I applaud the effort of tinkering, re-creating and sharing, but I think the name is misleading - it is not at all a "local GPT". The contribution is not to do anything local and it is not a GPT model.
It is more like an OpenClaw rusty clone
283 karma · joined December 10, 2020
It is more like an OpenClaw rusty clone
Do you have the GPU running all day at 200W to scan for wake words? Or is that running on the machine you are working on anyway?
Is this running from a headset microphone (while sitting at the desk?) or more like a USB speakerphone? Is there an Alexa jailbreak / alternative firmware as a frontend and run this on a GPU hidden away?
- processing: while there is no post processing, it needs scene depth information which requires pre computation, segmentation and depth estimation. Not a one-shot technique and quality depends on computational depth estimates being good
- no free lunch. The optical setup needs to trade in some light for this cool effect to work. Apart from the limitations of the prototype, how much loss is expected in theory? How does this compare to a regular camera setup with lower aperture? F/36 seems excessive for comparison.
- resolution - what resolutions have been achieved? (maybe not the 12 MPixels of the sensor? For practical or theoretical reasons? ) What depth range can the prototype capture? "photo of Paris Arc de triumphe displayed on a screen". This is suspiciously omitted
- how does the bokeh look like when out of focus? At the edge of an object? The introduction of weird or unnatural artifacts would seriously limit the acceptance
Don't get me wrong - nice technique! But to my liking the paper is omitting fundamental properties
While this methods has no post processing, it requires a pre processing step to pre-calture the scene, segment it, estimate depth an compute the depth map.
You try to compute something that supercomputer can't - by not computing it? Instead the formula is stored in a data structure. But once you need to access all the values you still have something that does not fit the memory and needs to be computed.
I can't judge on the Java side, but suggest to pick a better example on how this can be useful.
Rough estimate - with an excellent 0.5" angular resolution and 35km triangle this could work.
Did you buy all these colors and paint and scan them? Did you analyze the shopping images of the bottles and classify them into hex colors? Or maybe just group by the color names given in the storefront listing?
Vastly different efforts, different "accuracy", but still, each methods has its use. But knowing what to expect would be nice.
I think this software-only post is meant for IP cameras / surveillance cameras. Internet is the oposite of closed circuit.
Maybe CCTV is used as a synonym for surveillance now in some regions of the world, but certainly confusing for a non-native speaker.
Mobley plotted two graphs of the solar irradiation and a black body fit. The only difference is supposed to be the x axis with a plot over frequency or wavelength. This is a non-linear mapping and a different shape is expected. But the result is that the peak irradiation level is at either 501nm or 882nm (when converted back). That can't be right. The labeling of the x axis does not change the maximum.
What he meant to do was to plot either the solar irradiation in W m^-2 nm^-1 or as the number of photons. With lower energy photons (towards the red) the same irradiation level will consist of more photons. This shifts the maximum number of photons towards higher wavelengths. 880nm sounds plausible.
A few questions remain unanswered though: What can the current plant already do? It sounds like a multi-day sequential process per batch. How many batteries could that give?
The mixed metal product also contains nickel-manganese-cobalt. But certainly with a lot of other stuff and not in the exact ratio you would put in a battery. Even if we were to continoue with NMC batteries (LFPs are more common today). It looks like a first concentration step to get the interesting 10% of the rock. What separation process still remains? I expect a concentrate still to be much more useful than bare rock.
What are the overall economics? I understand that you won't need the separate mining as Olivine is considered waste and has already been piled up. But is that an economic benefit? (cheaper?) Environmental? Or time to market? (you don't need another mining permission for more capacity).
Is it just a more green but more expensive extraction from unused Olivine? Or will this replace all other dirty extractions mining soon? (too good to be true)
However, the examples indicate that if you have a loop that is executed over and over, the setup cost for configuring the fabric could be worth doing. Like a continuous audio stream in a wakeup-word detection, a hearing aid, or continous signals from an EEG.
Instead of running a general purpose cpu at 1MHz the fabric would be used to unroll the loop, you will use (up to) 100 building blocks for all individual operations. Instead of one instruction after another, you have a pipeline that can execute one operation in each cycle in each building block. The compute thus only needs to run at 1/100 clock, e. g. the 10kHz sampling rate of the incoming data. Each tick of the clock moves data through the pipeline, one step at a time.
I have no insights but can imagine how marketing thinks: "let's build a 10x10 grid of building blocks, if they are all used, the clock can be 1/100... Boom - claim up to 100x more efficient!" I hope their savings estimate is more elaborate though...
If you didn't catch the reference, this is referring to the recent vibe coding incident where the production database got deleted by the AI assistant. See https://news.ycombinator.com/item?id=44625119
There is no silver bullet - you can't just build a 10bn$ nuclear bomb programm and call it a day. All the other means are still needed to transition away from fossil fuel.
The earlier we start the better.
There is open distributed monitoring for all kinds of signals, like seismometer networks, weather, ads-b... Is there anything like this for the power network? Like a reference design or an esp32-shield?
How would it look like if we were serious?
I would make it three phase, with direct coupling to also see the exact voltage changing over the day. Sometimes we have issues with local voltage rising too high and PV inverters shutting off. I'd like to see and log this. An audio ADC should be good, but needs three channels.
For distributed sensing and logging, you would need a reasonable accurate time synchronization. Raw ntp over internet might not be good enough, at least not for localizing fault propagation issues over the whole continent. Better stick a 5€ GPS module on there.
Anybody seriously working on this..?
