A Prompt Cannot Hold a Good Design: Six AI Practitioners at Midyear 2024

Sarah Guo and Elad Gil open a midyear episode of No Priors by admitting the year is already half gone. The format is a reel of favorite moments from 2024 conversations, spanning research labs, hyperscalers, and startups, with links back to the full interviews. The first clip is Emily Glassberg Sands of Stripe. Guo asks her to set Stripe aside and name the biggest white space for AI in financial services, whether for a startup or an incumbent.

Sands will not predict what others will ship. She will name the gaps she can see. The first is identity, which sounds like a marketing problem and, she says, is actually load-bearing for fintech. Knowing who a business is and what it sells matters for credit. It also matters for whether the merchant meets a card network's rules or a BIN sponsor's requirements. Mapping "who is this, and what are they selling" onto a complicated regulatory environment is hard, valuable, and still being solved in fragments.

The second gap is integration. Stripe already has no-code payment links and invoicing, but a deep, user-specific integration has historically meant a bench of payments engineers. Large language models are proving they can write that code; Stripe already has internal cases that work. As the decisions get more complicated, automated integration remains unfinished. The third opportunity sits on top of the payments data itself. You can use it to make payments smoother. You can also use it to understand and improve the business, a health check rather than a clogged-artery test. The business model is unclear, and either a startup or an incumbent could own it. Stripe's own motive, she notes, is aligned: the more it helps users' businesses grow, the more Stripe grows, and the more the economy grows. Using financial data to grow the pie is powerful whoever does it.

Elad Gil frames Figma as a company sliding from human-to-human collaboration toward human-to-AI collaboration, a copilot for creativity. Dylan Field answers from first principles. What matters in design is the iterative loop, going back and forth with an agent and adding instructions over time. A prompt cannot communicate what a good design actually contains: the culture and ethos of the moment, the temporal feel of a sequence of interactions, the affordances people already know on a given platform, emotional state, videos the designer has watched, in-person research interviews, product requirements, and visual style. Fitting all of that into one box is hard.

If AI only unblocks brainstorming and a first draft, that is already interesting, because the designer can keep iterating from there. Asked whether AI will eliminate designers, Field splits the horizon. In the short term he thinks replacement is unlikely. In the long term, if superhuman intelligence arrives, he does not know what it means for anyone on the call. Assuming continued improvement instead, design looks well placed. Before any part of the role is replaced, he expects augmentation, access, and efficiency: designers get more done, and more people who are not designers can make something usable.

His tone changes on engineering. The abstraction level of coding is moving up. Copilot already exists, and many companies want to go further. It is not clear to him that humans will write every line of code in one, three, or five years, even if a human stays in the loop. He will not make multi-year bets in the current environment. As an optimist, he expects more software, and better designed software, not fewer designers or engineers. Visions in which a strong enough model erases the interface entirely, including gadgets like the Rabbit R1, strike him as cool. They can be true at the same time as a world that, in a year or five, simply contains much more software than today, because demand is there.

Brett Adcock, CEO of Figure AI, is building humanoid robots for dull and dangerous jobs humans should not do. Guo asks how you run a hardware-plus-software-plus-AI program at velocity. He says he believes in iterative design. The team does not spend a lot of time on research and analysis. It spends it building and testing, which shakes out problems, teaches, and feeds the next product in the continuum. Hardware and software should be updated forever. He does not think it will ever be good enough for them.

The process has an aerospace accent. First they figure out who the customers are and what the robot needs to do. Then they write requirements: how many pounds it must lift, how long it must run, where it charges, safety so the battery cannot burn down the building, IP ratings on actuators. Because the machine will work in the real world, safety sits at the same altitude as performance. Only then does engineering design start, gated by three reviews the whole company runs through the year: conceptual, preliminary, and critical. Adcock installed a similar practice at Archer, the electric-aircraft company he founded earlier.

Software companies also prototype, test, and use data to see whether an idea actually fixed the problem. Figure tries to do the same thing as fast as it can. The difference is the clock. Hardware timelines are an order of magnitude longer, so the team is more methodical up front. When they design an actuator from scratch in CAD, they run trade studies on the options first, so they do not spend months on something that cannot work. Software iteration can absorb a bad sprint. Hardware cannot absorb a bad drawing as cheaply, which is why the gates exist.

