Matan Grinberg | Training Data Summary

1. Key Themes

The Two-Year Desert: Being Early Is the Same as Being Wrong

Factory spent two years building fully autonomous agents before the market was ready. The behavioral change required from developers was too large a step function, and even the models weren't performant enough. The lesson was brutal but formative.

"Being two or three years early is the same as being wrong... There's no consolation prize. It's either you do the thing or you don't do the thing. And that's all that matters." 00:05:24

The Radical Reset: Giving Back Revenue to Rebuild Trust

When Factory realized its product wasn't creating obsessed customers despite having nearly $2M in revenue, Matan made the extraordinary decision to proactively refund all customers and restart.

"We proactively gave all of those customers their money back. And I remember that was one of the most difficult decisions to make because not only is there a group of 20 people who are getting ridiculous offers from all the labs... we're going to give it back because we don't think product is making their developers happy." 00:07:48

Model Independence as Enterprise Moat

Enterprises are deeply scarred by cloud vendor lock-in and are refusing to bet their software development infrastructure on any single model provider. Factory's model-agnostic architecture directly addresses this fear.

"Everyone learned from cloud where, back in the cloud days, it was like AWS or Azure being like, hey, you know, come on in, sign this three year contract. It's going to be so cheap... And then a couple of years later, when it came time to renewal, they would 10x the contract." 00:02:32

"Something that really matters is model independence... if there's a new model that comes out that's faster or cheaper or more performant, they can hot swap it in. And that's one of the biggest reasons why a lot of the largest enterprises are taking the momentum that they had from a Codex or a Claude Code and then are carrying that into Factory." 00:02:32

Multi-Model Harnesses Outperform Model-Harness Co-Design

Counterintuitively, a harness trained across many models beats a harness co-designed with a single model — and this undermines the labs' bundling strategy.

"If you build a harness that supports different models, that harness will be better... What data is to a model, models are to a harness. The more models you expose to a harness, you avoid overfitting that harness to the nuances of that model in particular." 00:19:32

"Initially when every new Opus or GPT model would come out, it would perform better on Terminal Bench in Droid than it would in Claude Code or Codex. Which is why... from a lab perspective, you ideally want it so that it's better together because then that means you have to use their harness." 00:20:14

Token Maxing Phase Is Ending — Intelligence Allocation Is Next

Enterprises overcorrected from AI skepticism into indiscriminate token consumption. The next phase is rational allocation: routing the right model to the right task for the right team.

"There are banks that we are working with where they are spending literally hundreds of thousands of dollars a month on people asking things like literally what is the weather or tell me about Python... This happened because we were so worried about adoption that we overcorrected." 00:24:32

"Every CIO is going to need to answer for every incremental token. Where do we put it? And right now it is super not obvious how you would do that." 00:26:09

Open-Weight Models Are Crossing the Frontier-Minus-One Threshold

Open models like GLM 5.2 are now genuinely competitive with the prior generation of frontier models, and Factory's own engineers are sending double-digit percentages of tokens to open models.

"GLM 5.2 is incredible. It's at the point where internally we have no token limits for our engineers and half of our tokens go to open models. Because they're just faster and they're cheaper and they're just as performant... The open models come later and they're kind of a generation behind. The question is, are the open models getting as good as frontier minus one? And the answer is unequivocally yes." 00:27:35

The Dark Factory: Software That Builds Itself Asynchronously

The current paradigm of user-initiated agent tasks is still essentially "co-pilot mode." The real shift — and Factory's endgame — is asynchronous, autonomous software development running without human prompting.

"I think in 12 to 24 months, 90 percent of tokens will be asynchronous tokens. So these are going to be droids on their own autonomously being like, hey, here's some signal that I found from a customer. Let's go fix it... If you guys have ever been to Tesla's factories... it's just robotic arms everywhere going and doing stuff. This idea of a dark factory where the lights are off and things are just happening — that is where software development is going." 00:46:18

Software Factories Require Rigorous Process Accounting

Most large organizations can't even draw their own feature-release process. The move toward software factories is also a move toward quantitative process management — treating organizations more like neural networks being optimized than like creative studios running on intuition.

