The corporate has been on the defensive following the departure of key security researchers, stories that strict NDAs are silencing former workers, and backlash towards a brand new model of ChatGPT.
The dramatic exits of Jan Leike and Ilya Sutskever final week even compelled OpenAI’s leaders, together with CEO Sam Altman, to make public statements defending their efforts to regulate AI threat.
When a Vox report about OpenAI’s tight off-boarding agreements emerged the next day, Altman responded by saying it was one of many “few occasions” he’d ever “been genuinely embarrassed” operating OpenAI. He added he wasn’t conscious the clauses have been being imposed on departing workers and stated the corporate was working to rectify the settlement.
It is a uncommon admission from Altman, who has labored laborious to domesticate a picture of being comparatively calm amid OpenAI’s ongoing chaos. A failed coup to take away him final yr in the end bolstered the CEO’s status, but it surely appears OpenAI’s cracks are beginning to present as soon as extra.
Security staff implosion
OpenAI has been in full injury management mode following the exit of key workers engaged on AI security.
Leike and Sutskever, who led the staff answerable for guaranteeing AGI would not go rogue and hurt humanity, each resigned final week.
Leike adopted his blunt resignation with a prolonged submit on X, accusing his former employers of placing “shiny merchandise” forward of security. He stated the security staff was left “struggling for compute, and it was getting more durable and more durable to get this significant analysis completed.”
Fast to play the function of disaster supervisor, Altman shared Leike’s submit, saying, “He is proper, we now have much more to do; we’re dedicated to doing it.”
The high-profile resignations observe a number of different current exits.
In accordance with a report by The Data, two security researchers, Leopold Aschenbrenner and Pavel Izmailov, have been not too long ago fired over claims they have been leaking info.
Security and governance researchers Daniel Kokotajlo and William Saunders additionally each not too long ago left the corporate, whereas Cullen O’Keefe, a analysis lead on coverage frontier, left in April, in line with his LinkedIn profile.
Kokotajlo instructed Vox he’d “progressively misplaced belief in OpenAI management and their capacity to responsibly deal with AGI.”
The Superalignment staff led by Leike and Sutskever, which had about 20 members final yr, has now been dissolved. A spokesperson for OpenAI instructed The Data that it had mixed the remaining staffers with its broader analysis staff to satisfy its superalignment objectives.
OpenAI has one other staff centered on security referred to as Preparedness, however the high-profile resignations and departures aren’t a very good look for a corporation on the forefront of superior AI growth.
Silenced workers
The implosion of the security staff is a blow for Altman, who has been eager to point out he is safety-conscious in the case of creating super-intelligent AI.
He instructed Joe Rogan’s podcast final yr: “Many people have been tremendous fearful, and nonetheless are, about security and alignment. When it comes to the ‘not destroy humanity’ model of it, we now have a number of work to do, however I feel we lastly have extra concepts about what can work.”
Some assume Leike’s claims erode Altman’s authority on the topic and have raised eyebrows extra broadly.
Neel Nanda, who runs Google DeepMind’s mechanistic interpretability staff tasked with “decreasing existential threat from AI,” responded to Leike’s thread: “Fairly regarding tales of what is taking place inside OpenAI.”
On Friday, Vox reported that strict offboarding agreements primarily silenced OpenAI workers.
They reportedly included non-disclosure and non-disparagement clauses that would take away workers’ vested fairness in the event that they criticized their former employer, and even acknowledge that an NDA existed.
Altman addressed the report in an X submit: “That is on me and one of many few occasions i have been genuinely embarrassed operating openai; i didn’t know this was taking place and that i ought to have.”
He added: “The staff was already within the strategy of fixing the usual exit paperwork over the previous month or so.”
“Her” voice paused
Regardless of OpenAI’s efforts to comprise the chaos, the scrutiny would not look like over.
On Monday, the corporate stated it was pausing ChatGPT’s “Sky” voice, which has not too long ago been likened to Scarlett Johansson.
“We imagine that AI voices shouldn’t intentionally mimic a star’s distinctive voice — Sky’s voice will not be an imitation of Scarlett Johansson however belongs to a special skilled actress utilizing her personal pure talking voice,” the corporate stated in a submit.
The voice, a key a part of the corporate’s GPT-4o demo, was broadly in comparison with Johansson’s digital assistant character within the movie “Her.” Altman even appeared to acknowledge the similarities, merely posting “her” on X throughout the demo.
Some customers complained concerning the chatbot’s new voice, calling it overly sexual and too flirty in demo movies circulating on-line.
Seemingly oblivious to the criticism, OpenAI appeared triumphant following the launch. The often reserved Altman even appeared to shade Google, which demoed new AI merchandise the next day.
“I strive not to consider opponents an excessive amount of, however I can not cease serious about the aesthetic distinction between openai and google,” Altman wrote on X, accompanied by pictures of the rival demos.
OpenAI did not instantly reply to a request for remark from Enterprise Insider, made outdoors regular working hours.