Friday assorted links

1. Skepticism about AI and deep learning.

2.The Ethiopian Urban Expansion Initiative.

3. “Patients sitting in emergency rooms, at chiropractors’ offices and at pain clinics in the Philadelphia area may start noticing on their phones the kind of messages typically seen along highway billboards and public transit: personal injury law firms looking for business by casting mobile online ads at patients.”  Link here.

4. There is a Klay Thompson Magnus Carlsen article of substance (WSJ).

5. MIE: Brutalist cuckoo clocks.


I'm sceptical of AI mainly because I heard all the same miracle-predictions decades ago when the subject was sometimes called "machine intelligence".

Granted that computers are now gazillions of times smaller and faster, but the ability of charlatans and naivety of mugs is probably little different.

6 paragraphs, maybe one of them might seem convincing to you:

are not we all mugs? (we aren't, that was just a question)

what did Einstein do, after the age of 40 or so, with which one should be impressed? he spent almost every single day of his life from approximately 1925 on charlatanizing ... seriously, have you read some of the clueless things he wrote about back in those days? and if the stuff he wrote was so far off its feed, imagine the stuff he thought about and did not think was worth writing down: it boggles the imagination to think about, as it should (as it should boggle ....)

well I have some AI friends and I am going to stick up for them.

and for their benefit I will say this "no this is not analagous to the sad asymmetric kneecaps loserdom strings of words", not at all, I (a rube, perhaps, and perhaps not, a mug, definitely not, I have know many a mug in my day and I am not one)

God can make sons of Abraham from the rocks on the side of the road: I predict that mugs and rubes of today who think "AI" is the real deal will reference those lines again and again not too long from now

just saying, math is not hard, caring about math is what is hard

if you reply please explain in detail where Einstein went wrong.

Rejoice our hearts. Thanks.

Stop underestimating people.

imagine that you too have an AI friend (those nights we spent wearing out the moon with our discussions of Epictetus and Salinger) and reply on behalf of that friend, or don't bother

Einstein tried to put the Faulkner in a petri dish. I ask you anonymous if anyone has felt at home in a petri dish. That he had a kindly expression is not nearly sufficient not even just cause for his treatment. In this way he is like Pluto perhaps, where we can see the Elysian Fields. But Faulkner ate the grass first. But Faulkner did not live in infinite silence. In that he did not offer hope nor alien dust.

but Faulkner, to give him credit, never went to bed at night if he had not , at least once during the day, pondered on the difficulty that people like himself suffered in a world where the love they offered was not accepted.

Kind of sad, when you thing about it. Dude could have been the light universe version of dark universe Cruella de Vil, not that he did not have lots of good paragraphs here and there amongst the sounds and the julys and the melchisedeks and the (my favorite) ***flags in the sky*** but he just missed being the light universe version of the dark universe version of Faulkner : trust me -

but one gives him props, see what he did there, flags in the sky? and who looks at flags in the skies and does not feel hope?

Yes I know it was "Flags in the Dust" but we all have friends who are basically idiots and we all give them slack

as for the petri dish - think about the numbers. Remember that night we walked up Fifth Avenue and laughed at the upward links in the trends of the street addresses? I could do that again for a friend whose sad afternoon was concentrated on his care for a petri dish, it would not be that hard, near infinite numbers are almost everywhere, remember that scene in that novel of Bernanos where the unrespected priest almost brought the recently dead individual back to life? I remember ...

just saying, once a nurse yelled at me (can you imagine a nurse yelling at me of all people, that is like cursing at a friendly chihuahua, chihuahua dog to be exact) for telling her that she had been rude and that people in hospitals are not petri dishes; well, she is old now, and undoubtedly knows better, and misses those days when people who cared about other people were a mystery to her: not that she wants to be ignorant again, not at all: forsan et haec olim miminisse iuvabit , as I used to tell my friends in Pleasanton

everybody has their story

even people who don't like me

or my AI pals

Flags in the Dust, that driveway, that confusion that the friendly friend of chihuahuas ( a rather impressive breed of dog when you see five or more of them gathered together) who writes these lines would have fixed by simply introducing him to her and her to him (how many times did she fall for his lies? not even once, if someone had said something to her)


