Netcompany Snippets #9

The Brussels-Sacramento Axis

The Brussels Effect is well-known: companies around the world are in effect forced to comply with EU regulations to gain access to its vast market. But Luke Hogg at the Foundation for American Innovation thinks that Californian regulations are not much different. Sure, the EU has GDPR, but California has the California Consumer Privacy Act, CCPA.

More regulation...

Speaking of GDPR and other regulatory regimes, Brian Chau notes that “If you live in the EU, expect to increasingly become a second-class digital citizen, just like China in the last decade”. Research suggests that indeed the GDPR regime stifles innovation, but the question remains: what values do we stifle if we loosen the regulatory stronghold?

Gemini will be significantly bigger than ChatGPT

Google’s ChatGPT competitor, Gemini, is coming this December. The model is pretrained on Google’s own TPUv5 chips, initially with 5x of GPT-4's flops, and projected to 20x by end of next year. In “Portents of Gemini”, Zvi argues that although this increase in compute looks extreme, it is “the type of progression you would expect”.

Fighting the New Digital Dark Age

Data grows exponentially, but Adrienne Bernhard believes “we have fallen from a golden age of preservation in which everything of value was saved.” The reasons are limitations on hardware longevity, as well as a lack of format accessibility and comprehensibility. “The first Dark Ages”, Bernhard writes, “weren’t actually characterized by intellectual and cultural emptiness but rather by a dearth of historical documentation produced during that era”. To combat the New Digital Dark Age, The Rosetta Project is a long-term archival method where “pages are microscopically etched and then electroformed in solid nickel, a process that raises the text very slightly - about 100 nanometers - off of the surface of the disk.”

Paul Christiano tries to explain neural network behaviors

Paul Christiano is one of the main people working on AI alignment, and he recently presented a new idea at Simons Institute. In short: Researchers would love to explain why neural networks work, but proving statements about how they work is nearly intractable. Christiano’s idea is to formalize heuristic arguments instead of proofs. His hope is that interesting neural network behavior will always have an explanation recognized by such a heuristic estimator.

The first global, AI safety summit

In June, UK Gov announced that it is to host the first global summit on AI safety. Now, the UK Government announced its goals for the conference, which will take place on November 1-2. The agenda includes a shared understanding of risks, the development of (inter)national frameworks, areas for collaboration, new standards of governance, and appropriate measures for individual organisations.

Remember to subscribe for more bi-weekly Snippets