Slashdot Mirror
Facebook Runs On AI - But 70% of Its Engineers Who Use AI Aren't Experts (wsj.com)
An anonymous reader shares an excerpt from a WSJ report: AI algorithms are inherently black boxes whose workings can be next to impossible to understand -- even by many Facebook engineers. "If you look at all the engineers at Facebook, more than one in four are users of our AI platform," says Mr. Candela. "But more than 70% [of those] aren't experts." How so many Facebook engineers can use its AI algorithms without necessarily knowing how to build them, Mr. Joaquin Candela, Facebook's head of applied machine learning says, is that the system is "a very modular layered cake where you can plug in at any level you want." He adds, "The power of this is just hard to describe." Pieces of that platform are performing all kinds of "domain-specific" tasks across Facebook's properties, from translation to speech recognition.
You might as well say that Facebook's AI runs on electricity and (generously) 99% of Facebook's engineers aren't experts in electricity generation and distribution either.
AntiFA: An abbreviation for Anti First Amendment.
Are there answers to these questions behind the paywall? I'm guessing no.
Expert these days has come to mean someone with a paper credential.
Even their head of machine learning is calling it an AI platform. He's clearly one of those "experts."
Isn't this as if I were using libc or, god forbid, libc++, boost even, while not being an "expert" there? I'm pretty certain it would take me an obscene amount of effort to even replicate some of the stuff in boost, for example.
Isn't this all that modern development has been trying to achieve since forever?
Judging from the posts I saw on Facebook it was running on "No Intelligence".
Sent from my TARDIS
So this is a triumph for the engineers that put that stuff together: it can be used by non-experts to meaningful effect.
It is prepping the data that is hard. The Machine Learning Algorithms have been established for a long time. The big limiters on it have been processing power and decent data sets.
How many developers understand encryption algorithms that they use for security... this is the point of libraries?
love is just extroverted narcissism
but I'm not a fucking mechanic.
... some as simple as the fact that i like to see my newsfeed with Recent Posts first, instead of what Facebook thinks I want to see first, is beyond the capability of the Facebook AI. Each time I go to Facebook,, I have to set the option to show Recent Posts first. If the Facebook AI can't get that right, what can it get right?
Facebook like many big-data/cloud companies who implore this technology operate and profit from the intelligence of their users. Their algorithms only mine and correlate it. They don't say this because it shows the emperor nor their algorithm has clothes. The algorithm being utilized by companies as of now is one which functions on large data sets, correlation, optimization, and brute force state space traversal w/ incremental combinatorics. It isn't intelligent. It doesn't embody intelligence and they by and large prefer it this way. They prefer this as opposed to Strong AI because strong AI will not have a dependency on big data nor on a significant amount of compute resources .. both of which many of the current tech titans have built their empires upon. Furthermore, several investors have stakes in such ventures and enterprises like Elon Musk. Fearing loss of wealth or having his holdings disrupted, he further convinces the public that Strong AI is dangerous. It is only dangerous in its sheer disruptive power of current tech companies. They know what it is in so much as its potential to disrupt and destroy aspects of their cash cows.
If any regular person was ever privy to the board room conversations as to how this cloud era was born, their stomachs would turn :
> Lease/Renter class
> Re-occurring revenue
> Build it and they will come.. then take all of their data and sell it
None of the algorithms at work in Weak AI are hard to describe. They are rather simple. It's a multi-layered mess because no one stopped to think about the nature of intelligence. What they have, although a convoluted mess works because they have tons of data and compute resources... They eventually realized this dependency was a feature not a flaw : No one can compete against them unless they have the massive data stores and the compute power necessary to wrangle in 70s era approaches. So, they then went on a campaign to make everyone believe no one else could ever compete who isn't an already established player w/ hordes of PhDs, data, and computational equipment. The PhD are elated because their specialized investment in deep learning isn't nullified. The entrenched players are elated because they maintain their monopoly and an antiquated paradigm. The various media outlets are elated because they get to play lip service to the common perception which heralds and lauds established players... No need to do real investigative work or commit resources.. just play the narrative. Having cornered media perception and the industr No one funds anything outside of the Bigdata/cloud centric algorithms because it is declared pointless and a dead end... You have your occasional pioneer who actually wrote seminole papers on these approaches speak out but it gains no traction. Corporations have restructured themselves and keep strategic capital available in case any new rising star pushes their head above water... In more convoluted attempts, people create (Open)AI groups that horde funding resources, compel a potential rising star to give their IP away like an idiot, or reform themselves into a regulator body that presides over someone else's technology (fear this tech.. here, let my Non-profit regulate/audit it).
