What You May Have Missed #31

Sam Altman's World Tour Highlights / Research (Orca, AlphaDev, MusicGen) / Products (Google, The Frost) / Articles (AI or nuclear tech?) / Misc (Images as QR, Generative Fill's fixing pics)

Midjourney

Sam Altman’s World Tour Highlights

• Sam Altman says he worries making ChatGPT was 'something really bad' given potential AI risks (Spriha Srivastava on Business Insider): "What I lose the most sleep over is the hypothetical idea that we already have done something really bad by launching ChatGPT," Altman told Satyan Gajwani, the vice chairman of Times Internet, at an event on Wednesday organized by the Economic Times. Altman said he was worried that "maybe there was something hard and complicated" that his team had missed when working on the chatbot.”

• Sam Altman says OpenAI won’t go public now because he may have to make ‘a very strange decision’ that investors will disagree with (Rachel Shin on Fortune): “In terms of decision-making, Altman said he wants full autonomy from OpenAI shareholders, saying in Abu Dhabi that ‘the chance that we have to make a very strange decision someday is nontrivial.’ Altman didn’t clarify what a ‘very strange decision’ could look like, but added, ‘I don’t want to be sued by…public market, Wall Street, etc.’”

• OpenAI CEO Calls for Collaboration With China to Counter AI Risks (Karen Hao on WSJ): Dialing in from overseas to a packed conference in Beijing on Saturday to widespread cheers in the audience, Altman emphasized the importance of collaboration between American and Chinese researchers to mitigate the risks of AI systems … ‘China has some of the best AI talent in the world,’ Altman said. ‘So I really hope Chinese AI researchers will make great contributions here.’”

• Intelligence as a “fundamental property of matter” (Bilawal Sidhu): “Q: ‘After doing AI for so long, what have you learned about humans?’ Sam Altman: ‘I grew up implicitly thinking that intelligence was this, like really special human thing and kind of somewhat magical. And I now think that it's sort of a fundamental property of matter...’”

• Regulate us, but not really (Luiza Jarovsky): “There was a disproportional time talking about the risk of a super powerful and perhaps dangerous Artificial General Intelligence (AGI). It felt like part of a PR move to increase the hype and interest in AI-based applications … there was no mention of OpenAI's role or plans in mitigating existing AI-related problems, such as bias, disinformation, discrimination, deepfakes, non-compliance with data protection rules, etc. It looked like talking about a future AGI was a way to distract from reality.” (About the live event at Tel Aviv University with Sam Altman and Ilya Sutskever.)

Research

• Orca: Progressive Learning from Complex Explanation Traces of GPT-4 (Microsoft): “Orca [is] a 13-billion parameter model that learns to imitate the reasoning process of large foundation models [instead of imitating just the style like self-instruct models, e.g., Alpaca, Vicuna, Falcon]. Orca learns from rich signals from GPT 4 including explanation traces; step-by-step thought processes; and other complex instructions, guided by teacher assistance from ChatGPT … [It] reaches parity with ChatGPT on the BBH benchmark and shows competitive performance (4 pts gap with optimized system message) in professional and academic examinations like the SAT, LSAT, GRE, and GMAT, both in zero-shot settings without CoT; while trailing behind GPT–4.”

• How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources (Allen Institute for AI): “Our evaluations show that the best model in any given evaluation reaches on average 83% of ChatGPT performance, and 68% of GPT-4 performance, suggesting that further investment in building better base models and instruction-tuning data is required to close the gap.”

• Reading Between the Lines: Modeling User Behavior and Costs in AI-Assisted Programming (MIT + Microsoft): “Our studies revealed that when solving a coding task with Copilot, programmers may spend a large fraction of total session time (34.3%) on just double-checking and editing suggestions, and spend *more than half* of the task time on Copilot related activities, together indicating that introducing Copilot into an IDE can significantly change user behavior.”