In summers during college, I worked at Raychem manufacturing under William Johnson, in tiny corner of engineering on a huge factory floor. In our lab and on the floor the activity was plastics extrusion, specifically for heat-shrink tubing. The most memorable project was hand-inflating the "balloon catheter" medical devices invented by Dr. Thomas Fogarty. (I later worked for a company Dr. Fogarty founded, Vital Insite). The most memorable location was the tall building with a 10 MeV electron gun for irradiating plastic (the "beam room" felt like a bomb shelter). That location is now home to Facebook headquarters, and the thought of factories in Menlo Park seems odd.
My first job out of college was working in Dan Murnick's lab (department 11131 in the Physics Division) on atomic/nuclear spectroscopy. I became a "laser jock," singlehandedly responsible for the entire experiment: a six-watt green Argon-Ion pump laser, a hundred-milliwatt red single-frequency, single-mode tunable dye laser, an intersecting-beam plasma chamber, and the lock-in electronics and PDP-11 computer necessary to gather data. The tremendous difficulty of optimizing the laser tuning (involving about a dozen degrees of freedom, counting all the angles of all the mirrors) became a lifelong lesson in the difficulty of high-dimensional optimization in general.
After Bell Labs I joined the Peace Corps, teaching physics in a rural town in Cameroon, West Africa. I considered it the best job in all of Peace Corps, and my students loved my teaching style...I still hear from them from time to time, thirty years later. It was the perfect place to invent physics demonstrations for an appreciative audience. I learned of John Hopfield there, and the talk he gave on "neurons" prompted me to re-invent the Hebb rule (and various other principles) while I was in scientific isolation in Africa. I thought I had understood a crucial part about the brain, even then, and the confidence I gained from finding those ideas actually in play made me think I could contribute to the field, and prompted me to apply to Caltech.
During my post-doc at NIH I worked on a software implementation of my brain-architecture project, presenting the results as Unsupervised Pixel-prediction at the Neural and Information Processing Systems conference ("NIPS 1995") and the Society for Neuroscience conference in San Diego. I also came up with a model of thalamus as a specific kind of signal processor very useful to a brain: a non-rectifying predictive comparator. The model was never published because I finished it just as I left NIH to move to Silicon Valley.
In Silicon Valley (home to my family home in Menlo Park) I worked at a new startup company every year or two (the list is on my resume), starting with Senior Engineer (i.e. programmer) and ending (?) with Chief Algorithm Officer. As I specialized in understanding the flow of numbers, inferences, and statistics behind profitable machine intelligence, I discovered that few computer scientists understood continuous systems, and few physicists and mechanical or hardware engineers understood software. So that intersection became my niche.
I loved my one bio-physics job (at Vital Insite), doing the simultaneous jobs of chief theorist and chief experimentalist. It was a bio-tech startup hoping to create a non-invasive device for measuring inter-arterial blood pressure (which is like measuring the pressure inside a champagne bottle without opening it). The stakes were at least hundreds of lives and billions of dollars a year, and the technology was brilliant: injecting vibrations into the radial artery near the wrist, and measuring their phase-velocity and dispersion as they traveled along the artery toward the heart. I got to work on a small team of physicists, nurses, mechanical engineers, programmers, and machinists. And I enjoyed a privilege rare these days: both discovering and explaining a new albeit very arcane phenomenon. (For the curious: the vibrational phase velocity in the artery slowed with the dorsiflexion of the forearm, which I proved with ultrasound and autopsy occurred because the artery was flattened by tendons passing it. One can understand the physics of this effect as follows: in an ordinary cylinder, the phase velocity depends on the elasticity of the cylinder's wall. But in a squashed elliptical cylinder, it also depends on the spring constant pushing the ellipse out of round, and that additional compliance slows the waves).
Because I always insisted on writing my own code (not in MatLab, but in "real" languages like C++ or Java), I was able to play with algorithms and tune them by hand, and ultimately to design the algorithms to do the tuning, i.e. to make the system informationally self-sufficient. I learned that the best ways to understand the mathematical constraints on autonomous, real-time systems is to build them.
My two-year return to academic science at the Redwood Neuroscience Institute revived my interest in the problem of the brain, and reminded me it had not yet been solved. Starting with the job-talk I presented there ("Brain-like software architecture: confessions of an ex-neuroscientist"), I advocated a brain-architecture of prediction modules I called "compressors" which might solve the brain's problem of representation in a scale-free way. That software architecture was adopted by Numenta, but to my knowledge there is little understanding of the specific signal-processing each separate module actually must do. I still believe in that design, but now understand each module to be a truly continuous simulation engine.