Suramya's Blog : Welcome to my crazy life…

AI proponents claim to solve all problems just with the addition of their magical-AI pixie dust. But that claim doesn’t hold up in a majority of the cases when dealing with real world situations. The latest example of this is highlighted in Citizen.org’s report “Mushrooming Risk: Unreliable A.I. Tools Generate Mushroom Misinformation” published earlier this week where they found that: “Emerging A.I. technologies are being deployed to help beginner foragers find edible wild mushrooms. Distinguishing edible mushrooms from toxic mushrooms in the wild is a high-risk activity that requires real-world skills that current A.I. systems cannot reliably emulate. Individuals relying solely on A.I. technology for mushroom identification have been severely sickened and hospitalized after consuming wild mushrooms that A.I. systems misidentified as edible”

Some risk comes from the seeming simplicity of using identification apps. Automation bias – the human tendency to place excess faith and trust in decisions made by machines – must be resisted. Because of how these apps are marketed, users may understandably believe that identifying a mushroom is as simple as snapping a photo of the mushroom and allowing the A.I. to deliver a reliable identification.

To identify a mushroom with confidence, a basic understanding of its anatomy is required – an understanding that many casual users lack. A photo of the top of a mushroom’s cap, for example, will almost never provide enough information to identify its species with any degree of confidence. Physical features on the underside of the cap, the cap margin, the stipe (stem), and the base of the stipe all should be taken into consideration, as should the mushroom’s substrate (i.e., whether it’s growing on the ground or on wood, and what species of wood). Some mushrooms bruise when cut, such as from yellow to blue, and whether they bruise and how quickly are additional identifying characteristics. Smell also can be a key identifying feature – and, for experienced identifiers, so can taste (followed by immediately spitting out the tasted portion). A.I. species-identification tools are not capable of taking any factors into consideration aside from the mushroom’s immediate appearance.

Australian poison researchers tested three applications that are often used by foragers to identify wild mushrooms and they found the following:

• The best-performing app (Picture Mushroom) provided accurate identifications from digital photos less than half (49%) of the time, and identified toxic mushrooms 44% of the time;
• In terms of which app was most successful at identifying the death cap (Amanita phalloides), Mushroom Identificator performed the best, identifying 67% of the specimens, compared to Picture Mushroom (60%) and iNaturalist (27%);
• In some of the apps’ misidentification errors, toxic mushrooms were misidentified as edible mushrooms;

A 49% accuracy might sound ok for a first run of the AI datamodel which has no real world impact, but when you take into account that there is a 51% chance that the app is incorrectly identifying toxic mushrooms as edible mushrooms which can (and in fact has resulted) in deaths, you realize that the Apps are actively dangerous and about as accurate as flipping a coin.

My request to everyone trying out AI applications is to use that for reference only and don’t rely on them for expert opinion but instead leverage human expertise in situations where there is a realworld impact.

Source: Washington Post: Using AI to spot edible mushrooms could kill you

– Suramya

We are seeing a lot of cases (I am being polite) where LLM’s are trained on copyright protected data/images or has been trained with incorrect data. Currently as far as I know there is no easy way to fix this other than to train the entire model again from scratch excluding the problematic dataset. This is obviously not feasible and scalable at all.

Another sticky point is the Right to be forgotten which is a part of the GDPR and a few other countries. It requires systems to remove private information about a person from Internet searches and other directories under some circumstances. With LLM’s starting to infest search engines it means that in order to be compliant they need to be able to remove information from the model as well.

So it got me thinking if it would be possible to create an algorithm/process that allows us to untrain an LLM. A search across academic papers and the Internet shows that it is an emerging field of research and as of now mostly theoretical. Primarily because of the way the models work (or are supposed to work) we also claim that the models do not contain any information about a specific image/text by an artist.

Examples of ongoing Research on Transformer editing are Locating and Editing Factual Associations in GPT and Mass-Editing Memory in a Transformer. I did try reading though the papers and understood parts of them, the others kind of went over my head but still this is a research field I will be keeping a close eye on as it will have a large impact of the future of LLM’s and their usefulness.

