I did for years. Custom roms and all. Then over time there was just less and less of a need for me to do so, and more and more of a hassle if I did (banking apps breaking, etc).
I'd love to hear how people write tests for ML. When I'm doing a greenfield project with a new model, a lot of issues are very statistical, e.g. incorrect downsampling - the model will run and train, just less optimally than normal.
I can't put optimality bounds because I don't know how well the model _should_ train, and when it doesn't train, that's not necessarily because of an incorrect implementation. And, actually training a model for a test is quite resource and time heavy.
* the tools that operate the model.
* to make sure the fitness function calculates fitness correctly,
* simulation runs right.
* the storing and recovering of the model operate correctly - that is, data is saved and then recovered correctly and consistently.
* that the engine that runs the training operates correctly
* that shutdown and startup work
* the a crash can be recovered from usefully
The bit that's not obvious is how the low weight and short leverage of the headset affects lag.
When you are wearing a heavy headset that extends far from your face, it's not just the rendering latency and screen latency that affects the disconnect between your head movements and what you see. The headset physically lags behind your head motion because it has inertia. The total lag is the sum of the digital and physical lag. So, improving the frame rate can only get closer to the physical lag.
And, that's on top of the practically-instant pixel response of OLED vs LCD.
All that is to say that there are physical explanations for why 90 and even 75 Hz is better in practice than people would reasonably expect on the BSB. I can confirm first-hand. And, so have many reviewers.
I never did as much thinking or testing of dropout on transformers as the author, but it didn't seem to help with my "baby" (~10 million param) transformer models. IIRC the latest Llama models don't use dropout either.
Same, I was never able to debug why dropout > 5% really hurt convergence speed for my toy LLMs. I chalked it up to the models not having enough parameters to fit fineweb and just stop using it.
My intuition is very undeveloped on this, but it makes some kind of sense to me that dropout would make convergence slower, because you're ignoring a bunch of parameters in every batch. The goal seems to be to get a better, more general model by trading off some training time.
Is there a good alternative with high-bandwidth, high-quality video? I just tried Discord, Telegram, and Element - they all compress their video quite noticeably into a blurry mush.
In principle, I agree. However, I haven't had a tasty apple in ages - they're all way too sweet for my taste.
I've been thinking about how modern fruit optimize for (among numerous other things) sweetness, and whether modern fruit are actually healthy in terms of glycemic index / glycemic load / etc.
My SO is basically like "fruits _are_ filled with sugar" and they're not wrong. It feels pretty hard to make strong qualitative judgements on this stuff. Feels better than a snickers bar, surely?
The only real thing that feels kind of easy to say is that any sweet drink is probably worse for you than just drinking water. Easiest diet in the world is to just never buy soft drinks, and the extra trick is to also not replace it with orange or apple juice.
While fruits are packed with sugar, eating fruit also comes with fiber. This extra digestion slows the intake of sugar into the bloodstream, and getting fat/unhealthy from sugar comes from to much of it going through the liver. Because you may have too much sugar at a time, the liver needs help from the pancreas, which secretes insulin to store the sugar, which makes you fat.
This is biologically confused. Fructose is processed by the liver. Glucose stimulates insulin release from the pancreas. Sucrose is one glucose and one fructose, but as you might guess from the name, fruits often contain free fructose.
Fiber doesn’t slow sugar absorption by very much. It is better than, say, HFCS, but mostly because you can’t ever eat as many calories as you can drink.
Do you know if there's a good rule of thumb for how different that ratio might be? I do like having some quantitative ballpark to go along with the qualitative texture
You may be interested in the glycemic index [1] which represents how much a particular food causes your blood sugar to spike compared to pure sugar. Based on a cursory search, the GI for an apple is somewhere in the 30s which is way less than a candy bar which can be 70+.
Fruits have sugar but unless you're on a diet that heavily restricts sugar intake it's not an alarming amount. A normal sized apple for example is larger by volume than a Snickers bar but has 1/4 the calories and half the sugar.
Eating too many fruits can provide too much sugar, but fresh fruits have low sugar content in comparison with any artificial sweet food.
Only dried fruits, like raisins, dried figs, dried dates, dried prunes and so on, have high sugar content, well over 50%, so they are comparable with chocolate or candy bars.
Most fresh fruits contain only around 10% sugar, with a only a few, like grapes or fresh figs exceeding 15% of sugar (but less than 20%).
This means that for most fresh fruits you can eat a half kilogram (or a pound) per day, while still avoiding an excessive sugar intake.
Where we shop (in western Europe) there are these shiny huge apples, often each of them has a separate sticker. All ultra sweet because, well, people are often stupid and our instincts are too strong for some situations that we didn't yet evolve to handle better. Sugar addiction from early age is one of them. Human liking of sugar developed during times where sweet fruits were rare and no ultra cheap refined sugar or HFCS was discovered. Companies deliver what people buy more.
Then at the side there are big bags (~3kg) of these not so appealing small apples with various flaws. These are the ones we buy, either bio or not, and they are much less sweet. They last less, presumably less chemistry within to keep rot away for longer, which is a good (even if annoying) sign.
If you like tart apples, try Granny Smith. They are usually harvested a little bit too early, so they are often very sour and very green (the ripe version has a very slight shine of red). Their skin is thick and requires a little effort to chew. I love them.
1400 Watt of LEDs is the kind of lighting you'll find atop a tall pole in a sports stadium. That's ~200,000 lumen, or about 10x as much as you'd need to light a large room really brightly. If you put that next to a skylight, it would make the sun look dim. It's certainly not impossible, but that's a lot for a single point light source.
What are your rooms like? Do you live in a castle?
I have 2x SmallRig RC 350D [1] and Godox M600bi [2]. These are medium-spec videography lights that draw their rated power from the wall. Lux @ 3m is noticeably (10x) dimmer than the sun.
I have tripped my breaker when running the setup, so I run from two outlets on two breakers. For my current (quite large) room, I'd love to upgrade to the 5000W lights (Nanlux Evoke 5000B or the Aputure STORM XT52), but electrical wiring would be a hassle. For a standard room, I find 700W to be sufficient.
The sun is really bright. My outdoor Hue sensor regularly reads 50k+ lux in sunlight. A room in my house with 100 watts of LEDs reads ~300 lux from the sensor on my dresser.
Yeah it’s truly astonishing how bright the sun is when you start trying to recreate it at home. But my room is brighter inside than an overcast winter day outside! That was my goal, and it’s substantially improved my mood. I’d do a write up but my strategy has been “keep buying lights until it feels bright enough and distribute them around the room”. I should probably get a real lux meter I’ve just been using my phone which seems a bit off.
reply