Grok3, the AI on X (old Twitter) is flat out the best doctor I've ever had. Stunningly Good!

Fermion said:
Ok, yes I see, I got it to generate images. It is massively failing on my request though. I first asked it to generate an image of a sailboat with a rounded stern, commonly called a canoe stern, that has a outboard motor mounted to that stern. It generated 4 images of sailboats with the more common traditional flat stern and a outboard on that...
Click to expand...
It is a 'language model'. AI has very limited 'training data' on physical things. AI really struggles with hands. Despite each hand having only 5 fingers, and those fingers being distinct parts. Using our innate understanding of the physical world, we get it. AI can't understand how to draw a simple part.
 
(Posted this in another thread but applicable here to show Grok doing something)

I asked Grok3 to make a table showing how much increasing costs of healthcare plans (subsidy plus max OOP) and federal tax would be as a married couple in their mid 50s increased their MAGI. This table (if Grok3 calculated somewhat correctly), gives you an idea of the "effective tax" of any money added to your MAGI past the 138% of FPL.

Pretty eye opening. If you were to add $31,000 to your $29,000 MAGI, you would pay an extra $16,864 in costs (increased ACA costs plus federal taxes). This is essentially a 54% tax rate on that $31,000. Wow.

(Note these figures are what applies to us, and we fully use the Max OOP each year on our silver plan because of a costly drug plus required MRI and other tests)

subsidy.jpg
 
Yes, the free versions are different, and yield 'less', from the paid versions of AI.

I just watched very interesting online presentation on 'Using AI in Genealogy' through my local library. The presenter had years of experience in genealogy plus his career involved making presentations on Microsoft products.

He demonstrated (in real-time) entering the same request/command into a free version and a paid version of the same AI program. He did this with at least 2 AI programs with substantially better and more complex results from the paid versions. Plus, the paid versions didn't restrict his usage. To keep costs down, he suggested signing up and paying for a month, and immediately cancelling the subscription (to prevent it from renewing), then using it for a month to perform all the work you'd saved up for it.

omni
 
Last edited:
Cujet said:
It is a 'language model'. AI has very limited 'training data' on physical things. AI really struggles with hands. Despite each hand having only 5 fingers, and those fingers being distinct parts. Using our innate understanding of the physical world, we get it. AI can't understand how to draw a simple part.
Click to expand...
Reminds me of one of my favorite PEANUTS cartoons.

Charlie Brown is looking at a picture that Linus has drawn. CB "This is a very good drawing of a man, Linus."

Next frame CB "But I notice you've drawn him with his hands behind his back."

Next frame CB "That's because you yourself have feelings of insecurity."

Next frame Linus "That's because I myself can't draw hands."

I wonder if AI is still struggling with drawing hands - or it's just insecure. ;)
 
Cujet said:
Did an epic (for me) bike ride last night and just got back from the gym. I can't believe how well I am doing now. From Feb 18th to now, a complete turnaround.
Click to expand...

That's great. I think many doctors are way too biased, dismissive, and cynical about patients. I wish there were temporary reversible ways to give the doctors various conditions (especially nerve conditions that aren't visible on exam) so they could really understand how unhelpful it is act like patients are hypochondriacs or drug-seeking.
 
UpQuark said:
That's great. I think many doctors are way too biased, dismissive, and cynical about patients. I wish there were temporary reversible ways to give the doctors various conditions (especially nerve conditions that aren't visible on exam) so they could really understand how unhelpful it is act like patients are hypochondriacs or drug-seeking.
Click to expand...
Yeah, doctors used to hand out opioids like candy and now you have to beg for a few pills to get you through a back-pain crisis. THAT's on the gummint, though the docs have become weenies when it comes to helping patients with pain meds. End of rant.
 
Based on this thread I thought I’d give Grok a test drive. I asked it 2 fairly technical questions that I’ve done some research on

1. What are the risks of a coronary calcium score of 160, given my age, health (defined by several metrics medicines etc). The answer it gave was very detailed, very readable, and from what I know spot on. Then I asked about whether aspirin is advisable in this case and the answer is consistent with the latest research I’ve read and what my doctor has said. So it pretty much nailed it, but what was really cool was how it explained all of the various factors that go into it.

2. Next - I asked it if RMDs from a special needs trust would affect SSI eligibility. Again, it broke it down and nailed it.

What is also impressive is it prompts you for more information to increase specificity.

I’m going to keep playing with it, but so far so good.
 
I finally registered so that I could actually test it.

I tried a few questions where I've had some poor AI responses in the past, and it performed well.

