Sam Altman on OpenAI Board Crisis, Compute as Future Currency, and AGI

Sam Altman, CEO of OpenAI, recently offered extensive insights into the tumultuous events surrounding the OpenAI board, the future trajectory of AI, and the company's vision. He underscored the growing importance of compute, predicting it will become the "currency of the future" and the world's most precious commodity by the end of the decade.

The OpenAI Board Saga and Its Aftermath

Altman characterized the November 2023 board events as the "most painful professional experience" of his life, marked by chaos, shame, and upset. Despite the difficulty, he was deeply moved by the "unbelievable support" he received, describing it as a "great deal of love and very little hate." He believes this challenging period ultimately strengthened OpenAI, building resilience for future obstacles.

He acknowledged that the path to Artificial General Intelligence (AGI) is likely to involve significant power struggles, and the board incident highlighted the necessity of a robust organizational structure and a board capable of functioning under immense pressure. Altman reflected on OpenAI's unique structure, where the non-profit board held substantial power without direct accountability to shareholders, a departure from typical corporate governance. He expressed a desire for the OpenAI board to be accountable to the world at large.

The new board, still in formation, aims to include more experienced members. Altman explained that the initial appointments of Bret Taylor and Larry Summers were made under extreme duress during the crisis. Future board members will be selected based on diverse expertise, including non-profit governance, company operations, and legal knowledge, with a focus on forming slates rather than individual appointments. While technical savvy is important for some board members, it's not a universal requirement, as the board needs to consider the broader societal impact beyond just technical understanding.

Personal Reflections and Trust

The board saga, Altman admitted, has made him more hesitant in trusting people, a change he dislikes given his naturally trusting disposition. He expressed concern about becoming less of a "default trusting person" but clarified that he is not worried about becoming cynical.

He also addressed the lawsuit filed by Elon Musk against OpenAI. Altman stated he doesn't fully understand Musk's motivations but clarified that OpenAI's initial vision was purely research-focused, without a clear path to productization or the massive capital requirements that later emerged. He believes the company made reasonable decisions at each step, leading to its current hybrid structure.

Regarding Musk's departure from OpenAI, Altman explained that Musk chose to leave after his proposals for total control, merging with Tesla, or making OpenAI a for-profit entity under his control were rejected. Altman also commented on Musk's criticism of OpenAI's "closed" nature, suggesting that Musk's recent decision to open-source Grok came after accusations of hypocrisy. Altman expressed sadness about the public nature of the conflict, calling it "unbecoming of a builder."

The Meaning of "Open" in OpenAI

Altman stated that if he could go back, he would choose a different name for the company. He emphasized that the most important aspect of OpenAI's mission is making powerful technology available to people for free as a public good. He believes providing free or low-cost powerful AI tools is crucial for fulfilling their mission, regardless of whether the technology is open-source. He advocates for a nuanced approach to open-sourcing, releasing some things while keeping others closed.

Sora and the Future of Visual AI

Altman described Sora, OpenAI's text-to-video model, as "truly amazing" on both a product and philosophical level. He believes these models understand the world more deeply than often credited, citing examples like Sora's ability to handle occlusions and represent underlying physics. While acknowledging current limitations, such as "cats sprouting extra limbs," he believes these will be overcome with larger models, better technical details, and more data.

He confirmed that Sora's training involves a significant amount of self-supervised learning on "internet scale data," but also incorporates "lots of human data." OpenAI is cautious about releasing Sora due to potential dangers like deepfakes and misinformation, and the need to improve its efficiency for widespread use.

AI and Copyright

On the issue of AI training and copyright, Altman believes that creators of valuable data deserve compensation for its use. He suggested that artists, for example, should have the option to opt out of their style being used for AI generation or have an economic model associated with it if their style is used. He drew a parallel to the music industry's transition from CDs to streaming, where the model changed but creators still get paid.

