AI Collated Comments
On the Dangers of Stochastic Parrots
[One] category of risk involves bad actors taking advantage of the ability of large [neural-net based language models -LM-]s to produce large quantities of seemingly coherent texts on specific topics on demand in cases where those deploying the LM have no investment in the truth of the generated text. These include prosaic cases, such as services set up to ‘automatically’ write term papers or interact on social media..
Yet another risk connected to seeming coherence and fluency involves machine translation (MT) and the way that increased fluency of MT output changes the perceived adequacy of that output. This differs somewhat from the cases above in that there was an initial human communicative intent, by the author of the source language text. However, MT systems can (and frequently do) produce output that is inaccurate yet both fluent and (again, seemingly) coherent in its own right to a consumer who either doesn’t see the source text or cannot understand the source text on their own”
"Debunking the great AI lie - Noam Chomsky, Gary Marcus, Jeremy Kahn"
Machine intelligence is increasingly being linked to claims about sentience, language processing, and an ability to comprehend and transform natural language into a range of stimuli. We systematically analyze the ability of [an "AI" NLP system] DALL·E 2 to capture 8 grammatical phenomena pertaining to compositionality that are widely discussed in linguistics and pervasive in human language: binding principles and coreference, passives, word order, coordination, comparatives, negation, ellipsis, and structural ambiguity. Whereas young children routinely master these phenomena, learning systematic mappings between syntax and semantics, DALL·E 2 is unable to reliably infer meanings that are consistent with the syntax of the prompts
"It has happened again. Microsoft and Nvidia have built [NN] three times the size of GPT-3, the former holder of the title. However, in contrast with GPT-3, this new model hasn’t caused any commotion, neither in the press nor in the AI community. And there’s a reason for that...
Do we really need yet another biggest neural network?.. What are the consequences of following the path of scaling models ad infinitum — for the AI community in particular, and the world in general? Wouldn’t it be better if we gave more space to other approaches to AI and artificial general intelligence (AGI)?..
[NN based Large language model] LLMs have been recently entitled 'foundation models' by dozens of Stanford researchers. They argue that these ever-larger neural networks comprise “an emerging paradigm for building artificial intelligence systems.” But not all AI experts agree with this fancy title.
Jitendra Malik, a professor of computer science at Berkeley says that 'the term ‘foundation’ is horribly wrong.' Adding, 'The language we have in these models is not grounded, there is this fakeness, there is no real understanding.' Mark Riedl, a professor at the Georgia Tech School said on Twitter that “branding very large pre-trained neural language models as 'foundation' models is a brilliant .. PR stunt.”..
But 'foundation models' isn’t the only title the AI community has given to LLMs. In a paper dating March 2021 Emily M. Bender, Timnit Gebru, and others called these models 'stochastic parrots.'"
AlphaGo Zero shows that corporate R&D has starved basic research in favor of safe bets and tinkering at the margins. Deepmind's AlphaGo Zero — which taught itself to play a remarkable game of Go in just 72 hours — is an ironic poster child for this phenomenon. AlphaGo is part of a long-term shift in AI research from generating machine comprehension to "machine learning" that is just a fancy form of statistical analysis, a brute-force approach that relies on ingesting lots of human decisions and making statistical observations that can be used as predictions about the future
"Despite the hopeful attitude during the 1950s and 1960s, it was soon acknowledged that Artificial Intelligence was a much harder problem than initially assumed. Today, AI capable of thinking like a human is referred to as artificial general intelligence (AGI) and still firmly the realm of science-fiction. Much of what we call ‘AI’ today is in fact artificial narrow intelligence (ANI, or Narrow AI)"
Andrew Ng, Wired
Andrew Ng builds [ML] systems for a living. He taught.. at Stanford, built .. at Google, and then moved to the Chinese search engine giant, Baidu, to continue his work at the forefront of applying artificial intelligence to real-world problems [..].
“For those of us shipping AI technology, working to build these technologies now,” he told me, wearily, yesterday, “I don’t see any realistic path from the stuff we work on today—which is amazing and creating tons of value— [..] for the software [..] to turn evil.”
The bigger worry, he noted, was the effect that increasingly smart machines might have on the job market, displacing workers in all kinds of fields much faster than even industrialization displaced agricultural workers or automation displaced factory workers...
There’s been a lot of fear about the future of artificial intelligence. [Some] worry that AI-powered computers might one day become uncontrollable super-intelligent demons [..] But Baidu chief scientist Andrew Ng—one of the world’s best-known AI researchers and a guy who’s building out what is likely one of the world’s largest applied AI projects—says we really ought to worry more about robot truck drivers than the Terminator.
One of the big problems I see is that, over the years, we've seen not only the periodic collapse and growth of AI, but also the collapse of the science of AI - the collapse of AI as a science. Eric Horvitz who is the head of Microsoft Research said what he thinks ML is is not a science but a kind of alchemy, that is, try stuff, see if it works on the benchmark datasets, if it does you are good. End of story.
A lot of this comes from money; I heard the following comment at a conference, alchemy brings in a lot of money. Harry Shum is the head of AI products at Microsoft, he said "if a kid knows how to train five layers of neural networks, the kid can demand five figures. If the kid knows how to train fifty layers, the kid can demand seven figures".
This is what all my graduate students are doing now. They are tuning parameters on neural networks, and they are not doing science, but getting seven figures. But I think that's been very harmful for science.
"If you work in AI you are most likely collecting data, cleaning data.. evaluating with data. Data, data, data. All for a model to say: It’s a cat. The marketing power of AI is such that many companies use it without knowing why. Everyone wanted to get on the AI bandwagon. I liked the magical world AI promised and I’ve found a shadow of what could’ve been [he quit since]. We’re not even aiming at creating general intelligence anymore. We’ve settled for stupid software that knows how to do extremely specific tasks very well"
Alive
The reason AI come alive story is used more in movies these days is not bcz Holywood necessarily knows what the phuck is going on on the tech side.. They are basically repurposing the Pinoccio story for a new domain. The "inanimate object coming alive" angle.. that's all. Pinoccio for the good side, Frankenstein for the bad. Old story adapted to something new, wout any change in semantics.
Storyline preference becomes clear around the issue of identity. They usually portray AI as a unique, non-copyable thing... When it is sent somewhere it ceases to exist at the source. Why? That's not how software works!
Spoon
Cld say around topics which under weak AI category (speech recog, computer vision -basic pattern recognition-) some milesones were reached, sure. Such tech could be used in negative ways, of course, but factory automation has been killing people for centuries now. Any tool can have adverse effects. A spoon could kill you. If you fall on it certain way, it is standing pointing up, on the ground.. That's not what the phuckers are talking about. Phuckers are all about strong AI, singularity. That is the big delusion.
TIME article
"Independent of whether you believe progress is slowing or not, increases in the speed and performance of computers do not necessarily imply that we will attain strong AI soon"
Memes