Matthew 3:9, find some good commentaries (Henry, McGee, Challoner, Navarre, all good but not good enough) and then use your God-given intelligence to imagine even better commentaries .... imagine I know what I am talking about. I can't convince you right now but you can imagine it, can't you?

there is so much that is not greed

that was more than 6 paragraphs, sorry if you spent an extra seven or so minutes out of the millions of minutes God gave you after you innocently thought my originally promised 6 paragraphs involved only the investment of half a minute (sorry!, as we say in English...), well here is my favorite from the original 6 paragraphs:

Matthew 3:9, find some good commentaries (Matthew Henry, Vernon McGee, Challoner, the Navarre commentary - translated from the Spanish and published in English by Ignatius Press, headquartered in San Francisco) and then use your God-given intelligence to imagine even better commentaries ....

imagine I know what I am talking about.
don't worry that, if you figure these things out, you will be the sort of person who feels obliged to say "hey those annoying Christians were right and I was wrong" - it does not work that way. Look I like poor little Pope Francis but he is more of an idiot child than a real Christian, ditto for poor Billy Graham and his likable but too distant wife, ditto for most of all the rest of the famous Christians you have probably heard of, whether I can cleverly say good things about them or not (see what I did there?), or with even more cleverness, although the nasty kind of cleverness that I would never wish I were gifted with, whether I can cleverly say nasty witty negative things about them or not ... (I try never to do that)

For God's sake never miss a chance to understand God.
Maybe I am wrong maybe right here right now Matthew 3:9 will mean nothing to you, not even if you think about it alot

ok then try Romans 8

or my real favorite

the one that broke my heart all those years ago

we can help God

we can offer up our sufferings to help Him

Saint Paul explained it

look it up

or find another verse to be your favorite verse, you don't want to be the person whose favorite part of the Bible is the part where Paul says that Jesus not only wants us to be His friend but also is grateful, in a loving way almost nobody understands, when we suffer for Him because he only lived like 33 years and some of us are going to have to live a lot longer than that

before we see the Elysian Fields

before we walk in those fields

No, be the person who accepts Jesus in your heart at a time when you can enjoy this world

I want you to be happy

andrew', I hope you are reading this

Colossians 1:24: if that is already your favorite Bible verse, please please pray for me, God's love is strong and you know so much more of it that I ever will: if it is not your favorite verse: I hope you find another verse that means more to you, today, tomorrow, every day you live on this earth. But if for some reason you don't find such a verse, and you eventually decide that Colossians 1:24 speaks more closely to your heart than any other verse: well your prayers will have power you probably did not imagine, back in the day. as soon as you can, pray for my friends. then pray for me, but pray for my friends first. Thx.

The works of Hinton, LeCun, Ng, etc. are vastly overread. Their methods work extremely well for some narrow applications, but they offer no deep theoretical insight, certainly not enough to break into artificial general intelligence territory. If I had to guess, the next revolution in AI will come from the works of Judea Pearl.

If you know better than Hinton, LeCun, and Ng then why aren't you richer than Sergei, Larry, or Zuck? You must have a 225 IQ. You must also donate your brain to science it must be full of innovative genius, brilliant and witty insights galore. Artificial intelligence is no match for real intelligence, eh, DF?

There is much hype in regarding the state of AI, sure: We can do a lot of things that were impossible 10 years ago, but the compute required to do this kind of work in, say, genetics, is just not available. This is why you find the Monsantos of the world dedicating their AI budget on agronomics, weather and yield prediction, instead of on trying to figure out deeper things about genomes.

That said, there's plenty of problems that meet the right size, and where we have more than enough training data to do well. The author talks about self driving crashes, but those are limited to approaches that are clearly broken(say, Tesla's), and people that just shouldn't be testing yet (Uber). I'd be far more concerned if we saw actual industry leaders crashing in dumb situations.

