←
JanBosch793 posts
JanBosch avatar

JanBosch

@JanBosch

Software engineering professor in industry working on open innovation, architecture and software reuse

◷ Joined March 2008▧ 138 Photos and videos
JanBosch avatar
JanBosch@JanBosch·Nov 15

Detection is unwinnable: your opponent improves, knows your classifier, iterates for free. So we pivoted to provenance — a chain of custody signed in hardware at the sensor, defeated by pressing three keys. The screenshot problem is unsolved. janbosch.com/machines-that-t…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Nov 8

The people you’re automating weren’t only executing the process. They were compensating for the fact it has never once worked as written. Automate them away and you don’t get the same organisation cheaper — you get one with nothing left that adapts. janbosch.com/machines-that-t…

JanBosch avatar
JanBosch@JanBosch·Nov 1

When iteration is expensive, engineering is a sequence of choices. When it’s nearly free, engineering becomes a search. And a search optimises exactly, and only, what you wrote down. The objective function has quietly become the design. janbosch.com/machines-that-t…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Oct 25

Almost all industrial sensing was installed to close a control loop. A loop needs the present value, repeatability, low latency. It never needed absolute accuracy, retained history, aligned timestamps or context. All four are mandatory for learning. janbosch.com/machines-that-t…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Oct 4

A decade of "two years away" wasn't a perception problem. Perception got solved roughly on schedule. Validation cost scales with how wide your domain is. So the response that actually shipped was shrinking the domain. The ODD is the product. janbosch.com/machines-that-t…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Sep 27

In industrial AI the product is the normalization layer, not the model. Every plant is out of distribution with respect to every other plant. The number that matters: your second deployment cost divided by your first. janbosch.com/machines-that-t…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Sep 20

Part 4 of Machines That Think takes on the best argument in robotics: the world is a brownfield built for a bipedal primate with two hands, so a human-shaped robot is the universal adapter that needs no retrofit. But is it superior and is it certifiable? janbosch.com/machines-that-t…

JanBosch avatar
JanBosch@JanBosch·Sep 13

There is no internet of grasping. Across 1,228 VLA papers, rates fall from 95% to 30% under modest perturbations or from 90%+ to 0%. Meanwhile 4.66 million industrial robots work superbly by not generalizing at all. Robots work but general-purpose ones? janbosch.com/machines-that-t…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Sep 6

The comforting story is that the edge solves the data center's problem. The constraint at the edge isn't compute; it's memory. A model has to fit in RAM So the edge isn't an escape hatch, it queues up behind the same constraint. janbosch.com/machines-that-t…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Aug 30

New series: Machines That Think, on AI leaving software for the physical world. Part 1 starts at the bottom of the stack. The bottleneck isn't the leading edge of fabrication; it's packaging and memory. Allocation of those has become a political act. janbosch.com/machines-that-t…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Aug 23

Code is becoming a byproduct. The real artifact is contract-and-evaluation. But the sharpest pushback hit the weak half: AI gives us conformance by evaluation only. New post on why the fix is a runtime evaluation loop, not a better test gate. 👇 janbosch.com/the-evaluation-…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Aug 16

Back from the break with one idea I can't shake: in a few years, nobody will write software and nobody will care. My bet: the durable artifact stops being code and becomes contract-and-evaluation. That is intent plus the evidence that it holds. janbosch.com/field-notes-is-…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Jul 12

We’re watching a structural transition in software-intensive business that ranks with the largest such transitions in the history of the discipline. Navigate this transition deliberately and define the next era of the technology economy! janbosch.com/blog/index.php/…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Jul 5

The AI era has made pricing and sales more important, not less, because the pricing models are in flux and the companies that choose their value metric well will capture structurally more of the value they create than those that use yesterday’s models. janbosch.com/blog/index.php/…

Post attachment
JanBosch avatar
JanBosch@JanBosch·Jun 28

The current organizational models are designed for a world in which humans did the work. In an AI-first world humans increasingly orchestrate it and organizations need to redesign themselves around that reality to gain a decisive advantage. janbosch.com/blog/index.php/…

Post attachment