svantana

10 ms·

121.
by svantana 2y ago
They forgot to add "...in the US". Some jurisdictions have more protection [1]. In practice though, it works as it always does: do what you want, you probably won't get sued, but if you do, it totally sucks - even if you do h
122.
by svantana 2y ago
That sounds like something one of Plato's cave dwellers would say.
123.
by svantana 2y ago
Yes that does seem worrying, and might explain why they've only (publicly) drilled a few inches here & there. Maybe they could give the waveguides some outer grid or fins or whatnot to give extra support?
124.
by svantana 2y ago
I think the idea is that the heat from the radiation turns the walls of the hole into a very hard glass structure, which should be hard enough to withstand the pressure.
125.
by svantana 2y ago
It's the same on my 2024 MBP. Looking at a gamma calibration image [1], I estimate the built-in screen to have a gamma of ~1.4 (where the stripes and filled areas have the same brightness), way below the standard 2.2. [1] http:/&
126.
by svantana 2y ago
No. Trump had the largest share, not the majority.
127.
by svantana 2y ago
Yes, but o1, o3 and sonnet are not necessarily pure language models - they are opaque services. For all we know they could do syntax-aware processing or run compilers on code behind the scenes.
128.
by svantana 2y ago
This sounds backwards. All the big players in AI are getting sued while most of the minor players are flying under the radar.
129.
by svantana 2y ago
This is one of the themes of the classic 1979 book "Godel, Escher, Bach".
130.
by svantana 2y ago
Don't you also need to add to that the time it takes for chickens to get to egg-laying age? Which seems to be 4-5 months.
131.
by svantana 2y ago
Networking, in the broadest sense of the word. Gone to tech meetups, gotten to know a lot of people in my particular niche (sound and image processing).
132.
by svantana 2y ago
Besides that, AI training (aka gradient descent) is not really an "embarrassingly parallel" problem. At some point, there are diminishing returns on adding more GPUs, even though a lot of effort is going into making it as parallel
133.
by svantana 2y ago
It's not really clear from the blog post, but it seems like it was a side project. That's why it took so long. Personally, I've gotten a lot of mileage out of doing freelance work 2-3 days/week and working on my own proj
134.
by svantana 2y ago
Wait what? Human programmers produce incorrect results all the time, they are called bugs. If anything, we use automated systems when correctness is important - fuzzers, static analyzers, etc. And the "AI" systems are improving by
135.
by svantana 2y ago
True, but the jury is still out whether text generators help out with either of those.
136.
by svantana 2y ago
Yes, but T5 is at least a normal acronym: Text-To-Text Transfer Transformer (albeit a bit forced)
137.
by svantana 2y ago
That's R1-preview released a while back - the real R1 is even better.
138.
by svantana 2y ago
Looks nice, but I don't get why it's marketed specifically towards autonomous vehicle development? Couldn't this be a viable platform for lots of other types of dev as well?
139.
by svantana 2y ago
It might be solid science, but not delining data from projections is such a basic mistake that I'm disinclined to take any of it seriously.
140.
by svantana 2y ago
That's a good point. Similarly, there is resistance against cochlear implants in the deaf community. Side note: your writing is full of homophones (berth, grate, all though). While not a big deal, I would assume that modern NLP tools c
141.
by svantana 2y ago
I assumed that was money he earned from this spy business, but it's not really clear.
142.
by svantana 2y ago
Presumably the guy is distinctively not rich, or he wouldn't have gotten involved in such risky business to make money. Most of the hit songs he's been involved with are based one or more older songs, so he may not make much fro
143.
by svantana 2y ago
Couldn't you implement a bitnet kernel, and use that as a co-processor to a PC? Or is the I/O bandwidth so low that it won't be worth it?
144.
by svantana 2y ago
I think the basic premise of this paper is wrong. Very few natural signals are bandlimited - if images were, they would be no need to store in high resolution, you could just upsample. Natural spectra tend to be pink (decaying ~3dB/oct
145.
by svantana 2y ago
The training data in this case is feedback from users - reported responses. It's only logical that as that dataset grows and the developers have worked on it for longer, the 'toxic' answers will become more rare. And of cours
146.
by svantana 2y ago
The raw language models will always have strange edge cases, for sure. But chat services are systems, and they almost certainly have additional models to detect strange or harmful answers, which can trigger the "As a chatbot" type
147.
by svantana 2y ago
Yup you're right, surprising for a site so dedicated to benchmarks.
148.
by svantana 2y ago
Pretty much guaranteed. They explicitly said there would be 3 announcements, and the remaining one is almost surely MBPs. https://www.macrumors.com/2024/10/28/apple-promises-two-more...
149.
by svantana 2y ago
Geekbench clang benchmark: M4: 21k lines / (core-second) https://browser.geekbench.com/v6/cpu/8495624 M2: 16k lines / (core-second) https://browser.geekbench.com/v6/cpu/8546977
150.
by svantana 2y ago
> none of these were actual computer output Not surprising since real-time, high-def color CGI hadn't yet been done in 1968.

More ›