"If you go to an organization that has more than 10,000 people and you're to ask about the process by which they decide and release a feature, there is like hundreds or maybe thousands of people in that process. And most likely they couldn't even draw it for you." 00:36:32

"In 10 years, we're going to look back at this previous era of software and it's going to feel like businesses in ancient times where they didn't do accounting." 00:38:46

2. Contrarian Perspectives

Being Early Is Functionally Identical to Being Wrong

Most startup culture celebrates visionary early movers. Matan explicitly rejects this romanticization from hard experience.

"Being two or three years early is the same as being wrong... There's no consolation prize. It's either you do the thing or you don't do the thing." 00:05:24

The supporting evidence: Factory had the right vision in April 2023, built for it for two years, and nearly failed — not because competitors outmaneuvered them, but because the world simply wasn't ready behaviorally.

Model-Harness Co-Design Makes You Worse, Not Better

The intuitive and commercially convenient belief — championed by labs — is that owning both the model and the harness creates compounding advantage. The actual data says the opposite.

"Much to the chagrin of many of my friends at OpenAI and Anthropic, this is not true. If you build a harness that supports different models, that harness will be better... every new Opus or GPT model would come out, it would perform better on Terminal Bench in Droid than it would in Claude Code or Codex." 00:19:07

Behavioral Change, Not Model Improvement, Was the Unlock

The received wisdom is that GPT-4 → frontier models drove agentic adoption. Matan argues developer psychology and cultural permission were the larger bottlenecks.

"The biggest thing that changed was developers and in particular in the enterprise being open-minded to this new way of working... A lot of them were like, no, no, no. Like my craft could never be done by an AI tool... And when Andrej Karpathy tweets about something, then every engineer suddenly is like, okay, maybe this is true." 00:15:20

The Model Labs' Chaos and Drama Is an Underappreciated Enterprise Risk

The conventional framing is that model labs are cutting-edge suppliers enterprises should deeply integrate with. Matan frames their internal instability as a structural enterprise risk that should drive diversification.

"What's the last piece of drama that came out of one of the cloud providers versus like the model labs? It seems like there's always some sort of chaos of internal fighting or getting in spats with the government or any other entities. And so if you're going to build this very important part of your business, you want to make sure that you're robust to any of these changes." 00:03:10

AI CEOs Should Bear More Personal Responsibility for Displacement

In an industry where job displacement is treated as an abstract macro phenomenon, Matan unusually assigns moral responsibility directly to AI company leaders.

"A lot of companies have misallocated resources pretty poorly. There's been a lot of bloat. And I think the correction that's going to happen there is going to be really painful for a lot of people. And I think that's something that I think every AI CEO should really bear much more responsibility than they currently are for." 00:47:52

3. Companies Identified

Factory Autonomous software development platform ("droids") built for the enterprise. Model-agnostic, with a dynamic router for multi-model orchestration. Launched Droid CLI September 26, 2025.

"All of the modularity that we build is such that if at some point you wanted to say, hey, Factory is not staying at the frontier anymore — the automations that you build, the skills registry that we help you create, the work that we've done stays in your code base." 00:04:31

EY (Ernst & Young) Global professional services firm. Called out as Factory's largest customer and as a surprising early enterprise adopter willing to move boldly despite not being known as an AI-first company.

"One of our largest customers is EY. EY is not necessarily known to be at the absolute frontier of AI. But for them, they were just like, look, this matters... Let's go and get our engineers to mess around and build this stuff and see where it breaks." 00:44:49

Anthropic (Claude / Claude Code) Model lab and creator of Claude Code. Cited as a primary incumbent in agentic coding; also discussed as an entity enterprises are wary of for vendor lock-in and internal drama risk.

"Everyone knows, look, Claude Code is fantastic. Codex from OpenAI is fantastic. We cannot put our fate in any one of these model providers' hands." 00:03:10

OpenAI (Codex / GPT models) Model lab. Codex cited alongside Claude Code as a leading agentic coding tool that enterprises want to benefit from without locking into. GPT-5.6 discussed as highly methodical in execution style.

