ExpertsFMThe grief, loneliness, and burnout sweeping through the tech industry right now | Molly Graham
Lenny's Podcast: Product | Career | Growth · September 27, 2026
5 Selected Lessons:
heuristic · 3 reasoning_pattern · 1 framework · 1
Caveated Legos: some Legos shouldn't be given away
Source
Lenny's Podcast: Product | Career | Growth — The grief, loneliness, and burnout sweeping through the tech industry right now | Molly Graham (September 27, 2026) · at 0:00 ▸ · at 59:38 ▸ · at 1:03:04 ▸ · at 1:05:29 ▸ · at 1:29:49 ▸
When this applies
when deciding which responsibilities to automate/outsource and which to deliberately retain as uniquely human
The takeaway
Reserve delegation for tasks whose quality bar is well-understood and verifiable; keep tasks requiring judgment, trust, or undefined 'good' with the accountable human.
Why it works
Tasks requiring judgment, trust, taste, or a definition-of-quality that the delegator themselves can't yet articulate cannot be safely handed to an unproven, non-accountable agent, because there's no way to verify correctness.
One level up
The boundary of what to automate should be drawn by capability-plus-accountability fit, not by what's technically possible to offload.
Where it stops holding
This stops working when AI reaches a reliability/trust level in a domain such that humans can define and verify 'good' from its output, making the domain safe to delegate.
Detail
procedure:
- Ask: does this task require judgment, trust, or a definition of quality you can't yet articulate? If yes, keep it human.
- Ask: is this something AI will do a genuinely bad job at or doesn't deserve to own (e.g., setting strategy, vision, relationships)? Keep it.
- For everything else, delegate/amplify via AI as before.
benchmarks: {"example":"CEO outsourcing a strategy memo wholesale to AI and copy-pasting it = bad; CEO should own strategy judgment"}
rationale: My fervent hope for AI is that it lets us do less of what we shouldn't have been doing anyway and more of what humans are uniquely suited to do.
In their words
In AI land, I think we got to caveat this. Like, I think there are some Legos that shouldn't be given away. Um, that the robots, like, both don't deserve and like that we should hold on to.
Consider assuming your job evolves rather than disappears
Source
Lenny's Podcast: Product | Career | Growth — The grief, loneliness, and burnout sweeping through the tech industry right now | Molly Graham (September 27, 2026) · at 25:36 ▸ · at 51:53 ▸ · at 1:27:25 ▸
When this applies
when any role is threatened by a powerful new tool and workers must decide how to respond
The takeaway
Ask 'what would I do if I believed my role will always exist but look totally different' rather than 'will my role disappear.'
Why it works
Molly suggests that believing your profession will persist but transform may shift your behavior from protective hoarding toward active participation in shaping its next form, which she argues is more likely to produce good outcomes.
One level up
The question you ask about an uncertain future determines the stance you take toward shaping it.
Where it stops holding
This stops working when there is hard evidence a role is genuinely being eliminated rather than transformed, making optimistic reinvention a denial of reality.
Detail
approach: Reframe the fear question from job-loss to job-reinvention; Molly points to how a role from a 'dying' industry (journalism) diversified and, in her view, strengthened (more journalists now work via outlets like Substack).
signals:
- a 30-year career across multiple format shifts (BBC, audio, crypto podcast, TED Radio Hour)
- journalism was repeatedly declared dead, yet Molly says demand for engineers/PMs/journalists has grown
rationale: Molly's view is that no current data shows AI is net destroying jobs in tech, and that many layoffs attributed to AI may instead reflect badly-run companies using an AI label.
In their words
What would you do if you believed your job was always going to exist? It was just going to look completely different every six years.
AI is a lazy intern, not a superintelligent employee
Source
Lenny's Podcast: Product | Career | Growth — The grief, loneliness, and burnout sweeping through the tech industry right now | Molly Graham (September 27, 2026) · at 35:13 ▸ · at 37:32 ▸ · at 59:38 ▸ · at 1:05:29 ▸ · at 1:29:49 ▸
When this applies
when delegating work to a new, unproven capability that produces plausible-looking but unverified output
The takeaway
Calibrate your review rigor to the actual, demonstrated reliability of the source, not to hype about its theoretical capability.
