CHAPTERS
- 0:05 – 3:43
Image generation hits another “wow” moment (anime, style, fidelity)
Sarah and Elad kick off by reacting to the latest leap in image generation quality, sparked by nostalgia-driven “Studio Ghibli/anime” aesthetics. Elad frames this as one of several recurring inflection points where the public suddenly realizes how far the tech has advanced.
- •Historical waves: GAN-era novelty → Midjourney/Stable Diffusion breakthroughs → latest step-change
- •Quality/fidelity and cohesive style imitation are now dramatically better
- •Image gen is increasingly native inside frontier model product suites (e.g., OpenAI)
- •Expectation that another major leap will arrive within a year or two
- 3:43 – 4:05
From horizontal tools to vertical workflows: design, editing, and controllability
They move from “cool outputs” to the practical trajectory: image tools becoming seamless for real work. The emphasis shifts to controllability and interactive editing as the next unlock for mainstream use.
- •Verticalized products will emerge from today’s general-purpose generators
- •Interactive manipulation and real-time editing (e.g., Krea-like workflows) matters
- •Untapped needs: text rendering, logos, and integrated design systems
- •Controllability is positioned as the real creative power multiplier
- 4:05 – 4:42
Natural-language control over video and voice: emotion as an interface
Sarah highlights how generative media is expanding beyond images into highly controllable video and speech. She points to products that let users specify emotion and delivery style in plain language, collapsing complex production into short prompts.
- •Natural language as a control layer for emotion/tone/voice (e.g., “whisper ASMR”)
- •Creative production shifts from tooling expertise to intent specification
- •Video generation is implied as the next major consumer-facing step
- •Augmented/immersive possibilities are teased (living inside an “anime/manga world”)
- 4:42 – 6:07
Public markets jitters vs. startup reality: why early-stage keeps shipping
Sarah asks whether markets and consumer confidence declines are worrying; Elad largely shrugs. He argues that most software startups see limited day-to-day impact unless they’re late-stage or directly exposed to capital market tightening.
- •Market drawdowns create uncertainty but rarely change early-stage execution
- •Main impact channel: venture liquidity and valuation compression
- •Late-stage/pre-IPO companies feel the valuation/public comps pressure most
- •Software startups are less exposed than hardware-heavy businesses
- 6:07 – 8:08
A 2008 flashback and a “who cares?” mindset for builders
Elad recounts attending Sequoia’s famous 2008 “RIP Good Times” talk while running a tiny startup—and being told not to worry. The anecdote underlines his belief that even big macro shocks often don’t derail tech’s long-term trajectory.
- •2008 crisis context: dramatic warnings vs. minimal relevance for a 6-person startup
- •Tech kept “humming along” despite broader financial collapse
- •Perspective: today’s volatility feels minor in comparison
- •Long-term: major tech incumbents grew massively after prior downturns
- 8:08 – 9:06
Tariffs and tech: when protection helps, when it harms, when it’s leverage
They dig into tariffs as a more nuanced, sector-by-sector policy tool rather than a blanket good or bad. Automotive is used as the key example where protection could preserve industrial capacity against increasingly competitive Chinese manufacturers.
- •Tariffs may be beneficial in select strategic industries (e.g., autos)
- •Some tariffs are negotiation tools rather than permanent policy
- •Potential downsides: higher costs passed through and economic drag
- •Call for item-by-item analysis instead of one catch-all narrative
- 9:06 – 9:36
Industrial policy reality check: rebuilding capabilities takes investment
Sarah agrees tariffs can be productive when paired with an affirmative industrial policy. The conversation emphasizes that catching up in key supply chains (defense, automotive, components) requires sustained investment in skills and cost competitiveness.
- •Tariffs alone aren’t sufficient—need broader industrial policy
- •US competitiveness gaps exist in multiple domains
- •Strategic areas: defense-related components and automotive supply chains
- •Balance of “support + protection” to rebuild capacity
- 9:36 – 11:33
Foundation model market convergence: benchmarks, product sameness, distribution
They pivot to what’s happening in frontier models: rapid capability gains and visible convergence across top labs. Sarah notes similarity across product surfaces (search, research tools, reasoning), shifting differentiation toward distribution and consumer surplus.
- •Benchmarks show increasing competitiveness and clustering near the top
- •Google/Gemini remains a serious contender due to talent and infrastructure
- •Product surfaces converge: search/research experiences + reasoning features
- •Differentiation likely shifts toward distribution and go-to-market
- 11:33 – 13:16
Beyond LLMs: specialized models where economics and attention diverge
Elad broadens the lens: many valuable models live outside the mainstream LLM spotlight—physics, materials, robotics, and domain post-training like healthcare. He argues attention/funding can be misaligned with real economic value, creating openings.
