I contend that is impossible to make an unbiased AI. I did an AI image recognition project several years ago. It used yolo to categorize rust into grade 1, 2, and 3 for offshore platforms. When creating our training dataset, we had different rust inspectors from different parts of the world drawing different lines in the sand between what was category 1, 2, and 3. We had to eventually pick which bias we wanted to roll out worldwide. The advantage for a giant corporation was that now the same consistent bias was being used worldwide and fewer people had to be safety trained to go on the offshore platforms. If that incredibly dull and basic application can’t be unbiased, I don’t think it is possible to avoid bias in anything produced with a training dataset. The very word “training” implies it. Someone somewhere decides A is in the training and B is not, and a bias is born, intentionally or not.
So the task is really to find the AI with the bias that works best for your application, not to try and remove bias.
Apple A series CPUs and now M series CPUs have consistently been top of the benchmarks in single core performance for most of the last decade. This even holds true when pitted against desktop Intel and AMD chips. For someone who works with workloads that struggle to be very multithreaded, I do watch this. I must be that 0.1% of the audience
Vindication! I’ve spent over a decade of my life putting physical interactives into museums. I have preached (sold) many museums on the stance that they should put unique experiences into museums that can’t happen on an iPad at home, to varying degrees of success. The museums that have listened are the ones that continue to be wildly successful to this day.
They are hard to do right though. I used to compete in combat robotics and the stresses put on museum exhibits is higher. I tell my new engineers that if their exhibit can be dropped into a gorilla enclosure and survive, they are about half way strong enough. Little makes up for raw experience in the art of building bomb proof exhibits, and many companies have failed before getting good. The amateur hour exhibits from the low bid newcomers that inevitably fail and/or need a lot of expensive maintenance has left a sour taste in a lot of museum’s mouths. A lot of those museums have knee jerk reactioned the opposite direction to touchscreen exhibits, only to see their ticket sales slowly drop. Thankfully, i’m seeing the pendulum of the industry swinging back towards physical interactives again.
They are probably referring to the much larger driver facing curved touchscreens for carplay/android auto that merges with the screen used for the instrument gauge. Also, the driver assist tech on the newer kia/hyundais is very good, especially for highway traffic, needing very little driver intervention.
Microsoft knew they would never get significant market share unless they offered open source alternatives that let you circumvent the telemetry in the early days of VScode. Embrace. The acquisition of github was part of this strategy. They made an ecosystem that sucked a lot of plugin developer talent into their ecosystem. Extend. Now the market share is firmly in their grasp and competitors have become weaker. Extinguish.
GPT 4.5 seems to get it right, but then repeat the 700 pounds
"A woodchuck would chuck as much wood as a woodchuck could chuck if a woodchuck could chuck wood.
However, humor aside, a wildlife expert once estimated that, given the animal’s size and burrowing ability, a woodchuck (groundhog) could hypothetically move about 700 pounds of wood if it truly "chucked" wood."
Thanks for pointing that out. I thought it was some sort of variation on an E-ink display because of the black and white limitation. Nothing about transflective tech limits it from full color other than price. I guess that leads me to an evolution of my question: Why are no* tablets using full color transflective displays?
*I did find the HannsNote2 [1] does, but it only came out last year, and this tech has been around for donkey's years.
Whatever happened to the transflective lcds that were popular in carputers in the 2000s? They seem to be a perfect fit for a tablet and I have been puzzled that no one has jumped on using them in one.
from the transflective wikipedia page [1]
"A transflective liquid-crystal display is a liquid-crystal display (LCD) with an optical layer that reflects and transmits light (transflective is a portmanteau of transmissive and reflective). Under bright illumination (e.g. when exposed to daylight) the display acts mainly as a reflective display with the contrast being constant with illuminance. However, under dim and dark ambient situations the light from a backlight is transmitted through the transflective layer to provide light for the display. The transflective layer is called a transflector. It is typically made from a sheet polymer. It is similar to a one-way mirror but is not specular."
