Interview
Eiso Kant, CTO @Poolside: Raising $600M To Compete in the Race for AGI | E1211
- The current era is viewed as the foundational turning point for AGI comparable to the mobile and internet revolutions, with a predicted multi-decade or multi-century trajectory of continuous intelligence gains driven by algorithms and hardware efficiency.
- Economic abundance will emerge from human-level capabilities in specific high-value sectors before full AGI is reached, with software development expected to be the primary domain where machine capabilities first close the gap with humans due to its simulatability and scaling potential.
- Speech recognition, Full Self-Driving, and image generation are predicted to become commoditized rapidly, whereas software development will maintain a large intelligence gap, with Tesla expected to dominate FSD by leveraging non-simulatable real-world data.
- Future models will require magnitudes more data than humans for complex reasoning and planning, necessitating the creation of datasets containing intermediate thinking steps, while synthetic data will only be valuable when paired with an "oracle of truth" for validation.
- Training efficiency and hardware costs will see significant improvements over the next 12 to 24 months driven by a competitive price war and the vertical integration of silicon among hyperscalers like Amazon, Google, and Microsoft, who will hold a cost advantage over those reliant on NVIDIA or AMD.
- While the industry will move past experimental phases into developer-led AI-assisted work, end users will face prohibitive costs for training massive models, leading to a strategic shift toward distilling intelligence into smaller, economically viable models.
- A capital investment of $600 million is currently sufficient to enter the capabilities race but will be insufficient long-term, as the entry price for hyperscaler infrastructure players is estimated at $100 billion with multi-year spending far exceeding that threshold.
- GPU supply currently lags demand despite a recent six-month shift, and physical constraints plus algorithmic challenges prevent unlimited money from buying unlimited compute advantages, with data center infrastructure requiring radical changes in size and power to support massive training clusters.
- Three primary chip players plus possibly AMD are expected to dominate high-volume silicon production, with delays in NVIDIA's Blackwell generation potentially extending the competitive value of current H200 hardware.
- Regulatory trends are predicted to target end-user applications rather than compute or data limits, likely becoming an expensive bureaucratic overhead that disadvantages young startups while favoring entities with massive capital.
- Talent, proprietary research, and strategic viewpoints are identified as the most scarce resources rather than capital, with the race to AGI requiring teams capable of enduring significant sacrifices and navigating a "race" without stumbling on capabilities or go-to-market strategies.
- China is assessed to be at an incredible level of AI capabilities, not significantly behind the U.S., leading to a recommendation that the West attract Chinese talent to accelerate the Western capabilities race, with global conflict identified as the only potential cause for progress to halt.
- Value accumulation is expected to occur at the end-user level rather than the model layer, prompting a strategy of building vertically integrated intent businesses, with a prediction that only a select few companies including OpenAI, Anthropic, and Poolside will achieve escape velocity alongside hyperscalers.
- The exponential nature of current technological progress is noted as distinct from historical precedents like the 1996 internet boom, with expectations that the industry will not repeat history but rather rhyme, as the company aims for independence rather than acquisition.