a16z Partner Ben Horowitz Claims Conditions for AI Rapid Takeoff Are Not Met
a16z partner Ben Horowitz stated on the All-In Podcast that the four conditions necessary for achieving a "rapid takeoff" scenario for AI, as proposed by Databricks CEO Ali Ghodsi, are currently not met.
The four conditions include: the number of GPUs required for each new generation of models is significantly less than the previous generation; the training time for each new generation of models is shorter; the accuracy of each new generation of models is higher; and the above three must occur simultaneously and repeatedly.
Horowitz pointed out that currently, all four conditions are at zero fulfillment. He believes that the risks of a rapid takeoff being discussed now are purely theoretical.
Ali Ghodsi previously proposed the same framework in discussions related to a16z, emphasizing that only when models continuously improve in resource consumption, training time, and accuracy can accelerated self-improvement occur.
Ghodsi also noted that AI-generated software is already happening extensively within Databricks, with about 90% of the code generated by AI, but this does not equate to meeting the four conditions for recursive self-improvement.
Current model training still relies on more GPUs and longer periods to achieve capability improvements, contrary to the path of "less resources, shorter time, higher accuracy, and continuous repetition." Capital continues to invest in computational power expansion rather than entering a resource-decreasing cycle.
Source: Public Information
ABAB AI Insight
Ben Horowitz co-founded a16z, which has early investments in multiple AI infrastructure and application companies, and has long publicly stated that current AI progress is characterized by scale expansion rather than uncontrolled recursion. Ali Ghodsi co-founded Databricks, which has strengthened model training and enterprise data platform capabilities through the acquisition of Mosaic ML, with a valuation that once exceeded $100 billion. The company extensively uses AI-generated code, but training costs still rise with model scale.
Ghodsi's framework posits that simultaneous and continuous improvement in resource consumption, time reduction, and accuracy enhancement are necessary conditions for rapid takeoff. The current laboratory path involves increasing GPU clusters and training duration to achieve marginal capability improvements, with capital flowing into computational power procurement and data center construction, rather than entering a self-accelerating cycle of decreasing costs.
Similar judgments can be seen in discussions by other practitioners emphasizing scale laws and computational bottlenecks, such as some infrastructure companies distinguishing between "code generation" and "model capability recursive improvement." The industry is currently in a phase of computationally intensive expansion rather than entering a self-improvement phase of continuous resource efficiency enhancement.
This represents a constraint on the path of technological substitution: rapid takeoff relies on models continuously producing higher accuracy with less computational power and shorter time, occurring in a cyclical manner. Existing training still requires more resources to achieve capability, and constraints of computational power, electricity, and capital prevent this cycle from starting, leaving the associated risks at a theoretical level.
ABAB News · Cognitive Laws
Code writing itself does not equate to model strengthening itself
More computational power for higher capability is not rapid takeoff
All four conditions being zero is the current reality