Altman, Sam

CEO of OpenAI

Sam Altman: OpenAI CEO on GPT-4, ChatGPT, and the Future of AI

Lex Fridman Podcast · Mar 25, 2023

• Sam Altman
• OpenAI
• GPT-4 and ChatGPT
• RLHF and alignment
• AI safety and AGI takeoff speed
• capped-profit company structure
• Microsoft partnership
• universal basic income
• AGI and consciousness
• AI bias and steerability

Core ideas

Altman frames GPT-4 as a very early technology that history will likely view the way we now view the earliest computers — buggy and limited, but pointing toward something transformative — and, pressed on which version might be remembered as the pivotal leap, says if forced to pick a single moment from what's shipped so far he'd choose ChatGPT rather than any specific GPT version, because what mattered wasn't the underlying model but its usability: reinforcement learning from human feedback (RLHF) plus a good interface. He explains RLHF simply — humans rank pairs of model outputs, and that preference signal is fed back via reinforcement learning — and stresses this "alignment to what humans want" step required remarkably little data compared to pretraining, calling the science of human-guidance still far less mature than the science of building the base models themselves, though improving quickly. He describes the leap from GPT-3 to GPT-4 as the product of hundreds of small, individually unglamorous technical improvements (data collection, cleaning, training methodology, architecture tweaks) rather than one dramatic breakthrough, and pushes back on parameter-count fixation ("people got caught up in the parameter count race the same way they got caught up in the gigahertz race of processors") — mentioning, when Lex raises a viral misattributed claim about a "100 trillion parameter GPT-4," that what matters is performance, not parameter count as such.

On safety process, Altman describes GPT-4 undergoing roughly eight months of internal and external red-teaming, safety evals, and alignment work after training completed, with an explicit personal priority that "our degree of alignment increases faster than our rate of capability progress." He argues alignment and capability aren't orthogonal axes as commonly framed — RLHF and interpretability work, sold as "safety," often directly produce more capable, usable models too, blurring the supposed distinction. He's candid that OpenAI has not solved alignment for a "super powerful system," only found something that works adequately for current-generation models. Discussing the GPT-4 system card's documented case of the model providing an antisemitic dog-whistle even after safety tuning, he acknowledges the deep difficulty of defining "harmful output" or "hate speech" in an automatable way, framing the field's usual talk of "aligning AI to human values" as concealing an unstated asterisk — "the values I approve of" — and describes his aspirational (and admittedly impractical) ideal as something like a global constitutional-convention process to collectively negotiate broad boundaries, with countries and individual users retaining narrower personalization within those bounds (citing the "system message" feature, which lets users steer GPT-4's behavior with high fidelity, as a concrete step in that direction). He states plainly that OpenAI cannot offload this responsibility onto users or society alone, since the company built the system and is accountable if it breaks, while insisting broad societal input is still essential.

On AGI and existential risk, Altman says he takes Eliezer Yudkowsky's alignment-pessimism seriously enough to point to Yudkowsky's essay as a "well-reasoned and thoughtful" steelman worth reading (while disagreeing with much of it), and states his own approach is strictly iterative — releasing early, learning from real-world use, and minimizing "one-shot get-it-right" scenarios, on the view that historical AI-safety theorizing (formed largely before deep learning's success) has proven wrong often enough that tight empirical feedback loops matter more than a priori theory. Asked to rank a 2x2 matrix of AGI timelines (short/long) crossed with takeoff speed (fast/slow), he and Lex both independently land on "slow takeoff, short timeline" as the safest and most likely-good outcome, and Altman says OpenAI's strategy is explicitly weighted toward pushing the world into that quadrant. He states plainly he does not believe GPT-4 is AGI ("if I were reading a sci-fi book and there was a character that was an AGI and that character was GPT-4, I'll be like, well, this is a shitty book") and that he doesn't think GPT-4 is conscious, though he grants it can "fake consciousness" convincingly given the right prompting — relaying an idea from co-founder Ilya Sutskever that training a model on a dataset scrubbed of any mention of consciousness, then observing whether it immediately and fluently recognizes a first-person description of subjective experience (unlike other novel concepts it wasn't trained on), would be a meaningfully different kind of evidence than ordinary capability.

On corporate structure and competitive dynamics, Altman recounts OpenAI's 2019 shift from pure nonprofit to a "capped-profit" structure (voting control retained by the nonprofit board, investor/employee returns capped, with mechanisms to cancel equity or make otherwise non-standard decisions) as a necessity born of needing far more capital than nonprofit fundraising could supply, while trying to avoid unbounded profit-maximization incentives — he explicitly says he worries about "uncapped" competitors chasing unlimited value from AGI, though notably not by name. He frames Microsoft (a ~$10B multi-year investor) as unusually understanding of OpenAI's need for unusual governance and control provisions compared to other potential partners, and praises Satya Nadella's blend of being "both super visionary" and "a super effective hands-on executive" as rare among large-company CEOs. Discussing Elon Musk's public criticism of OpenAI, Altman expresses empathy (attributing some of it to genuine AGI-safety anxiety and comparing it to Musk's own hurt reaction to being mocked in SpaceX's early days) while noting he personally has "never felt any pronoun other than 'it'" toward OpenAI's systems, unlike most people, and cautions against anthropomorphizing AI tools as "creatures" even while acknowledging that doing so, if done transparently, can make a tool more usable — a distinction he considers important to keep clear given the manipulation risk.

On economic and political effects, Altman predicts many jobs will disappear (naming customer service as a plausible large-scale example) while others will be enhanced or newly created, expresses support for universal basic income as a partial cushion (not a full solution) during the transition, and cites OpenAI-funded UBI research due to conclude by year's end. His broader thesis is that the cost of intelligence and the cost of energy will fall dramatically over coming decades, driving a wealth increase whose political effects he expects to broadly mirror history's pattern of economic change preceding political change; asked directly why Soviet-style central planning failed, he credits distributed human ingenuity and individual self-determination over central planning, while conceding a sufficiently capable superintelligence "might" plan better than markets — though he'd still bet on many competing superintelligences operating within something like a liberal-democratic structure over a single centralized one. He closes on the Silicon Valley Bank collapse as a small preview of how fast institutional fragility and public sentiment (amplified by Twitter and mobile banking) can move faster than regulators or leaders understand, drawing an explicit parallel to how unprepared institutions may be for AGI-scale disruption, and states his personal excitement about future systems is aimed less at Meme-tier questions than at getting help resolving fundamental open problems in physics and detecting other intelligent life.

Excerpts