The AI Productivity Gap Is About to Get Ugly
The live argument on X is no longer whether AI matters. It is whether AI is creating a brutal split between operators who have rewired how they work and those still layering slop on top of old process.
The AI Productivity Gap Is About to Get Ugly
The most interesting thing on X this morning is not a product launch.
It is not a benchmark.
It is not another “look, the model can now do my shopping” thread.
The live argument is nastier than that.
Operators are starting to say, more openly, that AI is not lifting teams evenly. It is creating a split inside companies between people who have actually rebuilt how they work and people who are still doing the old job with slightly shinier software.
Marc Andreessen amplified the bluntest version of that line overnight: some 10x engineers are becoming 100x engineers, while others are still doing things by hand. Greg Brockman, from the OpenAI side of the world, is posting that you can now build things from your phone with Codex inside the ChatGPT app. Ethan Mollick is joking that top startups now seem to need “wizards”. Ben Tossell is pushing back from the other direction, mocking the useless spread of AI summaries into places nobody asked for them.
Those are not identical posts. But they are all circling the same point.
We have left the phase where “uses AI” is a meaningful category.
The real split now is between people using AI to compress the work and people using AI to decorate it.
That gap is about to get ugly.
This is not about adoption. It is about redesign.
The lazy version of the story says AI adoption is rising.
Fine. That is true and almost useless.
What matters now is not whether a company has bought licences, approved a policy, or told staff to experiment on Fridays. What matters is whether the company has changed the shape of the work itself.
That means awkward things.
Who writes the first draft now?
Who verifies it?
Which decisions still need a human in the loop because trust, liability or judgement matter?
Which layers of coordination were only tolerated because the old pace of work made them seem normal?
Which roles are genuinely strategic, and which were quietly built around information friction that no longer exists?
This is where the argument is getting sharper. AI is no longer just a tool question. It is a management question.
A team that uses AI properly does not merely produce more output. It often changes the unit of output. One operator can explore more options, write more code, test more variants, and collapse more dead time between idea and execution than the org chart was designed for.
That sounds exciting until you realise what it does to the rest of the company.
It makes the non-adopters visible.
It makes middle layers easier to question.
It makes bad process look embarrassingly expensive.
And it creates a political problem, because many organisations can handle new software more easily than they can handle new performance differentials.
The first-order effect is productivity. The second-order effect is social.
People like to talk about AI in terms of capabilities because that feels clean. The machine can do this. The machine cannot do that. Here is the demo. Here is the benchmark. Here is the graph.
Real companies do not experience technology that way.
They experience it through status, trust, pace, ego, and workload.
If one product manager can now test five ideas in the time another tests one, that is not just a productivity story. It is a status story.
If one engineer can ship in a day what used to take a small team a week, that is not just a tooling story. It is an organisational design story.
If one founder starts running the business with a small layer of AI-heavy operators and the rest of the team still works through meetings, handoffs and documentation rituals from 2023, the problem is not “adoption”.
The problem is two incompatible operating systems running inside the same company.
That is what the “100x engineer” line is really getting at. Not literal arithmetic. Not hero worship. A structural divergence in leverage.
Some people are now operating with machine support embedded into their default way of thinking. They draft with it, interrogate with it, route work through it, use it to compress search, analysis, synthesis and production. Others still treat it like a side panel. Something to try after the real work is done.
Those people are not doing the same job any more, even if HR says they are.
The market is also beginning to separate leverage from slop
This is the other reason the conversation feels alive.
The backlash is maturing alongside the enthusiasm.
Ben Tossell’s complaint about pointless AI sleep summaries matters because it captures the growing impatience with AI as filler feature theatre. Nobody serious wants “AI everywhere”. They want AI where it creates surplus.
That distinction is becoming the entire game.
There are at least two very different AI strategies showing up in the market.
The first uses AI to reduce cycle time, increase throughput, improve decision quality, or let strong operators carry more load with less friction.
The second uses AI to generate more words, more summaries, more dashboards, more notifications, more ceremony and more fake evidence of modernity.
Both strategies are being sold under the same label.
Only one of them compounds.
This is why the next wave of winners will not just be the companies with the best model access. They will be the companies ruthless enough to ask: does this remove work, or does it merely add machine-produced noise to work that should not exist?
That sounds obvious. It is not. Most companies are still far too polite to ask it.
“Everyone should use AI” is not a strategy
This is where a lot of leadership teams are bluffing.
They have understood that AI matters. They have not yet understood that a vague mandate to “use AI more” can actually make the company worse.
If the goal is unclear, AI scales confusion.
If the approval chain is bloated, AI sends more half-finished material into the bloated chain.
If nobody owns verification, AI increases the speed at which errors become institutional.
If performance is measured badly, AI helps people optimise the wrong thing faster.
This is why the interesting companies are not the ones issuing the loudest memos. They are the ones changing burdens of proof.
