Interview, Fireside Chat
Better AI Models, Better Startups
Y CombinatorGary, Jared, Harj, Diana, Melanie Warrick, Mark Mandelmann, Mark Blythington, Joel Morton, Jordan, Francesc Campoy Flores, Carrie Nordlund
- Startups are advised to focus on building anticipation based on product announcements rather than fearing direct competition from major incumbents, as history suggests survival depends on non-bundled software, specific verticals, or use cases big tech underestimates due to complexity or liability.
- New model releases generally benefit the ecosystem by adding capabilities and modalities, though startups risk obsolescence if their core features can be easily replicated in the next major release of a dominant provider like OpenAI.
- Upcoming model generations are expected to reach intelligence levels of 100, 110, 120, or 130 compared to the current approximate 85, with next iterations including speech, video, and "true mixture of expert" architectures that improve energy efficiency.
- Context windows are projected to expand to one million tokens for commercial use and ten million tokens in research, potentially reducing the necessity for RAG systems in consumer applications but maintaining their critical role in enterprise environments requiring privacy, fine-tuning, and specific data retrieval.
- The competitive landscape is expected to diversify with the emergence of Meta's Llama 3 (400 billion parameters) and other alternatives, fostering a non-monopoly pricing environment that could support a thousand successful startups.
- OpenAI is predicted to launch a desktop-based "HER" digital assistant within the next two to five years, capable of accessing local files, executing transactions, and monitoring user tasks minute-to-minute, potentially extending into B2B through a GPT Store.
- Significant revenue growth opportunities exist in B2B AI applications, particularly in fintech and healthcare, driven by regulatory compliance, data privacy needs, and the automation of high-complexity human workflows that incumbents may overlook.
- The market is anticipated to shift from "tools plus people" to fully automated "tools," potentially converting billions to trillions of dollars in transactional labor revenue into software revenue over the next 10 to 20 years.
- While massive consumer software from big companies is predominantly B2C, creating a large opportunity for B2B applications, startups should target "unsexy" utilities or "edgy" consumer spaces involving legal/PR risks where incumbents are hesitant to operate.
- Robotics and custom silicon are expected to become viable sooner rather than later, driven by unified models, a halving of model running costs, and affordable hardware like the $16,000 Unitree biped, enabling internet-independent devices.
- Character AI and Replica AI are forecast to achieve deep retention through long-term user interactions with virtual entities, while opportunities in deepfakes for famous likenesses face significant regulatory and shutdown risks, especially during election seasons.
- Advancements in voice generation are expected to produce more human-like and emotional output compared to existing models, and live translation capabilities are predicted to have transformative consequences for global communication.
- As the industry matures, future product demonstrations may shift focus from groundbreaking technologies to minor feature improvements, with the primary risk for startups becoming competition within their own specific niche rather than from generalist giants.