Risks of Large Language Models (LLM) 

Thank you for taking these notes and sharing 🎉❤

Thank you!! Wish you all the best!

What do you mean?

A lot of these models fall over when applied in the real world because almost all of them assume stationary data... So even things like the shift in word use frequency when ingesting the Google books corpus are problematic. As are the attitudes towards other sexes and races, etc... In those writings. Watson did well on the game show, but didn't do so hot at their other hopes for it. It largely failed to make headway on biomedical problems and pharmaceuticals and was unable to generate the profit IBM had hoped it might. The people that work in these domains understand how their data works and the potential for sophisticated models to fail in actual application, hence why linear modeling is still so common for them. It isn't just them being Luddites or dinosaurs, it's that it isn't their rodeo. They've been through building sophisticated models and watching them fail to improve real diagnosis rates on new incoming patients because they have more systemic but shifting (non-stationary) differences in their mean and standard deviation, etc... that the models spuriously relied on. I see the same thing all the time now. It goes like this: 1) They gather a nice curated data set 2) They split into test and train 3) They use train to build the model, test on test and results are great, seems like it generalizes to test data. They say, "let's use this for real in the clinic!" 4) They try it in the clinic and results are not necessarily awful,... but they're below the benchmark heuristic or standard measure previously in use. The new data is messier and had systemic differences from the training/testing data, and again, these characteristics drift over time, so the longer they try to keep it going in the clinic, the worse things seem to get. 5) The publish that good part from before they went to the real world use case and the real case is a failure so they don't even try to publish it. People get more hyped and gassed up on the sophisticated new methods. 6) Eventually people become disillusioned, but not until failing themselves. In the macro scale you get things like the "AI Winter." It's difficult to scrutinize the model data because it's so large. Some are also being scraped without appropriate permissions, etc... I absolutely agree that transparency and auditing are crucial in this process.

Just going psycho.. because they all left behind. Google Brad itself atleast 10 years behind the GPT-3.

​@@bibinkunjumonYou haven't got a clue.

Watson was a very, very different sort of model. Most of these risks didn't apply to Watson because it was far more limited in every aspect. The model didn't use attention or trasnformers, was on a much smaller training data that was much more curated, and if I recall correctly the natural language phrasing of Watson responses was separate from the part that actually generates an answer. Watson wasn't trained to output probable text, it was trained to output factual answers that were then converted to a natural language representation. It's been a long time since I read much on Watson though, so I may have some errors in my memory regarding it.

Chat-GPT and similar Language models tend to let the language model define the answer. It doesn't have true comprehension of what its saying, it's instead just a very fancy autocomplete keyboard. Look up NLU vs LLM for better understanding Chat-GPT has an enormous LLM, and until they shed more light on the data that trained it, the prevailing understanding is Chat-GPT leveraged just having an enormous volume of training data with likely an enormous blend of high and low quality data; relying on frequency statistics to hopefully get to the correct answer at the end of it. Watson and other trained ai's are more train with only a source of truth that is factually correct, and therefore will give answer from those sources of truth. They are more NLU's. Lately Watson and other trained ai's use much smaller LLM's mostly to generate responses that way it feels less like an Alexa answer and more like an ai. My tinfoil hat theory is that Chat-GPT always planned to be bought out OR buyout a good NLU and integrate it into their system eventually eliminating hallucinations.

@@NerfThisBoardGames "very fancy autocomplete keyboard". Thank you! I was looking for the great analogy for ChatGPT and you gave it to me 😄

Hmm interesting

Explain for dummies

This information is already elsewhere and I don't think IBM is going to be hurt by that scenario lol

Cynical

See ibm.biz/write-backwards

🤣🤣

That's fine but what's the problem with this video?

Dude. Decades ago the tech was digital watches and calculators. Now you have chats posing as humans and being used as educators while simultaneously working like a drunkard on drugs. So yeah IBM needs to speak of ethics, law and politics of tech. That's what a responsible company does. Unlike the OpenAi joke.

No, China does not take an objective approach while others can. China doesn't make a good faith effort to figure out facts or uncover bias. Note that she also brings up audits as well, ideally by external unbiased auditors.

This!

Most people don't meet the bar for being reasonably intelligent. It's sad, but that's the way it is.

It seams you didn't understand what she meant with bias.

Haha, if it would auto improve you would know it by know. World domination in 3 days.

That's the thing, LLMs as stated in the video are huge statistical models that do not posses any actual understanding. A simple example would be feeding it with large amounts documents of false (or even partially true) information (or even grammatically/syntactically wrong information). The AI model consequently would spit out gibberish or false information at best. And that's the main problem with statistical models: 1) They are as good as the data that are being fed, 2) There is no actual way for any AI (statistical model) to distinguish truth/accurate from false/inaccurate data. And that's because they are not even close to possessing anything like understanding or reason. As for a GPT-x that gets better by itself still falls to my 2nd point. One major factor in making all these assumptions is the anthropomorphization of such models as a result that it is being fed with human data. On the other hand with game AI like Deepmind's AlphaZero, where the goals and the environment is relatively simple, the models surpass human capabilities and the models "act more like bots/computer" (meaning that certain actions cannot be understood by humans). Now back to LLM, until the time comes where an LLM or any derivative model has like reasoning, we are safe from a world ending AI that can turn itself to a terminator. We should be far more worried with the use of LLMs as a tool rather than an AI overlord.

Moreover, there’re some recent papers showing it’s very easy to “spoil” the results with relatively little malicious data. There’s even a suspicion this has already been abused in the updated models used by GPT. The trustworthiness issue also gets only harder with more data, so this is not going to get any better, it will only become harder to detect.

Nice word soup 🍲