Veedrac

74 ms·

781.
by Veedrac 7y ago
I published some minor corrections/clarifications to this on Reddit. HN won't let me edit, so here's a link: https://www.reddit.com/r/MachineLearning/comments/e453kl/d_a...
782.
by Veedrac 7y ago
Sigh. > The Economist [...] said that GPT-2’s answers were “unedited”, when in reality each answer that was published was selected from five options > [Erik Bryjngjolffson] tweeted that the interview was “impressive” and that “the ans
783.
by Veedrac 7y ago
There are often conspiracies like this around high-profile mistakes. There's a similar one that says New Coke was a PR stunt to increase sales of the original Coca-Cola. “The truth is we are not that dumb, and we are not that smart.” O
784.
by Veedrac 7y ago
Putting memory closer to the CPU, through dense and short-distance interconnects, allowing for much more efficient communication. Eg. https://www.youtube.com/watch?v=-besHp8HLxo .
785.
by Veedrac 7y ago
A major improvement to the energy cost of data movement will come from closer memories, like HBM and on-die stacked DRAM.
786.
by Veedrac 7y ago
I don't understand what's being defended here. The guy already admitted what's on the computer. Like, the EFF saying > "We store a wealth of deeply personal information on our electronic devices. The government simpl
787.
by Veedrac 7y ago
The article is less awful than the title. In short, the thesis is that ML seems only able to learn associations, rather than stronger, causal models. It should be fairly obvious that ‘curve fitting’ is a misleading category—these models are
788.
by Veedrac 7y ago
No, not particularly. But bear in mind that I'm making a distinction between Muslim terrorist propaganda (censorship OK) and people arguing on behalf of the Quran (censorship generally not OK).
789.
by Veedrac 7y ago
“it would eject some very prominent politicians from their platform. Maybe that wouldn't be such a bad thing. <crowd cheers>” I don't know how to read that as anything other than approval of political censorship. Child abuse
790.
by Veedrac 7y ago
I don't think adverts need the same freedom of speech protections that general users do. Banning factually inaccurate adverts is fine IMO.
791.
by Veedrac 7y ago
I think terrorist propaganda is a clear case where censorship is reasonable, and I believe all major platforms already do so.
792.
by Veedrac 7y ago
My opinion is based off the fact I just saw a room full of people cheer for the censorship of politicians. I'm not of the opinion that reduced visibility and annotations will solve the problem. I just don't trust people like the a
793.
by Veedrac 7y ago
From the speech: > Twitter could deploy an algorithm to remove more white supremacist hate-speech, but they have reportedly haven't because it would eject some very prominent politicians from their platform. > Maybe that wouldn&#
794.
by Veedrac 7y ago
The response to this is unfortunate. Consider the run-on effects it will have on other companies that the primary response to YC pulling out of China is negativity. Requiring companies to stake everything on an action like this only serves
795.
by Veedrac 7y ago
> Figure 2 shows this dramatic slowdown, where we have gone from doubling general-purpose CPU performance every 1.5 years (1985 through 2003) or 2 years (2003 to 2010) to now being in an era where general purpose CPU performance is expec
796.
by Veedrac 7y ago
IIRC the price is ‘millions’.
797.
by Veedrac 7y ago
As someone who has posted an GPT-2 excerpt and accidentally had people confuse it for a human comment—I thought the context would make it obvious—calling a language model ‘pathetic’ for only occasionally getting math right hardly strike
798.
by Veedrac 7y ago
A synapse is responsible both for storing its weight, and for operating on its weight. Therefore, model size and throughput are interrelated numbers. If a synapse in a bee brain fires at 10 Hz on average (my guess), total events per second
799.
by Veedrac 7y ago
WASM doesn't need a JIT, it needs a compiler. Unlike a standard JIT, you don't need decompilation and instrumentation. WASM is also designed so that browsers can reuse their JS JIT as a WASM compiler. Even from a JIT perspective,
800.
by Veedrac 7y ago
Graal is a better target for these runtimes than WASM is, true, but Python-on-Graal is still a terrible target for the web. Huge runtimes with large memory footprints and startup times don't make for a pleasant user experience. Polyg
801.
by Veedrac 7y ago
You'd lose a lot and I'm not sure what you gain. Sans a few niggles here and there, WASM is pretty great for the web. It's small, lightweight, sandboxed, verifiable, fast to JIT, has fallbacks to ASM.js/stock JS, mostly
802.
by Veedrac 7y ago
You can, it just makes for a useless result.
803.
by Veedrac 7y ago
I'm fairly sure the consensus from the hardware side, especially considering how companies like NVIDIA, Intel and Arm are reacting, is that we still have a lot of exponential growth to go, especially for AI. A factor 1000 increase in m
804.
by Veedrac 7y ago
Moore's law never slowed. At worst, one can rightly say Intel had issues with 10nm. https://docs.google.com/spreadsheets/d/1NNOqbJfcISFyMd0EsSrh...
805.
by Veedrac 7y ago
Lakefield is a 1+4 core part, it's tiny and won't need chiplets. Chiplets are necessary only once you get to core counts that won't fit on a single square of silicon. Intel have announced co-EMIB, which seems to be their solu
806.
by Veedrac 7y ago
This conjecture sounds wrong to me. 3D stacking and chiplets are complementary; going only 3D has thermal and cost issues, so it makes sense to combine both. One shouldn't replace the other any time soon.
807.
by Veedrac 7y ago
Yes, at the same rate it always has, to a T. https://fuse.wikichip.org/news/2207/tsmc-starts-5-nanometer-...
808.
by Veedrac 7y ago
They say they removed pages, not websites. Having false positives isn't a problem when you're still left with 750GB of data—quality matters more than slightly higher quantity at that point.
809.
by Veedrac 7y ago
According to Scott Aaronson this claim is incorrect. > Next he said that the experiment is invalid because the qubits have to be calibrated in a way that depends on the specific circuit to be applied. Except, this too turns out to be fal
810.
by Veedrac 7y ago
> What can a human do that a machine cannot? The ability to think like a human. The ability to see and recognize the world around you. The ability to change your behavior. The ability to communicate. The ability to be free in a society d

More ›