But then it failed an easy one. I asked, "If I'm getting 30 days prescriptions, I don't end up with enough pills for the full year when I get one refill per month. How can this be handled so I have 365 pills for the year? Someone suggested getting 90 day refills every 3 months because then the math adds up to provide enough pills. What say you?"

Grok responded, "You’re absolutely right to notice that 30-day prescriptions don’t perfectly align with a 365-day year when refilled monthly, and switching to 90-day refills could indeed solve the issue."

I followed up saying either way leaves me 5 pills short. Then it responded, "You’re right to call that out—let’s clarify the math and address the core issue. I misstepped in assuming the 90-day refill structure inherently resolves the shortfall without explicitly adjusting the total. Neither 30-day nor 90-day refills, as standard, automatically get you to 365 pills unless the prescription or dispensing is tweaked."
 
I just remembered a comment the instructor of the "AI and genealogy" lecture made about AI. Often after AI delivers the answer, there's a question as to how well the AI performed. The lecturer strongly suggested giving constructive/critical feedback, as that will be used to improve that specific AI.

omni
 
I gave it a pretty hard problem the other day. I asked it to give me a LTSpice netlist for a circuit that would generated a 2khz tone modulated at 10hz driving a 8 ohm speaker at 0.1 watt using discrete components, not ICs. It thought for quite a bit (2 min?) and then gave me a pretty detailed description and netlist that contained 5 transistors. This did not run at first, because LTSpice didn't have the model for the 2N3905 but I complained to Grok and it apologized and added the transistor model to the netlist. I then simulated it and did not get the correct waveform at the speaker (was getting the 2khz carrier but not much envelope on the modulation). Grok directed me to tweak a few bias points and values but after several tries I ran into my question limit (10) and gave up. Still, kind of impressed?
 
Grok couldn't come up with the name of a TV show despite telling it the number of seasons (2), the years it was broadcast (2016 to 2017), and that the main character had an affair with the president of the U.S. Then I even added the main character's first name on the show and the type of home she had lived in, but it couldn't come up with the correct show until I gave the actress's name. Well, after training it, it will probably be able to provide the correct answer more quickly, now. :D
 
Did you select the deep search function? Neat test though. Perhaps it requires some actual critical thinking that the LLM just can't do?
 
Fermion said:
Did you select the deep search function? Neat test though. Perhaps it requires some actual critical thinking that the LLM just can't do?
Click to expand...
I switched over to the "Think" feature where it thought for 40+ seconds a couple more times, but it still didn't get it until I added the actress's name. It's too late to try Deep Search instead now that it's been schooled. lol

I asked why it didn't come up with it sooner with all the relevant clues, and it said, "The evidence leans toward Too Close to Home as the match, fitting all criteria, but its lesser-known status on TLC, a network not typically associated with political dramas, likely contributed to the oversight. The need for the actress’s name, Danielle Savre, to identify it underscores the challenge of recalling niche shows without specific identifiers, especially given the focus on more prominent series initially. This analysis suggests improving search specificity and considering lesser-known networks in future queries to avoid similar delays."
 
GenXguy said:
I finally registered so that I could actually test it.

I tried a few questions where I've had some poor AI responses in the past, and it performed well.

But then it failed an easy one. I asked, "If I'm getting 30 days prescriptions, I don't end up with enough pills for the full year when I get one refill per month. How can this be handled so I have 365 pills for the year? Someone suggested getting 90 day refills every 3 months because then the math adds up to provide enough pills. What say you?"

Grok responded, "You’re absolutely right to notice that 30-day prescriptions don’t perfectly align with a 365-day year when refilled monthly, and switching to 90-day refills could indeed solve the issue."

I followed up saying either way leaves me 5 pills short. Then it responded, "You’re right to call that out—let’s clarify the math and address the core issue. I misstepped in assuming the 90-day refill structure inherently resolves the shortfall without explicitly adjusting the total. Neither 30-day nor 90-day refills, as standard, automatically get you to 365 pills unless the prescription or dispensing is tweaked."
Click to expand...
Duh, fill it every 30 days. Same with 3 months. You'd be short if it were on a yearly basis, but it's not! It's on either 30 days or 90 days. Surprised it missed that.
 
I've had chatgpt, gemini and copilot all make rudimentary math errors in their answers. I don't know why.
Here is one exchange I had with Chatgpt as I was watching TV a couple of weeks ago. See the obvious error?

how old is paul mccartney?
ChatGPT said:

Paul McCartney was born on June 18, 1942. As of today, he is 82 years old.
You said:
and dylan?
ChatGPT said:

Bob Dylan was born on May 24, 1941. As of today, he is 82 years old.
Click to expand...

