This YC Paper Club explored the growing trend of specialization in AI hardware and software, covering multi-GPU kernel optimization, intelligence per watt metrics for local AI inference, the implications of AI writing systems code, heterogeneous hardware designs for inference, and GPU-accelerated game engines for reinforcement learning.
Vinoo Ganesh, CEO of Kepler, discusses building verifiable AI for financial services. He argues that while LLMs are powerful probability machines, they are unreliable for deterministic financial tasks. Kepler's solution involves a deterministic substrate that wraps AI, based on three tenets: atomic provenance (AI references, not creates, numbers), scope determinism (AI decides what to compute, but deterministic tools perform the math), and derivation chains (traceable calculation paths for derived figures). This ensures numerical accuracy, verifiability, and traceability, moving beyond simple citations to true proof of work.
Ishan Anand discusses synthetic personas for market research, drawing parallels to weather forecasting. He outlines three key failure modes—latent confounders, prompt sensitivity, and difficulty predicting actions—and explores techniques like fine-tuning and calibrated prompting to overcome them. Anand stresses the importance of validating personas against human data using distributional metrics and establishing a "noise floor" based on human-to-human consistency, advocating for their use as economic actors and complements to, rather than replacements for, human research.
Alexandr Wang discusses his journey from Scale AI to Meta's superintelligence lab, emphasizing the importance of conviction, systems thinking, and identifying exponential growth opportunities in AI. He highlights Meta's vision for 'personal superintelligence,' the strategic role of open-source and affordable models, and the immense potential of agentic looping for driving innovation and outcompeting incumbents. His core advice for young entrepreneurs is to develop an unshakeable internal compass for the future, embracing vision and ambition as the new scarce resources.
Nubank, serving 135 million customers, uses AI agents for support. The talk reveals how simulated data for evaluations (evals) has enabled them to ship AI agents 20x faster. By addressing the bottleneck of multi-turn, stateful eval data, Snowglobe's grounded simulations create realistic customer interactions, allowing rapid testing, derisking, and significant improvements in customer satisfaction and self-service rates, even for open-source model experimentation.
Morgan Stanley's AlphaLab is an open-sourced multi-agent system designed to automate quantitative research. Initially, AlphaLab 1.0 automated code generation, backtesting, and experimentation. Facing challenges, AlphaLab 2.0 evolves to prioritize building robust, verifiable environments, which serve as reinforcement learning signals, enabling the system to meta-optimize itself. This shift redefines the human role from performing research to designing these critical environments.