doublekill

These systems are interested in probabilities, not in finding the causes of shoplifting. If not correlation, then what else?

I do not follow you, so I think you do not follow me. Robbers or mobility are not protected classes.

Nearly everything is correlated with race or protected social status. For instance, it is farcical to be able to target advertisements on Facebook Likes, but not on race. These systems will discriminate on racial proxies, and the question is if you are ok with that, not if those features happen to be legally allowed.

Age discrimination is also illegal, but do you really want to track grannies the same as teenagers for shoplifting?

Same group that complains when a TSA agent frisks a five year old white kid. All in the name of fairness.

It is when the quest for fairness goes too far. Predictive policing is close to this, and also causes strong reactions: complaining about negative feedback loops, but seemingly ignoring that these problem neighborhoods need more police attention, if we want to combat violent deaths and youth gangs.

Police has a limited capacity. I want that capacity, paid for with taxes, to be optimized. If that means racial profiling, because some races are more prone to crime, then so be it. That is another form of fairness. Humans employing these techniques make use of common sense, a long sought after feat that AI barely mimicks.

You do not fix societal problems, such as racism or racial crime statistics with algorithms anyway. These communities should put the blame elsewhere, starting with themselves, before they point the finger at "racist" algorithms. It is not statistics fault when certain groups are more or less likely to commit crimes.

Especially when using words like "illegally", probably without consulting a lawyer. May now be open to countersuit for accusing them of illegal behavior.

There is still creativity in the architecture, the training data, and the sample selection. None of the art is randomly created, but carefully curated by a human.

I think the 450k GAN painting was a bargain. There is no doubt that neural art will increase in quality and quantity over the years, maybe even surpassing us in stylistic insight. That painting is the first of its kind, much like owning a first-in-style classical painting from the middle ages. Either that or it becomes a worthless parlor trick (but I deem the chance of this low).

Letters are not uniformly distributed.

When poker players need a random source (for instance, deciding when to bluff) they can look at the current or previous tabled cards (or their own cards).

You can have organizational breath with smaller teams. You pair a great programmer with an exceptional programmer and it helps if they are specialized in roughly the same topics.

Admit it to yourself. Then come prepared (to the job interview or inception meeting).

Another thing that has not been discussed yet: OpenAI does not want to be responsible for the output of this model. Can you imagine the headlines? "OpenAI released a text generating AI and it is racist as hell". People should have learned their lesson after the Tay fiasco.

I am ambiguous on the issue of release vs. non-release, but the mockery and derision they faced makes me ashamed to contribute to this field. No good faith is assumed, but projection of PR-blitzes and academic penis envy.

Perhaps AI researchers are simply not the best for dealing with ethical and societal issues. Look at how long it took to get decent research into fairness, and its current low focus in industry. If you were in predictive modeling 10 years ago, it is likely you contributed to promoting and institutionalizing bias and racism. Do you want these same people deciding on responsible disclosure standards? Does the head of Facebook AI or those that OK'd project Dragonfly or Maven have any real authority on the responsible ethical use of new technology?

I am not too sure about the impact of a human-level text generating tool. It may throw us back to the old days of email spam (before Bayesian spam filters). It is always easier to troll and derail than it is to employ such techniques for good. Scaling up disinformation campaigns is a real threat to our democracies (or maybe this decade-old technique is already in use at scale by militaries, and this work is merely showing what the AI community's love for military funding looks like).

I am sure that the impact of the NIPS abbreviation is an order of magnitude lower than that of this technology, yet companies like NVIDIA used Neurips in their marketing PR before it was officially introduced (made them look like the good guys for a profit). How is that for malaligning ML research for PR purposes? Would the current vitriol displayed in online discussons be appreciated when there was a name change proposal for the betterment of society?

Disclaimer: this comment in favor of OpenAI was written by a real human. Could you tell for sure now you know the current state of the art? What would these comment sections look like if one person controls 20% of the accounts here?

If you have two great engineers in a kitchen, discussing a relevant problem, you want there to be synergy. That synergy is broken when one of the engineers has to Google how to reverse a binary tree.

It is also a safe place to work for the really exceptional engineers. They can talk freely about complex computer science, without getting blank stares or having to dumb it down. Otherwise it gets frustrating fast.

I thought this was about World Models paper. There maybe you can give them a little leeway: AI has more slack equating humans and machines.

This paper reeks of antromorphization for the hype of it. It gets you headlines like this, but is probably detrimental to the field as a whole.

