Interview, Fireside Chat
How to get Washington and Silicon Valley to tame AI | Mustafa Suleyman
- Models are expected to autonomously operate online businesses with minor human oversight, turning $100,000 into $1 million within a couple of months, with full autonomy for such tasks projected for two years, while recursive self-improvement and runaway intelligence explosions are not anticipated for approximately 10 years.
- In 18 months, Inflection AI plans to train models 100x larger than current frontier models, with global model sizes projected to reach 1000x current scale within three years; however, GPT3-like capabilities are expected to become trainable at 1.5 to 2 billion parameters, a trajectory expected to hold for three to five years.
- Inflection AI targets having 22,000 H100 GPUs operational by December and adding 1,000 to 2,000 units monthly to train models larger than GPT-4 by the spring or summer of next year, while noting that a $10 billion training run is at least five years away due to chips delivering 2 to 3x more FLOPs per dollar.
- Open source models are expected to trail frontier capabilities by two to three years for at least the next five years, whereas export controls will likely deny China access to next-generation chips like Hopper Next, slowing their progress by 30 to 50 percent even as they utilize daisy-chained H800s.
- Risks include the rapid proliferation of power allowing individuals to achieve unprecedented one-to-many impact similar to the 1990s social media trajectory, with models potentially reaching 2 to 4 orders of magnitude beyond current levels within a short timeframe, though immediate misalignment risks at Inflection AI are viewed as low due to a lack of autonomy.
- Regulatory responses are expected to focus on preventing AI use in electioneering, implementing scale audits, and shifting voluntary commitments to legislation at the upcoming UK AI Safety Summit, with open sourcing restricted soon due to bioweapon risks anticipated within 10 to 15 years.
- Political discourse is expected to see continued difficulty in engaging with opposing views, and academic access to frontier models may become unsustainable as researchers move to commercial labs within a couple of years, contrasting with the view that recursive self-improvement debates are currently a speculative distraction.