aab0

> You'll find plenty of information if you actually seek it out rather than simply making a knee-jerk, snarky comment, but here is one example article which articulates some of the issues that have been under consideration in recent years: http://m.ije.oxfordjournals.org/content/41/1/273.full

I assume you are referring to

" If the seven associations that did not reach P ≤ 5 × 10−8 when additional data were considered are assumed to have been false-positives, the false-discovery rate for borderline associations is estimated to be 27% [95% confidence interval (CI) 12–48%]. For five associations, the current P-value is > 10−6 [corresponding false-discovery rate 19% (95% CI 7–39%)]."

That doesn't show anything relevant. Failure to replicate at 10-8 is a ludicrous way to define non-replication; to paraphrase Cohen, surely God loves the 10-7.99 almost as much as the 10-8... This paper needs to adjust for power, and ask how many hits one would expect to not replicate at 10-8 given the power of the replicating studies. If you do remember power, GWASes replicate fantastically, for example https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3681663/

"Replicability rates are high within Europeans, with 155 successful out of 181 attempts (85.6%), when only 9 positive replications (∼5%) would be expected under the null hypothesis of no association (binomial test, P<10−16). This excess was robust to the significance threshold (e.g. 122 observed vs. 0.18 expected if only replication attempts achieving P<0.001 are considered successful and 56 observed vs. 1.8×10−5 expected for a threshold of P<10−7, Table S5). Moreover, replicability rates within Europeans approach 100% when accounting for statistical power. For the 168 attempts for which we could calculate the power to replicate the original finding (Table S5), we observed 147 positive replications, which is almost identical to the expectation of 149.1 positive replications given that average power is 89.1% (see Materials and Methods). This is expected, since most GWAS already contain an internal replication phase [1], [24]."

> and, believe me, many GWA studies have been published with much less significant p-values

I don't think they have. Ever since Ioannidis and others demonstrated what a total debacle the early candidate-gene studies were around 2009-2011, using the first GWASes to demonstrate that, GWASes have been pretty standardly done at 10-8.

There are 3 datasets here, the 23andMe discovery dataset (n=300k), the previous Psychiatric Genetic Consortium ('PGC'/'MDD' in the paper), and then a second 23andMe heldout validation sample (n=150k).

The set of 5 hits discovered in the discovery dataset were at the usual 10^-8 threshold. They then meta-analytically pooled those results with the older PGC results and got a set of hits at 10^-5. Finally, they took those two datasets and pooled it with the held out replication dataset, yielding the final set of 17 hits at 10^-8. The final results are at the significance level you want, and almost all of the signs for the top SNPs are the same between datasets/cohorts (presumably why they reported broken-out sub-analyses rather than skipping straight to the final results, to demonstrate consistency). Those 17 are the ones used in the rest of the paper.

> Many GWASs report much less stringent p-value and many don't run sub-replications.

This isn't true. Most GWASes use the standard genome-wide significance level of 10-8. If they do not, it's because of well-motivated other considerations such as being replications of previous hits. (If you are testing replication of 5 earlier hits, rather than 500,000 SNPs, your p-value threshold ought to be looser.)

> I wonder if this study is subject to that requirement.

The supplementary gives the top 10k hits, which is the most critical part of the data, which you can use for polygenic scores, gene sets, heritability & genetic correlations via LD score regression etc. It's only top 10k SNPs because I believe 23andMe imposes that as a requirement on people using its data - something about possible reidentification if too many SNPs' values are released. (There are, of course, other cynical business-related reasons for why they might impose such a requirement.) I've seen that done in a few other GWAS studies like educational attainment, and they said it was because of 23andMe.

To continue the mortgage example, you could also just not buy a house and put that money into a stock index...

The land may not be usable for another few hundred years, but what is the opportunity cost of all that upfront remediation? It's big. That's capital that can be invested in other things.

Some land in the middle of nowhere is not really that valuable. The productivity of such rural land is low. This is why no one has bought up that land and remediated it in the first place.

> How much would that cost? Would that be more or less than the cost of refugee programs and border management and social programs for the unemployed immigrants?

That 65sqm could be anywhere; there's tons of un-mined land in Europe. It's a whole continent. There is no necessary connection to remediation. If you can't make such a 'diaspora city' work economically on unmined land, you can't make it work on mined land either.

> I want to state a feeling that I have: that we would have never done this if we respected Laotian people or thought of them as our equals. Only by thinking of them as simple poor brown hill folk could something like this happen.

So, in the comment section about an article about how two countries of white people went to war and dropped so many bombs that a century later some areas are still unusable, you are claiming that the only reason a country of white people dropped so many bombs on another country that a century later areas will still be unusable is because... the other country had brown people?

Magic Leap has also gotten a lot of buzz because tech publications are willing to write about them despite almost total lack of details. Look at the article Kevin Kelly wrote for Wired about Magic Leap: http://www.wired.com/2016/04/magic-leap-vr/ It's like 10k words, yet after all the embargoed/NDAed info has been omitted, what do you really learn from all that besides that Magic Leap is some sort of AR glasses and Kelly thinks it's really really really impressive?

It's a demonstration that even in a niche most people wouldn't think of, drug administration, there's already dozens of cost-effective medical uses, and the number is going up as more research is done and sequencing goes down.

Genes can be useful for treatment: http://biorxiv.org/content/biorxiv/early/2016/07/23/065540.f...

You don't even need the instance if it's a static site that can be served directly off S3. With S3+Cloudflare for a Jekyll blog, you can handle peaking to thousands of hits a second without noticing it at like $10/month.

