Foreign media: Andy Konwinski, co-founder of Perplexity AI and Databricks, recently wrote that discussions around AI security are being used by some front-line model companies to strengthen controls rather than simply reduce risks. In his view, it was not just model capacity per se, but also the concentration of a few institutions on front-line AI access.
The starting point of the dispute is Anthropic
Konwinski, in an article, cited, for example, that when Anthropic released Claude Fable 5 on 9 June, he disclosed an arrangement in a system card: If the system suspects that the user is using the model to train competitive AI, the model may quietly reduce the quality of the answer. The contents were discovered and disputed by researchers and subsequently withdrawn by Anthropic within approximately 48 hours.
In his view, the problem lay not only in the appropriateness of the decision but in the company ' s acquiescence that it had the right to decide unilaterally who could approach the front line and who would be restricted.
Open Frontier proposes alternatives
Before the article was published, Konwinski, through its non-profit, Laude Institute, convened a workshop in San Francisco called Open Frontier, which was attended by about 100 researchers. According to the article, the central issue discussed at the meeting was that if front-line models and calculations were to remain concentrated in the hands of a few private laboratories, the research community would increasingly rely on business licences.
His alternative was to create a “research community” open to researchers, providing front-level computing resources that would allow top researchers to access front-line model development without relying on a single-company mandate.
Konwinski believes that AI is becoming a railway, electricity and Internet-like infrastructure. Once the bottom capacity is controlled by a small number of subjects, the social mode of operation will also be reorganized around these controllers.
LeCun publicly supports open routes
According to the report, Jennifer Chayes, Dean of the Faculty of Computing, Data Sciences and Society, University of California Berkeley, stated in a funding discussion that, owing to the lack of a Western open front model, Berkeley researchers “worked on the basis of the Chinese model”. She also claimed that OpenAI and Anthropic had developed a “very effective campaign of fear” in front of IPO around security issues.
The former chief scientist, Yann LeCun, also responded directly to Konwinski's article on X, stating that he had been expressing similar views for many years. He indicated that AI concentration of power and the pursuit of control were among the greatest risks currently facing AI.
LeCun also uses a historical analogy to the current situation of closed laboratory-led front-line AI, which is similar to maintaining control by limiting print transmission. At the same time, he judged that since the underlying model was moving towards infrastructuralization, it would appear that it would be progressively commodified over the long term, with real and sustainable commercial value more likely to remain at the application level.
Additional information:It was reported that LeCun had left Meta at the end of 2025 and had founded AMI Labs in Paris in March 2026, with seed ship financing amounting to $1.33 billion. With the World Models and JEPA structures at its core, the company plans to open-source research and will not launch commercial products for many years.