- •Specialized domains: physics, materials, robotics, scientific modeling
- •Biology is receiving outsized visible momentum relative to other fields
- •Opportunities emerge where commercial value is high but attention is low
- •Key strategic question: what gets absorbed into general LLMs vs. stands alone?
- 13:16 – 16:17
“One ring” vs. fragmented landscape: data engines decide who can win
Sarah reframes the “model company vs not” debate and focuses on proprietary data generation as the real moat. In areas like robotics or chemistry, the key constraint is building a data collection/experiment engine—not just model architecture.
- •Many “model companies” start from shared foundations (language/video pretraining)
- •Core differentiator: how you collect/generate the missing domain data
- •Physical-world data (labs, robotics experimentation) is harder than digital RL loops
- •Standalone model opportunities increase when big labs won’t build the data engine
- 16:17 – 17:51
Research approaches that can justify new companies: SSMs, formal methods, agents
Sarah outlines technical theses that may still support independent model efforts: state-space models for efficiency, formalism-based reasoning gains, and action-taking agents in software. A major open problem is reliable, generalizable RL environments for agents.
- •State-space models as an efficiency/compressibility-driven alternative in some settings
- •Formal methods (e.g., Lean-style translations) to improve math/code reasoning
- •Agentic models focused on taking actions on the web/software
- •Need for stable RL environments: plan → call models → verify → retry loops
- 17:51 – 21:06
Speed/cost vs. capability: the model “quadrants” and orchestration layers
Elad introduces a practical two-axis framing: inference cost/speed versus reasoning power. He connects this to product design—general, expensive “smart” models vs fast niche models—and explains why orchestration and routing across models is central to today’s “agentic” systems.
- •Two-by-two: speed/cost/performance vs reasoning/fidelity
- •High-capability slow models fit high-value analyses (e.g., complex legal docs)
- •Fast models often win in narrow vertical tasks
- •Agentic systems commonly rely on orchestration layers and model routing
- 21:06 – 22:23
What expertise will matter next: infra, co-design, and agent evals
They jokingly propose a “visa program” for the talent they most want—then get specific. The valued skill clusters span research, infrastructure/scaling, hardware-software co-design, and applied agent-building (evals, RL environments, orchestration).
- •High-demand talent: scaling/efficiency infrastructure and model researchers
- •Hardware/software co-design for next-gen accelerators and sparsity/MoE
- •Applied agent engineering: orchestration patterns, verification, evals
- •Domain + product engineering hybrids are especially valuable
- 22:23 – 24:14
Stack clarification and consumer comeback: winners emerging, but uncertainty remains
Elad argues the AI stack is entering a “business as usual” phase where layers (models, infra, apps) feel more defined, even if temporarily. He also sees early signs of consumer experimentation returning beyond prosumer staples like ChatGPT and Midjourney.
- •AI stack solidifying: model layer + infra (RAG, evals) + applications
- •Services/vertical software is a major near-term impact zone
- •Consumer is re-emerging after a lull; experimentation is picking up
- •Caveat: the field repeatedly shifts—clarity now may scramble again soon
- 24:14 – 26:17
Anthropic’s Model Context Protocol (MCP): standardizing tool/data connections
Sarah explains MCP as an open standard for connecting models to existing enterprise tools and data sources via a consistent interface. With broader ecosystem support (including OpenAI signaling support), they expect it to accelerate agent development, though it’s not a complete solution by itself.
- •MCP defines a standard interface to expose data/tools (docs, logs, IDEs, business apps)
- •Open standard, not proprietary; intended for broad adoption
- •Developers still must describe tools cleanly for reliable use
- •Expected to speed up practical agent integration in enterprises
- 26:17 – 27:44
What’s next for consumer applications—and why 2026 could look different
They close by contrasting enterprise momentum with the still-unclear shape of breakthrough consumer agents. Sarah expects new consumer experiences beyond “search/research wrappers” to appear soon, while both emphasize that the current calm is temporary.
- •Enterprise agent ecosystem is getting more fertile via standards + better models
- •Consumer winners are not yet obvious; most tools still resemble search/research
- •Expectation of new consumer agent breakthroughs “this year”
- •Closing note: enjoy the moment of clarity before the next destabilizing leap
