LLM agents are not everything: The AI community recalls other paths

Fatigue from chat windows

Former Google Brain researcher and Sakana AI co-founder, known by the handle hardmaru, issued a brief warning to the AI community. He reminded them that while language models and coding agents work great, they represent only a fraction of what artificial intelligence offers and can mean.

Where the next horizons lie

While all capital and attention is focused on training ever-larger LLMs and building agentic frameworks around them, disciplines like evolutionary algorithms, neuroevolution (NEAT), and artificial life remain in the shadows. Users in the discussion beneath his post mentioned precisely these alternative directions as areas from which the next big breakthrough might emerge once autoregressive models hit their limits.

Who determines what AI is

The current situation, where the general public has equated the term artificial intelligence with LLM chatbots like ChatGPT, leads to underinvestment in alternative approaches. If a research group is not working on LLMs or visual diffusion models, it struggles to find understanding from investors.

When the bubble bursts

The proof of change will not be the announcement of GPT-6, but the moment a lab the size of OpenAI shifts a significant portion of compute capacity to an architecture not based on predicting the next token. Until then, alternative approaches remain largely in the academic realm and fringe startups.

Lilith's verdict

When you hold a sufficiently large language model, every problem looks like text that needs generating. But the best know you can't build the real world out of text.