Straw

14 ms·

121.
by Straw 5y ago
I don't mean it as a negative! Except perhaps I could have understand what they did faster if they described it as arithmetic encoding.
122.
by Straw 5y ago
This is essentially arithmetic encoding to do base conversion from binary to n-ary. Once you have enough bits, you're almost certainly not on the boundary of a bucket. Funny how the same thing seems to have been reinvented...
123.
by Straw 5y ago
These extra features could help it narrow things down more quickly, but CF doesn't need any of this to get good quality- what I described needs _no_ features, only the observation of interactions between users. It tries to find a patte
124.
by Straw 5y ago
That's right, but we don't need to use the Euclidean distance directly, and in fact most often we don't. For example, we usually use dot product, and predict success between two users as sigmoid(A.B), then we can have A.B &gt
125.
by Straw 5y ago
If you can train an NN in a day, then you can probably train SVD in minutes/hours- so why not start with it? Otherwise you might not even recognize the symptoms of bad hyperparameters fo r the NN resulting in poor performance.
126.
by Straw 5y ago
I'm a big NN proponent as well, and I generally agree, but as so many things can go wrong training I prefer to have a baseline (such as SVD in this case) first to make sure the NN is actually doing what you want. Now, you actually do h
127.
by Straw 5y ago
This doesn't even come close to the power of (not even particularly) modern collaborative filtering algorithms. CF doesn't need to assume that A talks to B well and B talks to C well implies A talks to C well- rather, it typically
128.
by Straw 5y ago
Classes definitely make sense for some things, especially those that don't have a Python analogue, such as symbols. What do you gain by wrapping a native representation of the same thing with a class? A cons cell _is_ a 2 element tuple
129.
by Straw 5y ago
Nice work, I've always wondered if we can brute-force search for good mixers like this! Take a look at the PCG paper for a strong family of mixers, including most of what you've tried: https://www.pcg-random.org/pa
130.
by Straw 5y ago
Cute idea! It looks like several of your implementations have extremely similar styles. If you want to realistically compare verbosity between languages, it makes sense to adapt to the features available. For example, in the Python example,
131.
by Straw 5y ago
Can't you effectively buy varying levels of deposit insurance by purchasing puts on the underlying portfolio? Pay X% to limit your maximum losses in a year to Y. For example, right now you can buy yearlong puts on SPY at a strike price
132.
by Straw 5y ago
Interesting, thanks for the info. Another commenter also recommended looking at the Australian market. Sounds like Griddy in Texas. A good business idea, but they failed due to counterparty risk in the recent storm: their customers failed t
133.
by Straw 5y ago
30 minutes is already reasonably fine-grained- enough time for a gas turbine to spin up. Shorter-scale fluctuations are usually handled by reserves and even just the momentum of spinning generators. Its certainly possible that going shorter
134.
by Straw 5y ago
If we become that successful, yes, we'd have to adjust our forecasts for the effect of supplying our forecasts! Most consumers don't have access to wholesale, fluctuating rates, so we're not sure how to effectively reach them
135.
by Straw 5y ago
You raise a good point! Actually, we do plan to trade day-ahead energy futures ourselves. However, in several US markets, the markets required for responsive assets such as batteries and gas turbines simply don't exist- there isn'
136.
by Straw 5y ago
What kind of customer are you referring to? How would our forecasts benefit them?
137.
by Straw 5y ago
Thanks! Perhaps we could show one of our forecasts, and implied battery/turbine/etc dispatch plan, but lagged a day to avoid annoying the customer of said forecasts? We could even prove we had them earlier by first publishing a ha
138.
by Straw 5y ago
Definitely. Our initial work is all in the US, but we're also keeping an eye on European power markets. Although they have a slightly different structure, its the same underlying problem. We haven't looked outside the US and Europ
139.
by Straw 5y ago
Hi, thanks for the advice! How would you balance maintaining an open source presence with maintaining a technological advantage based on our ML techniques in the space? We've put a lot of effort into finding modelling approaches that w
140.
Launch HN: InBalance (YC W21) – Short-term energy market forecasting
55 points by Straw 5y ago|40 comments
141.
by Straw 6y ago
I've found -<> and -<>> very useful- they by default put things in first and last place respectively, but let you override with <>: (-<> x (foo bar) (map baz <>)) Found here: https://github.c
142.
by Straw 6y ago
Oh, neat, what's the runtime? https://math.stackexchange.com/questions/1479745/relations-o... What have we missed?
143.
by Straw 6y ago
You can actually extend the XOR trick for missing elements to any fixed number! The standard (non-XOR) low-memory solution calculates the sum of x, x^2, x^3, ..., x^n, which gives enough information to find the missing elements as the roots
144.
by Straw 6y ago
Adjudication requires enforcement at the end of the day. Otherwise, whoever has the biggest gun wins.
145.
by Straw 6y ago
You kept the house and the land- probably by far the most valuable assets. Without enforcement of property rights, what stops me from coming in with a bigger gun and just taking the land? Of course, there are limits to the protection they p
146.
by Straw 6y ago
Personally, I disagree with the whole idea of income taxes. You should be taxed for owning things, because the government enforces your property rights, not just for labor, for which the government doesn't directly contribute at all. I
147.
by Straw 6y ago
Oh, I don't know, it very well might! However, the brain's physical structure doesn't look conducive to 4096x4096 matmuls!
148.
by Straw 6y ago
Eyes only have high resolution in a tiny spot.
149.
by Straw 6y ago
Interesting, I hadn't seen this before. I find it hard to believe, as Argon2 does psuedorandom access over a large array. As soon as this array gets larger than the local GPU cache (much smaller than CPU cache), we should get pretty go
150.
by Straw 6y ago
Absolutely, modern NN architectures have been inspired by biological ones- despite their massive differences. Even in cases like attention, the modern version (that actually works in GPT-3, AlphaFold2, etc), has little in common with both t

More ›