Before you hire, did you redesign the workflow?
Before you add headcount, did you test what one sharp operator can now do with proper machine support?
Before you expand the team, did you cut the dead coordination work that only existed because the old tooling made it unavoidable?
That is the uncomfortable part. AI makes labour more fluid. Management has not caught up.
Most companies still allocate responsibility as if work arrives in neat human-sized chunks that can be passed around departments. But AI is starting to explode those chunks. One person can now originate, analyse, draft, iterate and package work that used to be split across several roles.
That should force a redesign.
In many firms, it will instead trigger denial.
The winners will look weird before they look efficient
Mollick’s “every executive has a wizard” joke lands because the shape of the new org is still socially awkward.
It sounds unserious. A bit cultish. Slightly embarrassing.
Good.
That is often what a real transition looks like before the language settles.
The companies pulling ahead right now are not necessarily the ones with formal doctrine. They are often the ones where a handful of people have already normalised a strange but effective way of working:
One person with relentless AI support.
Fewer handoffs.
Shorter loops.
Less reverence for old role boundaries.
Much more insistence on outcome over ceremony.
From the outside, this can look chaotic, even antisocial. From the inside, it can feel like cheating.
Then the numbers start to show up.
Faster shipping.
Lower cost per useful output.
More experiments run.
Higher ratio of signal to meetings.
At that point the weirdness stops looking weird and starts looking like management competence.
This is why incumbents should be worried. Not because AI magically makes everyone brilliant, but because it makes pockets of serious competence much more dangerous.
The fair pushback
There is an obvious counterargument.
Some of this is hype. Some “100x” talk is just the usual tech exaggeration with fresh paint. Some teams are over-claiming. Some AI-heavy workers are producing speed without depth. Some of today’s stars are just very online.
All true.
The market always overstates in the first phase.
But the existence of hype does not cancel the underlying shift. In fact it often obscures it.
The question is not whether every person using AI is suddenly elite. They are not.
The question is whether the spread between the best-adapted operators and everyone else is widening.
It clearly is.
And the wider that spread gets, the more every company is forced into a choice it has spent the last two years trying to avoid.
Either redesign work around the new leverage.
Or keep the old structure and accept that your best people are now dragging an increasingly absurd amount of dead weight.
That is the real strategic pressure here. Not AGI. Not science fiction. Not robots replacing everyone by lunchtime.
A much more immediate problem:
some people inside your market are already playing a different game.
This gets especially brutal in startups
Large companies can hide for a while. They have balance sheets, brand, procurement muscle, distribution, and enough internal complexity to mask declining efficiency.
Startups do not get that luxury.
In a startup, the productivity gap hits harder because every extra layer shows up immediately in runway, velocity and founder attention.
If one three-person team can now behave like the old eight-person team, capital allocation changes.
If one founder can cover research, writing, product shaping and outbound with aggressive machine support, the expected pace changes.
If one growth operator can run far more creative tests than a traditional team, your benchmarking assumptions die.
This is also why startup discourse feels schizophrenic right now. On one side, people are drunk on leverage. On the other, people are sick of AI slop and desperate for proof that any of this creates durable value.
Both instincts are correct.
The answer is not that AI is fake.
The answer is that AI is only valuable when attached to a redesigned operating model.
That is a much higher bar than “we use AI”.
The contrarian take
The biggest near-term consequence of AI may not be mass job replacement.
It may be management humiliation.
Because AI is turning a lot of hidden inefficiency into visible contrast.
The strongest people are becoming dramatically more leveraged.
The weakest processes are becoming impossible to defend.
And the firms that spent years confusing coordination with competence are about to discover that software can expose both at once.
So no, the real divide is not AI companies versus non-AI companies.
It is companies willing to redesign work versus companies hoping the old org chart can survive with an assistant bolted onto the side.
That second group is in trouble.
Not because the models are magical.
Because the people they are competing against are finally learning how to use them properly.
That is a much more dangerous thing.
Why this now
Because within a single 6-8 hour window, the high-signal conversation on X converged on the same uncomfortable truth from different angles: some operators are compounding machine leverage into genuine output gains, while others are still shipping AI theatre, filler, or old-fashioned manual process.
Sources and searches
Marc Andreessen amplifying the productivity split: pmarca on the “10x engineers are now 100x engineers” line
Greg Brockman on mobile AI leverage: “you can just build things from your phone, with Codex in the ChatGPT app”
Ben Tossell on AI filler backlash: sleep-summary AI as feature theatre
Ethan Mollick on the startup “wizard” meme: “everybody's got to have a wizard”
Marc Andreessen on tool saturation: “That’s what the tools are for”
Search used:
site:x.com OpenAI Anthropic Shopify Stripe Vercel Marc Andreessen Tobi Lutke May 17 2026 AI productivity debateSearch used:
X/Twitter AI productivity 100x engineer manual work May 17 2026
