toxik

14 ms·

211.
by toxik 2y ago
The average Western man burns less than 2000 kcal/day, I would say optimistically 1500 kcal/day for the people around me, most of them don't exercise much. The problem with your computation is that the body is not stupid, whe
212.
by toxik 2y ago
Toasts on smartphones are worse, it happens so incredibly often that YouTube obscures part of the video that I want to see with an unnecessarily large toast to tell me I did a thing. Most notably, switching quality. Or just now I enabled a
213.
by toxik 2y ago
Aside but -- it is also aligned to not generate profane text, right?
214.
by toxik 2y ago
Yeah, it should be sentences that have low next token distribution entropy. Where an LLM is sure what the next word is. I bet people do real well on those too. By the way, I also had 5/15.
215.
by toxik 2y ago
Another little tidbit about RLHF and InstructGPT is that the training scheme is by far dominated by supervised learning. There is a bit of RL sprinkled on top, but the term is scaled down by a lot and 8x more compute time is spent on the su
216.
by toxik 2y ago
IIRC concrete aggregate should be jagged so that it interlocks with itself. You actually can't just use any old sand, you need sharp sand, which is an increasingly scarce resource.
217.
by toxik 2y ago
Have to say I agree with this, 4o is dumber in my subjective experience.
218.
by toxik 2y ago
I disagree, Python is exactly the glue language that Ousterhout was describing.
219.
by toxik 2y ago
ILP is instruction-level parallelism, if you had a hard time remembering like me.
220.
by toxik 2y ago
Eh, it’s just some sass.
221.
by toxik 2y ago
Writing fleshes out thought because writing is like talking to yourself with automatic history recording. I suggest trying to skip the middle step (writing) and just talk to yourself via voice recording or something else. It works and it ta
222.
by toxik 2y ago
You can’t take a song photo, there is no permanent one unique recording of a song that has to play forever for it to exist.
223.
by toxik 2y ago
It is pretty hard to subtract a signal like that, if it’s really just signal C=A+B then maybe if you have exactly B you can get A back from C, but with microphones picking up radio or something like that? No shot
224.
by toxik 2y ago
Transformers have finite context, RNNs don’t. In practice the RNN gradient signal is limited by back propagation through time, it decays. This is in fact the whole selling point of transformers; association is not harder or easier in near&#
225.
by toxik 2y ago
Surely you heard about surfels, sounds like a very similar approach.
226.
by toxik 2y ago
Would the climate model take its own climate impact into account?
227.
by toxik 2y ago
I disagree with the idea that notebooks and McDonalds are somehow inherently and fundamentally bad. They fit a certain niche, and you should use them accordingly.
228.
by toxik 2y ago
Fair point but I was talking about cars in general, which are all increasing in weight. EVs just helped make it palatable to have 2000+ kg tanks for personal use. Not singlehanded but certainly contributed.
229.
by toxik 2y ago
Not only are EVs going up in weight, all cars are. That does increase brake dust. I also don’t think it’s even close to offsetting the road damage and tire dust caused by increased axle pressure, which is famously a cubic relationship.
230.
by toxik 2y ago
Writing Mandarin with pinyin is way easier, you just type the first letter of each sound in a sentence. “I like to eat ice cream” is 我喜歡吃冰淇淋, written entirely by the letter sequence wxhcbql.
231.
by toxik 2y ago
That resonates with my impression too. The problem is that the two world views have a conflict in what are appropriate climate action policies. If you believe we should protect nature, that humanity is the bad thing, you want to ultimately
232.
by toxik 2y ago
Good thing we’re moving to PHEVs and other cars that have more axle pressure and generates more tire and brake dust. Wait, that’s not a good thing at all.
233.
by toxik 2y ago
Notably, humans also cannot track this to infinite depth. Somehow we know this and reach for algorithmic solutions (like pen and paper).
234.
by toxik 2y ago
This problem formulation has been around for a while, it’s kind of the holy grail of modeling. What is new compared to PixelCNN and related is this position embedding idea.
235.
by toxik 2y ago
That is indeed an issue. Their sampling method rejects impossible combinations.
236.
by toxik 2y ago
# shuffle the combined batch to prevent the model from learning order indices = torch.randperm(combined_images.size(0)) combined_images = combined_images[indices] combined_labels = combined_labels[indices]
237.
by toxik 2y ago
Seems like an odd thing not to state prominently. ChatGPT 4o solved it every time for me. Seems to work OK even with the "hands tied behind back" prompt. Wonder if this paper is just inaccurate or if OpenAI adjusted the model -- s
238.
by toxik 2y ago
Then they should have. It isn’t clear to be that this change is measurable when the code is optimized. Even in the presented numbers the difference seems meh at times.
239.
by toxik 2y ago
When my company makes 10% more profit from my work, my pay should increase 10% for it to be fair. This does of course not happen, which is my point. When the company makes 10% less, though, pay cuts are not unusual.
240.
by toxik 2y ago
Are you really suggesting paracetamol causes dementia?

More ›