#Fine tuning

1 messages · Page 1 of 1 (latest)

mellow rock
#

Nice idea @amber merlin , what you're talking about is fine-tuning. So chatgpt is so good because of its training data which was given to its base model, similarly if we can generate sufficient training data on MacB, and train OpenAI's base model text-davinci on it, we can have our very own chatgpt for MacB(MacBPT¿)

glacial tundra
#

I think it will for sure be and could possibly already do.

What a doctor does in practice is just to collect data.

For example:

  • What's your diet
  • Whats the injury
  • where does it hurt
  • describe the pain
  • Check blood levels.

Then once all that is collected the doctor can determine what is the most likely sickness (and if there is one)

Something like OpenAI could easily do the same if it had dataset of every clinical test and with right filtering it could know what to collect, and with valid input it could give a valid output (ie diagnosis)

Naturally it's worth noting that Open AI and Chat GPT right now is just taking knowledge developed by humans and wouldn't go on and find cures to diseases on its own, but if there is the data that Symtpoms X, Y and Z exist in a client then its most likely sickness Q and how it's most effectively treated.

Another thing worth mentioning is that it will probably still have a problem with self diagnosis, the AI can easily be tricked might not a have a "bullshit filter" like a doctor might have.

short jetty
#

This is definitely possible, but I can tell you from experience (in the US) that it will be a while before it happens. Getting software certified for use in a hospital is an incredibly long and rigorous process. You have to comply with HIPAA as well as a dozen other safety and general regulations to ensure its safe for use in a medical environment

glacial tundra
#

Yeah it will most likely take a very long time before it has real official use.

However i see emerging markets with limited access to medical assistance to be utalizing it first.

Like for example how certain parts of african nations have never built a bank as they moved directly into digital banks.

icy lily
#

I'm in the mental health space and looking at using psychologist reports to support doctor decision making. I haven't wrapped my head around what problem is really being solved (if doctors see it as a problem).

The HIPAA concern is mainly for personal/privacy info, like DOB. Summarized knowledge shouldn't be a problem.

I think if AI stays away from prescribing or diagnosing, then it would be easier to get buy-in for AI.

mellow rock
icy lily
#

Thanks @mellow rock I’ll check out what Stanford did too. The more ways people attempt this, the more this moves the AI industry forward.

barren sail
#

Hi @icy lily I am also interested with mental healthcare and running some experiments with test data. I am trying fine tune chatgpt with some synthetic mental health patient data. However after the fine tuning chatgpt is able to detect but it is not able to combine the personal data with generic data. Like % of patients living CA. It knows how many people are living in CA and patients living in CA from my test data. Any thought how to combine the chatgpt response with my personal data?

#

By the way my data is tabular data (csv)