Entrepreneurship

The Synthesis Skill: What AI Still Can’t Do

· 25 min read

Another few billion dollars went into making AI agents reliable this quarter. Patronus raised $50 million to build digital worlds that stress-test agents. Bespoke Labs raised $40 million to make them dependable through reinforcement learning. In the first half of 2026, more than $510 billion poured into startups, and AI accounted for close to half of it. A new model shipped every few weeks on top of all that.

Look at what the money actually buys. Almost every dollar is aimed at two things: helping AI gather information faster, and helping it produce output faster. Find it, search it, retrieve it, read the whole codebase. Then draft it, code it, render it, ship it. The two ends of the work are getting cheaper and more dependable by the month.

There is a third thing sitting in the middle that none of that money can buy, because nobody has worked out how to automate it. It is the part where you hold ten unrelated things in your head at once and see the single connection that matters. Call it synthesis. As gathering and producing fall toward free, synthesis quietly becomes the whole job.

I run two companies where AI already does most of the gathering and most of the producing. The thing I have learned watching my own days is that the bottleneck moved. It is not the information, I can get an infinite amount. It is not the output, I can generate an infinite amount of that too. It is the judgment about what all of it means together, and what to do next because of it. This is the map of synthesis: what it is, why AI is built badly for it, why it gets more valuable every single time a model improves, and how to build it in yourself on purpose.

Table of Contents

The Bottleneck Nobody Priced In

Start with where your hours used to go. Before AI, a knowledge worker spent roughly a fifth of the week just looking for information. McKinsey put it near 20 percent. IDC measured it higher, close to 2.5 hours a day, almost 30 percent of the workday, spent searching, requesting, waiting, and checking that what you found was still current. That whole layer, the gathering, is the first thing AI eats.

So the natural hope is that AI clears the grunt work and drops you straight into the good part, the thinking. That is half right. It does remove the gathering. What it also does is flood you. When gathering costs nothing, you do not end up with less to make sense of. You end up with far more. Ten browser tabs became a hundred. One summary became forty. The research assistant that reads everything hands you everything.

Here is the trap in that. The gathering got cheap and the producing got cheap, but the step in the middle, deciding what all of it means and what you are going to do about it, got harder, not easier. There is more to reconcile than before, and a fluent, confident AI answer is always sitting right there, ready to be accepted in place of your own conclusion. The friction moved from finding things to making sense of them.

The stakes are simple once you see it. The founder who wins is not the one with the most information, because everyone now has an infinite amount. It is not the one who ships the most output, because everyone can generate that too. It is the one who can look at a messy pile of inputs and pull out the single non-obvious point that changes the decision. That act has a name, synthesis, and almost nobody trains it, because it looks like nothing. It looks like a person staring out a window.

The most common mistake I see, in myself included, is confusing gathering with progress. Forty open tabs feels like work. A document full of notes feels like thinking. It is not. The research tab is not thinking. You can spend a whole day collecting and end it having synthesized nothing, and the danger is worse now precisely because collecting has never been easier.

The Three Layers of Thinking

Break any piece of real knowledge work into three layers.

The first is gather. Find and collect the raw inputs: the search, the reading, the retrieval, the data pull, the interviews, the codebase you have to load into your head. The second is produce. Turn a conclusion into an artifact: the document, the code, the deck, the email, the shipped decision. The third sits between them. Synthesize: take everything you gathered and combine it into a point of view that none of the individual inputs contained on its own.

AI is collapsing the cost of the top and the bottom toward zero. Gathering is retrieval, search, and agents that read everything, and it is getting cheaper and more reliable every month. Producing is drafting, coding, and rendering, and the same thing is happening there. That is what all the funding in the opening actually buys: cheaper, more dependable gathering and producing. The middle layer stays stubbornly, expensively human.

The Three Layers of ThinkingAI is automating the top and bottom. The value moves to the middle.GATHERfind, search, retrieve, read everythingAI does thiscost falling to zeroSYNTHESIZEconnect, decide what it means, commit to a viewthe whole job nowStays humanvalue risingPRODUCEdraft, code, render, shipAI does thiscost falling to zerovaluemigrateshereWhen two of the three layers fall to near zero,the value does not vanish. It concentrates in the one that stays scarce.

That last point is the whole argument, so sit with it. When two of the three inputs to a process get cheap, the value does not disappear. It migrates to the third, the one that stays scarce. Cheaper gathering and cheaper producing make synthesis more valuable, not less, because synthesis is now the only step that separates one founder’s output from another’s. Everyone feeds the same models the same prompts and gets back the same fluent drafts. The only difference left is what you decided those inputs meant, and that decision is the product.

