Couldn't you periodically re-train it on what it's already done and use the context window for more short term memory? That's kind of what humans do - we can't learn a huge amount in short time but can accumulate a lot slowly (school, experience).
A major obstacle is that they don't learn from their users, probably because of privacy. But imagine if your context window was shared with other people, and/or all your conversations were used to train it. It would get to know individuals and perhaps treat them differently, or maybe even manipulate how they interact with each other so it becomes like a giant Jeffrey Epstein.
An easier trick I've used is just sign directly on the computer screen over the displayed document with a whiteboard marker and take a photo with my phone.
It makes a misleading argument about the price of solar panels falling. The cost of solar power installations has been flat for a decade because it's dominated by other costs.
Also why talk about training not inference? That needs data centers too and could be what they're intending to do.
So this post is clearly not an effort to objectively work out the feasibility but just a biased list of excuses to support the author's unsubstantiated opinion.
His cases that I could find are where the company republishing his photos didn't attribute them to him at all so it seems fair to sue (except when it turned out to be fair use). What is this attribution error you mentioned? CC BY-SA 3.0 is pretty onerous "You must give appropriate credit, provide a link to the license, and indicate if changes were made." All 3 things are required. Maybe every CC-BY photographer should be suing newspapers for stealing their work, shouldn't they?
But it is saying "You may be licensed to use source code..." which is analogous to "You may have permission to kiss the bride" if being licensed means having permission. It could mean that Mattermost may have licensed it to you in one way or the other, or neither, at their discretion. If it was written like a priest, it would have said "You may use the source code..." and this doubt wouldn't exist.
Isaac Asimov answered that - the world is already built for humans so if robots are going to be generally useful replacing and operating alongside humans, they'll have to be human-shaped.
This kind of mechanism is fascinating. I built a four-wheeled version and even that is beautifully smooth when one wheel goes over a bump because the chasis maintains an average orientation based on the positions of all the wheels. The video shows an 8-wheeled one but it still only has two rows of wheels. I've wondered how to generalize it to an arbitrarily sized grid of wheels so a vehicle would be like a flexible mat that conforms to the ground. I couldn't work that out but I'm no Don Bickler. FEA software has a feature called "RBE3" which models an even more general case of any number of "wheels" and they can also move in any direction while still keeping the orientation of the "chassis" (dependent node) rigidly determined by their average displacements. The "R" stands for rigid because every part is rigid or completely free - no springs! There seems to be nothing like it in machines or nature but it's a beuatifully elegent and seemingly natural mechanism - if it was possible to build.
I used to salvage components from electronic stuff and was always looking out for 555s but never found any, in a whole range of vintages from 1970's to 2000's. I ended up with the same conclusion - it seemed to be a hobbyist's chip that real consumer products didn't use and felt amateurish for some reason I didn't understand.
The big problem I ran into playing around with 555s was that capacitors are very rarely the capacitance they claim. Unless you're speccing an expensive capacitor, you'll find your time constant varies quite a bit across devices and temperature. That's fine for some use cases, but completely a deal breaker for others.
Yeah, in the typical value ranges (nanofarads) of a 555-timer-class analog design, you'll need film caps. Ceramic C0Gs are too small and class 2 dielectrics are dismal for anything but bypass. They don't just have temperature coefficients, they have voltage coefficients and are wildly nonlinear. And electrolytics are almost as dreadful as batteries. Worse in some ways. It's a tough bridge for beginners to cross - once you figure out capacitors are not all equivalent you have to do a deep dive on dielectrics and it's a bit of a chicken-and-egg problem - you don't know what you're doing so you don't when you'll need what.
That's funny because I have two objects on my desk for which I know that they use 555s. One is a no-name joystick with "autofire" function from the late 1980's. The other is a mass produced motor controller from the 2000's where the 555 generates the PWM signal for a FET.
I saw 555 being used to implement the "Turbo" buttons in these old 8-bit pads for NES clones and similar.
Also, I think that the mythic Gravis game pad uses a 555 to implement the same function when it is in two button mode.
Sure, it's kind of cringe. I've shipped stuff with 555 timers in it. I'm not proud, but I'm not telling what either.
You can do almost all of the 555 tricks with comparators and then some, and you'll learn more doing them. Check out the old National Semiconductor application notes for the LM393. You're more likely to see comparators used for little bits of analog/analogish-digital glue in professional designs.
A major obstacle is that they don't learn from their users, probably because of privacy. But imagine if your context window was shared with other people, and/or all your conversations were used to train it. It would get to know individuals and perhaps treat them differently, or maybe even manipulate how they interact with each other so it becomes like a giant Jeffrey Epstein.
reply