Part of the Troubling Trends in Machine Learning Scholarship

> In the first avenue, a new technical term is coined that has a suggestive colloquial meaning, thus sneaking in connotations without the need to argue for them. This often manifests in anthropomorphic characterizations of tasks (reading comprehension [31] and music composition [59]) and techniques (curiosity [66] and fear [48]). A number of papers name components of proposed models in a manner suggestive of human cognition, e.g. “thought vectors” [36] and the “consciousness prior” [4]. Our goal is not to rid the academic literature of all such language; when properly qualified, these connections might communicate a fruitful source of inspiration. However, when a suggestive term is assigned technical meaning, each subsequent paper has no choice but to confuse its readers, either by embracing the term or by replacing it.

http://approximatelycorrect.com/2018/07/10/troubling-trends-...

Yes, you can create confidence estimates for both large neural networks and gradient boosting (see for instance the thesis of Yarin Gal). This covers the majority of commercial and academic applications.

ML is actually a field with very high standards for replication, in part because emperical results are currently the focus. If certain methods don't generalize to other datasets, then all bets are off: you are dealing with data that violates the IID assumption. No statistics, bean counting, or ML is going to help you get significant results.

ML is different from statistics. If you want to learn more read Breiman's Statistical Modeling: Two cultures.

AI is real and is a legit field of study, of which ML is currently very popular, so some articles conflate these two. But it is far from bullshit. If you want to learn more read Artificial Intelligence: A modern approach.

Purely from a business perspective you are correct.

Can not put a price on accessibility though (screenreader, forced IE use, etc.).

Especially for content websites I don't think there is a non-economic reason to justify non-accessible content.

If you believe the market is 100% efficient, then it follows it always has been this way. Do you really think the market is equally efficient as it was a 100 years ago? If not, then it follows that market efficiency is a gradient, and gets more efficient over time. I do not believe the market will ever hit 100% perfect efficiency. There is room for improvement and room for longer term alpha.

Another argument against EMH is to assume it is true. Market wages for quants and hedge fund employees would reflect their true value. But wages for quants are a lot higher than the cost of using a random number generator. If all trading success can be attributed to luck or chance, it would make no economic sense to hire expensive quants, hence EMH being true leads to a contradiction.

I do believe in a no free lunch theorem for market pricing. Averaged over all traders and all stocks and all strategies, there is no perfect approach that beats the market consistently. But that is of little practical value (in inference and search the no free lunch holds, but we can still use prior information to limit search ranges, focus on the profitable markets, and use approaches that worked on similar problems).

You are an expert on ad targeting and ML, because you look at (but do not click) internet ads once in while?

This is just another meaningless "me too!".

Do you know how matrix factorization works?

They need that list for max profit. Just serving ads without an intelligent platform behind it to know where and when to serve them, will hurt the CTR (which hurts both the ad tech platform and the advertiser).

Why would ad tech companies gather data and invade your privacy, and then not use it to sell more profitable ads? That makes no economic sense, but is a costly form of voyeurism. The "myth" you refer to sounds like a poorly thought out conspiracy theory.

Which advertising company oversold you their targeted ads? And how did you find out (how much did it cost) it was not true to the extend you believed?

Advertising was not made for OP. It was made for people who do not see the difference between ads and search results. It was made for people who impulse buy to improve their mood. For people who care about pictures, not stats.

Personalization tech for ad tech is top of the line. Really pushing that part of ML forward. It drives the internet with billions worth of profit. Ad tech companies can know more about you than intelligence agencies, and sometimes they are one and the same.

Targeting is changing how people vote. It is influencing social mobility. It can turn startups into money printing machines. It is not something you can debunk in a single blogpost, just because it does not apply to you.

You are the vocal minority.

X is a meritocracy, and everyone with skill and imagination may aspire to reach the highest level.

Where X is [machine learning, tennis, chess].

I think the intent behind question 3 is to find people who want to work on something that defines their life enough it makes it on their obituary.

Selecting for life passion, not temporary dreamers.

You could create a readability score from normal non-malicious code.

Eventhough I do not know what above code does, at a glance, I can tell that this is not normal code. A machine could too.

I've read discussions between Wikipedia editors and people the article was about, and they were told to not edit the article, or produce a "no original research" authoritative source. On facts about their own life. So either a kind soul should edit it for them, or TechCrunch should pick up on this story.

And then hope they don't call you something outrageous like: the grandfather of X. Or: some people call her a Y. Because that will be in the article lead as a cold-hard machine readable logical fact for the ages.

A random number generator or binary enumerator could output results that were not inherent in its programming.

Or if you pose they don't, then you need to explain better what is so special about human brains that they can produce results not baked in by natural programming.

For a more recent working definition of intelligence, see the work of Marcus Hutter and Shane Legg on Universal Intelligence.

That we don't understand or can't define intelligence is a popular trope not grounded in reality. There are entire scientific and well-established fields that study digital and biological intelligence.

DL is a subset of ML is a subset of AI. No fallacy or conflation.