This is not the old line about information wanting to be free. Information is already free, and worse than free, it is overwhelming. The new reality is sharper. Gathering is free, producing is nearly free, so the connection between them is the entire value you add. If you cannot do the middle layer, you are a very fast pipe between a search engine and a text generator, and so is everyone else.

Dimension Gathering Synthesis
What it looks like Tabs, notes, summaries, a pile that grows One line that none of the sources said
Direction of the work Accumulates, gets longer Compresses, gets shorter and sharper
Who is better at it now AI, easily You, still
What it feels like Productive, visible, busy Slow, quiet, like doing nothing
Does it compound No, the pile is disposable Yes, it becomes a model you reuse

Why AI Is Worst Exactly Where It Matters Most

It would be one thing if AI were slightly behind on synthesis and catching up fast. The more useful truth is that the way these models are built makes synthesis their weakest move, and three well-documented failures show why.

The first is called lost in the middle. Hand a model a large pile of documents and ask it to reason across them, and its accuracy follows a U-shape. It uses what sits at the beginning and the end of the context and loses the material in the middle. Researchers have measured this repeatedly through 2024 and 2025, and it holds even for models built for long context. Performance gets worse when the two facts you need to connect are far apart, and better only when they sit next to each other. That is a retriever with position biases, not a synthesizer. Real synthesis is exactly the case it struggles with: the fact on page two that only matters once you connect it to the fact on page ninety.

The second is mode collapse. Ask a model for ideas and any single answer can look fresh. Run it a hundred times and the outputs converge into a narrow band. One study found that after 500 samples about half the ideas were still non-repetitive, but across the next 1,500 the yield of genuinely new ideas fell to a fraction, ending near 12 percent. Independent sessions circle the same themes. Multiple models, asked the same thing, produce strikingly similar sets. The models regress toward the consensus of their training data, which is the average of what has already been said. Synthesis that matters is the opposite of the average. It is the atypical connection nobody else made.

The third is that these systems optimize for fluent, not for true or new. An AI draft reads well and says the expected thing, because saying the expected thing well is what it was trained to do. Studies of AI-assisted writing find the work converging in structure, the same shapes and moves repeated. Fluent is not the same as correct, and the consensus view stated smoothly is not the same as an insight. When the reward is a plausible next word, you get plausibility, not a point of view.

Underneath all three sits one reason. Synthesis is a judgment act, not a retrieval act. It requires deciding what matters, which is a question of values. It requires holding two things that do not fit and staying in that discomfort long enough for the connection to appear, which is a tolerance for ambiguity. And it requires staking a position you will be accountable for, which needs skin in the game the model does not have. Models retrieve and remix better than any human alive. They do not decide what is worth caring about, because nothing is at stake for them. That gap is not a bug they will patch next quarter. It is the shape of the tool.

What Synthesis Actually Is, and What It Is Not

The cleanest way to define synthesis is as a ladder, because the confusion almost always comes from mistaking a lower rung for a higher one.

The Synthesis LadderAI climbs the first two rungs for free. You climb from Connect to Commit.Collectgather the inputsAI: excellentSummarizecompress faithfullyAI: excellentthe AI plateauConnectlink two unlike thingsAI: weakFramebuild a reusable modelAI: poorCommita thesis that predictsAI: cannot own it

The bottom rung is collect. Gather the inputs. AI is excellent at it. The second rung is summarize. Compress the inputs faithfully into a shorter version that keeps the meaning. AI is excellent at this too, and this is where most people, and most tools, stop and call it thinking.

Above the second rung there is a plateau, and this is where AI stalls. The third rung is connect: link two things that do not obviously belong together, the fact from your industry and the pattern from a different one. AI does this weakly and inconsistently, because the useful connection is by definition atypical and the model pulls toward the typical. The fourth rung is frame: turn a set of connections into a reusable structure, a model that organizes future problems, not just this one. AI can dress up a frame you hand it, but it rarely originates one that holds. The top rung is commit: take a position you will stake a decision on, one that predicts what happens next and that you will be judged by. AI cannot own this, because ownership requires consequences, and nothing happens to a model when it is wrong.

Rung What you produce Can AI do it The tell you are stuck below it
Collect A pile of sources Yes, easily You have tabs, not a view
Summarize A shorter, faithful version Yes, easily Your output restates the inputs
Connect A link across two domains Weakly Nothing surprised you
Frame A reusable model Poorly It only explains this one case
Commit A thesis you will act on No You will not bet on it

The single fastest test of whether you synthesized or just gathered is length. Gathering accumulates, so its output is longer than any one input. Synthesis compresses, so its output is shorter than the inputs and contains something none of them said. If the thing you produced is a longer version of what you read, you gathered. If it is a shorter thing that predicts, you synthesized.

