I run a company, and I’ve now been using ChatGPT Plus daily for a month, GPT-4 in particular. I stand by what I said in my previous post about the mind-blowing impact GPT is going to have on my sector. The idea of this post isn’t to be alarmist this time, but to take a bit of stock of what I’ve tested and practiced, and also of the reactions I’ve seen around me and on my team.

Prompt Engineer

Right, first, hot take: creating prompts is far from simple (this is one of the points I’ll go into in the Contempt section). I now see that it could become a job in its own right.

Here’s what I’ve done with GPT-4 over the past month:

• drafting terms and conditions of sale for a SaaS product [legal]
• drafting administrative cover letters [administrative]
• drafting beginner/intermediate/experienced/expert technical tests for hiring devs [technical]
• drafting training materials [technical]
• drafting email [sales]
• drafting a new prospecting approach, with wording variations depending on parameters [sales]
• an internal mobile app* for Coddity, from A to Z (currently being deployed to the stores), in React Native (I’d never done React before) [technical]
• doing a coding game for the hell of it [technical]
• refactoring and improving some of my old code [technical]
• writing unit tests on a project that didn’t have any [technical]
• formatting data [technical]
• a new feature integrating OpenAI into one of our internal products (ginto)

As you’ll have noticed, my use of it is very varied. It has become a tool for me, like any other, built into my personal “workflow”.

The truth is, the results of these different tasks were, at the start, not great. I had to experiment a lot (but really a lot) to get to results that were effective and usable. The good news is that whether the subject is legal / admin / sales / technical, any progress you identify in crafting prompts in one domain serves you in the others. And I now get very strong results on the first try for any new task. The data in the prompt, the context you have to build, the examples you have to give: that’s what guarantees effectiveness.

My day-to-day productivity has really improved as a result.

Contempt

Now, the reactions, particularly in my sector: software development, infra, and other such joys, in which you get a recruitment pitch a day because the market is short 40% of the people it needs. I see 3 categories of people:

• Those who are convinced: they’ve used it, they’ve experimented, they understand how it works, and they fold it into their tasks to varying degrees
• Those who reject it: they’ve tried it, they haven’t experimented, and they reject it with, at a minimum, indifference, but very often contempt
• Those who… don’t give a shit: they haven’t tried it, they’re not curious, but they’ll get on it when they no longer have a choice

I think the tidal wave I imagined at the start is going to be slowed by group 2, until the individual “wow moment”. That “wow moment” comes, I think, when you take the time to experiment (individual training or external training) and to understand how to use the tool. And that’s where the difficulty of producing a good prompt comes into play (not to mention trying version 3 on the corner of a table — you see, it’s shit)

Do a code review of this for me is not a good prompt => it’ll flag points a junior would find Write me the unit test for this code is not a good prompt => it’ll write tests, but they’ll be incomplete Write an email to meet Carrefour’s CIO is not a good prompt => the email will be completely worthless, with no style other than the GPT one you see everywhere on LinkedIn

The other brake: the data. Yes, right now you have to be careful about everything put into ChatGPT or sent to the API. Still, from what I’ve heard, the salespeople at Crosoft are doing the job, going around the big groups to pitch GPT integration into VS Code or Office 365. In any case, from what I’ve read, we’re a year away from a chipset packing the power of an individual GPT-4, which will neutralize this point. When every company can run LLMs on premise for cheap, that’ll be an endgame.

Impact

GPT, and Bard for that matter, are not going to replace devs. To pick up an analogy Chris, my partner, and I made for the rest of the team: we (devs + tech) are tailors at the dawn of the ready-to-wear wave. We are going to produce fast, at a quality level that can be quite high.

And we will always need tailors for specific, exacting cases, or to create the ready-to-wear collections.

And the price of software dev? I’ll let you follow the analogy with ready-to-wear. I read a tweet (lost, of course) about something I’ve found pretty true for a month now: we’re entering the era of “disposable” dev, meaning a specific piece of development answering a very precise need, which would never have been built before because it was too expensive for that limited scope.

Training

At Coddity we warned everyone very quickly. We cover the individual OpenAI licenses ourselves, and we’ve run training workshops (on a tool that isn’t even 3 months old!). A few numbers: only 40% took the license (which we pay for, I’ll remind you) and 25% wanted to join training workshops. You can pretty easily find our 3 groups mentioned above.

On the other hand, even if the impact inside our own team isn’t what we expected, all this experimentation work lets us build two training offers:

• one specific to developers,
• one specific to business leadership: training in promptology (I love it), and integration with prototyping, with a design sprint approach (and we know that for these people the “wow moment” is going to be radical)

AD BREAK: if you don’t know how to prompt and you think getting trained is an interesting personal and collective investment, and one you can capitalize on, CALL ME.

(*) For the anecdote, the mobile app I built started out as an app I wanted to dedicate to town halls: a simple news feed to display, with data stored on Airtable. I think it was done in 1 hour (most of the problems coming from Expo and the wonderful JS ecosystem and its incompatibilities). I kept evolving it as I went: adding a calendar, read tracking, practical info. It was after that that I evolved it into my company’s app, adding the internal wiki, the internal rules, and above all OAuth2 authentication. All told I think I spent 2.5 days on it. If I’d had every feature in mind at the start of the project, 4 hours would have been enough, I think.