The Impact of AI on Jobs and Creativity

Altman reframed the discussion about AI's impact on jobs, suggesting it's more productive to consider what percentage of tasks AI will perform rather than what percentage of jobs it will replace. He envisions AI as a tool that will allow people to operate at a higher level of abstraction, making them more efficient. He believes that while AI tools will be used in the production of many videos, human creativity and direction will remain central. He also expressed confidence that humans will continue to create and find ways to be useful, as it is "pretty hardwired."

GPT-4 and Beyond: Smarter Systems and Human-AI Collaboration

Altman humorously stated that GPT-4 "sucks" when compared to the future potential of AI. He expects the leap from GPT-4 to GPT-5 to be as significant as the jump from GPT-3 to GPT-4, describing it as "smarter across the board." He finds GPT-4 most impressive as a "brainstorming partner" and for its potential to handle longer, multi-step tasks, though this is still an infrequent occurrence. He likened this advancement to a "magical moment" where the AI "gets you," understanding the deeper intent behind even poorly formulated prompts.

He highlighted the importance of both the underlying model and the post-training steps (like RLHF) in making models like ChatGPT useful and compelling. He also discussed the expansion of the context window in GPT-4 Turbo, envisioning a future with "billions" of tokens, allowing AI to integrate all of a user's information and history, leading to a more personalized and helpful experience.

Regarding the issue of "hallucinations" or factual inaccuracies in AI, Altman acknowledged it as an area of intense focus for OpenAI, expecting significant improvements in future versions. He also discussed the company's exploration of memory features in ChatGPT, aiming for models that "get to know me and get more useful to me over time." He emphasized user choice and transparency regarding privacy in this context.

The Future of AI Development

Altman believes that future AI systems will need to incorporate "slower thinking" or sequential reasoning, allowing them to allocate more compute to harder problems. He hinted at ongoing research in this area, without disclosing details about the rumored "Q-Star" project.

He reiterated OpenAI's strategy of "iterative deployment," releasing models like GPT-1, 2, 3, and 4 to allow the world to adapt and consider the implications of AGI. He expressed a desire for even more iterative releases to avoid "shock updates." While he couldn't confirm a release date for GPT-5, he stated that OpenAI will release "an amazing model this year" and has "a lot of other important things to release first."

The biggest challenges for future models like GPT-5 involve a combination of compute, technical advancements, and continuous innovation across many "medium-sized things."

The $7 Trillion Compute Vision

Altman clarified that his "meme" about needing $7 trillion was a response to misinformation, but he genuinely believes compute will be the "currency of the future" and the "most precious commodity." He envisions an "unusual market" for compute, where demand will be immense if prices are low enough to enable widespread applications like curing cancer or personal AI assistants. He identified energy as the hardest part of this puzzle, advocating for nuclear fusion and a renewed embrace of nuclear fission.

AI Safety, Bias, and Competition

Altman acknowledged the potential for AI to be politicized and the risk of an "arms race" in AI development. He expressed concern about the "theatrical" nature of some AI risks, where dramatic but less probable scenarios receive more attention than widespread, slow-burn issues. He hopes AI can help humanity understand true risks and foster balanced perspectives.

Regarding competition, he sees the pros as leading to better products and faster innovation, but the con as potentially fueling an unsafe arms race. He emphasized OpenAI's commitment to prioritizing safety and aiming for a "short timelines, slow takeoff" scenario for AGI.

Addressing concerns about bias, particularly in light of recent issues with Google's Gemini, Altman stated that OpenAI works hard to avoid such problems. He proposed a public declaration of a model's "desired behavior," including edge cases, to clarify whether unexpected outputs are bugs or intended policy, allowing for public debate on the latter. He believes OpenAI is less susceptible to ideological leanings than some other tech companies.

He stressed that safety is a company-wide effort, encompassing technical alignment, societal and economic impacts, and protection against state actors attempting to steal models.

The Future of Human-AI Interaction and Embodied AI

Altman is excited about AI becoming "smarter" across the board, leading to a deeper "intellectual connection" and understanding between humans and AI. He envisions a future where programming might be done entirely in natural language, changing the skillset required but not necessarily the underlying problem-solving predisposition. He anticipates that the best practitioners will utilize a variety of tools, seamlessly switching between natural language and traditional coding languages like C when necessary.

