AI summary by TwinMind
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
Lex Fridman
1h 55m 10s
OpenAI Board Crisis November 2023
- Most painful professional experience of Sam's life - chaotic, shameful, upsetting
- Felt love and support from people during crisis, like attending own eulogy
- Happy it happened early rather than closer to AGI - builds resilience for future challenges
- Expects AGI road to be giant power struggle, needed to iterate on governance structures
- Board fired Sam Friday afternoon, he initially accepted it and got excited about focused AI research
- Saturday morning 2 board members called wanting him back, most painful period was weekend of back-and-forth
- Sunday evening board appointed new interim CEO - felt really bad
- Dominant emotion was love, not hate, despite pain
- Fugue state for 45 days after - psychologically down, hard to work normally
- Eventually decided to focus on mission rather than feel like victim
Board Structure and Governance
- Old board shrank from 9 to 6 to couldn't agree on additions
- Lacked experienced board members
- Nonprofit board has unusual amount of power - don't answer to shareholders like typical corporate boards
- Brett Taylor and Larry Summers chosen during tense weekend negotiations
- New approach: hire board members in slates not individually, need diverse expertise
- Board needs technical expertise, nonprofit experience, legal/governance skills
- Learned about structure, incentives, need for robust governance under pressure
Elon Musk Lawsuit and Relationship
- Elon wanted total control when he thought OpenAI would fail
- Wanted Tesla to acquire OpenAI or merge with Tesla
- OpenAI chose different direction, Elon decided to leave
- Sam respects Elon as great builder, finds lawsuit "unbecoming of a builder"
- Hopes for amicable relationship and friendly competition in future
- Elon's "change name to ClosedAI and I'll drop lawsuit" comment shows lack of seriousness
- Grok hadn't open sourced anything until people pointed out hypocrisy
Sora Video Generation Model
- Amazed by occlusion handling - person walks in front of object, object remains when person moves
- Physics representation over many sequence steps impressive
- Models understand more about world than given credit for, but also have clear limitations
- Cats sprouting extra limbs example of current weaknesses
- Approach feels different from human thinking but will improve with scale
- Uses visual patches like LLMs use text tokens
- Internet scale data with human involvement but not internet scale human labeling
- Still needs efficiency improvements before release
- Concerns: deepfakes, misinformation - need thoughtful deployment
GPT-4 and Future Models
- GPT-4 "kind of sucks" relative to where they need to get
- Expects delta between GPT-5 and GPT-4 similar to GPT-4 vs GPT-3
- Best current use: brainstorming partner, longer horizon tasks when they work
- Chat interface and RLHF post-training as important as base model
- Context window expansion to 128k tokens, eventual goal of billion+ tokens
- People use as default start for knowledge work tasks
- Hallucination remains issue, will improve but requires fact-checking currently
Technical Development and Approach
- 200 medium-sized innovations multiplied together into one giant thing
- Individual teams try to keep high-level picture while working in weeds
- Compute, technical advances, data all important bottlenecks
- Slower thinking/reasoning important future direction - allocate more compute to harder problems
- Q* project exists but not ready to discuss details
AGI Timeline and Capabilities
- Avoids specific AGI timeline predictions due to definitional problems
- Expects "quite capable systems" by end of decade, possibly sooner
- Key milestone: systems that significantly increase rate of scientific discovery
- First AGI won't solve grand unified theory but might guide needed experiments
- Iterative deployment strategy to avoid shock updates to world
- Amazing model coming this year, uncertain what to call it
Safety and Alignment
- Not top worry currently but could change
- Theatrical risks like "escaping the box" get disproportionate attention
- Eventually most of company will think about safety aspects
- Includes technical alignment, societal impacts, economic impacts, security
- State actors trying to infiltrate/hack - intensity increasing
- Need robust governance structures and processes
Competition and Industry
- Pros: better products, innovation, speed, cost reduction
- Cons: potential arms race acceleration
- Prefers short timeline, slow takeoff as safest quadrant
- Collaboration on safety beneficial vs silos
- Competition for talent should be friendly
Business Model and Future
- Hates ads aesthetically, prefers subscription model
- Users aren't the product when they pay directly
- Sustainable business without ads appears achievable
- Compute will be currency of the future, most precious commodity
- Need massive energy infrastructure - nuclear fission and fusion
- Helion doing best fusion work, nuclear fission underutilized due to fear
Philosophical and Personal Views
- Sora somewhat increases simulation hypothesis probability
- Deeply wants alien civilizations to exist, finds Fermi Paradox puzzling
- Hope from humanity's trajectory despite flaws
- AGI might be more like societal scaffolding than single brain
- Not afraid of death, would feel grateful for life lived
- Grateful for awesome human creations including ChatGPT
- Doesn't capitalize tweets due to internet culture background, not philosophical stance
