In Response to Dario's Position on Open Weight Models

Link: https://www.anthropic.com/news/position-open-weights-models

Lengthy legal battles did not kill piracy. What "killed" piracy (or made it significantly less attractive) was products like Spotify (i.e. providing a legal product that is easier, safer, reliable, and reasonably priced).

Same thing happened with strong encryption software in the 90s - the US govt tried to control it with export controls to try to prevent bad actors from getting access. But bad actors were never going to respect those restrictions, and limiting access mainly made it harder for law abiding companies and individuals to defend themselves.

Same principle applies to AI models and their cybersecurity applications. Restricting access just makes it harder for law abiding American companies to protect themselves more than it slows down bad actors who won't think twice about sidestepping regulations. Give everyone access to the same tools and the asymmetric advantage is nullified.

We just saw a real world example of this with the OpenAI - Hugging Face incident. Hugging Face's defenders initially tried using frontier models to investigate the attacks, but the models' safety guardrails blocked the actual exploit code and attacker commands. They instead had to use GLM 5.2, an open weight model, on their infrastructure. If we try to block open models - it would have only further limited their ability to contain it and understand what happened.

In his post, Dario argues for safety regulation, which sounds understandable and innocent. The problem is in the long run regulation often favors the larger better funded incumbents. The biggest labs already have legal teams, government connections, security infrastructure, compute, etc; smaller companies don't have the same level of resources, so satisfying regulations puts a much bigger strain on them.

To tie it to ad tech - there are some similar examples in the history of our industry.

• When Apple introduced ATT, the market expected Meta to be one of the biggest losers. Initially it was - they forecasted a $10B impact and their stock fell 26% in one day. However - they were able to leverage their existing relationships and spend billions of dollars rebuilding their targeting and measurement systems around server side advertiser signals (via Conversions API) and leverage their massive volume of first party data. Meanwhile - smaller advertising companies had far fewer options to adapt and were disproportionately impacted.
• Google Privacy Sandbox presented a similar issue. Rebuilding a programmatic ad stack around a new browser API is a significant undertaking for DSPs, SSPs, publishers, and measurement companies; and this new architecture was expected to yield significantly worse outcomes. Meanwhile, Google has a vast trove of first party data to learn on and practically infinite resources to invest. Thankfully this never came to fruition, but the industry already spent significant resources preparing for the transition.

I believe the same dynamics are at play in the discussion on model regulation. Whether or not you believe safety regulations are effective at reducing risk, they would likely have the effect of entrenching the large incumbents and making it substantially harder for smaller competition to exist.