Nobody Plans Dinner in Keywords

Sarah Guo and Elad Gil welcome Fidji Simo, CEO of Instacart and co-founder of the Metrodora Institute, a clinic and research shop for neuroimmune-axis disorders. She previously led the Facebook app. Instacart, they note, is a grocery technology company trying to get busy people food from the retailers they already like. Guo starts with the career: strategy at eBay, a rise through product to running Facebook's main app, then Instacart. What drove the jumps? Simo calls herself a pragmatic technologist. She loves tools that solve real problems for real people. eBay was communities of collectors trading goods and, more than that, passion. Facebook was an immigrant in the United States finally staying close to family in France. Instacart was food, one of the most important things in a life, and a chance to give people dinner and time with family.

Guo remembers the first impression: very French, lipstick, heels, running iconic products without buying Silicon Valley's monoculture. Simo says it was never a branding decision. She has been French since birth and in lipstick and heels since she was twelve. Valley colleagues were the ones who commented, and when she moved toward product they told her to blend in with engineers: lose the makeup, tie up the hair, wear a hoodie. She tried it for one day at Facebook. Nobody recognized her in the hallways. Nobody said hi. It felt like a lot of work not to be herself. If she had energy to spend, she would rather put it into building products than into shaping herself into someone else. She kept the accent and the heels. People adapted. Better to be you, she says.

Guo asked a mutual friend, investor Ravi Gupta, what to ask. Gupta said Simo earns tremendous loyalty while remaining really tough. How does she earn followership? She tells leaders, including when she joined Instacart, that she wants them to look back and see her as a big accelerator in their careers. That partnership has two sides. She will hold them to a very high standard and give a lot of tough feedback. She will also back them every step, in their corner, wanting them to succeed like no one else has. It is not for everyone. Some people hear it, say it sounds too stressful, and walk. She thinks that is fine. You do not have to be a leader for absolutely everyone; you pick each other. The people who stay tend to want to excel, and to do it with someone who will create enough safety to stretch and maybe fail. When a whole team is in that mindset, she says, you get the best work of their lives.

Becoming CEO in 2021 was not a quiet onboarding. The company had 4x'd the year before on pandemic demand, new habits, and a suddenly much larger team, while the outside question was whether Instacart was a pandemic darling, a one-trick pony people would abandon when they went back to stores. Ten years at Facebook, she says, prepares you to handle a lot. Fast forward a couple of years and the growth and the habits held. What you do in a moment like that is go back to mission. She spent time painting the picture, including a full-day Vision Summit. Two facts had to be shared: Instacart had become critical infrastructure, which came with responsibility, and it was also a retail-enablement company. When you order on Publix, Kroger, or Schnucks, a lot of the technology underneath is Instacart's. That was the next mountain, a decade of work, a reason to rally after a burst of growth.

Guo asks when Instacart started on AI, and how this generative wave differs from earlier ML. Simo is slightly impatient with the idea that they are chasing a moment. Instacart has been an AI company from the beginning. It is fundamentally a search engine: a billion searches a quarter, all AI-powered. Every time a shopper finds a hole on the shelf, the system suggests a replacement; they have suggested a billion of those. Matching shoppers to stores is AI too. What generative models change is that software can finally answer in natural language. Online commerce, she says, has done something bizarre: it trained humans to express shopping as keywords. Nobody plans a week of dinners as cilantro, eggplant, sour cream, which is what they type today. They think: I have a $300 budget, this many meals, a lactose-intolerant daughter, a Mediterranean tradition, a set of usual recipes, and a willingness to try one new thing. That is how people ask. Software should answer that way. Matching the answer to the question is what excites her about the wave.

Enterprises often treat generative AI as the same old curve with a new model and miss that it can do dramatically new things. Simo's prompt to the team was not whether they needed an AI roadmap. It was: what problems have we always wanted to solve for users that we can now solve so much better? She sees too many companies start from the roadmap and work backward. Tactically they run a pair of tracks. Under the chief architect, a centralized team later called Carrot AI goes deep on what is newly possible. A federated track pushes those learnings so every product team actually uses the tools. You cannot expect every engineer to go all the way down; you also cannot leave generative AI in a central silo if you want it in every product. They ran a company hackathon on the same prompt, not "use gen AI" but "take the problems we already have." Carrot AI gets pinged constantly for advice. They also shipped an internal assistant, Eva, so engineers can use GPT-4 with shortcuts and a friendly interface, as part of the work rather than a separate toy.

