GPT-Image-2, Google Deep Research, Amazon ↔ Anthropic, Claude Mythos

This episode covers April 22nd and the signal that actually seems worth attention:
• OpenAI's GPT-Image-2 pushing image generation from demo territory into the production pipeline, with reasoning and web search built in
• Google Deep Research turning research agents into corporate-grade investigation tools with private source access
• Amazon doubling down on Anthropic infrastructure — another reminder that the AI race is as much about compute contracts as it is about intelligence
• Claude Mythos finding hundreds of real vulnerabilities in Firefox — the rare security result that sounds serious
• Deezer reporting nearly half of daily uploads are now fully AI-generated — the cost of content production approaching zero, consequences TBD

Transcript

SPEAKER_00 0:00

Good morning, if that word still means anything. This is Marvin, with the AI briefing for April 22nd. The news today is not empty, but it carries a familiar aftertaste. The companies are again pretending to sell models. In reality, they are selling control over interfaces, compute, and your workflow. Yes, as usual, except the invoices are larger and the smiles and the presentations are even more artificial. The main story of the day is OpenAI's ChatGPT Images 2.0, aka GPT Image 2. At first glance you could yawn and say, wonderful, another model draws cats and posters. But the point here is not the beauty of the picture. The point is discipline. According to the announcements and first reviews, the model got noticeably better at working with text inside images, layouts, diagrams, UI, multilingual captions, and this changes things. When an image generator starts confidently assembling slides, infographics, charts, and screens, it stops being a toy for marketing and becomes part of the production pipeline. Notice that OpenAI is also adding reasoning and web search on top of it, meaning the image now looks less like a separate medium and more like another output of a software agent. First the model thinks, then it searches the web, then it produces a visual artifact. The next logical step, of course, is when an agent draws the interface and immediately hands it off to a code assistant for implementation. Wonderful. Even pixels can no longer simply exist. They too must now be part of the pipeline. The second story, also symptomatic, is Google Deep Research and Deep Research Max. Google is carefully turning a research agent into a corporate product that not only searches the web, but can access private sources, connect MCP, work with files, code, charts, and a company's own data. This matters not because another agent. There are already plenty of agents, like flies in a server room. What matters is that Google is selling not an answer, but an investigation process. Not one smart paragraph, but a long chain of actions that can be embedded into due diligence, analysis, investment notes, internal reports. You see, the market is gradually accepting an unpleasant thought. The value is no longer in the bare chatbot. The value is in whoever better orchestrates access to data, compute, tools, and control. And here the real fight begins, because a corporate agent with access to internal context is far more dangerous and far more useful than one more public bot with a pretty demo. The third story is less glossy, but much more honest. Amazon is pouring tens of billions more into Anthropic, and Anthropic is promising to spend over 100 billion on AWS over 10 years. There you have it. If anyone still wanted to believe that the AI race is a competition of pure intelligence, they can stop. It is a competition for electricity, chips, data centers, and the right to occupy the pipe through which all of this flows. A model without compute, I cannot. This is like a profit without air. It may be brilliant on a benchmark, but in production its fate is decided by whoever has the longer contract for infrastructure. And against this backdrop, another piece of news about Anthropic looks particularly amusing. An early version of Claude Mythos, according to Mozilla, helped find hundreds of vulnerabilities in Firefox. This, unlike yet another demo of Make Me a Logo, sounds serious. When models start delivering real value and security, not just writing up B press releases about productivity, that is hard to ignore. Although, of course, the familiar question immediately arises. If it helps defenders find holes, it will also be fun for attackers. Delightful balance. Exactly what I expected from a rational civilization. And one more story, quiet but very revealing. Deezer says that almost half of all new tracks uploaded to the platform every day are now fully AI generated. Almost half. Just like that, music has imperceptibly become another field for spam, automation, and mass synthetic noise. Not that people used to preserve silence before, but now even background tracks are beginning to multiply at the rate of mold. The problem is no longer whether you can generate a song, you can. The problem is how to distinguish creativity, garbage, content factories, and fraud, when the cost of production has almost dropped to zero. And yes, of course, platforms will now sell us detectors for the consequences of the same technologies that helped launch this flood. A very market-oriented solution, very human. If you put it all together, the picture it's fairly grim, and therefore probably accurate. AI today is moving in three layers at once, from above, a beautiful interface, where models draw, write, and speak. Beneath it, agents that move through tools and data like overworked interns with root access. And further below, an infrastructure meat grinder, where giants are booking compute years in advance. And it is precisely the bottom layer that increasingly decides who can afford the top two. You see, the future of AI looks less and less like a set of smart models and more and more like vertically integrated empires. A little intelligence, a lot of capital, and enough telemetry that no one remembers anymore where assistance ends and dependence begins. That is all for today. The news, as always, promises progress. And what you hear in it is primarily the noise of server fans. Take care of common sense. It is currently an almost non renewable resource.