Interview, Fireside Chat, Conference Presentation, Keynote, Panel
AMD, Starcloud, Coatue..10 Hot Takes From The Biggest Names in AI
Sourcery with Molly O'SheaMax Cook, Matan Grinberg, Pim de Witte, Philip Johnston, Mark Papermaster, Rodrigo Liang, Ramin Hasani, Robin Rombach, Chris Madden, Henri Delahaye, Molly, Philippe, Thomas
- Max Cook forecasts that keyboard and mouse usage will decline significantly, anticipating a definitive transition to natural language interaction via AI agents by 2030, potentially accelerated by an OpenAI "super app" for its nearly billion weekly active users.
- The industry is predicted to shift toward natural language interfaces and agentic workflows, with OpenAI expected to offer model routing and fine-tuning services using smaller, cheaper models, while the consensus suggests device interaction will evolve as historical tech eras transition.
- Open source model adoption is projected to surge, with token usage rising from less than 1% at the year's start to 90% within the next 12 months and enterprise open source usage potentially crossing 50% by the end of the calendar year, as cost-effective open models become primary for production while frontier models focus on discovery.
- Future foundation models are expected to co-evolve across intelligence, efficiency/cost, and substrate axes, with intelligence moving beyond data centers to diverse substrates like phones, laptops, airplanes, and cars to achieve enterprise scalability and value.
- Philip Johnston predicts the SpaceX IPO will be viewed historically as undervalued as the company generates over $10 trillion within the next couple of years, leveraging cost-effective launch vehicles to open industries like asteroid mining and enabling space-based compute for chips from various manufacturers.
- SpaceX plans to expand manufacturing in August for "StarCloud 3," a system designed to fit up to 10 megawatts of compute capacity per launch, which will support a chip-agnostic strategy including SambaNova, Cerebras, ARM, and NVIDIA components.
- SambaNova expects to ship its SM50 chip later this year for premium inference at a fraction of GPU cost and power, positioning itself to offer multiple services before the end of the year to support the industry's shift into the inference phase.
- Mark of AMD observes a rapid shift in agentic workflows where CPU to GPU usage has reached a one-to-one ratio in the last six months, predicting thousands of sub-agents will accelerate chip design and that agentic workflows are currently in their early stages.
- Ramin of Liquid AI warns that token-based revenue models may become obsolete as focus shifts to outcomes, while enterprises in Europe are expected to wake up to the necessity of treating AI with the same seriousness as traditional technologies to understand value before deployment.
- Robin of Black Forest Labs advocates for open innovation to prevent AI progress from slowing due to fear-mongering and closed systems, suggesting European tech must adopt a fundamental mindset shift beyond hosting events to drive meaningful AI innovation.
- Chris of Good Future Media identifies "clipping" as the dominant ad format for long-form content, with professional services charging around $200 per clip, emphasizing that the first three seconds must contain immediate visual or spoken hooks to achieve virality.
- Jesse Zhang notes that 90% of Decagon's workloads run on open source models due to latency and fine-tuning requirements, contrasting with an enterprise LLM spend on open source that dropped from 19% to 11% recently, though private workload adoption remains high.
- Matan from Factory forecasts that while frontier closed models will retain a place, their token share will shrink as open models become the primary choice for production-grade workflows, driven by their increasing performance, speed, and cost efficiency.
- Ramin highlights that 85% of current AI activity is dedicated to R&D code, which enterprises are only now addressing, and warns of potential government asset control and memory supply chain disruptions that will make efficiency and cost critical design priorities.
- Max Cook anticipates that "Thinking Machines" reaction models will integrate into every intelligence layer due to low-latency speech capabilities, while noting that wearables and Neuralink remain active areas of investment despite uncertainty regarding final interaction forms.
- The consensus indicates that as AI moves from discovery to production, open source will serve production workflows where parameters are known, whereas frontier models will remain best for discovering new use cases, fundamentally altering the open versus closed debate.