kolinko

13 ms·

301.
by kolinko 2y ago
If I'm not mistaken, SpaceX is already contracted by NASA to do on orbit refuelling, and one of the recent starships did initial trials of fuel transmission (internal for now). Update: - source article - $58M contract to do that - htt
302.
by kolinko 2y ago
It was a basic pattern matching, and a dictionary of patterns.
303.
by kolinko 2y ago
Plastic pollution in the oceans has nothing to do with climate change.
304.
by kolinko 2y ago
The math here is quite simple. A single cow can produce around 250 to 500 liters of methane per day through belching and farting. Let's take an average of 400 liters/day. Methane is a potent greenhouse gas. 400 liters/day ×
305.
by kolinko 2y ago
oh wow, TIL. I assumed Trump would be against any government control of Big Tech
306.
by kolinko 2y ago
Sometimes it ends up with US government splitting businesses - like Bell being split into AT&T and the others 50 years ago
307.
by kolinko 2y ago
Yes.
308.
by kolinko 2y ago
or onewheeling! https://youtube.com/shorts/kvFKIovjbWo?si=8pwlZfwtPZpHoQRu
309.
by kolinko 2y ago
IANAL, but precedents are generally not binding in Europe (Poland at least) - they just serve as an argument. From that perspective it doesn’t matter much if it’s the same or a different country. (may be wrong here)
310.
by kolinko 2y ago
I think Apple doesn’t communicate it because it makes some of the laptops with „the same” spec better and some worse. I remember that was the case with ssds some time ago - some of the macbooks had a better one, some had a slightly worse on
311.
by kolinko 2y ago
It would need to be two weeks, one week is how long Burning Man lasts, and there is very little range there :)
312.
Does Godwin's law (rule of Nazi analogies) apply in observable reality?(cyber.harvard.edu)
4 points by kolinko 2y ago|2 comments
313.
by kolinko 2y ago
What? 10 years ago we wouldn’t dream of what’s happening now. Models before 2017-2018 (first gpt/bert) didn’t produce any decent text, and before gpt2/gpt3 (2020) you wouldn’t get an essay-grade text. So you need to go back only 4
314.
by kolinko 2y ago
I would imagine these tools use exploits in the existing executables to deliver the payload, and then - if I understand correctly, use scripting/interpreted languages (Python) to run the operation. So, even if you’d need admin confirma
315.
by kolinko 2y ago
2016. Interesting read considering it was written right before the advent of llms
316.
by kolinko 2y ago
I bumped into that issue, and it caused a lot of friction between me and 3 young developers I had to manage. Ideas on how to overcome that?
317.
by kolinko 2y ago
Gmail was revolutionary at the start, but stopped innovating 10 years ago - why don’t we still have a good search engine within it? MapReduce would be invented anyway (I implemented it from scratch before learning of it’s existence). Chrome
318.
by kolinko 2y ago
Or perhaps it’s a meta comic about dual use :) A lot of the claimed s&r applications are really a thin cover for military gear. I remember trying to mentor a hardware startup on an event, and we just couldn’t find the common language wi
319.
by kolinko 2y ago
Perhaps if they are super specialised in one narrow niche. The moment they need to write code in a new language/framework/algorirhms/domain, they benefit tremendously from chatgpt. Or even just redacting their documentation e
320.
by kolinko 2y ago
The podcast about the comments in this thread :) https://notebooklm.google.com/notebook/7973d9a3-87a1-4d88-98...
321.
by kolinko 2y ago
Both have 120 hz, but iPhone has slightly higher pixel density, more nits, arguably better colors and stronger glass ("ceramic shield" vs guerilla glass). The price of the phones is similar - good point there, but it's entire
322.
by kolinko 2y ago
Kraken would cause no problems for 2.5-2.7 Taxing this is super simple, you just mark as "other" - akin to finding money on the ground. Arguing with bank - which specific bank would cause problems here?
323.
by kolinko 2y ago
Examples of such places? I did KYC/AML with multiple European banks when withdrawing crypto, and while they checked thoroughly, there were never any issues.
324.
by kolinko 2y ago
Depends on a chip architecture - etched claims 0.5M tok/s with weights that can be updated. The main constraint is with the model architecture, where it needs to be specific transformer-based model. But they claim the chip can do both
325.
by kolinko 2y ago
You can get away from that if you constrain it to a specific type of models (say attention based).
326.
by kolinko 2y ago
The chips are optimised for matmuls, but not for transformer architecture per se. With dedicated ASICS, and weights hardcoded (or stored in SRAM) we could theorically get 1 token per one cycle - so millions/billions of tokens per secon
327.
by kolinko 2y ago
What? Of course it is accessible. When you ask an LLM to estimate if a given answer was hallucinated, it converts internal probabilities into a token(s) that represent these probabilities. Here you have an output of chatGPT where it estimat
328.
by kolinko 2y ago
LLMs have no idea if they hallucinate ≠ LLM make no mistakes. The former statement is a bold claim that would require some sort of evidence. The latter nobody really argues with.
329.
by kolinko 2y ago
This has been disproven by multiple experiments, and also - there is nothing in the transformer architecture that would suggest this. On the contrary - even the most basic classifier neural networks by default produce not a binary answer, b
330.
by kolinko 2y ago
You can also buy this from Alibaba - just the real thing, not the LCD flat panels. (the real ones are 10-20cm deep). I got a few two winters ago, and they are amazing for dark and gloomy Polish winters. https://www.alibaba.com&#x

More ›