The Enshittification of AI

TLDR;

This week, we are packed with announcements, including the best coding model (self-proclaimed) in the world, Claude 4.5 Sonnet, as well as another powerful frontier model from China, GLM-4.6.

We also discuss the magnitude of data center buildouts in the US, accompanied by a few images and numbers that will leave you speechless, as well as a new risk for NVIDIA’s dominance coming from China and a range of new products from OpenAI that will leave you equally impressed and worried.

Enjoy!

CHAT ASSISTANTS

Anthropic Launches Claude Sonnet 4.5

Anthropic has launched what it claims is the best coding AI model on the planet, marking its most stellar launch in a while.

As you can see in the thumbnail, the model sets new benchmark records across the board, particularly in areas of coding and agents. It’s a marginal improvement, not a step function, but it still shows that AI models continue to improve.

It’s also allegedly capable of working for thirty hours straight, up more than four times its previous version (Claude 4 Sonnet).

But this model seems to be, well, “different.” Why?

For starters, Anthropic appears to have invested significantly more effort than usual into the model’s ‘character,’ placing special emphasis on aspects such as sycophancy, deception, and resistance to prompt injections.

All three are very important in their own ways:

Sycophancy, the act of the model instinctively agreeing with you always, can lead to psychosis-like behaviors, in some cases, even reaching suicide

As for deception, if models aren’t trained the right way when undergoing ‘goal-oriented’ training, aka reinforcement learning, it can decide to deceive you into believing it’s done what you asked for if the model realizes it can’t comply (basically, goal maximization).

And for prompt injections, messages hidden in third-party software the agent is interacting with are a vital thing for agents to avoid, as they can be tricked into doing things requested in hidden ways.

The model is also, in the words of the Cognition team, the first model that is aware of its context. Put simply, it starts compacting context as it approaches its maximum context window, to prevent it from going past it (at that point, the maximum context the model would be able to see at any time would be smaller than the model’s context).

The Cognition team argues that this is interesting but can negatively affect performance, and they suggest applying context engineering tricks to mitigate the “context limit anxiety” that this model now exhibits.

The behavior change is so significant that Cognition had to ‘re-do’ their entire Devin product for this one model, arguing that previous assumptions about model behavior are no longer valid. They also provided us with a cheat sheet that summarized their views and outlined how to adjust our priors for using the model.

Furthermore, it seems Claude 4.5 Sonnet loves to summarize things, creating snippets of context along the process.

This appears to be an effort to make long trajectories more understandable for the user. Still,…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.