Free-to-download AI models are already everywhere
Investor Sarah Guo says businesses want open models to control cost and capacity, so bans would only restrict law-abiding American businesses.
Invest Like the Best with Patrick O'Shaughnessy · Sarah Guo - Funding the Frontier - [Invest Like the Best, EP.489]Sarah Guo, a venture investor who runs the AI-focused firm Conviction, was asked what America should do about open AI models that rival the best private systems. Her answer was blunt.
"The cat is out of the bag, and these are in use everywhere," she said.
By open models, she meant AI systems anyone can download, modify and run themselves, as opposed to paying to use frontier models, the most advanced, largest systems from the top labs.
Guo argued that over the past three years, increasingly competitive open models have arrived from all directions, largely from China but also from the United States and Europe. She pointed to recent Western companies and model makers including Thinky, Poolside, NVIDIA and Mistral, and the widely anticipated Reflection, as evidence that powerful open systems already exist.
That spread will accelerate, in her view, for practical reasons. There are, she said, a huge number of cases "where it is too expensive, too sensitive, or too slow" for a business to send its data to a frontier provider. As people learn to do more with AI, she expects those cases to increase, because using the biggest models sold as paid services is objectively expensive. More broadly shared capability, she believes, will simply be used in more ways.
Restricting open models in the United States would, Guo argued, "basically just restrict law-abiding American businesses," slow them down or shift profits elsewhere, while attackers planning adversarial uses would ignore the rules. She added that the same frontier capability her companies use for defensive cybersecurity and biology work could obviously be turned to offensive or bioweapons uses, so the danger should be faced directly.
What she favors instead is testing of the most advanced models. For worries about hidden backdoor behavior in Chinese models, she said, the answer should be a rigorous set of safety tests, to learn what the risks actually are. There has not been nearly enough real research, she said, only speculation. She embraces a future of intelligence too cheap to meter, a phrase she credited to Sam Altman, OpenAI chief executive, supported by open models that let businesses control their own costs and capacity. No researcher in a frontier lab, she said, can imagine the diversity of uses every individual will find.
Questions people ask
What is an open-source AI model?
An open model is one whose inner settings, often called weights, are published so anyone can download and run it on their own computers. A closed model stays on the maker's servers and is used only through that company's paid service.
Why would a company prefer an open model to a big private one?
Using a giant private model can cost a lot per query, add delay, and require sending private data to someone else. Running a smaller open model in house can be cheaper, faster, and keep sensitive information inside the company.
What kind of safety testing is she calling for?
She means systematically probing a model to see what it can do, whether it hides unexpected behavior, and whether it could help with cyberattacks or biological weapons. She argues that kind of measurement would replace speculation with evidence.