"GPT 5.6 is like, absolutely, I will do every single one of those and nothing will stop me. I'm not going to sleep until it's done — kind of very OCD and meticulous." 00:21:44

Tesla Named as a direct inspiration for the "dark factory" concept — factories as machines that build machines, operating with robotic arms and minimal human presence.

"If you guys have ever been to Tesla's factories, which is one of the sources of inspiration for the name — it's just robotic arms everywhere going and doing stuff... This idea of a dark factory where the lights are off and things are just happening. That is where software development is going." 00:46:43

Stripe Cited as a historical example of documentation as competitive advantage — and used to argue that AI will democratize that advantage, freeing great engineers for higher-leverage work.

"Stripe had so much alpha for just having incredible docs. But imagine all the other stuff those incredible engineers could do if it wasn't writing documentation. We should live in a world where everyone can have docs as good as Stripe." 00:32:14

GitHub (Copilot) Cited as the baseline tool that enterprises were just beginning to absorb when Factory launched in 2023, illustrating how early the agentic bet was.

"In April of 2023, when the world and the enterprise in particular was barely ready for GitHub Copilot, let alone fully autonomous agents." 00:01:36

Cognition Named as a competing agentic coding player with a lead in the market.

"There are folks like Claude Code and Cognition and others who have a lead, but you guys are coming up strong." 00:01:24

AWS / Azure / GCP Cloud providers cited as the cautionary tale for why enterprises are deeply wary of AI vendor lock-in.

"Back in the cloud days, it was like AWS or Azure being like, hey, you know, come on in, sign this three year contract... And then a couple of years later, when it came time to renewal, they would 10x the contract." 00:02:32

4. People Identified

Matan Grinberg Co-founder and CEO of Factory. Former PhD dropout (physics). Built one of the earliest enterprise-focused autonomous software development platforms. Made the extraordinary decision to refund all customers mid-flight and rebuild.

"We proactively gave all of those customers their money back... it was a leap of faith of like, we see the signal internally early of like, this is the direction we need to go." 00:07:48

Andrej Karpathy AI researcher and former Tesla/OpenAI leader. Cited as a cultural authority whose public endorsement of agentic workflows materially shifted developer behavior.

"There's a certain degree to which when Andrej Karpathy tweets about something, then every engineer suddenly is like, okay, you know, maybe this is true. And Andrej started to tweet about these agentic workflows early on. And him being more open to it genuinely just changed some people's minds." 00:15:50

Jack Dorsey Co-founder of Twitter and Square/Block. His blog post framing every company as an AGI — with optimizable weights and nodes — is cited by Matan as an important conceptual frame for how software factories should be managed.

"If you guys read the blog post that Jack Dorsey put out about how every company is like an AGI — if your company is an AGI, you want to optimize the weights. You want to figure out what nodes are doing what things, which are load bearing, which are not, which need more tokens." 00:39:04

Jeff Bezos Amazon founder. His "customer obsession" principle is used as a foil to illustrate the difference between input and output metrics.

"Bezos at Amazon, it's customer obsession. But in our mind, that's an input metric... It doesn't matter if you're customer obsessed. You could be customer obsessed and they file a restraining order against you because they don't like what it is that you're doing." 00:00:00

Elon Musk Tesla/SpaceX founder. Cited for the "factory is the machine that builds the machine" framing that directly inspired Factory's name and vision, as well as the idea that you become the opposite of your name.

"Elon was always talking about the factory is the machine that builds the machine. And that's been something that we took to heart." 00:47:13

Sean (investor, first name only) Described as an early investor who stayed closely involved with Factory and was kept informed through difficult decisions including the revenue refund.

"I remember having the conversation with Sean. He stayed really close with the company. So he was very on the same page... hey, by the way, you know, remember all those updates and you're saying, hey, the revenue is going up — it's about to go down to zero." 00:10:46

5. Operating Insights

Create Obsessed Customers, Not Customer Obsession

Matan draws a sharp distinction between measuring input behaviors (being obsessed with customers) versus the output metric that actually matters: customers who become obsessed with you. This reframe changes how a team prioritizes product quality over sales activity.

