Matt Sotebeer

Founder, The gap between AI adoption and outcomes is where I work. Writing about what I see.

[Matt Sotebeer @MattSotebeer]

I called a college baseball coach in 2000. I'd never met him. Never been to the campus.

I asked him one question.

"What does it take to play at your school?"

He didn't ask how hard I threw. Didn't ask my height. Didn't ask my pedigree.

He asked one question back. 🧵

[Matt Sotebeer @MattSotebeer]

Strong reframe. The one thing I'd add is that work composition is a better number than adoption, but it's still an activity number. It tells you what got automated, not whether the automated calls came out better than the humans' did. That second read is the one I keep chasing. When you audit where the work happens, are you measuring whether the outcome changed, or just that an agent runs it now?

[Matt Sotebeer @MattSotebeer]

@rohanpaul_ai Half right. Analytical SaaS that just surfaces the answer is over. The layer that survives is the one that shows you the tradeoffs and hands the call to someone who can own it. The model got cheap; deciding what to trust didn't.

[Matt Sotebeer @MattSotebeer]

@sijlalhussain The readiness I keep seeing is external, though. Companies are betting AI halves the service model they buy. Internally, people already use it to do the job. But the money chases that external bet, not the internal redesign that turns adoption into value. Backwards.

[Matt Sotebeer @MattSotebeer]

@a16z @insitro @DaphneKoller 1000x more data aimed at the wrong objective just fails 1000x more expensively.

[Matt Sotebeer @MattSotebeer]

The build got cheap. The judgment behind it didn't travel with it.

I watched someone hold two AI-built reports off the same data, one his team made alone, one built with more care. To him they looked identical, two finished screens. What he couldn't see was what stood behind one: the hours of getting it right, and knowing the numbers actually were. When building is nearly free, that invisible part is the whole game, and it still walks out the door at 6pm.

[Matt Sotebeer @MattSotebeer]

@0xMovez Doing the work is the easy part but how good the work done is the receipt. I’m curious to what Andrew’s receipts would be or look like.

[Matt Sotebeer @MattSotebeer]

@SahilBloom Your right, though what changed for me is the fear or concern, is actually directly related to growth. The nerves are the actual growth signal and how can you get comfortable being uncomfortable.

[Matt Sotebeer @MattSotebeer]

@chamath @Benioff @satyanadella Everyone's racing to own company knowledge. Nobody's asking who owns the employee's. That's the real cascade.

[Matt Sotebeer @MattSotebeer]

@Codie_Sanchez ya i dont think most truly understand this until you have gone all in. i never would have described myself as crazy, but its a bit insane to do and you normalize it, which is really the crazy part.

[Matt Sotebeer @MattSotebeer]

This is a really great idea. We often hold ourselves to what NEW thing we are doing but the power is what we stop doing.

[GREG ISENBERG @gregisenberg]

i keep a running doc called “things i’m not doing anymore” and look at it monthly

really simple way to live a happier & more productive life

[Matt Sotebeer @MattSotebeer]

@lukepierceops There’s not enough success or failure yet, so positioning doesn’t exist yet. Just keep going deep on a problem, and create value.

[Matt Sotebeer @MattSotebeer]

@wizofecom Most cannot consistently post each day and that’s rule number one.

[Matt Sotebeer @MattSotebeer]

The system gets better through use, but only at the parts use can see.
Our biggest agent improvements came from the opposite event: someone rejecting a confident answer and writing down why. That log lives outside anything retrieval indexes, and it's the context that compounds judgment, not just routing. How does Glean capture the rejection, not just the use?

[Matt Sotebeer @MattSotebeer]

@emollick Or just use it every day and you find the edge. Guess there’s edge both directions.

[Matt Sotebeer @MattSotebeer]

This is worth a read. The model has the archive. Your company has the present tense. Rao makes the case that this gap is structural, not an engineering backlog.

[Paul Millerd @p_millerd]

one of the best AI essays ive read
https://t.co/lUQeOVqa5K

[Matt Sotebeer @MattSotebeer]

This reminded me of when I first read 7 Habits of Highly Effective People, having recently moved to NY and not knowing anyone, so I immersed myself in books. I couldn't tell you where I was more specifically or a certain moment, but this part never left. Covey paints a picture on a subway, kids running wild, a father doing nothing, whole car judging him. Someone finally says something to the father and he tells them how they just came from the hospital, and their mother died an hour ago and he didn't want to tell them yet. Nothing about the kids changed, one fact changed and every belief in that car updated itself in a second. The update seems like the hard part. It's the missing fact that is.

[Matt Sotebeer @MattSotebeer]

Agree on the sovereignty point, but it can't stop at the enterprise. If a company absorbs every employee's workflows and judgment into the company OS and the person owns none of it, that's just Karp's extraction one level down. The lab does it to the company. The company does it to the employee. Real sovereignty is the employee owning their own OS and the company owning theirs, working together, not one eating the other. Otherwise you've become the thing you're mad at.

[Matt Sotebeer @MattSotebeer]

@BraceSproul The wiki is the product, the repo's just the first place it's obvious. Only catch: docs that auto-update from diffs can codify a wrong read and hand every agent the same confident mistake. Someone still has to own whether the wiki's right.

[Matt Sotebeer @MattSotebeer]

@alliekmiller learn what was debated to the findings. Why goes a long way because its true until its not.

[Matt Sotebeer @MattSotebeer]

@SahilBloom its also okay to be maxxed out on your current friends and not looking for more.