Get your free personalized podcast brief
We scan new podcasts and send you the top 5 insights daily.
The current approach of building generalist models like ChatGPT, containing all human knowledge, is inherently unsafe. A safer paradigm involves creating specialized AIs for specific tasks (e.g., translation) that lack dangerous capabilities like bioweapon design, similar to how a Pentagon janitor is denied access to nuclear codes for security.
Related Insights
The AI industry is hitting data limits for training massive, general-purpose models. The next wave of progress will likely come from creating highly specialized models for specific domains, similar to DeepMind's AlphaFold, which can achieve superhuman performance on narrow tasks.
Simple refusal mechanisms in AI models are easily bypassed by motivated actors. Effective biosecurity requires deeper interventions, such as curating training data to exclude sensitive biological information or implementing strict access controls for the most powerful models, ensuring they aren't publicly available.
Instead of a single, generalizable AI, we are creating 'Functional AGI'—a collection of specialized AIs layered together. This system will feel like AGI to users but lacks true cross-domain reasoning, as progress in one area (like coding) doesn't translate to others (like history).
Instead of building a single, monolithic AGI, the "Comprehensive AI Services" model suggests safety comes from creating a buffered ecosystem of specialized AIs. These agents can be superhuman within their domain (e.g., protein folding) but are fundamentally limited, preventing runaway, uncontrollable intelligence.
Microsoft’s approach to superintelligence isn't a single, all-knowing AGI. Instead, the strategy is to develop hyper-competent AI in specific verticals like medicine. This deliberate narrowing of domain is not just a development strategy but a core safety principle to ensure control.
AI will not evolve into a single, omnipotent entity. Due to fundamental limitations like context windows, AI will be structured like human organizations: a fleet of specialized agents with distinct roles (e.g., content, research). This mimics how humans partition work to manage complexity.
A novel AI safety technique called gradient routing trains mixture-of-experts models to isolate dangerous knowledge (e.g., bioweapons, cyber exploits) into specific "expert" modules during pre-training. These dangerous experts can then be completely removed ("ablated") before deployment, creating an inherently safer model.
The pursuit of AGI is misguided. The real value of AI lies in creating reliable, interpretable, and scalable software systems that solve specific problems, much like traditional engineering. The goal should be "Artificial Programmable Intelligence" (API), not AGI.
The focus on AGI can obscure more immediate threats. Even narrowly capable AI tools pose existential risks. For example, an AI that only excels at biotechnology research could make it easy for malicious actors to develop dangerous pathogens, regardless of its general intelligence.
Yann LeCun posits that the goal of AI should not be to replicate the breadth of human intelligence (AGI). Instead, development should focus on creating specialized models that achieve superhuman depth in fields like physics and chemistry, as this is where true breakthroughs will occur.
