Summary

Image generation is entering another quality shock, but controllability is the commercially important unlock. Elad recalls the curve from a GAN artwork going to Sotheby’s in 2018 or 2019 through Midjourney and Stable Diffusion’s “everybody has seven fingers” phase to cohesive anime-style output. He thinks another leap will arrive in roughly another year, moving from today’s “horizontal version” toward vertical products that can handle graphic design seamlessly.
Public-market weakness should have minimal operating impact on early software startups unless the macro path becomes existential. Despite consumer confidence at a multiyear low and the Nasdaq down 8%, Elad calls it business as usual; even during the 2008 financial crisis, a six-person startup had little reason to react. Sarah sees ample capital for high-quality early opportunities and surprisingly deep funding for expensive foundation-model plays, though crossover investors and pre-IPO situations face greater caution and liquidity pressure.
Tariffs are an industry-specific policy instrument, not one undifferentiated macro verdict. Elad sees some tariffs as potentially useful, others as negotiation tools, and others as destructive cost increases. Sarah points to Chinese automotive competition as a case where Europe might protect its industrial base, and argues that a productive version of protection requires substantial investment in domestic skills, cost competitiveness, defense components, and automotive capacity.
Foundation-model capability and product surfaces are converging, shifting the strategic question toward distribution and consumer surplus. Sarah points to ArtificialAnalysis.ai charts showing capability convergence and says Google’s latest Gemini release shows it remains in the game. Elad adds that search, research, and reasoning are converging as product surfaces; benchmark clusters and xAI reaching a roughly SOTA model in about nine months show that outliers can still emerge. The opportunity set therefore extends beyond another general LLM into neglected models for physics, materials, robotics, science, health, and biology.
Specialized-model defensibility starts with proprietary data generation or a genuinely different technical thesis. Sarah’s key question is what the “data collection engine” looks like: chemistry, biology, and robotics may require new physical experiments that an existing large lab may not undertake, unlike code’s abundant digital corpus and testable utility functions. Elad’s plausible technical wedges include state-space models, Lean-based formal reasoning, and better RL environments for software agents.
The AI market feels like “maybe inning three instead of inning one,” with enough standardization to invest but no durable equilibrium. Model, infrastructure, evaluation, orchestration, and vertical-application layers are becoming legible, while MCP offers an open interface between models and existing data systems. Elad still warns that “the more I learn, the less I know”; this “moment of calm” may last only until the next release.

Deep dive

1. Image generation’s next unlock is control, not novelty

Elad places the Ghibli and anime wave on a recurring curve: he recalls a GAN artwork going to Sotheby’s for auction in 2018 or 2019, then Midjourney and early Stable Diffusion amazed users despite “everybody has seven fingers.” Today’s systems deliver cohesive styles with striking fidelity, making this “the latest version” of the public realizing how fast quality is compounding.
Sarah says users are good at sensing current quality and controllability, but the latest wave shows how much room remains in images, video, text, and logos. Demand is elemental — “people want more cute, they want more beauty.”
Elad thinks another similar moment will arrive in another year, followed by commercially seamless graphic-design products: “We’re doing the horizontal version of it, and soon we’ll have the vertical versions.” Sarah points to Krea-style live editing and HeyGen’s natural-language control — even specifying “whisper, ASMR” in a few words — as examples of tools responding directly to intent.

2. Public-market turbulence barely reaches an early software startup

Sarah frames the stress case precisely: consumer confidence is at a multiyear low, the Nasdaq is down 8%, and tariffs target Chinese imports and autos.
Elad’s answer is “not very stressed.” Barring something existential, software startups can still sell and raise if they are working; hardware is more directly exposed. During Sequoia’s 2008 “RIP Good Times” presentation, he asked why a six-person company should care. A partner agreed: “You shouldn’t worry about this at all.”
Sarah sees high-quality early opportunities remaining well funded and says capital markets for expensive foundation-model companies are deeper than she expected. The pressure concentrates in crossover and pre-IPO situations after years of constrained liquidity, although revived M&A and companies preparing to list could help.
Elad’s tariff framework is “item by item”: some protective tariffs may be useful, others may serve negotiations or impose net costs. Sarah uses automotive competition as an example, arguing that Europe might protect its industrial base against increasingly competitive Chinese cars. She adds that protection needs a positive industrial policy because rebuilding US capabilities in defense components or autos requires major investment in skills and cost competitiveness.

3. LLM convergence redirects attention toward neglected model markets

Sarah points to ArtificialAnalysis.ai charts showing convergence in capabilities and says Google’s recent Gemini release confirms that it remains in the game. Elad adds that convergence is also happening in product surfaces: search, research, and reasoning are becoming standard, making distribution and consumer surplus central questions.
Elad points to ArtificialAnalysis benchmarks showing clusters of models within striking distance, alongside spikes in coding or reasoning. Grok/xAI reaching a roughly SOTA model in about nine months was “super impressive,” evidence that convergence does not eliminate sudden outliers.
The undercovered opportunity lies outside core language models: physics, materials, robotics, science, health, and other specialized domains. Biology gets attention — “a new biology model every week” — but Elad sees funding and researcher interest as frequently divorced from commercial value, leaving potentially large markets untouched.
Sarah’s answer to the “one ring to rule them all” question is the data engine. Existing language and reasoning can seed specialized systems, but robotics, chemistry, and biology may require collecting or generating knowledge that does not yet exist; operating a physical laboratory may be much further afield for a general model lab than training code in RL environments.

4. Specialized models need a structural reason to survive the steamroller

Elad’s credible technical wedges include state-space models that are efficient on compressible data, translating math and code into Lean for formal reasoning, and models trained to act reliably across software and the web. The last category still lacks consistently generalizable RL environments for agents, making it a real research question rather than merely another wrapper.
Elad evaluates models across speed, cost, and reasoning fidelity. A slow, expensive but highly capable system can analyze a 100-page Supreme Court brief; a fast, specialized model can serve a narrow task or vertical. General models supply reasoning and language, while orchestration layers route tasks among models — effectively what today’s “agentic” products do across coding, customer success, and other domains.

5. AI may be in inning three, but the calm may last only a week

Elad describes a virtuous cycle: M&A is alive again, model development and test-time reasoning remain expensive, and companies solving data, scale, and latency constraints can improve the ecosystem. Sarah calls this a comfortable time to invest and says it feels like “maybe inning 3 instead of inning 1,” with some useful standardization.
The premium talent spans research, infrastructure efficiency, hardware-software co-design for sparsity or massive mixture-of-experts models, domain-aware product engineering, evaluations, and RL environments. Agent orchestration remains nascent: gather context, plan, parallelize model calls, verify, and retry.
Elad calls this the “business-as-usual phase of AI.” The model stack is becoming legible; RAG has moved from a new thing into the established stack; evaluation practices are solidifying; and some vertical winners are emerging. Consumer experimentation is still nascent. ChatGPT, Perplexity, and Midjourney may be viewed as earlier consumer forays, while newer consumer products are only beginning to appear. Elad expects today’s clarity to scramble again within a year.
Elad describes Anthropic’s Model Context Protocol as an open interface connecting model capabilities to documents, logs, business tools, IDEs, and other systems; OpenAI has said it will support it. MCP is incomplete and developers still must describe tools cleanly, but it could accelerate agent development substantially. Winning consumer agents beyond search and research remain unclear, though Sarah expects examples this year.

AI 2027: month-by-month model of intelligence explosion — Scott Alexander & Daniel Kokotajlo All-In Live from Austin: Colin and Samir, Chris Williamson, and Bryan Johnson