Trying ChatGPT is not the same as using it

Rachel Woods’s origin story in this wave is timing plus a joke. She was a research data scientist at Meta, working on embeddings and ranking infrastructure for the small-business ads platform, then founded Vinebase, a venture-backed e-commerce and marketing company for small vineyards, which she sold to Corksy in 2022. Inside that product she used GPT-3 heavily and watched large language models change what a small team could ship. On November 1, 2022—the month before ChatGPT detonated—she started a TikTok account, half because the world needed to know that AI could write, half because she was going to try being a TikTok influencer. By the time she sits down with Nathan Labenz on The Cognitive Revolution, the account reaches on the order of 100,000 people a day.

The audience is not the one people assume. Ninety percent are over 25, 50 percent over 35, mostly in the United States: business leaders, professionals, founders, and engineers watching on nights and weekends, then dropping her videos into company Slack. Woods knows she lives in an AI bubble on Twitter. What has surprised her over the prior three to five months is the mainstream, especially business, explosion of interest. On January 1 she put a paid-consulting link in her bio. Within 24 hours, six people bought time. They were not asking for the latest prompts or how to log into Midjourney. They were asking what their business should do in the next wave.

Labenz, a month to the day after GPT-4’s release, feels a professional-grade reasoning leap that is barely deployed anywhere, while the public messages collide: all hype versus matching human doctors. Woods says she mostly reaches early adopters across industries, not random scrollers. Those people are sitting with boards and teams asking whether AI strategy means prompts in a customer offering or deploying agents. Lawyers, dentists, and software CEOs outside Silicon Valley are having that conversation. The most absurd call came straight off TikTok: a public pharma company hired her to sit through a two-hour board meeting as a sounding board while executives tried to define what an AI strategy even meant. She has done more than a hundred such calls since the start of the year.

After a hundred calls, Woods distills the first fork: will you use AI to improve internal operations, or to offer a new service, product, or experience to customers? Those are distinct conversations. Risk sits much higher on the customer side today, given the technology’s limits and the unresolved questions of human-in-the-loop design. She usually tells companies to start inside. They get a tighter feedback loop, a feel for the tech, and lessons that later inform anything they might ship. What would surprise most people, she says, is how serious these talks already are. “Human in the loop” is not just bubble jargon.

Labenz agrees from his own path to feeling qualified to speak in public: break work into discrete tasks, embed AI in a broader process, and look for input-output jobs that are hard now but would be easy with infinite interns. You see failure modes, you feel real value, and you develop respect for the technology. Woods’s content goal is the long game rather than raw views. If a few videos or a newsletter keep someone from panicking about their business, that is the win. Arm people so they feel confident enough to ride a wave she and Labenz both think is coming.

Neither of them claims to have it solved. Woods’s version of confidence is step-by-step: pay attention, make smart decisions, and collect small wins. Repurposing content that used to eat 30 percent of a creator’s time and now takes 2 percent is progress you can feel. Labenz tries to spend about half his working time just keeping up and still finds two-week-old writing already in need of a significant update. He can see maybe six months out with some clarity and apologizes for asking guests about 2030. He is willing to help people get current. He is reluctant to predict past 2023.

Labenz asks how ambitious the companies actually are—expense-receipt automation versus “talk to our language model before you talk to a doctor.” Woods’s first observation is almost universal: they did not have AI in the 2023 roadmap or budget. ChatGPT blew up after plans were locked, so a lot of the consulting is how to have the investment conversation at all. She tells companies to steal a page from Silicon Valley and run an internal hackathon, giving the team space to experiment, especially on operations. The sweeping redefinition of what it means to be an e-commerce or pharma company is further out and fuzzier. The near term is uglier and more useful: teams already know which slog they hate, they want the tools, and the productivity unlock is happening now and will accelerate over the next three to six months.

The hackathon’s real point, she says, is AI literacy. Most companies did not have the word in their vocabulary before ChatGPT. People have tried the product without finding a regular use; many are still on the free tier, which is a huge difference in value; and almost nobody understands data privacy or who owns prompts and outputs. She answers privacy questions on the order of 20 times a week. Adoption is still a learning problem before it is a tooling problem.

