eutectic

Pans themselves don't contain significant levels of PFAs, but I agree about the environmental issues.

Teflon is inert unless you really overheat it.

Probably the real value is even worse than the expected value, because money has diminishing returns to happiness.

Usually activation functions are not a bottleneck.

I think you misunderstand what quantiletransformer does; it just transforms the distribution of the data be be more normal, it's not a binning technique. To many quantiles will just result in excess noise, not loss of information.

What an inane comment.

You can efficiently maintain count/max/sum/hash/etc. for each component. It's occasionally useful.

The people who insist AI is impossible seem just as (or more) ideological to me.

I can imagine the smell.

I thought proof checking tools mostly negated the need for verified sat solvers.

Container ships are hard to beat for efficiency.

The vast majority of changes in plant genetics are from selective breeding, nothing to do with GMO.

No, the UK. But completing the square only takes a minute.

You mean a smear campaign like the blood libel?

I never memorized the quadratic formula, and I did OK.

You have a very expansive definition of 'recently'!

How effective was it for you?

Most of these examples are fear-mongering without evidence.

Any energy source that people actually use. And most of them have near-zero energy input during active use.

I suspect there's a sort of uncanny valley, where beyond a certain point bots become too complex to predict but too stupid to be interesting.

With deep learning maybe we have a hope to reach the other side where AI can add value.

Purifying wastewater is much cheaper than desalination.

Q is just energy out over energy in. It is not an economic return.

The centroid, surely.

That depends on the tokenization scheme.

Seems right to me.

I calculate melting the antarctic ice sheet to be more like 2e15 tons (ignoring the heat required to bring it to the melting point).

If this can be turned into mechanistic theory, then implement a version of it that can solve CartPole (as an easy first test). Otherwise, it's just hand waving and conjecture.

I think using the expected logarithm of the number of solutions (entropy) makes more sense from an information theory perspective. I wonder if it works better in practice.

I think the statistical aspect is an important part of the algorithm; it's trying to match the statistics of the input data. I guess you can think of it as an approximate sampler for Markov Random Fields.

The naive dynamic programming solution doesn't work because it's possible to 'backtrack'. But I used Bellman-Ford which is almost as simple and very fast.