A high-level ex-Openai political researcher, Miles,, taken on social networks Wednesday to criticize Openai to “rewrite history” of its approach to deploy potentially risky AI systems.
Earlier this week, Openai published a document Describing its current philosophy on the security and alignment of AI, the AI systems design process which behave in a desirable and explanatable manner. In the document, Openai said that he sees the development of AG, largely defined as AI systems that can perform any task that a human can, as a “continuous path” which requires “deployment and learning iteratively” of AI technologies.
“In a discontinuous world […] Safety lessons come from today’s treatment of systems with disproportionate prudence compared to their apparent power, [which] is the approach we have adopted [our AI model] GPT – 2, “wrote Openai. “We now consider the first to act as a single point along a series of increasing utility systems […] In the continuous world, the way of making the system accordingly and beneficial is to learn from the current system. »»
But Brundage says that GPT-2 in fact guarantees abundant caution at the time of its release, and that this was “100% coherent” with the Iterative Deployment Strategy of Openai today.
“The release of Openai de Gpt-2, in which I was involved, was 100% coherent [with and] foreshadowing the current OpenAi philosophy of iterative deployment ”, Brundage Written in an article on X. “The model has been gradually published, with shared lessons at each stage. Many security experts at the time thanked us for this caution. »»
Brundage, who joined Openai as a researcher in 2018, was the head of the company’s political research for several years. Within the OPENAI “AG Readiness” team, it was particularly placed on responsible deployment of language generation systems such as the AI Chatbot platform of Openai Chatgpt.
GPT-2that Openai announced in 2019, was a progenitor of the food of AI systems Cat. The GPT-2 could answer questions on a subject, summarize the articles and generate text to a sometimes indistinguishable level from that of man.
Although GPT-2 and its outings may seem basic today, they were peak at the time. Citing the risk of malicious use, Openai initially refused to publish the source code of GPT-2, opting instead of giving new selected access to a demo.
The decision encountered mixed examinations of the AI industry. Many experts argued that the threat posed by GPT-2 had been exaggeratedAnd that there was no evidence that the model could be abused in the way Optai described. Publication focused on the gradient went so far as to publish a open letter By asking Openai to release the model, arguing that it was too important technologically to hold back.
Openai finally published a partial version of GPT-2 six months after the unveiling of the model, followed by the complete system several months after that. Brundage thinks it was the right approach.
“What part of [the GPT-2 release] Has it been motivated by or based on the thought of Ag as discontinuous? None of this, “he said in an article on X.” What is the proof that this prudence was “disproportionate” ex ante? Ex post, he is prob. would have been ok, but that does not mean that he was responsible for the Yolo [sic] Given information at the time.
Brundage fears that OpenAi’s objective with the document is to set up a burden of evidence where “concerns are alarmist” and “you need overwhelming evidence of imminent dangers to act on them”. This, he says, is a “very dangerous” mentality for advanced AI systems.
“If I still worked in Openai, I would ask why [document] was written as it was, and what Openai hopes exactly to achieve by cautiously cautious, “added Brundage.
OPENAI has historically been accused to prioritize “brilliant products” to the detriment of safety and Product sub-reports To beat rival companies on the market. Last year, Openai dissolved its AGS preparation team, and a series of IA security and policy researchers left the company for its competitors.
Competitive pressures have only increased. Chinese Ai Lab Deepseek captured the attention of the world with its openly available R1 Model, which corresponded to the O1 “reasoning” model of Openai on a certain number of key references. The CEO of Openai, Sam Altman, has admitted that Deepseek has decreased the technological example of Openai, and said This OpenAi “would draw versions” to better compete.
There is a lot of money at stake. Openai loses billions a year and the company has would have Projected that its annual losses could triple $ 14 billion by 2026. A faster product release cycle could benefit from OPENAI final ends, but perhaps to the detriment of long-term security. Experts like Brundage wonder if the compromise is worth it.
