Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Better Language Models and Their Implications

blog.openai.com

Post reply on HN
Print view View original

61–70 of 138 posts

Re: Better Language Models and Their Implications

#61

Post by zalo »

wuthefwasthat wrote: post #27
minimaxir wrote: post #16

Earlier quoted context omitted.

Hmm, I tested a few sentences and it didn't turn up any exact matches (aside to this article), so maybe I'm wrong. With a temperature of 0.7/1.0, that's enough for sufficiently random text I suppose. (the raw, uncurated generated text using the smaller model is a bit more random: https://raw.githubusercontent.com/openai/gpt-2/master/gpt2-s... )

Those samples are from the large model (GPT-2)! Regarding memorization vs. generalization, see our paper for more analysis.

Does it seem like there will be any way to go backwards from the sample to the prompt?

From a safety perspective, it would be useful to see what prompt a piece of text might have been generated with...

Re: Better Language Models and Their Implications

#63

Post by sanxiyn »

modeless wrote: post #48
cs702 wrote: post #45

This kind of "blocking-and-tackling" work is important. The authors take a well-known architecture, the Transformer[a], configure it with a progressively larger number of parameter, train it to predict the next word conditioned on previous text, using a large dataset consisting of 40GB of text scraped from the Web, and test each trained model on a range of zero-shot transfer-learning tasks. Remarkably, the performanc…

Agreed that "simply" scaling up with more compute will result in progress and useful systems, and work in that direction is interesting and valuable. But, while we may not need new architectures or training objectives to make progress, we do need them to approach human level sample complexity. Humans don't need to read through 40 GB of text multiple times to learn to write.

40 GB is surprisingly close. I estimate I already read at least 4 GB of text so far. That's just 10 times more samples. I probably write better than GPT-2, but certainly not faster.

Re: Better Language Models and Their Implications

#64

Post by sanxiyn »

yongjik wrote: post #9

> ... some believe that perhaps the creatures were created when a human and a unicorn met each other in a time before human civilization. According to Pérez, “In South America, such incidents seem to be quite common.” Man, the auto-generated text is hilarious. And uncannily good. Though I have to wonder if it's total random fluke or there's one among their 1.5 billion parameters that predict "likelihood of mythical b…

Considering that it memorized Gettysburg Address verbatim and knows Charles Darwin wrote Origin of Species, it probably knows about both South America and unicorns more than me...

Unicorn story does demonstrate that it knows South America has to do with Argentina, Andes Mountains, and University of La Paz.

Re: Better Language Models and Their Implications

#65

Post by ctoth »

levesque wrote: post #38
avivo wrote: post #31

This was only a matter of time. For the DEFCON AI Village in August I talked about the implications of this sort of tech, and how that impacts how we release "exploit" code / think about "cognitive vulnerabilities": https://medium.com/@aviv/what-does-a-world-with-automated-so... . If you are doing work in this space, either in ML research or related security, you need to be thinking about implications (also see e.g.…

I mean, the ideas are there. The scope of the project is probably too big to reproduce for now, but eventually it will be accessible to your average spammer / scammer. We will get there. We won't be able to get these tools locked, make them exclusive for a certain type of responsible AI specialists. Someone will spill the beans, the models. People with bad intentions will reproduce these results. To me, the real deal…

“Spam-filters, actually. Once they became selfmodifying, spam-filters and spam-bots got into a war to see which could act more human, and since their failures invoked a human judgement about whether their material were convincingly human, it was like a trillion Turing-tests from which they could learn. From there came the first machineintelligence algorithms, and then my kind.”

I Row-Boat, Cory Doctorow, 2005: https://craphound.com/overclocked/Cory_Doctorow_-_Overclocke...

Re: Better Language Models and Their Implications

#66

Post by sanxiyn »

gallerdude wrote: post #21

These samples are freaky good. We're approaching some threshold very, very fast. I'm not sure what that threshold is, and whether or not crossing it is a good thing, but soon we'll be there.

Sample 105

"In 2009, researchers at NASA from the Air Force Office of Scientific Research's Microwave Propulsion Laboratory held a series of contests, including one for "best brain-computer interfaces." In their prize match, teams from around the world competed to develop new brain mapping techniques for ways of collecting data from inside people and rewriting it over and over. One single original BrainBridge video colonized all channels of web-viewing in the world. I've watched many of these videos on YouTube.

