When Code Costs a Dime, You Throw the App Away

Nathan Labenz frames the episode around a take that is already going mainstream: the next leap is actions. Language models have to bridge a messy request in English to an action space of APIs, browsers, other software, and eventually robots, while the user's history is scattered across email, Slack, and calendars. People have imagined assistants from *The Jetsons* to *Her*. ChatGPT broke through partly because it looked like one. In AI safety, assistants are treated as a laboratory for alignment. Economically, an assistant that actually works is a virtual employee — a prize large enough that Alexa, Google Assistant, and a rumored Siri update will all show up.

Lindy, announcing the same day, is not a hackathon toy. Crivello has raised $50 million; the team had been building on GPT-4 for a month before launch. The product is still a tight beta. On a screen share, he types a request to find ten software engineers in San Francisco, watches a list appear, then asks Lindy to draft recruiting mail. None of the UI, he says, is hard-coded. The model mixes interface components at runtime. Recurring jobs are the other half of the demo: five minutes before a meeting, pull the Zoom link, the email thread, the recording, the LinkedIn, and drop a briefing in Slack. Lindy already joins his meetings, transcribes them, and can be told to update Salesforce or open Linear tasks from the transcript.

The training story is the part that matters for anyone building agents. They taught a large language model to use tools from documentation alone — feed it the Stripe API or the Slack API as a human would read it, then give it a goal. "Help me find half an hour with Nathan tomorrow" becomes a sequence: Google Calendar for free slots, Gmail to send them, Calendar again to write the hold once Nathan replies. A new user does not hand over passwords the way they would to a human assistant. OAuth, just in time, on the first request that needs a given account. Under the hood the agent writes code, and that code includes functions to call itself back: grab the Zoom recordings, grab the mail, then summon itself to summarize.

Crivello keeps reaching for the Industrial Revolution because the economic rhyme is the point. Once goods got cheap enough, single-use objects stopped being unthinkable: Bic pens, Solo cups, a Napoleon anecdote about aluminum tableware that would now be a La Croix can tossed in the living room. Software still costs on the order of ten to a hundred thousand dollars to build. It is rapidly approaching a world where an app costs on the order of a cent to a dime. When that happens, you write a program for this session and throw it away when you are done. That, he says, is what Lindy is doing every time you give it a use case.

Labenz has already felt a sliver of it. Dating the recording as GPT-4 plus seven, he has asked the base model for a chart or a bit of analysis and watched it work — without auth, without a runtime, without the recursive loops Lindy adds. The interesting split is one-off versus standing automation. Finding time tomorrow is disposable. "Before my meetings, send me the Zoom, their LinkedIns, and a summary" is a program you keep. Crivello's first jobs that already work well are the ones you would give an executive assistant: email, calendar, contracts, prospecting, recruiting. Morning inbox triage comes with drafts in your voice for that particular recipient, on the theory that you do not write to your spouse the way you write to investors.

He does not want to pick consumer versus professional as a forever split. This is, in his telling, a new kind of computing — an inverse operating system. A normal OS lives on top of hardware and underneath apps. Lindy lives on top of applications and data, patches them together, and becomes the thing you actually use. The computing experience of the future is not you doing work on a computer. It is a conversation with a computer that works with you, or does the work.

Labenz wants to know why this is the moment, and how a startup survives Alexa, Google, and Apple. Crivello has been close to AI for more than a decade, pulled in by the ImageNet-era moment when CNNs and RNNs started to work. He took Coursera courses as a software engineer, could not find cute enough use cases, almost started something around GPT-2, poked at GPT-3, and only felt the products become viable when ChatGPT made good results show up fast. When you are close to it, he says, it is obvious that something unprecedented is happening. That is why the time is now.

He tries not to spend minutes thinking about competitors. Jeff Bezos's line, as he remembers it: they are not the ones giving you money. The market will be large enough for many winners. Incumbent assistants have a structural handicap, and it is speed. Reid Hoffman described driving at night on uncertain terrain, in fog, with no one able to see far. Huge companies are even less strong in that weather. Startups can ship something slightly rough on the edges without repeating Google's Bard demo, which cost billions of dollars of market cap overnight.

The practical niche is not a secret feature. Google Assistant, Siri, and Alexa are consumer products that have to serve a universal audience from day one. Lindy is aimed at busy professionals drowning in meetings — putting order into chaos — which those assistants cannot do. Apple has FaceTime, including links in calendar invites, and Zoom is still a formidable company because it focused on the B2B case Apple structurally cannot chase. The startup advantage is the ability to focus on any slice instead of casting a net that wide, plus the freedom to change direction, run experiments, and take risks a public giant cannot afford.

