18:20): And right alongside, a news item on a completely different scale. The Chinese developer Moonshot.ai, which makes the Kimi chatbot, raises $2 billion from Meituan, DragonBall, and state giant China Mobile. The company valuation jumps to $20 billion. Growth of almost 7x in just 16 months if you count from January 2025. How is this even possible? One builds a working product on a shoestring in 2 months, and the other in a year and change skyrockets to dozens of billions.

Host B (18:54): These 2 examples are just 2 sides of the same coin showing exactly how AI is capturing the world right now.

Host A (19:02): So, explain.

Host B (19:03): The Max messenger case perfectly illustrates the unprecedented lowering of barriers for developers. The infrastructure has become so modular and accessible that a small team no longer needs to write their own neural networks from scratch.

Host A (19:19): They just take ready-made ones via API.

Host B (19:21): Yes, they can use open APIs and focus exclusively on user experience, on bringing these 10 different AIs right to where people are already used to chatting—into the messenger interface. 54 days to a working product with monetization is the new normal speed of building value.

Host A (19:42): And the Chinese startup?

Host B (19:44): The Moonshot.ai case exposes a different macroeconomic layer. A $20 billion valuation and participation from China Mobile and the CITC fund, which is closely tied to the state. This tells us that at this level, the rules of free market take a step back.

Host A (20:02): So it's geopolitics?

Host B (20:05): Absolutely. Beijing isn't just funding a startup; it's implementing national strategy. AI has become the foundation of technological sovereignty and national security. States understand that losing this race will mean dependence on foreign intellectual infrastructures in the future, so national champions will be flooded with billions without regard to classic, you know, return-on-equity multipliers.

Host A (20:32): Let's try to tie this all together. We started with ruthless layoffs at Meta, where living people are essentially sacrificed for $145 billion in server purchases. Then we saw Anthropic gobbling up tools like Stainless to cut off competitors' lifelines for integration. We watched courts between former OpenAI partners destroy industry trust so badly that the Vatican has to write an encyclical to remind corporations about humanity.

Host B (21:08): Very accurate summary.

Host A (21:10): And in parallel with all this drama on the tech Olympus, technology is routinely changing the world. Codex protects corporate secrets, Sandbox AQ accelerates biologists, the Max messenger in a couple months gives people access to all the world's models, and China is building its own AI giants in a year.

Host B (21:30): And you know, if you look at all these threads simultaneously, we'll see a picture that's preparing us for the next absolutely inevitable stage.

Host A (21:40): Which one?

Host B (21:41): Let's connect 2 facts from what we've detailed today. Fact one: an enthusiast proves that a tiny 270-million-parameter model can autonomously learn in a virtual environment and start brilliantly controlling a complex physical robot, adapting to reality.

Host A (21:59): That very Neuropunk?

Host B (22:01): Yes. And fact two: industry giants are right now flooding hundreds of billions into massive compute clusters.

Host A (22:09): Like Meta?

Host B (22:09): Like Meta. And the question we need to ask ourselves: what will happen when these 2 trends intersect? What happens when a data center's computing power worth $145 billion is thrown not at generating pictures on the internet or text, but at simulating physical worlds? We'll find ourselves on the threshold of a moment when these systems can generate a million sim iterations per second, training entire swarms of autonomous machines.

Host A (22:40): That sounds both frightening and wonderful.

Host B (22:44): Robots, drones, industrial mechanisms will be able to get perfect behavioral algorithms before they even come off the assembly line.

Host A (23:32): So cloud giants will become like schools for legions of physical machines. Listen, that's an amazing thought. It turns everything we discussed today, from chip costs to ethics, into a very concrete reality of tomorrow. On that we conclude our exploration. Let this thought about the inevitable merger of boundless cloud budgets and autonomous physical robots stay with us for our own reflection. See you in our next deep dive!