It's price is it's biggest drawback, but it is also replacing any electronics used to run the switches at 12 or 24v which gets you above the noise floor if you are operating next to something noisy like a VFD. from the 6818 data sheet: "Robust switch inputs handle ±25V levels and are ±15kV ESD-protected" [1]
I guess I should rephrase. It saves all the interrupts except the one triggered at 40ms delay. For every button press without hardware debouncing, you can have 10s - 100s of 1to0 and 0to1 transitions on the microcontroller pin. This is easily verified on a oscope, even with "good" $50+ honeywell limit switches. Every single one of those transitions triggers an interrupt and robs cpu cycles from other things the microprocessor is doing. The code in the interrupt gets more complex because now it has to do flag checks and use timers (bit bashing) every time they are triggered instead of just doing the action the button is supposed to trigger. None of this is to say one way is the "right" or "wrong" way to do it, but putting the hardware debouncing complexity into hardware specifically designed to handling it, and focusing on the problem I am actually trying to solve in firmware is my personal preferred way of doing it.
I'm a big fan of debouncing in hardware with the MAX3218 chip. It will debounce by waiting 40ms for the signal to "settle" before passing it on. This saves your microprocessor interrupts for other things. It also will work with 12 or 24 volt inputs and happily output 3.3 or 5v logic to the microprocessor. It is pricey though at $6-10 each.
In a monkeys in front of a typewriter world, statistically, you are as likely to have a one off event that matches a specific bit pattern in the underlying format as you would the encrypted format. It would not be reproducible though since most encryption uses nonces
I really like all the caveats and the time taken to explain things in the first part of that document, but later it starts to rush and gloss over important details and caveats. On page 151 of that link, when it starts talking about using parallax to measure the distance to nearby stars, it says "However, if one takes measurements six months apart, one gets a distance separation of 2AU." This is obviously incorrect because the whole solar system is orbiting around the galactic core, which itself is moving with respect to the CMB rest frame. I did a quick calc based on the 552.2 km/s galactic velocity value from Milky Way wiki [1] and found that it moves an additional 0.97AU in 6 months. I am assuming that this has been accounted for by scientists, and is being simplified to make it more digestible for the reader, but it hides a rather large dependency for every higher rung on the cosmic distance ladder. A cosmic velocity ladder that seems to be based off of Doppler CMB measurements [2]. If we are indeed using measurements many months apart and under or overestimating our velocity through the universe, even a little bit, every higher rung of the ladder would be affected wouldn't it?
In the process of writing this, I thought "Surely we have launched a satellite pair that can take parallax measurements at similar times in different places!" They could range off of each other with Time of Flight, be positioned much further apart than a few AU, and take parallax star measurements at more or less the same time without atmospheric distortion, but it doesn't seem like we have. Both Hipparcos and Gaia were satellites that were deployed to measure parallax, but not as a pair. My reading suggests they used multi-epoch astrometric observations (speed ladder dependent) to generate their parallax measurements and it seems our current parallax and star catalogues are based on the measurements taken by these two satellites. New Horizons got the most distant parallax measurements by comparing simultaneous* earth observations, but it was limited to Proxima Centauri and Wolf 359, far from a full star catalogue.
I would love if someone more knowledgeable can steer me towards a paper or technique that has been used to mitigate the cosmic distance ladder's dependency on this cosmic speed ladder. Regardless of how certain we think we are of our velocity through the universe, it seems to me that sidestepping that dependency through simultaneous* observations would be worthwhile considering how dependency laden the cosmic distance ladder already is.
* Insert relativity caveat here for the use of "simultaneous". What I mean in this context is more simultaneous than waiting months between measurements.
I wish Supah Valves [1] were still available so I could recommend them. They use a smaller sprinkler valve as a pilot to move a much larger piston based valve _very_ quickly. That lets a lot more of your reservoir pressure get to the barrel before your projectile has shot out the end. The result? More range. More grins.
Thank you so much for the quick answer. That makes complete sense. I appreciate anyone putting in efforts in on Node-RED. I use it heavily in interactive museum exhibits and would spend a lot more time on software "plumbing" without it. Dashboard 1 has been very useful, but definitely starts to struggle on things like graphs with 3000+ points. I look forward to trying out this new one!
Impressive, but I can't foresee any project where I would pick this over the vastly more powerful PJRC Teensy 4.0 [1]. It is only slightly larger and only costs $3.85 more.