The next clip is OpenAI's Sora team, including Aditya Ramesh, Tim Brooks, and Bill Peebles. What has struck them is not only pretty frames, a cat's shadow in a pool of water, but what professional artists do once the model is in their hands. The team's own ideas of what to try are limited. People whose job is making content find more. One researcher points to Shy Kids, the Toronto collective, and their short *Airhead*, whose protagonist has a balloon for a head. The point was less any single clip Sora produced than a story that had been hard to tell and was now easier. Another researcher's personal favorite is Bling Zoo, a multi-shot scene posted on launch day: a New York zoo that is also a jewelry store, saber-toothed tigers covered in bling. For someone who loves making things but lacks the craft, iterating prompts to a piece they actually loved took less than an hour.

Gil asks about a timeline to professionally produced shorts and features, using Pixar as a comparison: years of shorts, then features, bottlenecked on world modeling as small as how hair moves, a prior generation of technology now about thirty years old. The Sora researchers refuse a date. They are more interested in uses that are not traditional film. Over the next couple of years people may make more movies with these models. They may also invent modes of interacting with content that do not look like current media, because you can tell the model what you want to see and it can respond. New forms, they say, are more exciting than an AI Pixar on a known schedule.

Gil then asks whether a video model becomes a physics engine, a wind-tunnel, a basis for robotics, or something else, noting OpenAI's earlier work on robot arms and self-play. One researcher says carrying out simulations inside the video model is something they expect to be able to do. Peebles takes the robotics question as the one that hits the nail on the head. Video contains what language does not: the minutiae of how arms and joints move through space, how legs in a Tokyo street scene meet the ground in a physically accurate way. Training on raw video, they believe, teaches a large amount about the physical world, and that will be essential for embodied intelligence. Half of the model's value sits on a screen. The other half points off it, toward robots.

Scott Wu, co-founder of Cognition, the company behind the AI software engineer Devin, is asked what will still matter for a human engineer five years out, a long time in AI. There is a meme that the hottest new programming language is English, and he thinks there is a lot of truth in it. The fundamentals remain valuable anyway. Everyone uses the internet; people who work with networks still benefit from understanding TCP. In the same way, we will communicate ideas in English to machines, and people who understand how computers work, logic gates, algorithms, and logical reasoning will still have an edge.

The job in five or ten years, as he draws it, looks like a mix of today's technical architect and today's product manager. You take problems the business is facing, think through what the solution should actually be, and break it down. More of the typing moves to the machine. Guo pushes the horizon further. Five years ago she would have told kids to study computer science and math; twenty years out she is less sure, if planning itself is done by machines. Wu says the education is worthwhile even if it were not practical, then repeats that the fundamentals will stay useful for a long time.

On superintelligence and the singularity, he is honest: everyone in AI has a guess, and it is hard to be high confidence. What he will claim is nearer. Concrete impacts on work, the economy, and people's lives will show up a lot sooner than those debates resolve. Even with tools available today there is a great deal that can change lives. The industry is still early. ChatGPT, he reminds the room, was about a year and a half ago, and there is a lot more to do on both the research side and the product side.

Alexandr Wang of Scale AI, the data foundry for the industry, is asked about trust now that systems are approaching and passing human abilities. Scale thinks about the whole AI life cycle: data abundance and quality going in, then measurement of the systems coming out. Measurement builds confidence and enables further development and adoption. Every AI company, in his telling, runs a loop of getting or generating data, training, evaluating, and going again. Evaluation is the hinge of that loop. It is also how society decides whether to trust the systems. Governments need to know whether models are safe enough to adopt. Enterprises need to know an agent will not create new consumer risk. Labs need a consistent read on how intelligent the thing they just built actually is.

What makes evals hard is that grading an approximation of human intelligence is not easy to automate. You are building IQ tests for models, which is already a philosophical problem: how do you measure the intelligence of a system? The practical problem is worse. Academic benchmarks the industry still cites are fraught. Many models are overfit to them; the items are in the training data. Scale's evidence is GSM1K, a new math evaluation held out so it could not have been in anyone's training set. Some models still performed well. Others performed much worse than their reported scores. How society is going to measure these models, Wang says, is an unsolved question.

His answer is to put the best human experts on the models, to see where they are strong, where they are weak, and what the risks are. He wants public visibility: leaderboards and evaluations that show performance in a rigorous, scientific way. He also wants platforms so governments, enterprises, and labs can evaluate constantly, not once. Data supply is Scale's original job. Building a layer of confidence through accurate measurement, he argues, is just as fundamental if the technology is going to keep being adopted. Guo and Gil close by sending listeners back to the six full episodes. The compilation's shared posture is the one Wu named: skip the end-state bet, keep iterating with what is already on the desk.