In a recent post, an author argues that large language models (LLMs) will likely continue to function as "cracked interns"—effective when guided by humans but incapable of full autonomy in most domains. The analysis posits that the inability of most firms to adopt fully autonomous AI is not due to skill gaps or slow technology diffusion, but rather to structural reasons inherent in current architectures.
Why LLMs May Remain "Cracked Interns" Due to Structural Architecture Limits
The author identifies three specific classes of firms that might successfully utilize LLMs. The first two are price-sensitive organizations that do not require the high reasoning capabilities of frontier models. These firms may benefit more from running cheap, open-source models on local hardware. For these users, success appears to depend more on "agentic swarm width"—using many small models to perform combinatorial searches—rather than the raw reasoning capacity of a single large model.
The third class includes firms involved in highly specialized fields, such as mathematics or security research. While these firms might use frontier models, the author suggests that the advantage of using wider swarms of cheaper models, such as DeepSeek V4.1 Flash, might outweigh the benefits of high-end reasoning. Additionally, the author notes that these firms often prioritize intellectual property secrecy, making them wary of sending sensitive data to providers like Anthropic or OpenAI, even with data usage agreements in place.
Ultimately, the author suggests that the future of AI compute may be driven more by massive "brainlet swarms"—AI acting as tools for human orchestrators—than by a move toward self-driving artificial superintelligence.
Sources
- Why I'm still bearish on LLMs after Navier-Stokes （Hacker News Frontpage, 2026-09-15）