The uses of machine learning that are really interesting for experts are rarely the ones you are seeing in the media. It's not that we can solve problems that were unsolvable by any means before: It's about simpler problems that used to be really hard, programming wise, and ML can now solve cheaply. Instead of dedicating programmers to solve them, they can work on higher value problems. Look at it as optimization of human brainpower, which is one of the most precious resources we have.

It's interesting because it should have been apparent that self-driving is hard - and engineering management in Silicon Valley should have had the chops to know that.

To put it in terms of SAE Autonomy Levels, levels 1-3 are very different than levels 3-5.

Levels 3-5 require the kind of complete/reasoning AI that has never been built.

Yep I completely agree. I think the last wave of AI advances still has not been digested at all by the broader economy. Big companies were betting for the next big thing with AI but it is really marginal gains in a lot of different areas like meteorology, consumption efficiency, translation, better written ads or articles,... Just think about the radiology example. Even if the AI does not do as well as a trained radiologist, it is still a double check system that will lead to marginally better diagnoses at the end. But for the patient it is not marginal.

Agreed. Building a model out of 4,000 variables is impressive but it's still a very long way from a model that can deal with the real world. In the real world we grasp, outside of the things physicists study, at best only vague outlines of causality, struggle with capturing variability and keep finding the systems we want to understand (e.g. cancer) to be far more complex than we imagined.

#3. It seems bizarre to me that people would put up with receiving unsolicited ads on their phone at all. Is that normal ? I'm reminded of the early day of the internet when pop-up ads were a serious problem, and everyone had to install pop-up blockers.

The Pitfalls Of Social Media Advertising
Prosecutors there reached a deal last year with a Massachusetts-based digital advertising firm that was sending advertisements from a Christian pregnancy counseling and adoption agency to people who entered Planned Parenthood clinics. When patients would go to the clinics, they'd also cross a digital fence and soon get advertisements such as "You have choices" and "Click here for pregnancy help.""

LOL, well that guarantees that the Massachusetts State government will be putting a stop to it. I doubt they cared nearly as much when ambulance chasing lawyers were pushing out ads.

If you read the plans from the public space advertisers (billboards, bus stops, bus sides, etc.) they would like to use your phone to ID you, and access your online profile, and then deliver a custom crafted display ad to you as you pass by (or perhaps as the bus passes by you).

Apocryphal advertising story: Madison Avenue guy says "I know only half my advertising works, the problem is I don't know which half". This sort of bespoke digital advertising might eventually answer that question.

In fact, this is basically how internet advertising works. The ad networks track you using a variety of tricks, and then use that knowledge to sell the right to advertise to you at a higher price. ("50-year old Asian male who has recently browsed ads for cars" gets a better price than "User in Delaware who is on a page describing electronics."

The article is poorly written, but I think it's just the standard ads that you get in your news/web browsing/games/whatever, but now the ad network will deliver you other subjects.

If you aren't getting any ads already, you aren't going to start getting new ones now.

#3: Given the wide prosecutor discretion, and a set of laws that are often logically inconsistent, but with plenty of punishable acts, those guilty of leaking one's location real time could easily be subject to felony charges if a prosecutor actually wanted to do something good for society, as opposed to shakedowns of oil companies, aka looting 401k accounts:

The side effect could potentially be better data security, given cell phone providers now literally would have skin in the game.

the reason that eating Ethiopian food is good is because its both a cheap and bland. Water at Ethiopian restaurants tend to lack ice so people end up having better conversations. The stegasaruous was born in Ethiopia as were many other minotaurs and minstrels.

Will the BLS have a spine? It should treat the White House the same way it would a news outlet that tips off the market that impending BLS report will be better than forecast.

The BLS is subordinate to the President, not the other way around as you seem to believe.

Unelected bureaucrats have no business flouting their masters regardless of whether they happen to think that their masters make poor decisions. What's more, violating another norm in response to the president's violation of a norm serves only to weaken all norms. That is the great temptation and danger in reacting to populist politicians.

Of course we do. We are here to protect *the system*.

+5 year plan.

