Nasdaq Down 8% and Ghibli Everywhere: AI Is Already in Inning Three

No Priors this week is just the hosts. Sarah Guo asks Elad Gil how he feels about Studio Ghibli-style image generation, and about the jump in quality around it. Gil has been an anime fan for a long time. Converting the world into anime or manga, he says, is "a very positive step for humanity," something he has been waiting for. He is more interested in the pattern: every year or two the image-gen world produces another wow.

The first wave, as he counts it, was GANs. Around 2018 or 2019 a GAN artwork went to auction at Sotheby's, one of the first AI-generated pieces to do so. The toolchains were clunky and the pictures look terrible next to today, but people were already saying art would change. Then came early Midjourney and Stable Diffusion: amazing, except everyone in the frame had seven fingers, and the same society-transforming speeches. The latest version is ChatGPT and OpenAI folding this into the 4o-class models, with cohesive style, aesthetic control, and text and logos that actually come out right.

Guo thinks users, not just researchers, are very good at projecting quality and how much room is left. The jump from eight bits of grayscale to images that might pass as photos of real people shocked people two generations ago. Midjourney, she says, had an aesthetic point of view and took a lot of user preference into the product; you can often tell a Midjourney image. End users had started to treat image generation as a bit more solved. This moment is a reminder of how much more is coming, including video.

Gil expects another wow in about a year, then the commercial versions: graphic design that works seamlessly instead of kind of working, plus vertical tools. He names Recraft on the design side and Krea for clicking into an image and modifying pieces in real time. Controllability is the other unlock. HeyGen has been showing natural-language control of a person's emotion and voice — whisper ASMR, the whole video in that register, from a few words of description. More control means more people who can create. Guo jokes she might just live in an ASMR world. Gil would take only part of that deal.

Are they freaked out about the macro? Consumer confidence is at a multi-year low. The Nasdaq is down 8%. There are tariffs on Chinese imports and on autos. Investors and companies are talking about how stressed they are. Gil is not very stressed. There is uncertainty, but for people building technology companies — barring something truly existential — it is business as usual. He has been through cycles in both directions. The path into the venture world is that money gets soaked out of VC, valuations come down, and the marginal startup finds less funding. Super-late-stage companies about to go public can get caught between expectations and the window. Day-to-day software startups, especially ones not doing hardware, should feel minimal impact if customers will pay.

In 2008 he sat through Sequoia's "RIP Good Times" presentation. Sequoia pulled founders together, brought in people with war stories from the dot-com collapse, and told them to batten down, do layoffs, and accept that the world would never be the same. Gil was CEO of a six-person company. He asked a Sequoia partner what this had to do with him. The partner said he was right: he should not worry, even as financial institutions were collapsing around them. That crisis did not do much to tech. Google may have done its first layoff. Otherwise the industry hummed, and the biggest tech companies are now 10 or 20 times larger. This, he says, is an even more minor blip from a long-term tech perspective.

It has almost no impact on Guo either, especially at the early end of the market. High-quality opportunities still have plenty of capital; she keeps discovering that the markets are deeper than she thought for expensive foundation-model plays. Crossover investors with public-equities exposure will get more cautious on pre-IPO rounds after years of starved liquidity. A return of M&A, and several companies ready to go public, should help. Tariffs, both agree, are not a catch-all. Automotive is a case where Chinese carmakers are getting so good that Europe, given how auto-dependent its industrial base is, might reasonably want protection. Some tariffs are a negotiation tool. Some will pass through as cost. The productive version, Guo says, is broader industrial policy: if the United States wants key defense or auto components at home, it is behind on skill and cost, and that takes real investment on both the positive and the protection side.

Gil points at charts from artificialanalysis.ai: a more competitive market for capabilities, amazing improvement over 18 to 24 months, and Google still in the game after the latest Gemini release — which should surprise no one given the infrastructure and the people. Guo notices convergence not only on capability but on product surface. Most labs now have search, a research product, and reasoning. A lot of it, she thinks, will end up as a question of consumer surplus and distribution: who reaches users, who gives more away.

On those benchmarks — reasoning, knowledge, tokens per unit time — a cluster of core language models sits in the same ballpark, with occasional spikes in coding or reasoning and a longer tail behind them. Outliers still matter. Grok, from xAI, coming out of nowhere with a roughly state-of-the-art model in about nine months was super impressive. DeepSeek and others keep rewriting cost assumptions. The site does not really cover image models, which live on other leaderboards.

What the leaderboards miss is a whole suite discussed a lot less: physics, materials, robotics, certain sciences, and specialized post-training such as health data on top of a core model. Biology is the one non-LLM area that already gets weekly new models. Other fields have large societal impact and, in some cases, large markets, but the interest of people who build models is often divorced from the economic value. Sometimes that is fair — a scientific application with no commercial path. Sometimes it is misaligned: things get funded while wide-open model spaces sit empty. Gil has been looking for alternative models that are interesting from a market perspective and a little ignored.

