"If you are interested in human-level AI, don't work on LLMs." -- Yann LeCun
Professor @ylecun made some crucial points recently. In my opinion, the most important one is: "If you are interested in human-level AI, don't work on LLMs."
Virtually every word in this presentation drives our work on "Forge" and "MALWIM" (Multi-Stage Agentic Loops With Inhomogeneous Models). We believe in an AI future that is specialized, democratized, affordable, and unique to each use case and company. We do not believe in a centralized AI landscape dominated by just a few massive players. The reason is clear: once you make models compact, efficient, and affordable, companies switch from being mere "wrappers" to actual model creators within their core application layer.
Beyond that strategic purpose, LLMs deeply lack the capabilities required for human-level AI, such as internal long-term memory, the ability to predict actions and consequences, and a holistic, global understanding of a topic.
The mainstream world may still be figuring out LLMs, riding the hype wave, and catching up to what visionaries like @ylecun view as a dead-end architectural path. Mainstream attention might not shift for another 24 months. However, once you see the reality, you cannot unsee it: no AGI, no ASI, and no major future breakthroughs can be achieved through sole-LLM architectures and basic agent designs.
We are currently training our first Forge/MALWIM model and agentic system, specifically designed for coding. Over the coming months, we will release benchmarks, results, and potentially a limited-access beta launch.View quoted post on X