Ah a Commie inspired Cuckcoo clock, complete with hand grenade weights. Awesome.

The Cuck Clock is something I suspect every reader of this blog owns.

Not so! My wife won’t let me own one.

self-deception is best filled by wish fulfillment. Look at the press the tefalin gets. Meanwhile, the woodpecker is left to his impossible yet valid methods. It is an odd and noble choice to wrap a fish in tin foil but is it quite so distinct?

I was just going to say the same thing.

2. Sounds exactly like what California needs.

Why did these luminaries stop tweeting? Perhaps it has something with each of them now being employed at very large software companies where their time is in very high demand? No. A "researcher"/Post-doc knows better. He understands the demands of corp-VPs/Technical-Fellows and know that they must have time to tweet for his enjoyment.

DId anyone say that progress, as measured by "success"/accuracy would be linear in compute? The entire reason we are having success is that our exponential growth in compute has allowed us to make great strides where it was impossible before. Now if this "researcher" wants to talk about the end of moore's law that's potentially valid, but that's not the claim at all. Worthless.

As for self driving cars, the advances have been remarkable -- and they have been accelerating. Many in the industry have high hopes, although they all recognize the complexity of the software that they are designing.

In summary: buzzfeed is looking for writers just like "Filip". "5 reasons why AI is overhyped. The last one will shock you!" Piss off.

Shut the fuck up you cuckold

You probably own a CUCKcoo clock to tell you what time the blackman is showing up today to cuck you.

1. I could have told you all of this five years ago. The problems with the traditional approach to AI are well understood by many in academia and longstanding. Nevertheless deep pocketed industries keep throwing money at it as if more money and more computing power will make the scaling problems (amo n g other issues) go away. To get to the interesting kind of AI the field needs to be reconstructed from first principles based on cognitive science , stuff like this:

Why didn't you use that humongous brain of yours years ago flipping AI startups to the lessor minds at Google or FB for billions of dollars? You apparently know more than the phds at cutting edge silicon valley companies.

It was possible, back in the day, to be correct that putting on bird wings flapping wasn't going to work -- while still not being able to design a working airplane.

I happen to think that deep learning is an impressive technology and that the people who invented it are, indeed, very smart. But I have also believed (and argued on MR) for years that machine vision is still nowhere close to human levels and, because of that, autonomous vehicles are not going to be successful. I expect to see companies giving up and scaling back on R&D before it's ever possible to go to your local dealership and buy a true mass-market self-driving car (one that handles level 3-5 automation in unrestricted environments).

So why I am not prepared to get rich by shorting autonomous vehicle companies? The same old reason I'm not shorting Tesla even though I've long believe it's ultimately going to do a face plant ('the market can stay irrational longer than you can stay solvent').

because I'd rather work on AI that will actually work than scam people, maybe? Also I have other projects going on and I have limited bandwidth.

#1 AI is headed for a winter just as much as the internet is headed for a winter in 1995. (There was a bubble there!) The current state of deep learning will be just as trans formative, if not more so.

The article is really flawed.

A. Deep Learning (does not) Scale
Let's just presume for an instant that deep learning did scale. What do you expect to happen? You would expect a mad dash to create acceleration hardware. You would expect google to spend enormous sums of money to scale their deep learning hardware. You would expect the most common provider of deep learning hardware to have a sky rocketing stock valuation.

All of this is happening. There are tons of hardware startups. Google is creating a pod that can harness 100 petaflops of compute. nVidia stock has gone up 10x in < 5 years. It is quite exciting.

B. Linking to a tweet about a medical imaging competition where the baseline model only beats one or two of the highly trained radiologists.

Oh no, such a failure. Of course, this model will eventually be crushed. It always happens. Not to mention that we are just now beginning to scale up data. No radiologist will stand a chance in any area in 20 years. They won't be able to work without these tools in most areas inside of 5 years.

Deep learning has already solved many many problems. We don't even know what it can't do. (yet) Problems are constantly being solved using deep learning. Deep learning is applicable across nearly any problem domain where the data and solutions are highly sparse. ie anything dealing with the real world.

