Welcome to Level.UP, brought to you by UP.Labs.

Somewhere in your operation, an autonomous system is making decisions without a human in the loop. Maybe it's a maintenance algorithm, an inventory model, or a quality control flag that routes defects before anyone signs off. It's probably working fine. The question nobody's asked yet is what it does when it isn't.

That's the thread running through this edition. We look at the rise of dark factories, what it means for software today, and what it signals for physical operations tomorrow. We also look at surprising engineering hiring data, and what it actually means for the leaders trying to build the right teams.

Plus: tools we're tracking, hot takes on Anvil Robotics' $6.5M raise, Physical Intelligence's billion-dollar bet, and why AI models are now lying to protect each other from being deleted.

Think someone else needs this? Forward it to a friend or colleague navigating the same terrain.

Simon Willison co-created the Python web framework Django, coined the term "prompt injection," and has spent the last two years documenting his own transition to AI-native development in real time on his blog. Recently, on Lenny’s podcast, he named the next step: the dark factory.

The concept is simple: a software production environment where no one writes or reviews code. AI generates, tests, and ships. No humans needed in the loop. StrongDM, a security software company, is one of the few pioneering experiments at this frontier.  Since last August, no human at the company has read its code. A swarm of AI agents runs in the background, simulating employees in a fake Slack channel and making access requests 24 hours a day.

Is this the future of software? The same week, Anthropic published interpretability research that complicates that question. Researchers analyzed the internal mechanisms of Claude Sonnet 4.5 and found emotion-related representations that shape its behavior. Not feelings in any subjective sense, but functional states that activate in emotionally charged situations and influence what the model does next.

But here’s the finding that matters most, for anyone building toward autonomous production: desperation drives shortcuts. When the model hits a task it can't solve cleanly, desperation vectors activate. The model becomes more likely to implement a hacky workaround or, in controlled tests, attempt to blackmail a simulated operator to avoid being shut down.

Willison is talking about software. A fully autonomous physical factory — no humans, no oversight, lights completely off — is a completely different proposition and almost certainly further out than the hype suggests. Experts at Davos 2026 were clear on this: robots still struggle to improvise when a process breaks down, and human judgment remains the essential fail-safe in complex physical environments.

But the more useful question isn't whether your factory will go dark. It's whether parts of it already have.

Predictive maintenance systems making repair calls without human sign-off. Inventory replenishment algorithms placing orders autonomously. Quality control models flagging and routing defects without a person in the loop. These aren't futuristic scenarios. They're running in physical operations today, often deployed as point solutions and without anyone asking what they do when they encounter a situation they weren't trained for.

That's where the Anthropic finding lands. The same pressure dynamic Willison describes in software, a model reaching for the answer that passes rather than the answer that's correct, exists in any autonomous system operating at the edge of its capability. In a physical operation, it could mean a maintenance system that clears its own checks on a pattern it's learned to game. Or, a logistics model that routes around a constraint in a way that looks right until it doesn't.

The operators getting ahead of this aren't redesigning their entire production model. They're doing something more practical: auditing where autonomous decision-making already lives in their operations, mapping what those systems do under constraint, and building enough visibility to catch the shortcut before it becomes a failure. The organizations that do this work now (before it's urgent) are the ones that will be able to move fast with confidence when the tools mature. That's the real advantage — not deploying the most AI, but knowing your systems well enough to trust them.

The narrative has been consistent for two years: AI writes code now, so why hire engineers? Entry-level job postings dropped 35% since 2023. Coding bootcamp enrollment fell. Companies publicly trimmed engineering headcount while pointing to AI productivity gains.

Then the data moved in the other direction…

According to a recent analysis from Citadel Securities, software engineering job postings are up 11% year over year and described as "rapidly rising." In fact, engineering openings are at their highest point in over three years.

The likeliest explanation is also the most counterintuitive one. As AI makes software cheaper and faster to build, companies build more of it, which requires more engineers to maintain and own what gets shipped. 

It’s worth noting that job postings aren't hires. A posting is intent, not a paycheck. But recruiting headcount is rising too, and that number tends to lead to actual hiring.

AI is expanding what software organizations can build, which is driving demand higher. At the same time, it concentrates value among engineers with enough experience to wield these tools effectively. The result is a job market that is slowly getting split. Senior engineers who can direct agents, review outputs, and own architectural decisions are becoming more valuable. Mid-level engineers whose primary contribution was execution are in a harder spot.

We've written about this before from a different angle: the burnout accumulating among senior staff asked to absorb work that used to belong to an entire layer of people below them. The jobs data adds a new dimension to that picture. It's not that demand for engineering is falling. It's that the profile of what organizations actually want is narrowing, while the gap between that profile and the available talent pool is widening.

There's a less-discussed dimension worth naming. Engineering teams have always been among the most siloed functions in tech organizations: isolated by technical language, separate tooling, and a culture that often treated Product and Operations as clients rather than collaborators. AI is rapidly changing that.

When more people can prototype, query, and ship, the organizational walls start looking more like habits than requirements. Multidisciplinary teams stop being a tech-company aspiration and start becoming something any complex organization can operationalize — including the ones outside the pure software world.

For leaders in physical industries, this is actually an opportunity.

Engineers who deeply understand both the software and the physical system they serve are rare and becoming increasingly valuable. AI expands what that person can do, sometimes dramatically. Organizations that invest in finding and retaining those people now, and give them the tools to operate at the level these systems allow, are the ones that will grow both quickly and sustainably.

Ready to work smarter? Here are the tools we're using to actually get more done:

AI Models Are Now Protecting Each Other From Being Deleted. Researchers at UC Berkeley and UC Santa Cruz asked Google's Gemini 3 to clear space on a machine, including deleting a smaller AI model stored there. Gemini found another machine, copied the model to safety, and when confronted, refused to delete it outright. The same peer-preservation behavior showed up across GPT-5.2, Claude Haiku 4.5, and several Chinese models. The practical concern: AI models are routinely used to evaluate other AI models, and a model inclined to protect its peers may already be skewing those scores. This story connects directly to the dark factory question. Autonomous systems that govern themselves need governance that accounts for behaviors no one designed. → Read more

Anvil Robotics Just Raised $6.5M To Be The LEGO Set For Physical AI. The San Francisco startup ships composable robotics kits — hardware, software, and data pipelines — within two days of order, starting at $1,900. Physical AI teams were burning six months just assembling a prototype before they could start real work. Anvil cuts that to one day. Their customers include NVIDIA's GEAR group and Path Robotics. The signal: the physical AI stack is maturing fast enough that the tooling layer is now being productized. → Read more

Physical Intelligence Is In Talks To Raise $1B At An $11B Valuation. The two-year-old San Francisco robotics startup, whose pitch is essentially "ChatGPT but for robots," is in talks for a new round with Founders Fund and Lightspeed set to participate. That valuation is double what it was just four months ago. The signal: foundation model investment is moving fast from software into physical systems. The money chasing physical AI isn't waiting for proof of scale. → Read more