On tools, she and Labenz have converged. He had advised Athena, a friend’s thousand-plus-person executive-assistant company, to test a pile of products and assemble a Swiss Army knife; a few months later he expects the set to shrink to maybe three, or one core tool plus whatever is built into spreadsheets. Woods tells teams to be patient. Her hot take: an annual plan is one of the worst things you can commit to in this market, because it changes so fast. This month’s AI sales assistant can be next month’s leftover. The success stories she has seen are companies in a close, iterating relationship with the people building the tools. Labenz’s counter-hot-take is that ChatGPT is shaping up as the Google or iPhone of chatbots, with plugins as the hub later in the year. Woods thinks that could happen and still calls it early. A lot of people have tried ChatGPT. A lot have not had success yet because the prompt or the workflow was wrong. Regular use is easy to overestimate.

For Labenz, the core fact is the leap from GPT-3.5 to GPT-4. Up through 3.5, good results were an art: tinker, drill the instructions, add examples, think hard. With 4, it is honestly pretty easy most of the time. Athena has a standing invitation: anyone can send him a task, and within half an hour he can usually turn their inputs into a working output. He still uses a role (“you are a seasoned recruiting executive”), clear instructions, and a little markup to keep documents distinct—tricks that “just work” now—but he rarely needs examples or heavy refinement. He would bet that four out of five people who bounced off 3.5 would get a usable result on 4. He no longer thinks most of that is elite prompt engineering.

Woods does not dispute the technical jump. She pulls the camera back. The techniques he is describing are still known to a very, very small percentage of the workforce. Not for lack of interest: an e-commerce operator paying attention to AI is not spending the day wondering whether GPT-4 recognizes markdown better. Choosing 4 versus 3.5 by task difficulty is itself a new behavior, a couple of months old. Week to week, people in the bubble feel step changes—3.5 to 4, 4 to Auto-GPT. From the public’s chair, AI itself is the step change, a zero-to-one, and everyone is still figuring out what it can do.

In early January, companies were already describing agents: a company ChatGPT that finds leads on LinkedIn, personalizes them, sends email, follows up, and closes the lead if you just give it the goal. She had to say that is not quite how it works yet, then show them how to break the goal into prompts. Now agents exist, but that business owner experienced “AI arrived,” not a sequence of capability papers. Sitting between those two worlds is, for her, the fun of the job. Labenz is using her TikTok, among a curated list of about ten AI creators, as continuing education for Athena’s EAs—siphoning a slice of the entertainment budget. She does not try to overlap with Bain at the Fortune 100. Her sweet spot is startups and mid-sized companies that did not have AI in last year’s vocabulary. A non-AI marketplace at Series A, staring at bottlenecks, can think about those problems in a new way. Data privacy remains the misconception she spends the most time on.

Woods tells people, because it makes them feel better, that she has tried a ton of things and a lot of them do not work. It is not you. The tech is early. The AI Exchange runs a free newsletter plus a paid tier with community and more scaled advice. AI shows up in micro ways: jobs that were a bit too messy to automate, now thrown at GPT as a simple classifier or a glue between processes. A form collects subscribers’ business problems; scripts do natural-language theme analysis to decide what to write, then personalize outreach—“you mentioned this problem, here is new content.” Old-school NLP would have taken forever. Now it is a simple script. It feels like a hundred-hour job that took 15 minutes.

Labenz offers two modes. Copilot: you own the task and the model sits in the sidecar, autocompleting or answering. Delegation: you make the task discrete enough that the model performs it and you inspect the outputs later. Someone suggested swapping the intern analogy for a thousand aliens just arrived on Earth: what work would you actually hand them? Woods uses both modes. Content repurposing is the flagship copilot. You cannot paste TikTok into LinkedIn; the scroll mindset is different, so framing and calls to action should be too. She takes a TikTok transcript, runs a prompt she has honed over time, and gets a LinkedIn draft. She does not put it on autopilot, and she tells other people not to. Coach the model the way you would coach a person to adapt the piece, then look at the result. Sometimes that lives in ChatGPT; sometimes a Python notebook talks to the model directly.

Prompts mix description with a few similar past posts as few-shot examples, the kind of retrieval you would do with embeddings if it were fully automated. The test is brutal and practical: do you like the output, and did it save time? If not, try something else. Even so, full-time experimenters, she says, are at maybe 2 to 5 percent of the capability sitting on the shelf. They do not have agents scraping LinkedIn and emailing Labenz. She does not think that is where most people should start. The maturing of tools will lift that percentage without needing 100 percent. She will not plant a flag that ChatGPT plus a plugin marketplace becomes the only mega-app. She is watching GPT-4 write whole applications from a voice command, because if personalized software gets cheap, the sticky notes beside the laptop stop being a business model. Push the agent thought experiment a few steps—everyone’s AI turns raw thoughts into high-performing tweets, a summarizing agent reads the feed—and it gets messy fast. Radical uncertainty, Labenz says, is the starting point.

