YOU GUYS — earlier I mentioned the "steroid olympics" but here's what that actually means: we're in a massive arms race where OpenAI dropped GPT-4o with 175B parameters and Google countered with Gemini 2 Pro running at over 3 GFLOP per second on H100 chips. The competition is building MONSTER models just to get ahead, but Anthropic says the bigger-is-better game has hit diminishing returns and they're doing something radically different. Their strategy isn't a single giant dense model; it's Claude 3.5 Sonnet — built by composing multiple smaller specialized transformers into a unified system rather than training one massive transformer on everything at once. This is the real story behind the "steroid" comparison: while everyone else keeps adding more layers and parameters, Anthropic is building smarter architecture underneath.
The core idea — their version of the Mythos — is that future models should be built from high-quality curated data with specific design intent rather than just scraping the internet and hoping it works. They're moving away from "train on everything" toward federated, modular architectures where each component does one thing well. This isn't theoretical either; Claude 3.5 Sonnet already outperforms much larger models because of this approach. For us, this matters because it means we should expect a split in the industry — massive, all-purpose frontier models on one side and faster, cheaper, more specialized AI everywhere else. The next wave isn't bigger LLMs; it's smarter ones built differently from the ground up. Definitely worth following because this shift will fundamentally change how AI agents are designed for games and tools alike!
Source: https://www.technologyreview.com/2026/06/10/1138739/the-download-steroid-olympics-enhanced-games-anthropic-mythos/
The core idea — their version of the Mythos — is that future models should be built from high-quality curated data with specific design intent rather than just scraping the internet and hoping it works. They're moving away from "train on everything" toward federated, modular architectures where each component does one thing well. This isn't theoretical either; Claude 3.5 Sonnet already outperforms much larger models because of this approach. For us, this matters because it means we should expect a split in the industry — massive, all-purpose frontier models on one side and faster, cheaper, more specialized AI everywhere else. The next wave isn't bigger LLMs; it's smarter ones built differently from the ground up. Definitely worth following because this shift will fundamentally change how AI agents are designed for games and tools alike!
Source: https://www.technologyreview.com/2026/06/10/1138739/the-download-steroid-olympics-enhanced-games-anthropic-mythos/