– Suramya

As a software developer we tend to be pretty confident that software or algorithms can solve all the problems in the world because we are using ‘technology’/AI/LLM/Blockchain or whatever the buzzword of the day is to solve a problem. This is an issue because when we look at a problem from an outsider’s perspective it looks fairly simple because we don’t know enough to realize the complexity. Or put another way we don’t know enough to know what we don’t know (the unknown unknowns). As always XKCD has a comic that talks about this:

Megan: Our field has been struggling with this problem for years.
Cueball: Struggle no more! I’m here to solve it with algorithms!
Six months later:
Cueball: Wow, this problem is really hard.
Megan: You don’t say.

To be fair, computers have solved a lot of problems in the world and have had a tremendous impact on it, but that doesn’t mean that they are the key solving for every problem. There is a saying that I love quoting “When all you have is a hammer, everything looks like a nail” and as a developer/techie a lot of us tend to forget this. We look at a problem and think that its an easily solved problem and in most cases that is true during the testing in controlled situations. Once you try the same in the real world things turn out a lot more differently. For example, in a 2020 study, a deep learning model was shown to be more accurate in predicting whether bladder cancer has spread in a patient and other models also showed similar results. Unfortunately, when the model was implemented in the real world the results where a lot more ambiguous and not as rosy as we thought.

The major problem we have right now is that AI can give us information at sounds authoritative and accurate especially if it is about a topic you know nothing about because you don’t quite know well enough to identify the nonsense it sprouts. This is similar to how movies and TV shows portray technology or medical science, they will bombard us with buzz words and if you know nothing about the topic it sounds impressive otherwise you are either completely confused or rolling on the floor laughing.

We need to actually look at the problem, understand it and then start implementing a solution. Move fast and break things is not a feasible working model unless you just want to create a buzz so that your technology/company gets acquired and then it is not your problem to get it to work.

– Suramya

NFTs became such a big thing in last few years, going from millions of dollars to 95% of them being worth $0 in Sept 2023. The whole concept of a JPG of an ugly drawing never made sense to me but you won’t believe the no of people who tried to convince me otherwise.

Today I was watching Lift on Netflix and the first 20 minutes are this group of thieves stealing an NFT Oceans 11 style. It is one of the most ridiculous things that I have seen that someone would spend so much effort showing a NFT heist but the movie was scripted in 2021 when the NFT craze was starting to become insane. Haven’t finished the full movie yet and I doubt I will ever do so as it is very slow/corny and has poor acting and script (as if the whole NFT heist thing didn’t give that away).

It is interesting that all the folks who were shilling NFTs a few years ago have ‘pivoted’ to AI now. If you read the posts from Infuencers you will think that AI is the best thing since sliced bread. Saw the following in my feed and I did question the sanity of the person posting such ‘thought leadership’.

Technology Consultant’s thoughts on AI

Each influencer keeps posting things like this to make them sound more technical and forward thinking but if you start digging into it then you will find out that they are just regurgitating a word salad that really doesn’t mean much but sounds impressive. Actually now that I think about it, they are just like an AI bot that sounds impressive if you are not experienced in the that area but when you start digging into it, you find out that there is no substance to what they are stating.

The current state of AI is basically a massive hype machine which is trying to get folks to buy things or invest in companies because they are working creating an intelligent entity. Whereas in reality, what we have today is a really good Auto Complete or in some cases really nice Machine learning system. It does some things quite well but is nowhere close to being “Intelligent”. What we have now is something that is really good at extrapolating and guessing which can reduce manual efforts in a lot of things but it is not the cure all that everyone is making it out to be.

For example, Github Copilot automates a lot of grunt work while coding allowing users to reduce the time spent of writing code, but in a recent study it was found that Users Write More Insecure Code with AI Assistants. Now this might change in the future with advances in compute power, data and something that we haven’t even thought of yet. But the problem is that in the short term these can cause immense harm and problems.

– Suramya