Asked how Lindy chooses among GPT-4, Claude, open-source options, and in-house training, Crivello gives the team a single instruction: quality of responses, and only that. He does not want a conversation about cost in the room. Model prices, he says, have been divided by about 20 over the last year; this is not the bottleneck. Lindy currently costs them dozens of dollars per customer per month. That is fine. GPT-4 is head and shoulders above everything they have tried. It is also often surprising how much you can beat GPT-3.5 on a narrow use case by training your own model for that case.

The landscape moves so fast they had to become model-agnostic — swap and retrain quickly when a new open-source option starts working. For now they mostly run GPT-4, while fine-tuning their own models on data they have collected, which are getting better rapidly on internal benchmarks. Eventually he expects a mix, still purely for quality: the specialist catching and then passing GPT-4 on Lindy's job. Labenz recites the price history — six cents per thousand tokens down to two, then to 0.2, then GPT-4 reintroducing a higher tier, roughly a 20x gap — and Crivello's answer is to pay it. Never bet against Moore's law, and this curve is an order of magnitude faster. The question is whether chips get better faster than good gets cheaper. History, he says, says good gets cheap very fast, and cheap does not necessarily become good.

GPT-4 versus 3.5 is not a mild upgrade. The bigger context window alone killed a pile of hacks — summarize, embed, fine-tune, dance around the limit. The window is still a constraint, much less of one. Prompt engineering dropped too: 3.5 had to be coerced; GPT-4 often works with zero extra prompting, like an assistant you do not have to tell not to do X, Y, and Z. The downside is speed. Generation is slow enough that the demo pauses for a few seconds. Even a 30-second wait, which they do not actually have, would still beat a human assistant. The product is available 24/7 and answers in about ten seconds. That is already superhuman in the ways that count for this job.

A human teammate gets to know you. The default AI pattern is embed everything, retrieve at runtime, stuff the top chunks into context. Crivello's system mixes three moves, because the model needs preferences and data. Preferences were going to live in a vector store until GPT-4's 32,000-token window made them look again. How many preferences does a person actually have? Not 30 pages. They now dump them in the prompt — no meetings on Friday, nothing before 11 a.m. as focus time — and will revisit retrieval if the list ever gets too long. It works well.

Data takes two paths. A context injector sits upstream of code generation, reads the user prompt, figures out which questions must be answered and where the answers live, then hydrates the prompt. "Send an email to John" first has to decide which John, using contacts, calendar, and mail; the John you write every day wins. Separately, the code-generation model can fetch at runtime: a Google Doc linked in the request is probably a Docs API call, a summary, then an email. On top of that they build a profile in their own database from whatever sources the user is willing to connect, until the assistant knows you better than anyone. Privacy is not a slogan. Crivello notes that secretary and secret share an etymology. Lindy is not supposed to leak; the company will not sell the data or run ads. They charge for access, which lines the incentives up.

They published seven constitutional principles on the homepage, both as a public scoreboard and as something to use at every altitude: high-level strategy, product calls, and training. Labenz hears Anthropic's Constitutional AI in it — a short written standard, then the model critiquing itself against that standard. Crivello says it rhymes and runs deep. They fine-tune, they use RL from AI feedback, and the principles sit in each of those loops. Principle five, comfort with ambiguity, is the one that used to be impossible for computers and is now arriving almost for free with language models. It is also in tension with another principle: reliability. Lindy does not get to screw up.

The internal image for ambiguity is the old essay *A Message to Garcia*. A general needs a letter delivered through a war; the soldier who simply says "done" is the one you want, not the one who asks who Garcia is or where. "Send an email to John" should not produce a quiz if you meet John every day. Comfort with ambiguity still loses whenever it collides with reliability. The heuristic is two-factor: how certain are you that this is the right action, and how bad is it if you are wrong. Searching email is one thing. Sending a termination mail to the wrong person is another. A second cut is read versus write. Read-only — search, research cat-moving companies — can go. Sending mail, buying something, acting on someone else's behalf waits for a yes.

GPT-4 no longer returns log-probabilities for the top tokens, which is how you used to read the model's confidence off the API. Crivello's hypothesis is that OpenAI pulled them so competitors could not distill GPT-4 from GPT-4. It is another reason to train their own model. In the meantime they have a separate confidence model: prompt and context in, proposed action out, a score for whether the action really follows. Guardrails are not a single judge at the end. They compose agents and workflows, and the rails sit on every step. A user can define "send a contract" as DocuSign plus a template plus email, then define "onboard a vendor" as first send an NDA, and end up with a constellation of named workflows calling each other.