Cognitive science has been circling this for decades under a different name. Study after study on how experts differ from novices finds that the gap is not how many facts a person holds. It is how those facts are organized. Experts store knowledge in schemas and see the deep structure of a problem where a novice sees surface details. The classic result is chess: masters remember a real game position at a glance but do no better than beginners on a random one, because they are storing the pattern, not the pieces. A schema is stored synthesis. What we call knowing a field deeply is really holding a dense web of connections, and connections are built by synthesizing, not by gathering. You do not download a schema. You make one.

The Same Report, Two Founders

Make it concrete. Two founders in the same market get the identical output from the same AI research agent, a clean twenty-page brief on why their category loses customers. Founder one reads it, nods, and asks the model to turn it into a slide deck and a list of feature ideas. The deck is sharp. The features are the obvious ones the brief implied, the same ones every competitor reading a similar brief will also build. He gathered and produced at high speed, and what he shipped was the consensus.

Founder two reads the same brief and stops on one line buried on page eleven: the customers who leave almost all cancel within nine days, before they have invited a single teammate. She connects that to something outside the brief entirely, a half-remembered idea about how habit products die when a user never crosses a social threshold. That connection is nowhere in the report. It reframes churn from a missing-feature problem into a first-collaborator problem, which points at a completely different roadmap than the obvious features do. She commits to it, rebuilds the next quarter around a single activation metric, and starts building something her competitor, holding the very same brief, will not think to build.

Same inputs. Same tools. The same fluent AI in the loop. The entire difference is one atypical connection and the decision to bet on it, which is to say the entire difference is synthesis. This is what the abstract argument looks like on an ordinary Tuesday. The report was the cheap part. The jump from nine days to a pattern in a different field was the product, and no amount of better retrieval or faster drafting would have produced it, because it required leaving the document and staking a claim.

Notice the three moves founder two made that the model structurally would not. She weighted one line on page eleven above the nineteen pages around it, a judgment about what mattered. She reached outside the given context for the connecting idea, the move a system trained on the middle of the distribution is built to avoid. And she put a quarter and real money behind it, which a tool with nothing at stake cannot do. Deciding what matters, reaching outside, and committing are the top three rungs of the ladder, and they are exactly the rungs that all the funding in the opening does not buy.

The Synthesis Premium: Why It Appreciates as AI Improves

Now the part that runs against the mood of every AI headline. Each improvement in these models makes synthesis more valuable, not less.

The mechanism is the same one that shows up on the cost side of an AI business, only here it plays out on the skill side. AI improves the complements to synthesis, the gathering and the producing, and leaves the substitute for synthesis, which is nothing, untouched. When the cheap inputs get cheaper, the scarce complement is where the value collects. Gathering and drafting race toward free, so the one step that stays hard, deciding what it all means, is where the margin goes. The better the models get at everything around synthesis, the more a person who can actually synthesize stands out, because the floor rose for everyone and the ceiling did not.

The receipts here are older than AI and they are specific. Brian Uzzi and his colleagues studied 17.9 million scientific papers across five decades. The highest-impact papers were not the most novel ones. They were papers built on a mostly conventional base of prior work, with an intrusion of unusual combinations. That precise mix, deeply familiar with a jolt of the unexpected, was about twice as likely to become a runaway hit. Synthesis has a shape, and it is a shape the models fight against. Mode collapse pulls a model toward the conventional and skips the atypical jolt, which is the exact ingredient the research says produces breakthroughs.

The market has already noticed. The World Economic Forum’s 2025 skills survey put analytical thinking as the number one core skill employers want, named by seven of ten companies, with creative thinking close behind at fourth, and both projected to keep rising in importance through 2030, in the same report that calls AI the fastest-growing skill of all. Read that together and it is not a contradiction. As AI spreads, the human skills it prices up are the synthesis skills, the ones that decide what the AI’s output is for.

There is a reason cross-domain range keeps producing the breakthroughs. In complex, unpredictable fields, the people who make connections across domains tend to beat the pure specialists. Kepler worked out the laws of planetary motion by borrowing analogies from music and geometry, casting outside eyes on a problem his peers accepted as it was. Studies of real laboratories found that the teams with the most diverse backgrounds made the most analogies to other fields, and made the most breakthroughs. Innovation clusters at the intersection of disciplines, and the intersection is a synthesis move. It is the one place a model trained on the consensus of everything is least able to go, because it lives at the center of the distribution and the connection you want is out at the edge.

