A high-profile ex-OpenAI coverage researcher, Miles Brundage, took to social media on Wednesday to criticize OpenAI for “rewriting the historical past” of its deployment strategy to probably dangerous AI methods.
Earlier this week, OpenAI revealed a doc outlining its present philosophy on AI security and alignment, the method of designing AI methods that behave in fascinating and explainable methods. Within the doc, OpenAI mentioned that it sees the event of AGI, broadly outlined as AI methods that may carry out any job a human can, as a “steady path” that requires “iteratively deploying and studying” from AI applied sciences.
“In a discontinuous world […] security classes come from treating the methods of at the moment with outsized warning relative to their obvious energy, [which] is the strategy we took for [our AI model] GPT‑2,” OpenAI wrote. “We now view the primary AGI as only one level alongside a collection of methods of accelerating usefulness […] Within the steady world, the way in which to make the subsequent system secure and useful is to be taught from the present system.”
However Brundage claims that GPT-2 did, actually, warrant considerable warning on the time of its launch, and that this was “100% constant” with OpenAI’s iterative deployment technique at the moment.
“OpenAI’s launch of GPT-2, which I used to be concerned in, was 100% constant [with and] foreshadowed OpenAI’s present philosophy of iterative deployment,” Brundage wrote in a publish on X. “The mannequin was launched incrementally, with classes shared at every step. Many safety specialists on the time thanked us for this warning.”
Brundage, who joined OpenAI as a analysis scientist in 2018, was the corporate’s head of coverage analysis for a number of years. On OpenAI’s “AGI readiness” group, he had a specific concentrate on the accountable deployment of language era methods resembling OpenAI’s AI chatbot platform ChatGPT.
GPT-2, which OpenAI introduced in 2019, was a progenitor of the AI methods powering ChatGPT. GPT-2 may reply questions on a subject, summarize articles, and generate textual content on a stage generally indistinguishable from that of people.
Whereas GPT-2 and its outputs might look primary at the moment, they had been cutting-edge on the time. Citing the danger of malicious use, OpenAI initially refused to launch GPT-2’s supply code, opting as a substitute of give chosen information retailers restricted entry to a demo.
The choice was met with blended evaluations from the AI trade. Many specialists argued that the risk posed by GPT-2 had been exaggerated, and that there wasn’t any proof the mannequin might be abused within the methods OpenAI described. AI-focused publication The Gradient went as far as to publish an open letter requesting that OpenAI launch the mannequin, arguing it was too technologically vital to carry again.
OpenAI finally did launch a partial model of GPT-2 six months after the mannequin’s unveiling, adopted by the complete system a number of months after that. Brundage thinks this was the suitable strategy.
“What a part of [the GPT-2 release] was motivated by or premised on considering of AGI as discontinuous? None of it,” he mentioned in a publish on X. “What’s the proof this warning was ‘disproportionate’ ex ante? Ex publish, it prob. would have been OK, however that doesn’t imply it was accountable to YOLO it [sic] given data on the time.”
Brundage fears that OpenAI’s intention with the doc is to arrange a burden of proof the place “considerations are alarmist” and “you want overwhelming proof of imminent risks to behave on them.” This, he argues, is a “very harmful” mentality for superior AI methods.
“If I had been nonetheless working at OpenAI, I’d be asking why this [document] was written the way in which it was, and what precisely OpenAI hopes to realize by poo-pooing warning in such a lop-sided method,” Brundage added.
OpenAI has traditionally been accused of prioritizing “shiny merchandise” on the expense of security, and of dashing product releases to beat rival firms to market. Final yr, OpenAI dissolved its AGI readiness group, and a string of AI security and coverage researchers departed the corporate for rivals.
Aggressive pressures have solely ramped up. Chinese language AI lab DeepSeek captured the world’s consideration with its overtly accessible R1 mannequin, which matched OpenAI’s o1 “reasoning” mannequin on a lot of key benchmarks. OpenAI CEO Sam Altman has admitted that DeepSeek has lessened OpenAI’s technological lead, and mentioned that OpenAI would “pull up some releases” to higher compete.
There’s some huge cash on the road. OpenAI loses billions yearly, and the corporate has reportedly projected that its annual losses may triple to $14 billion by 2026. A quicker product launch cycle may benefit OpenAI’s backside line near-term, however presumably on the expense of security long-term. Consultants like Brundage query whether or not the trade-off is price it.