No Priors Ep. 109 | With Sarah and Elad

AI image and animation quality is improving in recurring, dramatic waves.

From early GAN art and seven-fingered Midjourney images to today’s polished animations, each new generation resets user expectations and exposes how much more room there still is for quality and control.

Macro market turbulence matters far less to early-stage software startups than people think.

For small, viable startups—especially pure software plays—cycles in the NASDAQ, tariffs, and sentiment tend to be a shrug unless funding in venture dries up dramatically or you’re on the cusp of an IPO.

Foundation language models are converging on capability, making distribution and product differentiation crucial.

Benchmarks show many top models clustered in performance, so advantages will increasingly come from distribution, user experience, verticalization, and how well models are integrated into real workflows.

There is large, underexplored opportunity in vertical and scientific models beyond language.

Domains like physics, materials, robotics, and specialized healthcare models may hold significant economic and societal value, but they’re underfunded relative to their potential because they’re harder and less trendy than generic LLMs.

Data collection and generation are the core bottlenecks for non-text AI domains.

Unlike language and code, where data is abundant and digital, robotics, chemistry, and other physical domains require expensive, bespoke data generation (labs, robots, experiments), which favors companies that can build those engines.

Model choice will be shaped by a speed–cost–reasoning trade-off matrix.

Elad frames a 2×2 where slow, expensive but very capable models power deep reasoning tasks, while fast, cheap specialized models serve narrow but high-throughput use cases, with orchestration layers routing workloads between them.

The AI stack is temporarily stabilizing, with standards like MCP accelerating agents.

With clearer layers—models, RAG, infra, evals, orchestration, and now Model Context Protocol to standardize model–data/tool connections—founders have a more predictable platform to build on, even though the next disruption is likely close.