American AI Self Sabotage with FLOP Restrictions
Americans and Europeans shoot themselves in the foot while China advances
How much can politicians sabotage the AI landscape with bad policies? We’re about to find out. It’s starting with arbitrary FLOP restrictions.
Both the United States and the European Union have rolled out FLOP restrictions. FLOP is shorthand for floating-point operation.
What exactly is a FLOP?
A FLOP is a simple math calculation that computers solve. Imagine a book of math problems. Every problem involves basic arithmetic. Every time a computer solves a problem - it does a FLOP.
The problem with FLOP restrictions is that it likely to fail to mitigate risk at all. See the following excerpt from Sara Hooker’s latest paper on compute thresholds with FLOPs as a proxy.
Both the White House Executive Orders on AI Safety (EO) and the EU AI Act encode the use of FLOP or “floating-point operations” as a way to identify more powerful systems. What is striking about the choice of compute thresholds to-date is that no models currently deployed in the wild fulfill the current criteria set by the EO. This implies that the emphasis is often not on auditing the risks and harms incurred by currently deployed models – but rather is based upon the belief that future levels of compute will introduce unforeseen new risks. A key conclusion of this essay is that compute thresholds as currently implemented are shortsighted and likely to fail to mitigate risk.
On the Limitations of Compute Thresholds as a Governance Strategy. by Sara Hooker
Models for different things have different relationships with scaling and FLOPs. Higher compute models with FLOPs as a proxy generally scale better with language models. This suggests to me that any single number for a FLOP restriction is completely arbitrary at best and doesn’t serve the claimed premise of safety guard rails at all.
Established AI companies with more resources can comply more easily with FLOP restrictions - that they directly dictate through lobbying resources. This creates barriers for smaller or new entrants who may struggle with the costs and complexities of meeting these regulations. FLOP restrictions become a form of regulatory capture by larger first movers. This is anti-competitive regulation under the guise of providing safety.
Unlike America and Europe, China is working to actively increase compute.
China plans to increase its computing power by 50% by 2025, the country’s key ministries said Monday, as it looks to keep pace with the U.S. in artificial intelligence and supercomputing applications.
The world’s second-largest economy wants to have computing capacity equal to 300 exaflops, according to a plan from six government departments, including the powerful cyberspace regulator. That would be up from the 197 exaflop computing power the country currently has.
“China has found that traditionally, every 1 yuan invested in computing power has driven 3-4 yuan of economic output,” Akshara Bassi, senior research analyst at Counterpoint, told CNBC via email. “The investments echo China’s plans to drive economic output through leadership in technology prowess and integrating AI with existing technologies and solutions across all industries and domains.”
This is a negative trend for Americans. We must undo these FLOP restrictions in my opinion.
If you enjoyed agree with me, please share with someone you know. It’s important that America does not fall behind in AI with stupid policies.
We need to undo this executive order on FLOP restrictions from last year’s executive order. I’m fairly confident that Biden had no idea what a FLOP is when he signed this. We can’t allow lobbyists to enact regulatory capture and sabotage American leadership in AI.
Have any thoughts or questions on AI or FLOP restrictions? Let me know. Also open to perspectives I may not have not considered.