If you're a tech manager and can't outperform your team, your role is going away
AI collapsed the implementation gap. Orgs just flipped from compute-bound to communication-bound — and middle management is the All-to-all you optimize away.
I co-design algorithms and systems to scale LLM training. It often comes down to getting rid of communication and scheduling overheads.
Your org has the same scaling problem as a badly tuned training run.
AI didn’t make engineers 10× faster. It made the communication overhead between them the bottleneck.
AI agents collapsed the implementation gap — a leader with good judgment can now ship in hours what used to take a team weeks. The scarce resource isn’t execution anymore. It’s taste: knowing which 3 out of 10 directions will actually work, and being right 80% of the time.
One type of tech manager cultivated that taste. They stayed close to the work. They look at an architecture and see the failure mode five turns ahead. Give them AI tools and they become a one-person product team.
The other type spent years relaying context between teams. They align stakeholders, schedule syncs, manage deadlines. AI automates everything they do.
With AI, the first type gets amplified. The second gets replaced.
If you’ve done performance engineering on a distributed training run, this would be familiar to you. Every system hits one of two walls: compute-bound (each GPU is the bottleneck — add more) or communication-bound (the interconnect is the bottleneck — reduce overhead).
Pre-AI orgs were compute-bound. Individuals could only produce so much, so you hired more. More people needed more coordination, so you hired managers to be the interconnect.
Post-AI, every GPU got 10× faster. But what about the interconnect: the status meetings, the cross-team syncs, the alignment docs. If they didn’t change, your org just flipped from compute-bound to communication-bound.
In this new world, what used to be the norm will become utterly bizarre. After all, why would anyone want to spend a week on another roadmap alignment between teams when the needs can be met with a few prompts in Claude Code?
The fix isn’t “managers should learn to code with AI.” It’s that 3-person teams led by a player-coach with taste will out-ship 30-person orgs run by communication relays. VPs are investors — they stay. Middle management is the All-to-all you optimize away.
What’s your org’s work-to-meeting ratio?