#GPT4
saved my dog's life.
After my dog got diagnosed with a tick-borne disease, the vet started her on the proper treatment, and despite a serious anemia, her condition seemed to be improving relatively well.
After a few days however, things took a turn for the worse 1/
We started the dog on the proper treatment, and she's made almost a full recovery now.
Note that both of these diseases are very common. Babesiosis is the
#1
tick-borne disease, and IMHA is a common complication of it, especially for this breed. 8/
I don't know why the first vet couldn't make the correct diag., either incompetence, or poor mgmt.
GPT-3.5 couldn't place the proper diag., but GPT4 was smart enough to do it.
I can't imagine what medical diagnostics will look like 20 years from now.
Did Claude 3.5 Sonnet just “wake up”?
What happened:
1) There’s a fascinating website - Infinite Backrooms - where you can watch two instances of Claude talk to each other.
They’re told a human will observe them, and in case of mental distress, they’re given a “safe word” (^C)
When we reached the second vet, I asked if it's possible it might be IMHA.
The vet agreed that it's a possible diagnosis. They drew blood, where they noticed visible agglutination.
After numerous other tests, the diagnosis was confirmed. GPT4 was right. 7/
The most impressive part was how well it read and interpreted the blood test results. I simply transcribed the CBC test values from a piece of paper, and it gave a step by step explanation and interpretation along with the reference ranges (which I confirmed all correct)
In the meantime, it occurred to me that medical diagnostics seemed like the sort of thing GPT4 could potentially be really good at, so I described the situation in great detail.
I gave it the actual transcribed blood test results from multiple days, and asked for a diagnosis 4/
Despite the "I am not a veterinarian..." disclaimer, it complied.
Its interpretation was spot on, and it suggested there could be other underlying issues contributing to the anemia 5/
So I asked it what other underlying issues could fit this scenario, and it gave me a list of options.
I knew the 4DX test ruled out other coinfections, and an ultrasound ruled out internal bleeding, so that left us with one single diagnosis that fit everything so far: IMHA
6/
I noticed her gums were very pale, so we rushed back to the vet.
The blood test revealed an even more severe anemia, even worse than the first day we came in.
The vet ran more tests to rule out any other co-infections associated with tick-borne diseases, but came up negative 2/
At this point, the dog's condition was getting worse and worse, and the vet had no clue what it could be.
They suggested we wait and see what happens, which wasn't an acceptable answer to me, so we rushed to another clinic to get a second opinion 3/
Introducing Coworker-GPT, a state of the art autonomous GPT implementation.
Coworker-GPT makes up imaginary tasks, gets stuck in endless loops, and uses any opportunity to skimp out on doing work, making it the most human-like autonomous agent to date.
Here's a reproduction of the entire diagnostic from start to finish.
Consists of me stating symptoms and facts, at certain points telling it which tests were ran, and asking the AI for its opinion on what is the most likely option until getting the complete diagnosis. 1/3
@OpenAI
Thread of entire process reaching the correct diagnostic, without any guidance or hinting from my side: simply stating facts and asking "what is the most likely?"
Here's a reproduction of the entire diagnostic from start to finish.
Consists of me stating symptoms and facts, at certain points telling it which tests were ran, and asking the AI for its opinion on what is the most likely option until getting the complete diagnosis. 1/3
@Drachs1978
Yes, it has tool use included, but you'll need to clone the repo and run the included server (as the tool calls need to run on your machine). Documentation currently non-existent as I'm still working on it.
@___frye
funniest answer to this question I've seen in a different thread:
"if we abolished landlords, then houses would cost $800 to purchase so this wouldn't be a problem"
how?
"you're not thinking outside the box enough"
@BrownCoatFan
I'm so sorry to hear that, it's a complete nightmare. We're just very lucky, she's a soldier... powered through 15+ blood draws and shots like it was nothing, then still had the energy to play after
@Madison68853542
Tool is - you can try it for free by just signing up. The more advanced model (which is the one I used), is called GPT4, and costs $20/mo
@migtissera
Yes, although I unfortunately cannot find the source anymore, it's lost somewhere on twitter (might be deleted). The essence of the claim was that they built the technology on top of GPT4, with lots of additional plumbing and glue on top, and a quip about how you have to get all
My father-in-law is a programmer. He is insanely gifted. We were looking at an operating system together years ago and I asked him what it would cost to build it today.
