4 ms·

Indeed, I don't think ChatGPT can do science[1] or be a doctor[2]. [1] https://archive.is/AgWkn https://archive.is/AgWkn [2] https://inflecthealth.medium.com/

by zvmaz 3y ago

Indeed, I don't think ChatGPT can do science[1] or be a doctor[2]. [1] https://archive.is/AgWkn https://archive.is/AgWkn [2] https://inflecthealth.medium.com/im-an-er-doctor-here-s-what-i-found-when-i-asked-chatgpt-to-diagnose-my-patients-7829c375a9da https://inflecthealth.medium.com/im-an-er-doctor-here-s-what...

m348e912 3y ago
>>https://inflecthealth.medium.com/im-an-er-doctor-here-s-what-i-found-when-i-asked-chatgpt-to-diagnose-my-patients-7829c375a9da https://inflecthealth.medium.com/im-an-er-doctor-here-s-what... Took a look at this article you linked to and I came to the conclusion I almost always come to when professionals critique GPT. - He's holding GPT to too-high a standard and not recognizing the criticisms and misses of GPT also could apply to many medical professionals in his field. - He is not taking into account that GPT can be quickly trained with new data and address all of the mistakes that he pointed out. So is GPT ready to be an ER trauma doctor? No it's not, but it isn't far from augmenting some of the skills, knowledge, and diagnosis capabilities that an ER doctor may have.
- rvnx 3y ago
  Counter-example, GPT beating real doctors: https://twitter.com/peakcooper/status/1639716822680236032 https://twitter.com/peakcooper/status/1639716822680236032
  - TuringTest 3y ago
    That's like saying that a Google search beat real doctors, because the results surfaced a recent study they didn't hear about. That doesn't make Google better than doctors at diagnosing.
    - rvnx 3y ago
      Yes, there are plenty of such cases but also many where Google gives correct answers and doctors are wrong unfortunately :/ It's very difficult to find a good doctor. Many people are stuck and suffering without diagnosis because of real-life accessibility to good healthcare. I'm talking about having practical access to a good, affordable, available AND knowledgeable doctor. You get access to doctors which have time, and are near you, and in your budget, and they may not know your disease. This is precisely where language models can help to sort and rank the information that is on the internet (if digested by a professional it's even better, but even raw, there is some value to it).
      - TuringTest 3y ago
        That's the reason why a doctor with Google is better than a doctor without Google. That doesn't mean that Google or ChatGPT without a doctor is a good idea.
kebman 3y ago
I can verify this through personal tests and use. It's quite bad at connecting scientific dots or tying together even simple mathematical logic.
vixen99 3y ago
Completely agree having fed ChatGPT with a number of PDFs of recent papers in the biological sciences. Of course chatGPT is a terrific achievement but I wonder how a journalist in a serious weekly can write the following. 'Nor, ironically, is the future overly bright for the scientists who made these machines: one recent paper on the scientific capabilities of the new AI concludes: ‘[this] Intelligent Agent system [is] capable of autonomously designing, planning, and executing complex scientific experiments. Our system demonstrates exceptional reasoning and experimental design capabilities, effectively addressing complex problems and generating high-quality code.’ So that’s the boffins gone, as well. They’ve boffined themselves into the bin.' Beats me utterly. Is it something I've missed?
- dragonwriter 3y ago
  If its talking about a paper concerning an “Intelligent Agent system”, its not talking about the capacity of a bare LLM like ChatGPT, which (impressive as it is in some conversational tasks) is, as a user-facing product, basically a toy demo. Its talking about an LLM as the central component of a system where the core the system responds to LLM responses by executing actions and reprompting, with the core prompt setup to instruct the LLM on the use of those actions. This can provide abilities like exact recall (either via a local stored database or, e.g., web search) of a wider corpus of facts (not all at once, but selectively as needed) than can fit into prompt space (which is a lot better fidelity than you get with training), the ability to do exact complex calculations (which bare LLMs are bad at), and…well, lots of other things.
- flangola7 3y ago
  Which GPT model did you feed them into?

by zvmaz 3y ago

m348e912 3y ago

>>https://inflecthealth.medium.com/im-an-er-doctor-here-s-what-i-found-when-i-asked-chatgpt-to-diagnose-my-patients-7829c375a9da https://inflecthealth.medium.com/im-an-er-doctor-here-s-what... Took a look at this article you linked to and I came to the conclusion I almost always come to when professionals critique GPT. - He's holding GPT to too-high a standard and not recognizing the criticisms and misses of GPT also could apply to many medical professionals in his field. - He is not taking into account that GPT can be quickly trained with new data and address all of the mistakes that he pointed out. So is GPT ready to be an ER trauma doctor? No it's not, but it isn't far from augmenting some of the skills, knowledge, and diagnosis capabilities that an ER doctor may have.

rvnx 3y ago
Counter-example, GPT beating real doctors: https://twitter.com/peakcooper/status/1639716822680236032 https://twitter.com/peakcooper/status/1639716822680236032
- TuringTest 3y ago
  That's like saying that a Google search beat real doctors, because the results surfaced a recent study they didn't hear about. That doesn't make Google better than doctors at diagnosing.
  - rvnx 3y ago
    Yes, there are plenty of such cases but also many where Google gives correct answers and doctors are wrong unfortunately :/ It's very difficult to find a good doctor. Many people are stuck and suffering without diagnosis because of real-life accessibility to good healthcare. I'm talking about having practical access to a good, affordable, available AND knowledgeable doctor. You get access to doctors which have time, and are near you, and in your budget, and they may not know your disease. This is precisely where language models can help to sort and rank the information that is on the internet (if digested by a professional it's even better, but even raw, there is some value to it).
    - TuringTest 3y ago
      That's the reason why a doctor with Google is better than a doctor without Google. That doesn't mean that Google or ChatGPT without a doctor is a good idea.

kebman 3y ago

I can verify this through personal tests and use. It's quite bad at connecting scientific dots or tying together even simple mathematical logic.

vixen99 3y ago

Completely agree having fed ChatGPT with a number of PDFs of recent papers in the biological sciences. Of course chatGPT is a terrific achievement but I wonder how a journalist in a serious weekly can write the following. 'Nor, ironically, is the future overly bright for the scientists who made these machines: one recent paper on the scientific capabilities of the new AI concludes: ‘[this] Intelligent Agent system [is] capable of autonomously designing, planning, and executing complex scientific experiments. Our system demonstrates exceptional reasoning and experimental design capabilities, effectively addressing complex problems and generating high-quality code.’ So that’s the boffins gone, as well. They’ve boffined themselves into the bin.' Beats me utterly. Is it something I've missed?

dragonwriter 3y ago
If its talking about a paper concerning an “Intelligent Agent system”, its not talking about the capacity of a bare LLM like ChatGPT, which (impressive as it is in some conversational tasks) is, as a user-facing product, basically a toy demo. Its talking about an LLM as the central component of a system where the core the system responds to LLM responses by executing actions and reprompting, with the core prompt setup to instruct the LLM on the use of those actions. This can provide abilities like exact recall (either via a local stored database or, e.g., web search) of a wider corpus of facts (not all at once, but selectively as needed) than can fit into prompt space (which is a lot better fidelity than you get with training), the ability to do exact complex calculations (which bare LLMs are bad at), and…well, lots of other things.
flangola7 3y ago
Which GPT model did you feed them into?