
Pinned Tweet
Frontier models weren't built for industrial work.
So we built our own.
Squint's industrial AI model builds a custom knowledge graph for every customer. It's the context layer our apps and agents run on top of, so Squint fits each customer's operation like a glove.
Industrial work demands accuracy and completeness across fragmented sources. Our model delivers, scoring 78% on cross-source benchmarks. Claude Code: 53%. OpenAI: 46%.
h/t @alexrkonrad for covering
The Opus videos are such a tease because they look amazing but there’s always too much happening on each frame and the pacing is awful. Hope to see a fast follow from @claudeai
6
132
It feels like cameras have gotten too good, and it's made software worse...
I can't think of a single system that actually uses or stores the raw capture. Every layer in the process spends compute downsampling video that was simply too good to begin with. We pour R&D into better sensors and then pour more compute into undoing it.
2
100
Tired: Stopwatch, clipboard, Matlab
Wired: Industrial AI solves an NP-Hard problem that's been in manufacturing for 100 years
No factory in America is meeting its true potential. Today we start fixing that.
The starting point for every improvement project on a factory floor is a time and motion study. It’s been that way since 1911 and it still takes months and $300K in consulting fees... for a single line.
If we’re serious about reindustrializing America, the math doesn’t work. Vision models need to move into the workflows factories already run and replace these projects outright.
@SquintAI spent the last few years capturing how work is designed to happen. Now, with our Lean Manufacturing agent, we’re performing continuous improvement on the line as work happens. This is going to unlock more throughput, reduce waste, and allow manufacturers to meet the demands of today.
It’s been a thrill for us to build this and our approach is inspired by manufacturing itself: attack the bottlenecks. With our new agent platform we’re picking off the highest-value workflows in the factory one by one, pulling them onto the Squint knowledge graph, and automating them with agents.
Let’s supercharge American factories. 🇺🇸
I talked to @Inc about where this goes next, see below.
This is exactly why the next 18 months of vertical AI are going to be bananas.
The promise of vertical AI, as opposed to vertical SaaS, is to build software that understands a job well enough to actually do the work — not just give a person better tools to do it.
There are 3 trends that have created a perfect storm for vertical AI:
1. AI software is capable of getting jobs done in more magical ways than ever before.
2. The first place that magic has landed is in AI software development itself, so there's a mad rush to build for the builders.
3. This makes for a frenzied loop: build AI tools to solve problems in the current AI tooling → uncover new problems with AI tooling → build different tools to solve the new problems...
Every day a new (and better, and faster) way of building emerges. But it also introduces a new set of problems that require a new set of products. Infra and tooling right now are on a runaway treadmill that just keeps going up in speed. The side effect is an insane tailwind of constant stack upgrades for building vertical AI.
It's a generational opportunity to literally invent solutions for problems that haven't been solved in 100+ years. A factory floor today has the same tribal-knowledge problem it had in 1926: the person who knows why the line really runs is about to retire, and nobody wrote it down. We are now finally getting to the point where we can take on massive real world problems and build solutions that get those jobs done in a magical way.
Process is great as long as it’s disposable and time-boxed. Especially now that we’re building so fast, sometimes it helps to have a way to keep people up to speed.
Note I’m still allergic to most process 😅
2
153
Q2 was our best quarter ever at @SquintAI. We added more ARR in Q2 2026 than Squint had in total 12 months ago.
Our customers are transforming their operations at an unprecedented pace, since January:
- MAUs are up 9x
- Usage is up 12x
- Tokens processed are up 16x
Last month, we launched the first model purpose-built for the field and the floor – scoring 25% higher than the frontier models on our industrial bench.
Last quarter, our team visited customers in 13 countries across 4 continents.
I said this to the team: this is the first and last time we celebrate these numbers. The world deserves Squint and we’re not close to done.
If this fires you up, come build with us!
The biggest misconception about data centers is that they don’t employ a lot of people.
The AI supply chain is massively understaffed right now. From family-owned HVAC companies to regional steel pipe producers, businesses can’t hire enough people to meet the demand for America data centers right now.
Arm CEO @renehaas237 says logic, memory, and other AI bottlenecks are a huge opportunity for the US to boost domestic manufacturing.
Not just for data centers, but the entire ecosystem around them.
"There's been a lot of backlash to data centers. Some people look at the data centers and say, 'They don't actually employ a lot of workers, they're just giant Costcos with whirring motors inside.'"
"But I would argue that, while that may seem to be true in terms of when the operation is completed, the ecosystem that gets created by building out these data centers is immense."
"And if you can have the entire supply chain, whether it's around motors, turbines, air conditioning, chillers, chips — if the US can vertically integrate that, that's huge, huge job opportunities for us."
"The data center backlash reminds me of when the US had most of the semiconductor fabs in the world 30-40 years ago. People were complaining about chemicals and things of that nature... 35 years later, every advanced fab, whether it's memory or logic, is not inside the United States."
Great few days at @reindsummit. Back to Squint HQ more bullish than ever on the future of American industry.
Tons of really sweet product launches — but the biggest takeaway for me: this future won’t be built by new machines alone. We need to invest in the people on the front line who actually power American industry today.
The future of American industry is human + machine — not human replaced by machine.
7
108
Can’t stress enough how dire the consequences are if we don’t address this problem. We need to hustle.
“The average age of an experienced tradesperson in the United States is 55. And there’s no pipeline.
This is one of the biggest constraints we have.”
-Doug Petno of @jpmorgan
4
169
Humanity is not incentivized for one party (ie one person, one model, one company) to take all the jobs and all the value, basic human nature just won’t let this happen.
As such, the better implementation of this would be to find ways to adopt the frontier models and supplement them with a suite of in-house tools and platforms that increase value on the few vectors that matter most to your company.
I think we’re going to see an explosion of more personalized benchmarks, more personalized models, and more personalized deployments of software.
The software ecosystem will build on the AI capabilities to make possible n=1 apps and that will be the real fuel for GDP growth over time.
“The Clearinghouse controls four things: memory (what your agents know), context (what they see and how it’s served), execution (what they’re allowed to do), and governance (who’s allowed to do what, plus the audit trail behind all of it).”
This is especially true in industrials; the clearinghouse can’t just be a generic layer.
For an agent to work on the factory floor, it needs to understand the work, the SOP, the operator, the site-specific context, etc.
10
236
















