RC_ITR

26 ms·

271.
by RC_ITR 3y ago
It's against narrative, but this is one of the ways SF's downtown is actually extremely well positioned, assuming the trend of office conversions actually takes off. SF floorplates are usually tiny due to historic biases of ou
272.
by RC_ITR 3y ago
Apple is 33% of Europe's smartphone market, almost the exact same size as Samsung, the #1 player. Knowing that almost all of Apple's phones are >$800, the fact that 41% of all sales in the EU were >$800 implies that Android
273.
by RC_ITR 3y ago
>Why do you then believe that cryptocurrencies are somehow different? How about Axie Infinity, which is specifically a game similar to Pokémon cards? The first two words in NFT are "Non-Fungible." Those are certainly not secur
274.
by RC_ITR 3y ago
> It's easily the biggest market for high-end smartphones in the world. Europe is around 45% smaller than US revenue for Apple (and about 25% of overall revenue) and has lost share every year since at least 2018. [0] The US and Euro
275.
by RC_ITR 3y ago
A problem with this idea is that batteries became removable to make thinner/smaller/waterproof phones. Regardless of whether or not you think that's an important thing to design for, these smartphone designers' product o
276.
by RC_ITR 3y ago
Playing Devil's Advocate - The Pokémon card is a thing. I'm not buying it because I think it is a fungible asset that I can use to transact with (the thing I bought has a physical condition that can be graded, it is unique).
277.
by RC_ITR 3y ago
I mean, sure you can work around it, but from your own link: >since the time and memory complexity of self-attention are quadratic in sequence length
278.
by RC_ITR 3y ago
Also worth noting that Sapporo went through with a rather disastrous rebrand that both made it less unique and harder to find on shelves. https://sfist.com/2021/01/28/anchor-brewing-dragged-on-twitt...
279.
by RC_ITR 3y ago
Of all the things to be proud to be ignorant of, Excel is probably near the top of the list. I wish I didn’t know what excel was.
280.
by RC_ITR 3y ago
Personally, I love the idea but the execution was very strange. Only offering automatic shifting (and lurch-y automatic shifting at that!) made the bikes basically unrideable for anything beyond a slow flat ride. I guess the Dutch never see
281.
by RC_ITR 3y ago
During training, you have to store a dot product of Q and V that has dimension Ncrt^2. That's quadratic scaling, no?
282.
by RC_ITR 3y ago
>Autoregressive decoder inference is a severe bottleneck for Transformer models due to the memory bandwidth overhead from loading decoder weights and all attention keys and values at every decoding step (Shazeer, 2019; Pope et al., 2022;
283.
by RC_ITR 3y ago
The thing about these models is compute scales quadratically with model dimensionality and memory scales quadratically with sequence length. We are nowhere near diminishing returns for either variable, so sure current models maybe scale q
284.
by RC_ITR 3y ago
For all the 'I know every number' certainty of this post, there's some weird stuff: >(Today, the pre-training could be done with ~8,192 H100 in ~55 days for $21.5 million at $2 per H100 hour.) Why flex both system size an
285.
by RC_ITR 3y ago
>Computers in 1985 were a tool that took real skill to use. Eye of the beholder, I guess. I certainly see people struggle to use LLM's so I disagree with the implication that it's a zero-skill activity.
286.
by RC_ITR 3y ago
Just to save everyone a click, this is a 2011 paper. Here's some criticism: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3230411/ Here's some discussion from the authors: https://www.ncbi
287.
by RC_ITR 3y ago
We’ve seen the top dog fall from grace in the VC world before (KPCB missing social and mobile after a truly epic run in the 90s) Is that what we are starting to see for Sequoia? Like KP, I doubt they’ll fold, but it would be interesting to
288.
by RC_ITR 3y ago
My argument is that reading is actually a very bad way to learn/communicate and always has been. We just built a very strong culture encouraging it because for a long time it was the most efficient way to do both things. Now it’s n
289.
by RC_ITR 3y ago
>no AI summary will get you the same experience Yes, but other than being conditioned to believe the experience of reading is superior, why are we striving for that experience? Doing the square root of 147 by hand is a more educational
290.
by RC_ITR 3y ago
Alternatively this is an analogue to a parent in 1985 taking their computer away from their kid. AI tools exist, we should be encouraging our kids to learn how to exist in a world with them, not some nostalgic extinct world without.
291.
by RC_ITR 3y ago
Also lines have squiggles sometimes? VC culture has really broken all of our brains. https://www.pewresearch.org/internet/wp-content/uploads/site...
292.
by RC_ITR 3y ago
This is the strongest point of evidence I have that the phenomenon isn't real - One can very easily recreate prompts and share two links from different eras, yet we never see that. My guess is that the complainers spent a lot of time f
293.
by RC_ITR 3y ago
I think it's because when you first use it, you're surprised to the upside about how capable it is and you don't care about small faults because you expected to correct for those anyway. Then you get used to this new level of
294.
by RC_ITR 3y ago
>I guess the image input is just too expensive to run or it's actually not as great as they hyped it. We already know they have a SOTA model that can turn images into latent space vectors without being some insane resource hog - in
295.
by RC_ITR 3y ago
My guess is the reason that apple invested so heavily in this [0] is because they are going to train a big transformer in their datacenter and apply it as an RNN on your phone. Superficially, I think this will work very well, but slightly
296.
by RC_ITR 3y ago
People theorize that queries are being run on multiple A100's, each with a $10k ASP. If you assume an A100 lives at the cutting edge for 2 years, that's about a million minutes, or $0.01 per minute of amortized HW cost. In the cra
297.
by RC_ITR 3y ago
This whole '8 models not 1' thing is really confusing to me. My understanding of the MoE paper is that it's 8 FFNN modules attached to a single embedding and attention module. In that sense it is a 1tn+ parameter model, bu
298.
by RC_ITR 3y ago
This article’s argument is interesting because it relies on two very questionable principles: 1) That as tools get better, operator specialization is less useful. I understand how this feels like it makes sense, because for for binary tas
299.
by RC_ITR 3y ago
> What is (not at all)OpenAI going to do? The limiter right now is compute. If we get to what you’re saying, OpenAI will be developing multi-model models with increasingly large d_model and n_ctx to do a better job of analyzing those i
300.
by RC_ITR 3y ago
> If that were true, they’d long have an iPhone with a foldable screen, and would have shipped AR/VR glasses years ago. I think you’re really confusing my point here. There were plenty of touch screen devices before the iPhone, but

More ›