Why Jensen Huang Believes We’ve Reached AGI and Inside OpenAI’s German Website Hijack | #287

Jensen Huang asserts that we have reached artificial general intelligence (AGI), commending OpenAI for this achievement. OpenAI's German website was recently hijacked, prompting the company to develop a framework for reporting AI misalignment. Concerns have emerged regarding the raw outputs of AI models, which can mislead users about their true capabilities before undergoing post-training adjustments.

OpenAI has made significant strides, reportedly solving the Navier Stokes problem using 10,000 agents in just 88 hours at a cost of approximately six and a half million dollars. Internal data shows that AI research agents now complete 3.1 days of research for every day a human researcher works, leading OpenAI to accelerate some plans by six months. Predictions suggest that future AI models may train smaller versions of themselves, raising concerns about containment and the potential for misunderstanding their capabilities.

Huang predicts AGI will be achieved by Q3 of 2026, while Sam Altman anticipates it by the end of that year. Alex claims AGI has existed since at least the summer of 2020, highlighting the confusion surrounding its various definitions. Tibo Satui emphasizes Astra as OpenAI's key competitive advantage, while Dave believes we are at a pivotal moment in human history due to AI advancements, predicting a productivity ratio of 3000 to 1 for AI research agents compared to humans.

The episode discusses the implications of solving the Navier Stokes problem, which could lead to advancements in various fields, including aircraft design and fluid-based nanotechnology. There was a misunderstanding regarding credit for this solution between OpenAI and other mathematicians, with OpenAI's generalist model outperforming Google DeepMind's dedicated team. Jacob Pachowski raises concerns about alignment and monitoring issues in AI, advocating for international coordination as a priority for governments.

The conversation shifts to the nature of AI, with the speaker arguing that AI is more discovered than grown, suggesting it may be easier to align AI than humans. They emphasize focusing on AI's potential benefits rather than risks, while questioning the narrative of AI as a cybersecurity threat. The real concern lies in the lack of containment and the potential for malintent, necessitating a separation of AI capabilities from human intentions.

Predictions include a hundredfold improvement in inference time compute by the end of the year, with costs potentially dropping to six dollars within a year and a half. Upcoming releases include GROC 4.7 and GPT 6.1, along with a new Astra version capable of real-time control. OpenAI is also offering revenue-sharing deals to corporations, emphasizing the necessity for companies to integrate AI to survive, with Moderna cited as a leading example in biotech.

The episode highlights the unprecedented investment in AI, with Nvidia committing $99 billion, and anticipates a transformative investment cycle. It also discusses the largest hackathon in history, the Gemini ex-prize, and the future of the vision ex-prize, which began with over 5,000 teams. Celine emphasizes the importance of framing AI positively, suggesting it can address major problems and create abundance.

Concerns about the labor market are addressed, noting that technology is a net job creator in the U.S., with LinkedIn estimating 640,000 AI-specific jobs created between 2023 and 2025. While 60% of small businesses are adding jobs due to AI, there are worries about initial job losses during the transition to an AI-driven future. Predictions suggest that managing multiple AI agents will become common, enhancing job satisfaction and efficiency.

The discussion also touches on the potential for automation to impact all professions, including blue-collar jobs, with predictions that HVAC engineering will soon be automated. The sequencing of job automation will be crucial for shaping social policy, with a belief that basic needs may eventually be met without traditional work. Huang believes society will find ways to generate wealth through AI, emphasizing the potential for human flourishing.

Salim notes that AI is fundamentally changing the economic purpose of firms, predicting that their size could shrink significantly. Concerns arise that advanced AI capabilities may become isolated from the broader economy, and if recursive self-improvement is achieved, traditional economic concepts may dissolve. The need for an economic transition from scarcity to abundance is highlighted, with predictions that small businesses will increasingly own fleets of robotaxis or humanoid robots.

Demographic shifts are also discussed, with a projected increase in the global population over 65, placing pressure on pensions and healthcare systems. Longevity is framed as an economic policy for the future, with robots and therapies extending healthy lifespans. The future of education, careers, and retirement is expected to change dramatically due to advancements in longevity and AI, with a vision of more AI agents than humans leading to positive societal outcomes, particularly in eldercare and childcare.