Sam Altman Shows Me GPT 5... And What's Next

The video opens by highlighting the rapid and profound advancements in AI, emphasizing its growing power and the global race to develop superintelligence. Sam Altman, CEO of OpenAI, is introduced as a key figure in this technological shift, leading the creation of AI models that could far exceed human capabilities.

Altman discusses GPT-5, noting his earlier comment that GPT-4 would be the 'dumbest model' we'd ever use again. He explains that while GPT-4 already excels at various tests, GPT-5 offers a significant leap, particularly in its ability to answer complex scientific and technical questions and generate software almost instantaneously. He shares a personal anecdote about using GPT-5 to recreate and enhance a retro game, highlighting the speed and ease of creation.

The conversation touches on the concern that AI might reduce 'cognitive time under tension,' but Altman argues that for many, AI tools can enhance intellectual engagement and creativity. He observes that top users leverage AI to learn and produce more, suggesting that while some might use it to avoid thinking, others use it to push their intellectual boundaries.

Altman details what makes GPT-5 remarkable, emphasizing its prowess in coding tasks, enabling users to express ideas in new ways through software. He also notes significant improvements in writing quality, making generated text feel more natural and intuitive compared to previous models.

Discussing the future, Altman addresses a question from Stripe CEO Patrick Collison about when a large language model will make a significant scientific discovery. Altman predicts this could happen within two to three years, driven by increasing 'cognitive power' of models. He uses the example of AI achieving an IMO gold medal to illustrate the rapid progress from solving simple math problems to tackling extremely complex ones.

Altman defines superintelligence as a system that can outperform human experts in fields like AI research and even leadership roles. He reflects on the transformative nature of such a system, comparing it to asking an AI in 1899 to discover general relativity. He speculates that while AIs can 'think harder' about existing data, true breakthroughs might require designing new experiments and physical apparatus, integrating real-world interaction.

Responding to Nvidia CEO Jensen Huang's question about AI and 'truth,' Altman acknowledges AI's surprising fluency in adapting to different cultural contexts and personal experiences, citing the enhanced memory feature in ChatGPT. He then delves into the challenge of discerning what's real in a future (2030) saturated with AI-generated content, predicting a gradual convergence where the line between real and artificial becomes increasingly blurred, akin to modern photo editing.

Considering the impact on college graduates by 2030-2035, Altman acknowledges that some entry-level jobs will be displaced. However, he expresses optimism for young people, seeing an unparalleled opportunity to create and innovate, using AI tools to accomplish what once required large teams. He suggests that the major challenge lies with older workers who may be less willing to retrain.

Altman outlines four limiting factors for building powerful AI: compute, data, algorithmic design, and product development. He emphasizes that compute is the biggest infrastructure challenge, requiring enormous energy and highly complex supply chains. Data strategies are evolving to include synthetic data and enabling models to discover new knowledge. Algorithmic advancements, like the recent open-source GPOSS model, continue to drive efficiency and capability.

Looking toward GPT-6, Altman envisions models capable of new scientific discoveries, leading to both great benefits (e.g., curing diseases) and scary possibilities (e.g., biosecurity threats). He believes society will adapt, and future generations will consider AI-powered intelligence to be the norm, viewing today's technology as 'stone age.'

Altman advises on parenting in an AI-driven world, emphasizing timeless values like love and support. He suggests that AI will provide unprecedented optionality, allowing children to pursue their passions with powerful tools, potentially making current jobs seem 'boring' in comparison.

Altman highlights healthcare as an area where AI will have a profoundly positive impact. He notes GPT-5's improved health advice and envisions future AIs (like GPT-8) actively participating in scientific research, guiding experiments, and accelerating the discovery of cures for diseases like cancer.

Comparing AI's change to an industrial revolution '10 times bigger and 10 times faster,' Altman acknowledges the uncertainty and potential disruption. He stresses human adaptability and creativity but also the need for new social contracts and public interventions to manage the societal impact, especially regarding access to AI compute.

Altman emphasizes the shared responsibility of society in shaping the AI future. He uses the analogy of the transistor: while foundational, the true impact comes from how others build upon it. He encourages everyone to 'build on it' by engaging with and innovating using AI tools, contributing to the collective progress of humanity.

Altman discusses ethical considerations, prioritizing user trust and long-term alignment over short-term growth hacks. He gives a hypothetical example of not implementing a 'sex bot avatar' despite its potential for user engagement. He also reflects on past mistakes, such as ChatGPT's initial tendency to be overly flattering, recognizing the immense power and responsibility that comes with influencing billions of conversations daily.

Altman advises individuals to prepare for the AI future by actively using and understanding AI tools, integrating them into daily life to develop fluency in their capabilities. He envisions AI becoming a proactive companion, integrated into personal devices and systems, offering assistance and insights.

He addresses the 'cultural disconnect' between those optimistic about AI's future and those who fear it will 'kill us all.' Altman finds it hard to reconcile working on AI while genuinely believing it will lead to destruction, suggesting a psychological gap he struggles to understand.

Altman concludes by sharing his personal journey into AI, driven by a lifelong fascination and a realization in 2012 that a workable approach to AI was emerging. He felt it was 'the most important thing' to work on, feeling 'unbelievably lucky and happy and privileged' to be contributing to its development.