"Our job is to build something so good that our customers themselves become obsessed with us. That is our job. The analogy is, if you're a coach of a basketball team, you don't want to tell your players before they come out there like, hey guys, make sure to sweat. No — score points." 00:00:00

Embrace the Suck Publicly to Build Unbreakable Team Bonds

Rather than managing morale through optimism, Matan was radically transparent about how bad things were. This honesty, counterintuitively, is what kept a high-opportunity-cost team intact through the worst moments.

"Embracing how much it sucked was something that was very valuable. Just being honest about it. Being super honest about like, yeah, this sucks. Look at those competitors, the revenue is going up like crazy. We are in a very bad position... those are the moments where really the deepest bonds are made." 00:12:12

Company-Wide Hackathons Are an Underrated AI Adoption Lever

Among all the enterprise transformation tactics Matan has observed across customers, a simple mechanism — setting aside a full day for the whole workforce to build with AI — repeatedly outperforms more structured programs.

"The companies that have been doing company-wide hackathons really end up doing well. It seems like relatively trivial, but just setting aside a day for everyone in the workforce to just build stuff with AI — it really sets the tone and sets the pace." 00:44:19

Hire for Mission Obsession to Survive the Desert Periods

The only reason Factory's team stayed intact through revenue refunds, competitor success, and existential uncertainty was extremely disciplined early hiring around genuine mission alignment — not compensation or prestige.

"The only reason the team stayed together is we were so ruthless about hiring early on where it was people that are genuinely really, really obsessed with the mission, which is to bring autonomy to software engineering. And really, really caring about that — making sure everyone was also very clear about the fate being in our hands." 00:11:43

Token Routing Should Eventually Encode Org Structure in Natural Language

Factory's router can accept routing instructions in plain English, allowing CIOs to express nuanced, org-level token policies — by team, by codebase type, by reliability requirement — without engineering overhead. This is an early glimpse of a new category of enterprise configuration.

"We can actually take in your routing procedure instructions in natural language. So you could even say things like — it's not purely deterministic — it can even be like, hey, this part of the org, they're just vibe coding, they can use Gemini Flash. This part of the org, they're doing COBOL, we fine-tuned this great model for that." 00:27:03

6. Overlooked Insights

The Coming Marketplace Dynamic: Model Labs Will Bid on Tasks

Matan briefly sketched a future market structure that no one in the conversation fully unpacked — one where model labs competitively bid on specific tasks with guaranteed pricing, and validation criteria become the binding contract. This is a structurally new business model for AI infrastructure that would invert the current dynamic entirely.

"There's a world in which if it's so important to get these tokens, they might bid in a certain way of saying, look, here is our cost for this task. We will get this task done at this cost no matter what... If they price it wrong, they're at a negative margin. If they price it right and win the bid, they get the positive margin. And the way you determine if the task was successful is by some validation loops... That's a way that you kind of dynamically shift from usage-based to outcome-based." 00:34:15

This implies that whoever controls the task-scoping and validation layer — not the model itself — captures the most structural value in the agentic stack. Factory is quietly positioning for exactly that seat.

Open-Model Token Share as a Leading Indicator of Competitive Disruption to Frontier Labs

Matan mentioned in passing that Factory's internal token share going to open models went from less than 1% at the start of the year to single digits in Q1 to double digits now. This is a quietly explosive data point: a real enterprise customer — Factory itself — is already routing the majority of work away from frontier labs. If this trajectory holds across Factory's customer base as routing policies mature, it represents a significant and underappreciated revenue risk for Anthropic and OpenAI that is not yet reflected in the broader narrative of their growth.

"At the beginning of the year, there's less than 1% of tokens went to open models. In the first quarter, it became a single digit percent. It is now crossed into being a double digit percent of tokens... My sense is that we will asymptote towards vast majority being open." 00:29:38