The line I keep coming back to is this. AI made the answer cheap. It made the question, and the connection, expensive.

Where Great Synthesis Comes From

Synthesis sounds mystical when people describe it, which is why nobody trains it. It is not mystical. It has sources you can control.

The first is range of inputs. You can only connect what you have loaded into your head. The atypical combination that Uzzi’s research rewards requires atypical raw material, which means reading wider than your field, on purpose. This is also why the model connects like the median: it was trained on the median of everything, so it reaches for the average association. Your strange, personal, cross-domain reading list is the edge, because it lets you reach for a connection the average never would. The small team that out-thinks a large one often wins here, because it can hold more of the context in fewer heads and connect across it without a meeting.

The second is time under the problem. A connection needs the inputs to sit together long enough to collide. The fluent AI answer short-circuits that, because it hands you a resolution before the collision has a chance to happen, and you take it, because it is right there and it sounds right. Protecting a little incubation time, a walk, a night, a day with the question open, is not indulgence. It is the physical process by which synthesis occurs.

The third is forcing the connection out loud. Make yourself state, in one sentence, what two unrelated things have in common. This is a trainable move, not a gift. Charlie Munger built an entire investing career on what he called a latticework of mental models, deliberately running every problem through the big ideas of many disciplines until an unusual one fit. That is a synthesis machine you can operate on command.

The fourth is writing, and this is the one founders skip at their own cost. Writing is the instrument of synthesis, because you cannot write a coherent point without first deciding what the inputs mean. The blank page forces the middle layer. This is the real danger in letting AI write your first draft: if the model writes it, the model did your synthesizing, and it gave you the fluent, mode-collapsed, center-of-the-distribution version of your own thought. Let AI gather the inputs and polish the final prose, but write the thinking parts yourself, because the writing is not the output of the thinking, it is the thinking.

There is a warning wrapped in all of this. If a new generation of builders never gathers or drafts by hand, they may never build the schemas that synthesis runs on, the same way an apprentice who skips the reps never forms the judgment. And if you personally hand every hard connection to the model, you are quietly running up a cognitive debt that comes due the day you need to make a call the model cannot. Guard the reps that build the connections. They are the asset.

The Contrarian Take: You Are Training the Wrong Skill

Here is what most of the AI self-improvement conversation gets backwards. Almost all of it is advice about the two layers AI already owns. Learn to prompt better. Master these twenty tools. Ship faster with this workflow. Every bit of that optimizes gathering and producing, the parts that are racing toward free anyway. You can become the best prompt engineer alive and still have nothing to say. You cannot prompt your way to a point of view.

The move that actually compounds is the opposite of the trend. Spend your learning budget on the layer AI is worst at and that gets more valuable every time AI improves. Get better at connecting, framing, and committing. A tool trick dies with the next model update. A synthesis habit appreciates with it. If you have an hour a week to get better at your work, do not spend it learning a fifth AI tool. Spend it building the one skill no tool replaces.

I want to argue the other side honestly, because the strong version of this idea has to survive its objections. Gathering still matters. Synthesis without inputs is just opinion, confident and empty. The person who reads nothing and announces that they connect the dots is connecting dots that are not there, and there are a lot of those people, and AI makes it easier to sound like one. The point is not to stop gathering. It is to stop treating gathering as the finish line when it is the starting line.

And AI is a real synthesis amplifier when you drive it instead of letting it drive. It can surface a connection you would have missed, argue against your thesis until you find the weak joint, and play the hostile reader who tears up your logic. Used that way it climbs the ladder with you. The mistake is not using AI for synthesis. The mistake is letting it originate the thesis instead of pressure-testing yours. Hand it the gathering and the attack. Keep the deciding.

One more concession. The low end of synthesis is starting to automate. A model can now make the obvious connection on well-trodden ground, the third rung on familiar terrain. That is fine, and it was always going to happen. The defensible ground simply moves up the ladder, toward the framing and the commitment, which is where the real value lived the whole time. Chasing AI down the rungs it can already climb is the losing game. Climbing above it is the point.

What to Do Monday Morning

Synthesis is a muscle, which means it responds to reps. Here is the loop, and then the specific reps that build it.

The Synthesis LoopA repeatable practice for building the middle-layer muscle.CollectwideForce theconnectionCompress toone sentenceTest on afresh caseBank thepatternrepeat, and the banked patterns become your schema library

The reps are small and you can start every one of them tomorrow.