Let's say you build single photon detectors and ultra precise time stamping. Would that get us near? Today, maybe we don't have femtosecond time stamping and detectors yet. But that is something I can imagine being built! Timing reference distribution within fs over 100s of km? Up to now, nobody needed that I guess.
The Cob array of leds is the problematic choice, while on paper you get a lot of light, you don't get a point source.
If you look at the large theater projectors, I remember laser sources (but no speckle, is this pumped phosphor?) or expensive xenon bulbs. At least 10years ago...
I remember a solution (car headlights?) of decoupling the phosphor (for converting uv to while light) from the UV Led. So you have multiple diodes pumping a small piece of phosphor for a nice bright point source. Not sure how this can be replicated open source.
Another approach (studio led lights) wasomething with glass mixing rods? You insert the light of multiple leds into a glass rod, and a uniform beam emerges. But I guess this was more to improve CRI, rather than to increase brightness/point source.
It will be hard to overcome a lot of gaps in education
- where is a phone (he seems to have signs "phone inside"), what kind of device am I even looking for, visually?
- is this operational? did someone forget this on the wall?
- how do you operate the dial?
- do you even remember a phone number that is useful now? When the smartphone suddenly stops working?
Sadly, I would probably score 2/4 and not rely on it.
Going to this kind of resolution, the only practical use case is with cameras. You record some frames and later want to check the exact timestamp. This is even useful for a single image, not just high speed cameras.
But taking images is not just a single point in time. There is an exposure time and it is hard to measure absolute timings on start and end of exposure. With a segment clock display, the numbers will blur. For this a row of LEDs is fine - just light up one at a time. In the image you will see multiple leds on at once, giving you verification of start and stop.
But again, I needed just one clock and built it already.
The description confused me, as it describes the use of a real Lidar measurements to detect "change" in the terrain. But certainly, it can't be a temporal change before and after... to detect medieval settings in the data. Is the area still changing differentlybetween scans over multi year's? I don't think so.
I think this is visualization code highlighting natural VS. human train structures, at known locations of old settlements? Showing different approaches on how to visualize the man-made heights in the terrain.
But still, I'm lost how this could help finding new ones..
I was surprised about using dilation. I would have expected music21 to support rendering to a certain resolution/dpi setting directly and avoid rescaling the images. But from the music21 documentation this is not obvious how to do it. Rendering music to a low dpi screen nicely (pixel perfect) could circumvent some of the hardware limitations in the mid term.
Downsides: you need enough magnetic knobs and store them when not in use. If this becomes popular, will all knobs be compatible? What if you loose some? If you buy a used module? This idea somehow creates a new set of problems.
Idea: Knob with a plug!
The outer dimension of a knob is limited by how easily you can grab it. Even if used as a socket, you need the sideways spacing anyway. But the knob face is not used. Why not have the socket in the center of the knob?
How to build this? You would need a hollow knob riding on a bearing surface around the socket. I vaguely remember magnetic position decoders that work off-axis? Or maybe have multiple magnets in the rim of the knob, in an orientation that creates a uniform magnetic field in the center? The current prototype would even be suitable for trying.
Benefits: no extra removable components. Knob and socket can be used at the same time. Can you still reach around a plugged-in cable? Is this even useful?
For sending a VGA signal even at 640x480 you will need a short h sync pulse of 3.8us at 31.x kHz. You would need an audio interface without filters with single low pulses at 260 kHz samplerate. Otherwise the monitor will just not detect a signal.
You could however use h sync and v sync from the VGA output and feed audio to the rgb channels. But thiswould give a mess of wires and is far from the beautiful idea shown here to just connect the white/yellow plugs differently.
Let's say the device has a "24 bit color display". What about eye protection color shifting? This limits the color space used could reduce the effective remaining bit depth. Or maybe they do temporal dithering to get more bit depth? Or maybe the 24 bits are already achieved with temporal dithering?
It does not need to be a calibrated display, but a cheap tablet in sunlight will be worse than a color grading monitor in a reference environment.
I hope they also register the devices used and analyze the statistics on that.
But how do you test this? While the DNG specification is open source, the implementation was/is(?) not. Do I really need a copy of Photoshop to test if my files are good? How would I find good headers to put into my files? what values are even used in processing?
Maybe the situation has changed, but in the old days when I was building cameras there was only a closed-source Adobe library for working with DNGs. That scared me off.
Typically you want to pack them to avoid storing 30% of zeros. So often the bytes need unscrambling.
Any sometimes there is a dark offset: In a really dark area of an image, random noise around zero can also go negative a little. You don't want to clip that off, and you don't want to use signed integers. So there typically is a small offset.
"there is more Co2" is valid for cars burning fuel. But as soon as you recuperate (even in a hybrid) you might only have a fraction of the losses any more. Adjusting the speed more aggressively is possible, without breaking, with little loss. I totally agree that stop-and-go is annoying, but looking into the future, Co2 should not be a reason for the vehicles in 5-10 years when the research can be rolled out.
Is "slamming the brakes" still happening? Around here you have dynamic speed limit signs on the highway. In high traffic everybody then goes a little slower, but smoothly.
I suspect that if a road is loaded beyond max throughput, this method will also fail, even harder. Let me explain: I remember a graph from communications theory. With improving error correcting codes in transmission, you can get a clean signal for even worse channel conditions. But once it fails you will not have a signal any more. The better the code, the steeper the cutoff. Whereas without in FM radio, the degrade in user experience is also gradual.
So the analogy goes like this: I would expect that you could possibly load the road with another 10% more vehicles. But if one day you have 15% more, the blockage will be even worse than before. Could be worth simulating throughput for various loading situations.
Nice, but the terms were new to me. Would have helped to explain them first.