Meet the team
When your direct reports are AIs
For much of my career I’ve been a manager/ lead. I’ve always loved the way a team can create so much more than any individual can. I’ve liked getting to know people and understanding what makes them tick. I’ve made some strong friendships.
But over time the role became increasingly administrative. And the admin pushed out the technical work. So a couple of years ago I decided to go back to being a developer. An Individual Contributor. And I’ve enjoyed it. Really enjoyed it.
But of late I’ve realised I’m back in the tech lead role. Except, this time, my team is made of AIs.
Meet the team
Let’s start with the newest member, o3-mini. They are a coding whizz, but prone to verbosity and showing off. If I need a chunk of code written quickly and reliably, they’re who I turn to. But they are not good for succinct answers.
Then there’s Claude. They are great at writing. They have a concise and clear style that requires little markups. They are supportive. I turn to them when I need advice. They’re also pretty good at coding. Not quite as fast or reliable as o3 - they need a bit more support.
Then there’s DeepSeek R1. They joined the team for a week and impressed. Unfortunately o3 joined the week after and overshadowed them. To be honest, I feel a little guilty that I’m neglecting them. But there’s only so much time…
There’s also Gemini. They’re great when I need someone who can reliably handle massive amounts of data. And they’ve got unique skills - they can take an audio file and generate a transcript. Or summarize a YouTube video. Or review an mp3 file.
Others make occasional appearances. I’ll reach out to o1 or Deep Research if I need a detailed analysis of code, or a technical area, or a market. They are slow and need careful briefing. But get it right and they’ll impress.
My team lead role
Many things I do are familiar to any lead:
I assign work and brief each team member.
I review their output. If they are struggling I change the brief or bring in additional help, or suggest different approaches.
I get to know them. What are their skills? How can I best make use of their abilities?
But it’s also different. There are things I no longer do:
I don’t provide feedback. These team members, while brilliant in their own ways, are unable to learn. So all feedback is directed at me. I’m the only one on the team who can learn, so if I want to improve the team it’s down to me.
There are no personal issues to resolve or career guidance to provide.
There aren’t any team-building or morale events.
And while the relationship is friendly, there’s no long term friendship. It’s only a matter of time before each team member disappears, replaced by someone superior. I’ve got no control over that. I lost o1-preview back in December when they were replaced by o1. And earlier Claude 3.5 was replaced by 3.6.
But, for now, there are lots of new candidates vying to join the team. Gemini 2.0 (released earlier this week) is a strong candidate. There’s Qwen. There’ll be a new Llama soon. A big part of my role is interviewing the candidates. They all come with impressive resumes. But I need to spend time with them to understand what they can actually do. How they’d fit into my team.
My home team
I’ve also got a home team now too. This is new. Claude takes the lead on this team - they plan holidays, help me answer emails, provide parenting support & advice, answer medical questions. They are joined by o3-mini who writes scripts and apps - nowadays I often find it quicker to build a custom tool rather than try to find and configure an existing one. Plus I don’t have to install software from unknown sources.
GPT4o makes an appearance here as well - but only because it’s the only model that provides real-time voice. Otherwise there’s little reason to use GPT4o these days.
The longer term
I like to think of skills in three camps (GDI):
General intelligence. The ability to learn, to solve new & novel problems.
Domain knowledge. Specialist knowledge and skills you have in a specific area.
Interpersonal skills. Your ability to work with people, to influence, persuade, support.
It’s quite clear that my new team outclass me on D. They are better, faster coders than I am. That battle is lost. There is nothing to be gained investing time in improving my coding; the rate of improvement in AI will outpace anything I can hope to achieve.
But, for now, I retain a lead in G and I. Although AI is advancing in both. I’ve found myself thinking about how we humans futureproof ourselves. And I think our strategy needs three planks:
Keep up-to-date with AI. Understand what the tech is capable of. How to use it. Where it is going.
Focus on developing skills that are future proof. Spend time coding with o3 or Cursor + Claude. Explore the models for your use cases. Conversely learning a new software language, or practicing leetcode - or taking a year out to find yourself - are poor uses of time right now.
Build your team. If you don’t already have a team then you need one.
Building your team
Who should be on your team? There’s not really one size fits all right now. But Claude comes close. It’s great for writing & advice. It has a very nice manner. I can’t quite put my finger on why I like it - but I know many people who feel like me. Pay the $20 a month and use it avidly.
Disadvantages? Its knowledge cut-off is April 2024 and it’s not connected to the web. So it doesn’t know about anything that’s happened since then - and can’t use the web to find out. But it will sometimes forget this and claim it knows about more recent events - and offer to search the web. For now you need to filter those hallucinations.
Gemini is also strong. I’ve spent a lot of time with it. It’s got a massive context window (up to 2M tokens). That’s pretty much enough for anything - even my longest sessions struggle to use more than 300k tokens (and those involve multiple MP3 files). Gemini is also very cheap and has a thinking model. Gemini 2.0 came out this week - I’ll be ‘interviewing’ it next week :).
DeepSeek-R1 is another candidate - although I’d use it via Groq rather than directly. It’s also free which makes it good for experimentation. But, if I’m honest, I prefer Claude.
Much of the world still uses GPT4o - but I’d avoid it. At this point it’s pretty outdated. Similarly I wouldn’t waste time on Microsoft Co-pilot - it’s quite far behind the leading edge (although that could change).
And so?
For me, leading an AI team feels surprisingly natural. Many of the fundamentals remain the same: understanding team members' strengths, assigning work appropriately, maintaining high standards. But there are differences. I can't develop my team members - I can only develop myself. And my team members will inevitably be replaced by superior versions, often with no warning.
The tools to build your own AI team are available now - affordable, powerful, and improving rapidly. Start small, perhaps with Claude as your first team member. Experiment, learn, and gradually expand your team. But don't wait too long - your competition certainly won't.
The future belongs to those who can effectively lead both human and AI teams. The question isn't whether to adapt, but how quickly you can master this new form of leadership.