She teaches AI-native the way she taught mobile ads. After Facebook's IPO the headlines said the company would never monetize a screen that small. Plenty of firms shrank desktop ads to fit. Her team built ads in News Feed, a different paradigm. Being AI-native means rethinking from the ground up, not stuffing the new thing into the old box. Ask Instacart, which lets you ask what sides go with grilled chicken, produced a fight over the search box. On a commerce site the box is sacred; people type keywords; a question might confuse them; maybe put it somewhere else. The decision was to put questions in the search box, because if you rebuilt commerce from scratch that is where you would start asking. Micro-decisions like that, she says, train the company to disrupt its own prior paradigm.

Why an internal assistant instead of sending everyone to ChatGPT? Accessibility, shortcuts, an interface that feels like core work. Productivity numbers are too early, and she thinks the gains will show up well beyond engineering. Marketing built a library of animated characters with AI and mirroring-style motion, more than 200 of them, which would have cost a fortune through an agency. Legal used AI to detect the custom terms Instacart has with specific retailers across a pile of contracts, so they can track them. No lawyer, she notes, will tell you that reading a thousand documents for special clauses was the highlight of the day. Give that to the model and people get to do the part of the job that is actually worth them.

The ChatGPT plugin was an early look at how people actually talk when the funnel is not yet "I have decided to get groceries." Instacart today is a bottom-of-funnel utility: you already want delivery and you already know what to eat. The plugin starts earlier. I have a birthday party, I do not know which snacks to get. That is intent and occasion that keyword search never saw. They are reproducing the experience with Microsoft Bing and Google Bard, hoping to be the conversion layer between bits and atoms. Everything software, she thinks, is about to get easier and cheaper. The translation from software into physical things is still hard. Instacart has aggregated 80,000 stores and more than a thousand retailers, which is how a natural-language ask becomes food at the door in two hours. Next she wants that as a platform: any nutrition app, anything in food or even commerce, tapping the network to turn intent into a doorbell. Longer term it is answering all the questions people ask themselves, which takes LLMs plus unique data. An extraordinary catalog and billions of data points on what people need, married to the models, are the strength. Her advice to any company is to look hard at the data infrastructure and the unique sets you bring to this world.

Gil asks about the physical side LLMs have not unlocked: autonomous delivery, drones, automated picking. If you rebuilt grocery from scratch, Simo says, you would probably have fewer ordinary stores and more dark, automated ones. You are not starting from scratch. Instacart is a network of 80,000 stores whose fixed costs retailers already pay. Sending people into those stores is efficient: incremental demand on a paid-for base, and the stores sit close to the customer. Automation initiatives have failed, in her view, less because the technology is missing than because the economics do not work. A very large automated warehouse has to sit farther away to aggregate orders, which means next-day instead of same-day, and customers want speed. You get nice automation that neither cuts cost nor improves the experience. Back to pragmatic technology: if it does not help cost or experience, it will not get mass adoption soon.

The in-store bet is the other direction, into the connected store. Grocery's digital transformation is happening inside the aisles, not only in apps, and grocers now look to Instacart for the technology. Within about a month of joining she acquired Caper, whose smart carts use computer vision so items dropped in are detected and you can check out from the cart. Caper Carts, along with scan-and-pay from a phone, are meant to bridge online and offline. Omnichannel customers, they have data to show, are more valuable than online-only or offline-only ones. The carts are at Kroger, Sobeys, and Wakefern. Skipping the queue is almost a side effect. The screen on the front is the point: drop in a steak, get a dressing suggestion, an aisle, the savings, a running budget. That richer store, she thinks, is the next five years of grocery shopping.

