No Priors Ep. 100 | With Sarah and Elad

Benchmarks show major models clustering closer in capabilities, while training and inference costs for GPT‑4–class performance have fallen sharply—Elad cites an ~180x cost-per-token drop in 18 months—suggesting intense commoditization pressure at the base-model layer.

Despite headlines about a $5.5–6M training run and NVIDIA’s stock drop, the hosts believe DeepSeek likely consumed hundreds of millions in prior experimentation and largely fits existing trends rather than resetting the economics of training from scratch.

Leading models can capture sticky user share, and—critically—can be used to generate synthetic data, label data, and build tools that accelerate the next generation of models, potentially creating a compounding advantage if capabilities keep bootstrapping.

Deep Research already rivals or surpasses median analyst-level work for many tasks, especially exploratory research in unfamiliar domains, meaning organizations should start benchmarking internal knowledge work against what these tools can produce.

As LLMs blend search, social feeds, and media into one primary interface, whoever controls model outputs wields enormous narrative power; this makes multi-model competition and open source critically important from a civil-liberties standpoint.

Regardless of algorithmic efficiency gains, any serious AGI lab would gladly accept the largest available cluster (like Stargate) if capital were free, implying that large-scale pre-training will remain a key driver of frontier capabilities.

The hosts expect 2025 to see consolidation in generic models, expansion of vertical AI ops (legal, customer support, codegen, medical scribing), more agentic workflows, technical robotics milestones, and domain experts designing smarter data-generation pipelines in fields like biology and materials science.