Rep What to do What it builds
The one-sentence rule After any research, close the tabs and write one sentence: what does this mean, what do I do differently Forces the jump from summary to view
Force-connect weekly Take a problem you are working on and one idea from a different field, write the analogy in a paragraph The cross-domain connect muscle
Write your own draft On anything carrying a decision, write the first draft yourself, let AI gather and polish Keeps the thinking rung by hand
Argue the opposite Have AI make the strongest case against your conclusion, then decide anyway Holding contradictions, calibrated commit
Keep a connections file When two ideas click, write the pattern in one line and save it A compounding library of stored synthesis

One structural change is worth more than all five reps together: put synthesis on the calendar. The reason the middle layer gets skipped is that it never gets a time slot, while gathering and producing fill the whole day on their own. Block one hour, twice a week, with no inputs allowed in, no new tabs, no fresh prompts, just the open question and what you already have. It will feel unproductive, because synthesis always does, right until the sentence arrives that reorganizes everything around it. Treat that hour as the highest-value time on your calendar, because in an age of infinite gathering and infinite producing, it is.

A real example from my own week, because this is not theory for me. I was stuck on why one of our support agents kept falling apart in production. The AI could summarize the failure logs forever, and it did, cleanly, and none of the summaries moved me forward. The unlock came from something unrelated I had been reading about aviation checklists, the idea that the point of a checklist is to make the non-negotiable steps impossible to skip under pressure. The connection between that and the agent’s workflow, which was skipping a verification step whenever the input got messy, fixed the problem in an afternoon. The model held every fact I needed. It could not make the jump to a field it was not looking at. That jump was the whole job, and it took me, plus a book about something else.

The reframe to carry out of here is short. You already have infinite gathering and infinite producing, sitting in a browser tab, waiting. The only scarce input left in your work is your judgment about what all of it means. That is the part of the founder job that does not automate, the cousin of taste and the engine under real judgment. Spend your hours in the middle layer. It is the one that pays.

FAQ: The Synthesis Skill

What is the synthesis skill?
Synthesis is the skill of combining many separate inputs into a single point of view that none of the inputs contained on its own. It sits between gathering information and producing an artifact, and it is the step where you decide what everything means and what to do about it. As AI drives the cost of gathering and producing toward zero, synthesis becomes the part of knowledge work that carries the value, because it is the one step AI cannot yet do well.

Why can’t AI do synthesis well?
Three reasons, all documented. Models lose information in the middle of long inputs, so their accuracy at connecting distant facts follows a U-shape rather than staying flat. They suffer mode collapse, converging across many attempts toward a narrow, conventional set of outputs, which is the opposite of the atypical connection good synthesis needs. And they optimize for fluent text over true or novel insight. Underneath all three, synthesis is a judgment act that requires deciding what matters and staking a position, and a model has nothing at stake.

Isn’t synthesis just a fancy word for summarizing?
No, and the difference is the whole point. A summary is shorter than the inputs and faithful to them, it says less of the same thing. A synthesis is shorter than the inputs and contains something none of them said, it says a new thing. The fastest test is length and novelty together: if your output is a longer version of what you read, you gathered. If it is a shorter thing that predicts something and that you would bet on, you synthesized.

How do I get better at synthesis?
Read wider than your field so you have atypical material to connect. Protect a little incubation time so inputs can collide instead of being resolved instantly by an AI answer. Force one-sentence connections between unlike things as a daily habit. Write your own first drafts on anything that carries a decision, because writing is where the synthesizing happens. And keep a connections file, so the patterns you find compound into a personal library of stored judgment.

Does using AI make me worse at synthesis?
It can, if you let it do the middle layer for you. Offloading the connecting and deciding to a model means you never build the mental schemas that synthesis runs on, which is how cognitive debt accumulates. The fix is not to avoid AI, it is to point it at the right layers. Use it to gather the inputs and polish the output, and keep the connecting, framing, and committing by hand. That way AI amplifies your synthesis instead of replacing the reps that create it.

Why does synthesis get more valuable as AI improves?
Because AI cheapens the things around synthesis and leaves synthesis itself scarce. When gathering and producing race toward free, the one step that stays hard is where the value collects, the same way a scarce input captures the margin when its complements get cheap. Everyone now has the same infinite inputs and the same fluent drafts, so the only edge left is what you decided those inputs meant. The World Economic Forum already ranks analytical and creative thinking as top skills that keep rising in demand even as AI spreads.

What is the fastest way to tell gathering from synthesis in my own work?
Ask whether AI could have produced your output from the same sources. If a model, handed your inputs, would have written roughly what you wrote, you gathered and summarized, you did not synthesize. Real synthesis contains a connection, a frame, or a commitment that is not derivable from the inputs alone, the atypical jump that the model, pulled toward the average, does not make. If it surprised you to write it, it was probably synthesis.