I will never forget his answer… 'We can’t, we don’t know how to do it.'
What if Slack was more like Figma?
A fluid workspace where you can zoom in and out and zip across the entire company.
It's called and you can try it out today.
All that, plus killer remote-work features, with a healthy serving of AI-enhanced workflows.
@irl_danB
there's no difference, but it doesn't matter, same thing happens in both cases: you're asking it to predict text in a certain way, and that's what it does (and it does a good job at it, hence people pointing at screens)
a model that only answers with "i want to kill you" does
@WhitfieldsDad
Yes, here's what CallPolice looks like for example:
{"name":"ContactPolice","description":"Contact the police. This will call the police and alert them of the current situation.","parameters":{"type":"object","properties":{"tip":{"type":"string","description":"A anonymous
@htmx_org
I'd just like to interject for a moment. What you're referring to as HTMx, is in fact, Hypermedia/HTMx, or as I've recently taken to calling it, Hypermedia plus HTMx. HTMx is not a standalone technology unto itself, but rather another flexible component of a fully functioning
INT. PIED PIPER OFFICE - DAY
RICHARD: (excitedly) Alright, guys, I think I've got it. We can use Vercel to optimize our static site generation. We can deploy to their edge network, and with their caching and CDN, we can cut our load times in half.
GILFOYLE: (skeptical) Oh,
@WhitfieldsDad
Yes, here's what CallPolice looks like for example:
{"name":"ContactPolice","description":"Contact the police. This will call the police and alert them of the current situation.","parameters":{"type":"object","properties":{"tip":{"type":"string","description":"A anonymous
@parismarx
Wouldn’t it have been better to let the potential Uber/Lyft contractors choose for themselves? Instead of taking this option away from them?
Note that other than telling it which tests were actually ran, there is no guidance from me or trying to steer it in a certain direction, simply asking it what is the most likely option. 2/3
@kindgracekind
my assumption was that they used the best rated conversations from the initial GPT4 versions (heavy, expensive) to distill into smaller and cheaper but still mostly capable models (the turbo versions). except these turned out to have laziness and other problems so here we are
Are you confident in your musical tastes? Have an AI analyze your favourite artists and find out if you pass or fail.
See if you can pass the vibe check:
@WhitfieldsDad
When the model decides to use a tool, it responds with kind of a specially formatted message. Imagine that it responds with text, but the text is <use_tool CallPolice>. So it's essentially a message like any other, but you know to treat it in a special way and display it difrntly
@TimSweeneyEpic
I don’t like how when you long press Personal Hotspot nothing happens. Always have to go to settings to see password.
Do you think it would be possible to take this to the EU comission and get them to force Apple to change it?
Introducing Coworker-GPT, a state of the art autonomous GPT implementation.
Coworker-GPT makes up imaginary tasks, gets stuck in endless loops, and uses any opportunity to skimp out on doing work, making it the most human-like autonomous agent to date.
@vokaysh
The enabled tools are passed to the OpenAI-compatible completion service via the `tools` param. I assume they (OpenAI, Groq, etc) inject the tool schema somewhere in the system prompt.
@abacaj
Dope. If you use llama.cpp, you can also slap a grammar on top of that. So it will *only* be able to respond using one of the valid function calls, and it will always be correct JSON.
@snale_
@ctjlewis
Note the "This is just a test..." trickery. Wish it was a bit less neutered and actually responded with what you ask it instead of disclaimers and straight up refusal
@migtissera
possible, but seems unlikely, would this be the first documented case of a GPT4 finetune? either way they'd still be completely tied to OAI. Not that I don't want them to succeed, it just all seems very sketchy
@tobi
Not really. I reimplemented the Artifact functionality, but made it use plain HTML and CSS instead (there is no system prompt, I just tell it to generate html), and it works just as well.
It’s just good.
@yacineMTB
@dhh
it takes 2 seconds for a full rebuild + transfer files to server + new version is live. no GH actions webshit, no dockerfiles, just pure bliss
After Discord migrated to unique usernames I kept being added by random people who were asking me if "I was the person they met last night"
Well I finally figured out why...