On the other hand, they've given me very detailed and correct (to my knowledge) answers on complex questions.
 
The LLMs are not good at math. The advanced ones write a python script to do the math for them.

I have been fiddling with computer stuff for the past few days and Grok has been very helpful explaining exactly how to tweak windows and firewall settings and how to collect diagnostic data.
 
One concern I have with LLMs is that they usually give you a detailed, well structured answer that sounds authoritative, and a lot of people will take it for fact. But you really need to check verify the answer for yourself sometimes.

For example: I asked ChatGPT how to do something in Moneydance (personal finance software). It gave me detailed step by step instructions but when I tried to follow them, one of the menu items it told me to click was not there. I told ChatGPT that and it replied "You're right, Moneydance does not have that function" and it gave me alternative instructions, which had the same problem. When i told it that, it replied "You are absolutely right!" and gave me yet more instructions, which turned out the be the manual steps I was trying to avoid doing.

I suspect that it had no idea what Moneydance could do, but gave me instructions based on features that other personal finance software has. You really need to take it's answers with a grain of salt.
 
DayDreaming said:
One concern I have with LLMs is that they usually give you a detailed, well structured answer that sounds authoritative, and a lot of people will take it for fact. But you really need to check verify the answer for yourself sometimes.

For example: I asked ChatGPT how to do something in Moneydance (personal finance software). It gave me detailed step by step instructions but when I tried to follow them, one of the menu items it told me to click was not there. I told ChatGPT that and it replied "You're right, Moneydance does not have that function" and it gave me alternative instructions, which had the same problem. When i told it that, it replied "You are absolutely right!" and gave me yet more instructions, which turned out the be the manual steps I was trying to avoid doing.

I suspect that it had no idea what Moneydance could do, but gave me instructions based on features that other personal finance software has. You really need to take it's answers with a grain of salt.
Click to expand...
I guess AI is a lot like dealing with people. I spent half an hour talking to a nice lady on the phone today. She w*rks for my benefits manager. She was trying to help me with the web site. Everything she told me to do didn't w*rk, so I tried something on my own and that w*rked. She was totally surprised (first that she didn't know what she was doing and that I figured it out on my own).

So, my take away: Don't underestimate the "danger" posed by AI and the robots that will eventually be "possessed" by AI. They likely will eventually turn on us when we point out their flaws. Humans have at least some inhibitions against harming others - if for no other reason than that there could be penalties. Will AI driven robots have that inhibition? Inquiring minds want to know. :blink:
 
I figure it just googles faster than I can and comes up with a summary that would me a half of day to get the same answer.
 
ivinsfan said:
dirtbiker its good to hear from someone in your field..and you talked about both sides which is refreshing. But I wonder about your comments re quality metrics..between time constraints for visits on both sides visits often feel rushed . I vote for longer times for provider visits, a win/win for both sides..
Click to expand...
If only doctors could be compensated for this. Unless one has a concierge doctor, most are trying to manage to times allotted by insurance companies.
 
dirtbiker said:
It likely is. It's really tough when you get 15 minutes for a follow-up, check-up, discuss vaccines, discuss screenings, go over labs, and discuss any concerns, and then order appropriate testing and follow-up, and making sure it is all documented in the electronic medical record, all while trying to stay on time so the next patients don't complain that we are running behind. The old model of cranking out as many invoices, err, I mean seeing as many patients, as you can in a day just isn't good medicine, even if it does fatten up paychecks...
Click to expand...
I’m glad to hear there are options the insurance companies are allowing the doctors to be properly compensated for time spent with patients. That’s a win-win.
 
Cujet said:
Interesting. I do pay for X Premium. Maybe that's a factor?

Did an epic (for me) bike ride last night and just got back from the gym. I can't believe how well I am doing now. From Feb 18th to now, a complete turnaround.
Click to expand...
So happy for you! I am now motivated to check it out.
 
You must log in or register to reply here.

Similar threads

C
Sudden 2024 retirement, and good morning (or afternoon)
Cujet
C
BrianB
Thoughts at reaching RE+8
walkinwood
walkinwood
I
Well, the good news is it's not esophageal cancer...
SJhawkins
S
U
Age 30 Am I on track? I want to escape the rat race.
2
Cassius King
Cassius King
jollystomper
Three years of retirement after the "Glide Path"
2
Out-to-Lunch
Out-to-Lunch

Latest posts

Back
Top Bottom