He also hopes for the development of "motor robots" to enable AI to interact with the physical world, stating that OpenAI will "return to robots in some way at some point." He finds it "depressing" to imagine a future with AGI where humans are still required to perform physical tasks.

Defining and Achieving AGI

Altman finds the question of "when will we build AGI" poorly formed due to the widely varying definitions of AGI. He prefers to discuss when systems will achieve specific capabilities (X, Y, or Z) rather than a fuzzy milestone. He sees AGI not as an end, but more as a "mile marker" or even a "beginning." He anticipates that by the end of this decade, and possibly sooner, we will have "quite capable systems" that will be considered "remarkable."

He acknowledged that earlier versions like ChatGPT 3.5, if shown to people in the 1990s, might have been considered AGI. However, he doesn't believe GPT 3.5 "changed the world" in the way AGI implies, though it did significantly alter expectations for the future and put AI on a new trajectory. For Altman, AGI implies a "major transition" in how the world operates, something beyond just a useful tool.

AGI's Potential for Scientific Discovery

When asked what an impressive feat for AGI would be, Altman highlighted its potential to "significantly increase the rate of scientific discovery in the world." He believes that true economic growth stems from scientific and technological progress. He also expressed excitement about the prospect of a system generating "really novel intuitions, like scientific intuitions."

Regarding his personal interaction with a hypothetical first AGI, Altman admitted it's surprisingly difficult to pinpoint what he would ask. He doesn't expect the first AGI to immediately provide a grand unified theory of physics. Instead, he might start with yes/no questions about the existence of such theories or alien civilizations. He also suggested that an AGI might require more data or the development of new tools and measurements before it could answer profound questions.

The Power and Governance of AGI

Altman addressed the immense power associated with building AGI and the question of trusting himself with it. He firmly stated that neither he nor any single person should have total control over OpenAI or AGI. He advocates for a robust governance system. He referenced the past board drama at OpenAI, acknowledging that while the board legally could have fired him, in practice, it didn't quite work out, which he views as a "governance failure." Despite this, he reiterated his disinterest in super-voting control and his belief that no single company should make all the decisions regarding AGI. He stressed the need for governments to establish "rules of the road," even if it leads to accusations of seeking regulatory capture. He believes that in time, the importance of this approach will become clear.

Existential Risk and AI Safety

Altman stated that the risk of losing control of AGI itself is not his "top worry" currently, though it has been in the past and may be again. He acknowledged that this is a "theatrical risk" that can dominate public discourse and that many smart, well-meaning AI safety researchers focus heavily on this problem. While he appreciates their work, he believes it has sometimes overshadowed other significant AI-related risks.

The Capitalization of Tweets

Altman explained his lack of capitalization in tweets as a habit from growing up "very online" and chatting on computers where capitalization was uncommon. He sees it as a "dumb construct" in general, though he still uses proper capitalization in formal documents. He doesn't believe it signifies disrespect and finds it amusing that people care so much about it.

The Simulation Hypothesis and AI

Altman believes that the ability to generate worlds, as demonstrated by Sora, should "increase everyone's probability somewhat" or at least their openness to the idea that we might live in a simulation. He noted that while he expected something like Sora, it arrived faster than anticipated. He drew a parallel to the concept of the square root of negative one, suggesting that seemingly simple operators can unlock profound new realities, and AI could serve as such a "psychedelic gateway" to understanding reality differently.

Hope for Humanity's Future

Altman expressed hope for humanity's future, citing the remarkable progress made over a relatively short period. He believes that AGI might be more akin to the "scaffolding in society" that we all contribute to, rather than a single brain. He highlighted how human capabilities have dramatically increased not due to biological changes, but due to this collective scaffolding of knowledge and technology. This collective creation, he said, fills him with hope.

The Fermi Paradox and Alien Intelligence

Altman finds the Fermi paradox "very, very puzzling" but is confident that a large number of intelligent alien civilizations exist. He speculates that travel across space might be incredibly difficult, or that humanity might be "blind to what intelligence looks like," a perception that AI could help us overcome.