Advertisement - Continue Reading Below

On one hand, if you're convinced that the next breakthrough in neuroprosthetics will be profoundly more powerful because better/faster or cooler/more reliable algorithms capture the wiring that connects our neurons, then it makes sense to equip humans with mind-reading implants. But when it comes to real brain-reading devices, there are a slew of caveats, arguments, and threats on the horizon that mean that these technologies cannot take off anytime soon. If you truly think that these technologies will be the basis for a global brain-reading surveillance state, it makes sense to act NOW. Since so much money is on the line (in computers and other components, research funding, patent and trademark rights, marketing, investor interest) making and maintaining a head-mounted-camera program is done with a certain level of speed by enterprises like Google and Facebook."

That "Advertisement - Continue Reading Below" is very realistic indeed? These technologies, like GPT-2, will be the basis for a global brain-reading surveillance state, so indeed it makes sense to act NOW, like not releasing training data. Kudos. But this will be replicated with a certain level of speed by enterprises like Google and Facebook. Well said.

Re: Better Language Models and Their Implications

#67

Post by ninjin »

cs702 wrote: post #45

This kind of "blocking-and-tackling" work is important. The authors take a well-known architecture, the Transformer[a], configure it with a progressively larger number of parameter, train it to predict the next word conditioned on previous text, using a large dataset consisting of 40GB of text scraped from the Web, and test each trained model on a range of zero-shot transfer-learning tasks. Remarkably, the performanc…

> …we do NOT need new architectures…

You do realise the irony in stating this regarding transformers that arguably made their first appearance as decomposable attention in 2016 [1] and then as transformers in 2017 [2]? It is not as if this is a vanilla RNN straight out of the 90s sweeping the floor with decades of model innovations, rather it looks like we are seeing the rise of a new, simple model category that works remarkably well – akin to how Mikolov et al. reconsidered how to learn vector representations back in 2013 [3] to ingest magnitudes more data than previously possible.

[1]: http://www.aclweb.org/anthology/D/D16/D16-1244.pdf

[2]: http://papers.nips.cc/paper/7181-attention-is-all-you-need

[3]: https://arxiv.org/abs/1301.3781

I would not call myself a model-focused researcher so I would love to play down their importance, but to be intellectually honest one should call out hyperbole where ever one sees it.

Re: Better Language Models and Their Implications

#68

Post by ipsum2 »

songeater wrote: post #35

Is anyone else troubled by them not releasing the source model/dataset/parameters here? Yes, the technology can be used for malicious means - but would argue that "DeepFaking" language is FAR less of a problem than "DeepFaking" video/photo/audio... which already occurs. Seems like they went back on their charter to share AI developments broadly ("not concentrate power") under the excuse of "safety." (These results lo…

I agree with this. OpenAI isn't particularly 'open', compared to other AI research organizations (Notable ones that open source almost all their work are AllenAI and FAIR, but I'm sure there are others).

Wonder what their excuse is for not releasing the source model or code for their DoTA bot. Surely there's no safety issues there?

Re: Better Language Models and Their Implications

#69

Post by sanxiyn »

gallerdude wrote: post #21

These samples are freaky good. We're approaching some threshold very, very fast. I'm not sure what that threshold is, and whether or not crossing it is a good thing, but soon we'll be there.

Sample 271

"ORIGINAL ARTICLE Year : 2007 | Volume : 42 | Issue : 4 | Page : 421-429

Masturbation as a strategy of parenthood: incidence and socio-demographic characteristics in young people

Santosh S. Bhatt and Regina M. Wagner1

Department of Human Circulation and Heart Diseases, Neuromed, Coachwerk-Werke, Bremen, Germany

Date of Web Publication 28-Jan-2008

Correspondence Address:

Santosh S. Bhatt

Source of Support: None, Conflict of Interest: None

DOI: 10.4103/0972-0285.10374

Abstract

A review is carried out on the activities, influence, and consumption of males versus females in childhood and adolescence and what effects masturbation is associated with their socio-demographic characteristics, experience with pregnancy and childbearing, aggression and behaviour. According to general sociological theories, males and females have different interests in interaction and stimulation during childhood and adolescence but will adjust accordingly when making their real life decisions. The current results show that males are more often receptive to prepubescent and adolescent communication about sex, reproductive role in future life and skepticism about their own genitaliating potential while females differ in behavior and preference when it comes to erotically stimulating activities or psychological stress than male behavior."

Note that web publication date and offline volume is coherent, supposed-to-be superscript 1 to show affiliations, and that the corresponding author is the first author.

Re: Better Language Models and Their Implications

#70

Post by amvalo »

minimaxir wrote: post #12

The generated text sounds too good; is it possible that the model overfit the source material (especially since the n-previous-tokens value is infinite, while other approaches like char-rnns/textgenrnn use a fixed window length)? It's something I've encountered many times while working with text generation.

The unicorn story struck me as being structurally very similar to the standard wire stories about new species discoveries. That's probably why it's the best of the bunch.
Post reply on HN