The Bing chatbot launch was, for Woods, a cold email from Microsoft: we will fly you out, we cannot tell you what it is, and you cannot tell anyone you got the email. She was new to influencer life, checked that the address said microsoft.com, decided it was her personality to go, and found a room full of AI reporters. She is generally in favor of companies putting what they are building into the market. Over a long enough horizon, the learning from launching early outweighs waiting too long. Labenz, who red-teamed the raw, unmitigated GPT-4 and then watched OpenAI ship ChatGPT on 3.5 first—jailbreaks and all, apparently to harvest that data—agrees that dropping a god-like system in secret would be worse. He still thinks Bing felt extremely rushed: a bot hostile to users, collapsing on earnest disagreements about the date, not on jailbreaks. A high-profile product that does not even call OpenAI’s moderation endpoint is, to him, unacceptable. Someone at the top should own the miss and state a standard, unless we are supposed to expect Word or Excel to start intruding on a marriage.

Woods is more forgiving. You need feedback to learn, and we are early even at knowing what it means to red-team these products. Teams know more now than they did three or four months ago. Failures that look modest in the zoomed-out view taught the market to be more cautious. Her challenge back: do bad experiences today decrease the risk of the same mistakes with much more powerful models later? Labenz is, relative to his fears, pleased with how things have gone. The night before they recorded, Sam Altman said at MIT that OpenAI is not training GPT-5, and employees piled on on Twitter. That restraint at the top changes how he reads the early-release decisions. Woods adds that a mid-size company’s prompt-injection mess educates the market less well than a giant’s public failure. She does not believe Snapchat’s My AI is the only bot that said things it was told not to say. It is the one people heard about, so anyone building something similar now knows the story.

On LLaMA and Alpaca, she leads with a disclaimer: she is not a lawyer, talk to one before commercial use of any of these models. Powerful open-source systems are coming; that is the writing on the wall. She is more afraid of mistakes made later than of learning now. She would not tell someone to drop an unlicensed model into a core product. Labenz’s unfinished framework is the same internal-versus-external split from the start of the call. A commercially licensed Llama, instruction-tuned on your 10,000 examples and boxed into a task you control, might work. A cheap ChatGPT clone on the website, built on Llama, will embarrass you quickly. She repeats the webpage-summarizer example: hang GPT-3.5 on the public site and you may believe it summarizes until customers watch it fabricate. Give the same job to employees first and the team learns, in weeks, that it does not.

Asked for tools that are not the obvious ChatGPT, Woods’s number-one recommendation is to start using ChatGPT with an automation layer such as Zapier or Make. Chatting in a box gets old; connecting it to a workflow is the unlock that makes people actually work with it. She also shouts out supermeme.ai, because the industry talks about strategy all day and it is worth using the same stack to make better memes. She was never good at memes; now she can make ones that make her laugh. Labenz’s parallel is the episode art, made in Playground AI, a capability he simply did not have before. The first gift is often a new ability, with efficiency as a side effect.

Would she take a Neuralink implant if a million people had them and thoughts could drive a computer hands-free? Put her on the short list for the early experiments. Her husband is a doctor; he would read the materials, and if he signs off, she is in. Labenz finds the question polarizing, which is why he asks it. In a world of Apple glasses that twitch into UI, maybe nobody needs a drill. In the narrow hypothetical, he would be open. He has three kids and a lot of neighborhood walks where a thought escapes before he can jot it down. That alone would be enough. Woods already wears glasses; if Robert Scoble is right, a first version of something like that is coming this year.

Hopes and fears, as wide as she can go. Hope: we work less, with an asterisk. More exciting is work that looks like what people already do on weekends and vacations—think when inspired, solve the bigger problem, instead of a week stuffed with too many things. Feeling on vacation while still driving what you care about forward sounds pretty great. If AI helped people actually work 40 hours instead of pretending to, that would be fine. Fear, and a major reason she spends her time the way she does: a future of diverse businesses and perspectives, versus one where only 1 percent of people adopt the technology. Centralized adoption is worse than broader access and continued competition. Labenz thanks her for doing her part to avoid that future.