At a Rise of AI Virtual Chat in February 2021, held in partnership with AWS, Fabian Westerheide interviews Ronnie Vuine, CEO of micropsi industries, about applying AI to manufacturing. Vuine explains how his company's software lets industrial robots learn skills, such as finding a logo on a screen in a live demo or locating features on refrigerators during factory testing, from four or five human demonstrations instead of precise pre-programmed coordinates: the robot watches a worker perform the task, uploads the recordings to AWS for training, and can execute the learned skill after about three hours of cloud computation. He traces micropsi's roots to a Minecraft bot project and to Dietrich Doerner's psi theory of intelligence, and describes why the company deliberately moved away from pursuing AGI toward a sellable product. The conversation then turns to why AI adoption in manufacturing remains low, with Westerheide citing a figure that only six percent of manufacturing businesses use AI, and Vuine arguing that bolting sensors onto old machines is not the same as redesigning production as a software-driven system. He closes with advice on building AI capability in-house, a note on data privacy between customers, and predictions about GPT-3's near-term impact.
Key takeaways
- Micropsi industries trains industrial robots to perform tasks from four or five human demonstrations rather than precise pre-programmed coordinates.
- Ronnie Vuine says pursuing AGI directly is commercially unpredictable, calling it "alchemy" rather than engineering or science.
- The company's skill-learning system runs on an Nvidia GPU compute box and uses AWS cloud training, evaluating its camera feed twenty times per second.
- Only six percent of manufacturing businesses currently use AI, according to a figure cited by Fabian Westerheide.
- Vuine argues most factories only bolt sensors onto old machines instead of redesigning production as a software-driven system, unlike Tesla's approach.
Detailed notes
From a Minecraft bot to robots that learn spatial movement
Vuine says micropsi industries grew out of an earlier project to build an autonomous Minecraft bot that learned its own motivations rather than being told what to do. The team realized they had taken a shortcut by treating the game world as discrete, letting the bot "teleport" between known locations instead of learning continuous movement from one point to another. Solving that problem of moving from A to B without precise measurement, he says, is what pulled the company into robotics. Vuine argues there is a deep, not fully understood connection between this kind of spatial reasoning and intelligence more broadly, since animals and people evolved to handle continuous three-dimensional movement. He traces the company's name and approach to Dietrich Doerner's psi theory, a theoretical-psychology model of how intelligence, emotion and failure work that Doerner developed independently of AI research; Vuine and co-founder Joscha Bach read Doerner's work in a reading circle at Humboldt University and began implementing early versions of it in Java.
Why micropsi chose product over AGI
Vuine says he avoids discussing AGI with investors because pursuing it is "not engineering and not science, it's more like alchemy": progress cannot be planned or budgeted in a systematic way. Running a venture-backed company means working toward a roughly ten-year exit horizon, which he breaks down as three years to product-market fit and four years to revenue. After about two years of early, more exploratory work at micropsi, it became clear that AGI-adjacent research, such as artificial emotion, was not predictable enough to build a business on, since a moody robot has no commercial value even if it made the system more human-like.
How the skill-learning system works
In a live demo, an off-the-shelf industrial robot fitted with a standard, low-resolution camera (VGA-level, comparable to a twenty-five dollar webcam) learns to find a logo on a screen after a worker demonstrates the movement four or five times over about twenty minutes; the recorded demonstrations are uploaded to AWS, where the skill is computed in roughly three hours. The compute box driving the robot is described as essentially a gaming PC with an Nvidia GPU running Linux, evaluating the camera feed twenty times per second and making a continuous, self-correcting approach rather than one precise measured movement, the way humans plug in a USB cable by feel rather than exact geometry. A live customer example is testing refrigerators on a production line, where the robot must find features that always look similar but are never in exactly the same place. Vuine says customer data is never shared or pooled across clients; other customers benefit only because the larger pool of data lets micropsi backtest algorithm changes for robustness and speed.
Why manufacturing AI adoption is still low
Westerheide cites a statistic that only six percent of manufacturing businesses currently use AI. Vuine attributes the gap to the difficulty of connecting a "fluffy" software world to machines that are sometimes forty years old and built to do one physical thing well. He says the common approach of sticking sensors onto legacy machines to gather vibration or temperature data for predictive maintenance is a partial fix, not real digitization. He contrasts Tesla, which he says built its cars and factories as digital systems from the ground up, with German carmakers that he says have roughly 250 individually "smart" modules but no unifying operating system connecting them. His vision of a fully software-driven "general factory" points to additive manufacturing and 3D printing as the model, and he notes that continuous, fluid production such as beer brewing is already closer to full automation than discrete assembly work such as car manufacturing, which still requires many workers to move and fit parts together.
Advice for manufacturers: build AI capability in-house
Vuine recommends that manufacturers build AI capability internally rather than outsourcing digitalization entirely to consultancies, arguing that companies need enough in-house understanding to select and deploy AI products without losing the ability to update their own production later. He compares this to the shift from mainframe computing to PCs: businesses did not need MIT-trained engineers to run Microsoft Excel, just an IT person who was "good enough." On ownership, he confirms that a customer such as BMW would own the skills its robots learn through micropsi's software, just as it owns other purchased production inputs.
Q&A highlights
Asked about legacy hardware, Vuine says manufacturers are less locked into old plants than they think, since production lines are typically updated every five to seven years regardless of AI; adoption tends to start with better-funded sectors like automotive and then trickle down. On brain-computer interfaces, he notes that while such interfaces could offer very high input bandwidth, language itself has lower throughput than typing, so professionals doing a task well often still prefer a keyboard. Asked whether micropsi combines classical symbolic AI with machine learning, he says the current system is almost purely machine learning, but predicts the next wave of AI will combine pattern detection with reasoning, likely including probabilistic or Bayesian methods, since deep learning alone "finds patterns" but "doesn't reason." On GPT-3, he predicts real market impact within two to three years and warns that once AI-written text becomes indistinguishable from human writing, all online text becomes "fakeable," which he expects to undermine the practice of using text as a signal of consensus and push people back toward judging whether an argument shows real reasoning.
Quotes
The thing about AGI is it is not engineering and it's not science, it's more like alchemy.
Ronnie Vuine
Build the capabilities yourselves, don't allow people to talk you into the idea that you need external specialists that do the digitalization and do the AI for you.
Ronnie Vuine
Questions answered in this talk
How does micropsi industries train industrial robots without programming exact positions?
According to Ronnie Vuine at a Rise of AI Virtual Chat in February 2021, a worker demonstrates the desired movement four or five times in front of a camera, the recordings are uploaded to AWS, and the robot's skill is computed in about three hours of cloud training.
Why is AI adoption still low in manufacturing?
Vuine explains that most factories only attach sensors to decades-old machines to collect data, such as vibration readings for predictive maintenance, rather than redesigning production as a software-driven system the way Tesla built its factories from the ground up.
Should manufacturers build AI in-house or buy it from vendors?
Vuine recommends buying packaged AI products, comparing this to buying a PC, but says staff still need to understand the underlying technology well enough to evaluate vendors and keep ownership of how their production is updated in the future.