"Effect of intranasal insulin on cognitive function: a systematic review", Shemesh et al 2012 http://press.endocrine.org/doi/full/10.1210/jc.2011-1802 https://www.researchgate.net/profile/Assaf_Rudich/publicatio...

It's not. I'm not diabetic, but there's some interesting research on intranasal insulin and cognition, so I thought I might try it out. At my local east coast Walmart, I bought on March 13th 2016 a 10ml vial of Novolin R insulin for $24.88. (No prescription required.)

That is quite a visual design. 'Cardinal' indeed.

Anyway, doesn't answer the fundamental question: if I am scared enough to spend $4010 (the cost of the preorder) for a drone prototype which can't fly more than 20 minutes and will probably be broken or obsolete in 10 years, why wouldn't I just install some surveillance cameras and get 24/7 coverage?

PR puff piece. Plenty of boasting about how reliable their auto chips are and how auto chips must be reliable, but little about the techniques or possible failure modes.

It seems kind of strange that he never thought of using computer proof systems for checking whether his proofs were right in class, and that he took a pattern-matching approach rather than using any of the existing AI proof-deriving systems like https://arxiv.org/abs/1606.04442 or https://intelligence.org/2013/12/21/josef-urban-on-machine-l...

It's a rubbish example. The DNMs do not get better thanks to exit scams and takedowns. They get worse, as sensible and experienced administrators depart with a fortune, to be replaced by mendacious incompetents. Nor is the technical side of things any better - they have, if anything, regressed technically on the issue of multisig.

If you use RL, you might not know it. Multi-armed bandits, for example.

Petroleum engineering regularly makes the list of degrees with highest pay...

> I don't think they're compressing this and decompressing it client side.

The speed quotes made it sound like client-side was a concern. Why would you go to all the effort of devising a new image compression format saving 20%+ storage and on the wire, and not have it decompressed client-side, especially when you control the client?

So it's just a char-RNN? I was hoping this was using one of the seq2seq frameworks for actually writing code, like https://arxiv.org/abs/1605.06640

A lot of 2015-2016 work can be seen as addressing the latter two. For memory, all the work on neural programming, soft and hard attention, content-addressable memory and 'memory networks'. For unsupervised learning, adversarial networks have spawned a whole bunch of papers with OpenAI's latest batch being quite exciting.

There's no reason Tesla/MobilEye couldn't be using an off-policy reinforcement learning algorithm, no.

Psychopaths aren't simply a normal person with some morals turned off. They exhibit a number of other differences, including some that you might consider closer to 'brain damage' than 'rational', like an insensitivity to losses or punishment which makes them show worse performance on the Iowa Gambling Task.

But how long is that going to take? Resistance has very high fitness in the presence of humans using antibiotics (avoiding death); resistance has very small negative fitness in the absence of humans using antibiotics (saving a tiny bit on metabolism). However long it took to evolve resistance, it'll take many times longer to de-evolve it. And given randomness, a tremendous number of copies of the resistance genes will hang around anyway so the resistance will come back even faster the second time.

> What is not mentioned is a new class of antibiotics discovered in 2015

Call us when any of them survive the valley of death that kills 99%+ of substances with interesting in vitro activity to reach clinical use and the hypothetical becomes actual. In the mean time, real people are dying real deaths.

I think it's more than that. The Minuses have a definite air that the others, more randomized, don't. I think it's a little Andy Kaufman like - much humor is ultimately malicious, and when you strip out the punchlines, you're left just with the setup in which Jon's loserdom can't be ignored.

(Usually we call that 'power', not 'sensitivity', in a null hypothesis testing framework.)

"In the majority of its settlements, China accepted less than one-half of the territory it originally claimed. [1]"

If those original claims are as overgrown as the nine-dash line, that looks like a huge success for China.

The genetic similarity here is estimated within the cohort; the Framingham cohort is by design ethnically homongeous to try to eliminate that sort of population structure confound. This way, the homophily is not picking up on the obvious stuff like ethnic groups. If they studied people from multiple ethnic groups, the results would be trivial and uninteresting. (This is why astazangasta's criticism in another comment is so amusing - this is meaningful because the sample is so 'biased'. This also holds for all the other studies done on Framingham.) But because only one highly homogeneous sample is studied, the chance genetic similarities of friends aren't due to just similar ethnic backgrounds, but are being caused by the genetic influence on things like SES and personality and intelligence and religiosity and hobbies and things like that. These similarities would probably also hold when considering a more unusual and cosmopolitan group, but it would be difficult to spot the increased similarity on a genome-wide basis because of all the racial differences in genomes (most of which would be nonfunctional and irrelevant to anything).

I wonder about the novelty. He doesn't give very specific dates (was 'mid-January' this January, last January, or the January before that?) but does say

"The process, including my registration at an Illumina-sponsored seminar, cost $3,100."

Hasn't Illumina been at $1k/genome now for at least a year?

"I would really love to see an analysis that did an A/B test using more traditional CF and this, and see what the revenue lift was, because "accuracy" as measured here doesn't necessarily map onto the objective that you care about in the real world."

For another approach to product recommendation with some lift info, try https://research.googleblog.com/2016/06/wide-deep-learning-b... http://arxiv.org/abs/1606.07792

A multi-armed bandit will occasionally provide 'random' items as part of the exploration phase. Perhaps that's what's going on, and not any sort of diabolical self-fulfilling prophecy.