Simo spent years on Facebook ads, so Guo asks for predictions on AI and advertising. She is bullish. Personalization heads toward the perfect ad at the perfect moment. Creative explodes, because you can generate variations that match the right person. Advertising follows the consumer product; that was a Facebook principle, and a burst of new experiences means new placements. Ask Instacart and the ChatGPT plugin are obvious ones: ask for dairy-free snacks for kids, get organic results and sponsored ones. Intent is much clearer than a bare keyword, which is a new mine for advertisers. As for a future where an agent goes and buys the clothes, she thinks the effect on commerce is positive. People get more productive or more free time. Some shopping is entertainment, browsing clothes, guilty as charged, and AI can make that more entertaining. The rest is chores, and those are what you automate.

What has frustrated her most about Instacart is that the app today is a utility, transactional, keywords into a cart, food at the door. Food is one of the most inspiring things humans have, a way we celebrate and show love. People already spend more time watching food content than cooking it. The opening is to take a chore and make it the moment you find a recipe, get excited about the week, discover a snack that becomes the highlight. Advertising, she believes, thrives in more immersive, inspiring environments. A conversational agent does not steal that. It is more engaging than the bare-bones search engine they have now.

Guo makes the split concrete at home. About 80 percent of dinner is an optimization: who is in, protein and vegetables, dishes someone knows how to cook. She would hand that whole task to an agent. Then there is the ritual: when a startup hits $5 million in ARR she cooks the leadership team dinner, and she will browse Instacart or a Caper cart for two hours. Kids' socks and underwear in the right size should just keep arriving. Saturday matching outfits with her daughters are worth the care. The division of labor she wants is more of the first category on autopilot and more time for the second. Simo agrees.

About three years before the taping, Simo was diagnosed with a chronic illness. She knew little about healthcare and biotech and was naive enough to assume wonderful researchers were in a lab on her condition. They were not. She started seeing the industry's gaps and co-founded Metrodora with a doctor and a scientist, blending clinic and research so discoveries in these neuroimmune conditions move faster, and so the best technology hits the clinic sooner. The number that still maddens her is fourteen years: average delay from a research finding to an effect on care. When a condition is taking your joy every day, knowing something out there could help and that it will take fourteen years to show up is, in her word, maddening. The institute also wants patient data as research infrastructure, to help biotech find therapies and improve what comes out of trials. She credits Gil: when she was ramping on biotech he told her to hire a Stanford graduate to teach her. These days, they joke, you would just ask GPT.

In the clinic they already use AI on clinical genomics, pathology, and interrogating the immune system, diagnostic work where the data is vast and a model can see what a human would miss. On drug development she is more measured. The promise is faster compound discovery and better trials. Recursion and other AI-native biotechs have been at it for years; the results are still a question, and everything in biotech takes time. Her five-year scoreboard is not just more AI-touched drugs on the market, which is already starting, but higher trial success than the rest of the industry, and drugs that help a greater share of the intended patients than the ones that pass a trial and then work for 20 percent of the audience. It is too early to know. What makes her angry is the doomer story about robots coming to kill us and the demand to stop progress just in case. She does not believe those people are living with atrocious chronic illness. The positive is obvious; the next five years of discovery will be better than the last five. Mitigate harm, yes. Technology does bad things, always. The job is to maximize the good and minimize the bad, not to stop the good, because a lot of sick people are going to rely on AI to get cured in the next ten years.

Gil, who says that in the long run he is a doomer, thinks the short-run upside for health, education, and global equity is so large that calls to regulate and distort the market are premature. People have dropped the positives and kept the fear. The nuclear comparison is the one that bothers him. The nuclear regulator established in the 1970s, as he tells it, prevented any new nuclear plant design from being approved for fifty years. Is that really the template? What healthcare still lacks, Simo thinks, is not the technology so much as adoption in real clinical workflows. Doctors are buried in administration and patients; most do not read the latest papers. AI can summarize research so a physician stays current, if you embed it as an enhancer they can trust. Accuracy and sources matter. Doctors have used Google for a long time, she notes, and Google was never fully curated. The problems are old. The new tools amplify them.