Bullet proof
Or so they thought....
The real problem with modern, "deep learning" AI is that usually not even the experts can tell you how such systems work.
The most they can tell you is:
* The model makes the choices we labeled on our training data set
* We add stuff to the training set as it makes detected mistakes
The weights in the neural network after training become an opaque fuzzy partition of the training set.
Does this inspire confidence in you? Me neither.
To a Lisp hacker, XML is S-expressions in drag.
The entire point of expert systems is to distill the reasoning process of experts so that you don't have to have one of those available to you at all times.
In artificial intelligence, an expert system is a computer system that emulates the decision-making ability of a human expert
Honestly, having as much as 30% of the users being experts kind of sounds like a waste to me.
Probably for the same reason that nuclear power stations aren't staffed with scientists. They want people to read gauges, push buttons, pull levers, etc. rather than attempting to solve every (seemingly) trivial issue that comes their way. Also (probably) costs less.
There is no XUL, only WebExtensions...
please; I do apologize (you appear to be new here; welcome, and beware of trolls) , but Slashdot's text formatting capabilities are still stuck in 1997. You must make all line breaks manually with a
tag.
There is no XUL, only WebExtensions...
That's the biggest issue with Facebook right now.
but only 1% (or some shit like that) are experienced at OS or compiler development. News at 11 (btw, 80% of all statistics are made up, including this one, turtles all the way down.)
It's about having good tools. You don't need to know the details of AI to go through the process of building a data set and making practical use of machine learning. It's somewhat like how many programmers don't know digital electronics or assembler but are still able to write software. We're far enough along with machine learning that you aren't starting from scratch for each project, it's more of a modular system and much of it can be setup and configured with GUI tools now.
“Common sense is not so common.” — Voltaire
More of the article:
'If you look at all the engineers at Facebook, more than one in four are users of our AI platform,' says Mr. Candela [head of applied AI]. 'But more than 70% [of those] aren't experts.'
How so many Facebook engineers can use its AI algorithms without necessarily knowing how to build them, Mr. Candela says, is that the system is 'a very modular layered cake where you can plug in at any level you want.' He adds, 'The power of this is just hard to describe.' Pieces of that platform are performing all kinds of 'domain-specific' tasks across Facebook's properties, from translation to speech recognition.
This implies of the 25% of FB's engineers who use company AI services, 70% invoke it via a simple API without delving into the infrastructure or tuning it themselves.
Therefore only 7.5% of FB's AI users (30% of 25%) pass the Turing Test.
Sorry to hear this. I also note that you can't edit a post. I would have made sure to correct this in future. However, I am just here to post this one thing. Take it for the fruit that it could potentially be.... formatting aside.
Also in today's non-news:
Most software runs on an operating system, but 90% of the software engineers who write applications aren't OS experts.
The real headline would be, then, something like "Facebook engineers work with Russian intelligence...
Even if it isn't Not likely to see real AI for another 100 years. What we have now are machines that have better decision matrices, but they are not intelligent. If it doesn't have asimov's 3 laws programmed in, then it will be incredibly dangerous.
...to show pictures of food and cats ?
Give me a break, FaceBook software is idiotically trivial. The single hardest task FB engineers face is how to distribute the data, and that is a problem that is mostly solved by hardware.