Does he trust it to book a flight? Quite a bit, because confirmation still sits in front of the irreversible step: Miami to SFO, a list of options, he picks number two, Lindy handles the rest. More of the world is reachable by API than people expect. Where it is not, they are building a browsing agent as fallback. The long-term surface is not a single text box and not quite *Her*. They want to be to large language models what the PC was to the CPU: packaging, tools, data, and containers around a miracle that is already usable as a prompt. Email, Slack, meetings, voice memos, eventually a phone call it answers. Erotic roleplay, Labenz asks, nodding at a prior Replika episode, is not a job they are trying to do. Jailbreaks will be an empirical question for early users.

Training their own models does open a can of worms that a frontier API vendor has already spent years on. Crivello is glad OpenAI takes safety seriously and notes that even they cannot stop jailbreaks, which happen all the time. Even a perfect OpenAI would not stop other people from training their own systems. Whether we like it or not, the models will be out there, doing basically anything a human wants. That is a broader social problem, not a product checkbox. The same week, he and Labenz note, Meta's LLaMA and Stanford's Alpaca had already shown how fast a release, a fine-tune, a "it's just like text-davinci-003," and a takedown can cycle.

Labenz points to a paper from researchers at Penn and OpenAI on language-model exposure of work. Read the charts one way or another, he takes a lower bound around 25% of all work that models could eventually take on, especially once they have the surrounding tools Lindy is building. Crivello splits time horizons. Over the next five years he is not very worried about job loss. People have been worried forever; it is a failure of imagination about infinite human wants. Free people from automated tasks and they do things only they can do. Something like 90% of the workforce were farmers in the early 1900s; today it is something like 5%, and U.S. unemployment is not a disaster. Gross job creation and destruction every quarter run to millions; the headline is a small net. Economies, he says, are more resilient, elastic, and quick to reconfigure than the anxiety assumes. He is excited about growing GDP and making people 20, 30, 50% more efficient.

Keynes's 15-hour week did not arrive on schedule. Crivello thinks the forecast was too early and too eager, not wrong. Hours worked per person in the U.S. have been falling for decades; he puts current hours on the order of 40% below a century ago. At some point diminishing returns set in and people choose more time over more stuff. That will be part of the adjustment. Labenz is hoping for leisure and for robot servants, digital first. AGI and ASI are the other horizon. All bets are off, Crivello says; no one knows.

The product is still a limited private beta, onboarding weekly, with general availability sometime this year — a window that could mean next month or nine months. Reliability is the blocker he names, and he reaches for self-driving: one fatality per million miles is a statistic you cannot read off a demo. You need the miles. They are also still learning what people want an assistant for, and they would rather crack a magical product than dump a rough one on the general population. Asked what he uses besides Lindy, he names GPT-4 for writing — ramble for 20 minutes, Whisper transcript, light style guidance, a blog post in half an hour instead of six hours — plus Whisper wrappers and a small Heroku app that embeds Y Combinator talks so you can ask how to think about hiring and get distilled advice in minutes.

Labenz has noticed that when he asks guests what they use, the answer is still mostly ChatGPT plus one or two other things, despite a daily rain of new apps. Crivello thinks a lot of AI applications are too narrow — one tool to search YC videos, hundreds of Product Hunt launches a week. Nobody wants hundreds of logins. They want one big tool. ChatGPT is the closest thing right now; Lindy is another attempt at something broad. He does not think it kills all software. Amazon sells everything and you still use a vertical when you buy a house, a car, or a dress. Lindy is Amazon in that picture: the default text field for getting things done, with specialist CRMs and the rest still in the stack when a vertical is truly specific.

A hypothetical Neuralink, thought-to-text into Lindy, does not tempt him soon. Privacy and security are too problematic; he would wait a very long time, probably ten years or more. You only have one skull. Zooming out, his hope for the next handful of years is to get people off menial work — data entry, the calendar ping-pong of finding time — work humans should not be doing because humans, as he puts it, are AGI. Robots should have those jobs. His fears are misinformation, and, over the very long term, existential risk: a low but non-zero chance that things go very wrong with AGI.

Asked for remedies, he says something he flags as uncharacteristic: we need regulation. He does not trust the private sector to self-regulate on safety. Airlines are extremely regulated. Buildings are extremely regulated when you put up a house, for safety reasons. We need to do the same thing with AI. Labenz leaves it there — an interesting few years, a waitlist, an assistant that is supposed to put a life on autopilot — and Crivello goes back to launching.