4 ms·

One big concern I have with this is that the state of LLMs because of issues in where it can go horribly wrong, requires someone who is familiar with the subjec

One big concern I have with this is that the state of LLMs because of issues in where it can go horribly wrong, requires someone who is familiar with the subject to evaluate their output. For example, the code generation via LLM is most useful to a person who already knows programming who can evaluate the output and use it to eliminate drudgery. Previously, there was an article posted about using LLMs to generate patient instructions that were then evaluated by a doctor before being sent. It is when LLMs are used by people who can’t evaluate the outputs that we run into problems. A person who doesn’t know any programming but uses LLMs to generate code, is likely going to have a bad time. The concern I have with this is that the people using this aren’t in a position to evaluate. Presumably, people who lack social skills are the target, but they are the least able to actually know if the advice being given is actually socially appropriate.

Shrezzing 3y ago
I think the benefit here is mainly in lowering the barriers to entry for this kind of conflict-resolution training. Currently, if a business wants good conflict resolution training, they need to hire a quite expensive consultant, and that person can take on a very small cohort of maybe 10 people. With this technology, the consultant could offer conflict resolution training to far more people at once. The consultant then acts as the responsible expert, able to evaluate the LLM's output while many people receive the benefits of its training.
kordlessagain 3y ago
Giving the underprivileged tools to work on themselves is a huge win for society, regardless of the fidelity of all interactions. There are a lot of bad therapists in the world, and not all of them are going to give good advice all the time. But, bad therapists can still provide some information that may be useful to the person learning. Why would we expect perfection from interacting with an LLM instead of a bad therapist? Why be overly concerned about the possibility of introducing "bad data" in such a session? It could be that bad data then gives way to an opportunity for a leap of logic for the one using the tool. Without critical thinking present, it is unlikely anyone is going to learn to be adaptive over the long term. Obviously, the use case matters. I'm not saying to give such a tool to someone who is mentally ill, or incapable of critical thought. In general, I think expecting perfection from LLMs is unreasonable. If someone wants to learn to code, the second they have any interaction with an LLM about coding is the moment they begin to learn. I'm not sure about y'all, but a lot of times when I'm working a hard coding problem, I find myself going through confusion, frustration, interest, and then aggressiveness (anger+interest) to get through the challenge. I frequently tell people who are starting out coding that coding is one of those things that you have to get used "being confused about" to over time. You have to get used to creating problems for yourself, then coping with the feelings that come, and pushing through to a solution. I do think you raise a profound point about ability to evaluate the fidelity. I would argue that some people will eventually develop critical thought to apply to the problem, but if we instead see a "digital tragedy of the commons" happening, things could go sideways pretty quickly.
- maroonblazer 3y ago
  > You have to get used to creating problems for yourself, then coping with the feelings that come, and pushing through to a solution. A great description of any creative act, really.
- FranklinMaillot 3y ago
  I think the reason why LLMs pose a much higher risk is scale. A bad therapist will impact a limited number of patients, whereas an LLM deployed at scale will affect possibly millions of people. In matters of health and safety, the smallest error rate could have devastating consequences.
Morelesshell 3y ago
The concern is real but you know what? The last company i was part of, had a TON wtf moments. Like database unprotected on the internet (very short password too!), md5 for passwords, api key in the apps...

Shrezzing 3y ago

I think the benefit here is mainly in lowering the barriers to entry for this kind of conflict-resolution training. Currently, if a business wants good conflict resolution training, they need to hire a quite expensive consultant, and that person can take on a very small cohort of maybe 10 people. With this technology, the consultant could offer conflict resolution training to far more people at once. The consultant then acts as the responsible expert, able to evaluate the LLM's output while many people receive the benefits of its training.

kordlessagain 3y ago

Giving the underprivileged tools to work on themselves is a huge win for society, regardless of the fidelity of all interactions. There are a lot of bad therapists in the world, and not all of them are going to give good advice all the time. But, bad therapists can still provide some information that may be useful to the person learning. Why would we expect perfection from interacting with an LLM instead of a bad therapist? Why be overly concerned about the possibility of introducing "bad data" in such a session? It could be that bad data then gives way to an opportunity for a leap of logic for the one using the tool. Without critical thinking present, it is unlikely anyone is going to learn to be adaptive over the long term. Obviously, the use case matters. I'm not saying to give such a tool to someone who is mentally ill, or incapable of critical thought. In general, I think expecting perfection from LLMs is unreasonable. If someone wants to learn to code, the second they have any interaction with an LLM about coding is the moment they begin to learn. I'm not sure about y'all, but a lot of times when I'm working a hard coding problem, I find myself going through confusion, frustration, interest, and then aggressiveness (anger+interest) to get through the challenge. I frequently tell people who are starting out coding that coding is one of those things that you have to get used "being confused about" to over time. You have to get used to creating problems for yourself, then coping with the feelings that come, and pushing through to a solution. I do think you raise a profound point about ability to evaluate the fidelity. I would argue that some people will eventually develop critical thought to apply to the problem, but if we instead see a "digital tragedy of the commons" happening, things could go sideways pretty quickly.

maroonblazer 3y ago
> You have to get used to creating problems for yourself, then coping with the feelings that come, and pushing through to a solution. A great description of any creative act, really.
FranklinMaillot 3y ago
I think the reason why LLMs pose a much higher risk is scale. A bad therapist will impact a limited number of patients, whereas an LLM deployed at scale will affect possibly millions of people. In matters of health and safety, the smallest error rate could have devastating consequences.

Morelesshell 3y ago

The concern is real but you know what? The last company i was part of, had a TON wtf moments. Like database unprotected on the internet (very short password too!), md5 for passwords, api key in the apps...