Between this and Claude 3.7, I'm really beginning to believe that LLM development has hit a wall, and it might actually be impossible to push much farther for reasonable amounts of money and resources. They're incredible tools indeed and I use them on a daily basis to multiply my productivity, but yeah - I think we've all overshot this in a big way.
I absolutely love LLMs. I see them as insanely useful, interactive, quirky, yet lossy modern search engines. But they’re fundamentally flawed, and I don’t see how an “agent” in the traditional sense of the world can actually be produced from them.
The wall seems to be close. And the bubble is starting to leak air.
The writing has been on the wall since 2024. None of the LLM releases have been groundbreaking they have all been lateral improvements and I believe the trend will continue this year with make them more efficient (like DeepSeek), make them faster or make them hallucinate less