Will specialized domains get subsumed into core LLMs — one ring to rule them all — or stay fragmented? Guo thinks "model company versus not" is too binary. Even research labs start from pre-training on existing knowledge and reasoning that is more and more available; robotics starts from video; a code-focused effort still wants language in the mix. The distinction she cares about is the data-collection engine. In physics, chemistry, biology, and robotics, the data you would want often does not exist yet. The interesting theory of those companies is: our job is to collect or generate it efficiently and train on it.

Would you expect an existing large lab to stand up a physical robotics lab that runs new chemical experiments? That feels farther afield than code generation in RL environments. Anytime you go into the physical world, data is harder. That is why language models, which collect the wisdom of the internet digitally, saw the first recent breakthroughs at this scale. Coding is even kinder: lots of data already resident, plus a clear utility function — does the program do what you think? Those will keep being the easy areas. Gil's pet peeve is software founders who made their money in traditional tech telling everyone else to go be hardcore in biology and materials. He remembers Bill Gates saying twenty years ago that if he were starting today he would go into biology. Guo needles him: he is the opposite case. He actually has a PhD in biology, which is how he knows what reality looks like.

The other filter is an orthogonal technical thesis, architecturally different. Cartesia's founders, Karan and Albert, have been on the show; state-space models can be highly efficient on compressible data. Several companies translate problems into Lean as a path to better math and code reasoning. Others train models that take actions in software and on the web — closer to the frontier labs' line of fire, but aimed at a question that still feels open: consistent, generalizable RL environments for agents. If the thesis is true, she wants a reason the company should exist. Otherwise it is standing in front of the OpenAI, Anthropic, xAI, and Google steamroller.

Gil puts model choice on a two-by-two. One axis is speed, performance, and cost, which for many models are roughly the same thing: inference time. The other is reasoning fidelity. Slow, expensive, and not smart is empty. Slow, expensive, and very capable is high value: upload a hundred-page Supreme Court brief and wait for analysis you can argue a case with. Super fast and performant tends to be specialized niche models; some SSMs live there. The last quadrant is where generalizable models supply reasoning and language, then you apply them to a domain. Almost everything branded agentic right now, he says, is an orchestration layer on top of that: gather context, plan, call models, parallelize, verify, retry, copied domain by domain through customer success and code.

Guo sees a virtuous cycle that dwarfs the macro: M&A is alive again, so there will be outcomes; the surface area these models can attack is exploding; research bets are proliferating; reasoning and test-time compute keep getting more expensive, as does training. People showing up to solve data, scale, and latency problems help everybody. Someone has asked whether DeepSeek researchers are not allowed to leave China. Gil does not know if that is true. Any country should want to hang on to its best talent, he says, but perhaps not restrict movement. The United States should be trying to attract them. They riff on locking AI researchers between San Francisco's Mission District and Dogpatch, daily check-in at Atlas Cafe, a joke visa named Gillingrow.

For the next two or three years, not for someone's kids in ten, the expertise they would sponsor is specific. Model researchers. Infrastructure scaling and efficiency. Hardware-software co-design: the next TPU, chips that handle sparsity or massive mixture-of-experts. People with deep domain and user understanding plus product engineering for applied ML orchestration — evals for agents, RL environments, that still-nascent loop of gather context, plan, make a bunch of model calls, parallelize, verify, retry. That list is the hiring market they are staring at.

Gil calls this the business-as-usual phase of AI. The stack is reasonably well defined. The last couple of months clarified what is short-term crucial. The model layer, with its agent and reasoning accoutrements, is on its own scaling curve. Infrastructure has solidified: RAG used to be the new thing; evals are falling into place with companies like Braintrust. Applications are eating into services and verticals, and in a handful of those he is starting to see winners. Consumer is stirring after two or three years when almost nobody tried, unless you count Perplexity, ChatGPT, and Midjourney as early forays. ChatGPT may be the world's biggest new AI consumer product; Google, in some sense, was the original. It is a brief consolidation. It is also, he says, the only industry where the more he learns about the market the more confused he is. This moment of clarity, he guesses, and in a year all bets are off.

Guo is comfortable investing because it feels like inning three rather than inning one: a little stability, some real goodness around standardization. MCP is the example she thinks will accelerate a bunch of development. She is already meeting companies that wrap a useful enterprise data source and say, as a reflex, MCP server. Model Context Protocol came from Ben Mann's group at Anthropic. It is an attempt to spec a standard interface from model capabilities to systems that already hold useful data — documents, logging, business tools, the IDE. Sam Altman said OpenAI would support it. It is an open standard, a two-way connection between data sources and AI-powered tools, not a complete solution, and it got popular with developers very fast.

Developers still have to describe their tools specifically and cleanly. Even so, she thinks it makes the work much easier and will accelerate agent development a lot. Combine that with more model choices and a clear pathway to automate certain work as orchestration, and the enterprise side looks fertile. Consumer is the open question. She has not seen consumer agents that do not look like search or research inside the large-model products and actually work. She expects to see them this year. Gil's sign-off is to enjoy the calm. Calm is relative — enormous innovation, new things every week — but there is a slightly clearer view of who the main players are and how the pieces fit. Enjoy it for the next week, or the next few hours, before the next thing drops.