Indeed, I don't think ChatGPT can do science[1] or be a doctor[2].
[2] https://inflecthealth.medium.com/im-an-er-doctor-here-s-what...
[2] 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.
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).
'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?
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.