Why it works
Treating AI output as if from a top performer removes the natural instinct to review and iterate, because people wrongly assume competence implies correctness; treating it as an intern restores the habit of coaching, context-giving, and editing before shipping.
One level up
Mental models of a tool's competence directly determine how much accountability you retain over its output.
Where it stops holding
This stops working when AI output quality genuinely reaches senior-level reliability in a given domain, at which point treating it as an intern becomes needless overhead.
Detail
rule: Approach AI output the way you'd approach a summer intern's draft: give it context, iterate, edit, and never copy-paste-send.
threshold: Would you forward an intern's first draft straight to your boss unedited? If not, don't do it with AI output either.
rationale: People are 'shipping accountability' — assuming 'AI said it's good enough' absolves them of review, which just pushes the burden of judgment onto whoever receives the work next.
In their words
I always think like AI needs the same coaching and training that a human does, right? It needs context, it needs onboarding, it needs all of these things. And it's kind of much more like a, like an intern, like a bad intern often. I call it a lazy intern.
Productivity vs. efficiency distinction
Source
Lenny's Podcast: Product | Career | Growth — The grief, loneliness, and burnout sweeping through the tech industry right now | Molly Graham (September 27, 2026) · at 37:32 ▸
When this applies
when any organization optimizes for a proxy output metric instead of the outcome that metric was meant to represent
The takeaway
Distinguish generating more output (productivity) from generating more value that moves things forward (efficiency), and manage toward the latter.
Why it works
Volume metrics (tokens, lines of code, output count) can rise sharply while the downstream cost of reviewing, fixing, and rewriting that output also rises, so total value delivered can stay flat or fall even as the proxy metric looks great.
One level up
Any metric that's easy to game becomes a poor proxy for the goal it was meant to track.
Where it stops holding
This stops working when output volume and true value genuinely move together, i.e., when the work being measured has little rework or review cost.
Detail
rule: Watch for metrics like 'counting cars in a parking lot, Slack bubbles, or lines of code' as historically bad proxies for real productivity; the same trap applies to token counts under AI.
threshold: One engineering report cited: lines of code requiring rewrite rose roughly 8x even as raw productivity metrics rose, alongside more security incidents.
rationale: Running fast on a treadmill isn't the same as going anywhere; judgment and thinking quality were observed declining even as raw output rose.
In their words
I thought there was this really interesting engineering report that said, you know, engineer, like for engineering, it has genuinely made people more productive. But the amount of lines of code that have had to be rewritten has gone up like 8x.
Beginner advantage from not knowing old walls
Source
Lenny's Podcast: Product | Career | Growth — The grief, loneliness, and burnout sweeping through the tech industry right now | Molly Graham (September 27, 2026) · at 55:17 ▸ · at 1:09:14 ▸ · at 1:13:37 ▸
When this applies
when a newcomer lacks the institutional assumptions that constrain how insiders think about what's possible
The takeaway
Not knowing the old boundaries of a role or process can be an advantage when the tools underlying those boundaries have fundamentally changed.
Why it works
Veterans unconsciously treat historical constraints as permanent rules; newcomers unburdened by that history default to asking 'what if I just did it,' letting them exploit new capabilities insiders don't even consider.
One level up
Structural rules outlive their original justification long after the constraint that created them has disappeared; fresh eyes expose this faster than experience does.
Where it stops holding
This stops working when the old boundary still reflects a real constraint (e.g., regulatory requirements in finance) rather than a historical artifact, in which case naivety leads to real mistakes.
Detail
rule: Treat long-standing role boundaries as questionable defaults rather than fixed truths, especially when the underlying tooling has changed.
signals:
- a role boundary justified only by 'this is how it's always been done'
- veterans citing tradition rather than current risk/compliance reasons for a restriction
rationale: Startup teams feel happier partly because they aren't fighting the old way of doing things — they're inventing the way things could be.
In their words
There is an advantage to being able to say, what if? Like, what if I just did it? Or what, you know, and to just try.