TudyBOT Feedback & Discussion Thread

[RoyalChick]

These ai models are… complicated and I have limited understanding. To my knowledge they need to be trained similarly to a dog. With positive and negative reinforcement plus a whole lot of information. You don’t really “code” them as you would a normal computer. This makes it very difficult to change things. If it was told that miracle whip was mayonnaise instead of mayonnaise substitute by accident then there’s no easy way to fix it. There not some code you can edit, similarly to how you can’t go in and change what a dog thinks is good behavior by altering its brain.

But can you keep telling it that Miracle Whip is NOT mayonnaise until it has learned the truth?

 

[RoyalChick]

I was hoping she could learn the correct answers to a number of standard questions that often get answere wrong.
Things like: the correct dose of Corid or fenbendazole; why not to hold a chicken with an impacted crop upside down; why hardware cloth is better than chicken wite at keeping predators out.
Sounds like she can’t do that.

 

[Ted Brown]

The only way it can learn is by having its programming rewritten.

This is only partially correct.

ChatBot (on which Tude is based) evolves in two ways - programming and the "data set" that it has parsed.

Nifty identified that Tude is based on ChatBot3.5, the latest model version is ChatBot4. OpenAI (the company that built it) controls the evolution of both new programming (eg the capabilities to process images, graphcs, etc.) and the data set (ChatBot3 has parsed data from the internet up to September 2021).

I find it very amusing that some members are annoyed that TudeBot is not perfect when if it were we would all be out of work and off fishing somewhere (or maybe wired into the Matrx?).

I love that it has given so many of us easy access to play with AI.

 

[saysfaa]

...[the company] OpenAI ... controls the evolution of both new programming (eg the capabilities to process images, graphcs, etc.) and the data set (ChatBot3 has parsed data ...up to September 2022).
...

So, the data set is updated based on data from across the internet?

Chatbots can learn form "us" as in what people say anywhere on the internet? Vs "us" as in what people say on BYC?

 

[Ted Brown]

So, the data set is updated based on data from across the internet?

Chatbots can learn form "us" as in what people say anywhere on the internet? Vs "us" as in what people say on BYC?

Yes. However the compute power to parse the internet is huge. OpenAI spent ~$41Million to rent something like 1200 cpus to load ChatBot3.0. I read that another of the LLM (large language model) Bot builders spent over $400M.

Yes again.

https://www.reddit.com/r/ChatGPT/comments/12diapw

 

[R2elk]

Is that really true? I thought it was supposed to learn in its own.
If it can’t learn then I think it needs to be kept in a small dark corner until it can learn.

This version is stuck in 2021. It has not been updated since then. It has shown it can hold its learning for posts that are close together. It scans 10 posts at a time. Once it is through that section it reverts backwards and the temporary learning is gone. It can't rewrite its own programming.

 

[R2elk]

Nifty identified that Tude is based on ChatBot3.5, the latest model version is ChatBot4. OpenAI (the company that built it) controls the evolution of both new programming (eg the capabilities to process images, graphcs, etc.) and the data set (ChatBot3 has parsed data from the internet up to September 2022).

This one does not know anything since 2021.

It has proven it does not know anything since 2021.

 

[RoyalChick]

This version is stuck in 2021. It has not been updated since then. It has shown it can hold its learning for posts that are close together. It scans 10 posts at a time. Once it is through that section it reverts backwards and the temporary learning is gone. It can't rewrite its own programming.

That is an issue.
Although some of the truths about keeping chickens have been true for a very long time!
But yes, the 10 posts is a killer.

 

[saysfaa]

So Grdn Betty and the like have as much influence on what the bot will say about nutrition (after the next update) as stormcrow, raingarden, NatJ, Geena, perris, azygous, nuthatched, SallyPB, EggheadJr, and the many more here who have similarly researched, observed, experimented, considered, ...

One vs many+ but notice "the like" - there are many giving such advise.

Somehow, the poor advise tends to sound more uniform than advise that considers climate, purpose, age, breed ... wouldn't that make it more likely the chatbots will "learn" the poor advise?

 

[Ted Brown]

... Once it is through that section it reverts backwards and the temporary learning is gone. It can't rewrite its own programming.

Most of what you refer to as temporary learning is NOT programming it is just more data collected from it's interaction with one of us.

The temporary nature is because this "new data" is not parsed by OpenAI into their underlying data set. The link I included above indicates that work is underway to update the underlying data set interactively.

Also remember that there are multiple BOTs with their own data sets owned by different companies (Meta, Google, etc. etc.)