Manus AI: Another "Revolution" Built on LLM Limitations
Manus AI is built on existing LLMs and inherits their inability to verify their own outputs, so it cannot be a truly autonomous agent. The workable approach is human-AI co-thinking, with humans as verifiers.
This article is more than a year old. Some of its findings may have changed since.
Without robust verification mechanisms, no LLM-based system can function as a truly autonomous agent.
The media is buzzing about Manus AI—a supposedly fully autonomous agent. The only way to cut through the noise is to anchor ourselves in fundamental scientific evidence. By now journalists (especially those claiming to be the best) should have learned some fundamental truths about LLMs.
The Fundamental Design Flaw
Manus AI is built on existing LLMs (Claude, Qwen) with a tool orchestration layer. Regardless of implementation details, this means it inherits all the fundamental limitations of current LLMs, most critically their inability to verify their own outputs.
By their very own probabilistic nature, LLMs have no inherent guarantee of correctness. Allowing them to generate a longer sequence of tokens at inference time doesn't solve this; it often amplifies the problem.
Why Enhanced Reasoning Approaches Work
Recent approaches to enhance the reasoning capabilities of LLMs, such as reinforcement learning or supervised fine-tuning on problem solutions and derivational traces, work because of the external verifiers that are in the loop. Yet after post-training, when the LLM is enhanced but left to its probabilistic nature, impressive-looking reasoning still lacks rigorous verification.
Predictable Real-World Failures
Emerging user reports confirm what experts would readily predict: Manus AI fails to plan, gets stuck in loops, and fails to complete tasks. These aren't isolated incidents but predictable consequences of its fundamental design.
Here is the inescapable scientific reality of today's LLMs: without a robust, external verification mechanism (human or algorithmic), no LLM-based system can function as a truly autonomous agent.
The Human-AI Co-Thinking Alternative
The only viable (and still awesome!) approach is human-AI co-thinking, where AI is the generator, and humans the verifier—a methodology that AI Swiss has been advocating for and demonstrating across society since its inception.
Let's stop confusing tool integration and computational resources expended on token generation with genuine autonomy. With existing LLMs, the revolution is not about autonomy, but cultural transformation through scientifically grounded human-AI co-thinking.