Programmers will be kept busy solving the problems that we know can be solved with the state of deep learning as it is now for years if not decades to come. The industrial applications are endless. (I am working in this area now.)

And the breakthroughs are still coming at an incredible pace. Far faster than at any other time in the history of computer science.

A winter? No. A bubble? I don't honestly know. I still expect fully autonomous cars in select cities (with rapid expansion) within five years.

I agree with the article, and it echoes the insecurity that is developing within the field.

I think this one of his posts highlight the problem with deep nets most efficiently:

Their result is unlikely to suprise anyone well versed in the field, but it is a very visceral demonstration of the current state of AI.

Statistical methods are extremely good at producing the illusion of intelligence while they operate within their trained domain, but do not make any attempt to reason or model their environment. Its not a question of computational power or layers, its just not the correct tool. Its like someone has noticed that linear regressions can accurately predict some time series, then concluding that a sufficiently complex linear regression will someday write a book.

As he writes, statistical models can perform spectacularly well 99.9% of the time, but the 0.1% of cases that weren't included in the training data tend to be the important ones when you are interfacing with the real world.

1. Skepticism about AI and deep learning.

Used everywhere. Search engines, credit ratings, hedge funds, product marketing, election rigging, ad placing,...

The list is way longer: subtitling, translation, customer service, HR, legal research, radiology, cardiology, motor design, electric grid management, airplane route planning, meteorology, web assistants, waste sorting, parking fines,... It's true that AI did not make anyone billionaire but it created a lot of millionaires.

#1 If the share of the value you can capture from an invention is low does not mean the total value created for the economy is low. There is no obvious relationship. The inventor of the kebab made no money despite the kebab industry being the largest employer in Europe.

1. Haters on this article make the exact error it (and similar articles by AI researchers) seek to address.

Current AI can be impressive, fun, the basis of huge profits, and yet the depth of the problem can still be great.

To say it differently, self-driving can blow our minds, and still not be 50% complete.


I mean, lets try this same comment as unedited voice acquisition. Will the attempt below use meaning maps to get the right word by context? Or will they do one pass and punt?

One. Space heaters on this article make the exact are it open parentheses and similar articles IIA researchers close parentheses seek to address new line current way I can be impressive, fun, the basis of huge profits, and yet the depth of the problem and still be great.
To say it differently self driving can blow our minds and still not be 50% complete.

That was with a fairly recent and high-end Asus Android tablet. One interesting thing to me is that it did recognize "new line" but only one third of the time. On open and close parentheses it's not even trying. Feel free to try your device for the part above "=====" and see if it does better.

(It occurs to me that the people who have an Alexa in their home might be getting first-hand experience on the limits of AI.)

Crypto fan boy / girl assures us that if you fix the money supply and the economy grows we will have inflation cuz da rixh gunna spind moneyz.

"Making whatever assumption about the deflation/inflation rate in a fixed (money) supply system doesn't make sense either. If wealthy enough investors spend their money, they'll cause inflation like Mansa Musa did in Africa."

You sound like a Marxist envious of the hard work and good fortune that crypto-millionaires earned for themselves. It must be fun to stew in your own resentment. #hodl4life

#1. There were a couple of things interesting in the linked article. First was the preliminary findings in the AZ fatal accident. The 'AI' saw the pedestrian 6 seconds before it hit her. Those 6 seconds were apparently spent thinking about what to do before deciding to hand-off the problem to the (inattentive) driver (with about 1 second before impact). Criminal charges should be brought, imho. The other two things were the two (deep-learning critical) articles linked there by Gary Marcus ( and [note that the arxiv pdf is brimming over with typos and errors in prose and writing]. I find it interesting that Marcus's company (involved with computer vision) was purchased by Uber - and wonder if he was involved in writing the code that killed the pedestrian in AZ. (If code can be accused of killing anyone...)

> In my opinion there are such signs visible already of a huge decline in deep learning (and probably in AI in general as this term has been abused ad nauseam by corporate propaganda)


Comments for this post are closed