5 ms·

It’s possible to now get plausible diagnoses through an LLM. For which you can further discuss with your doctor. 1. Get your DNA fully sequenced for $200 from

by aantix 2y ago

It’s possible to now get plausible diagnoses through an LLM. For which you can further discuss with your doctor. 1. Get your DNA fully sequenced for $200 from Nebula. 2. Download your .cram file. 50gb+ 2a. Or if you have a 23andme profile, just request to download your raw data. Use that output file. 3. Use WGS Extract to generate a datafile from the cram file of the SNPs and genotypes. 4. Zip that output file. Upload to a ChatGPT session. 5. This enables you to ask questions about your specific genetic profile. Bring your findings to your doctor. Discuss their plausibility. It may surface a solution that they hadn’t thought of.

FeepingCreature 2y ago
> WGS Extract Good god that install script. This is it, this is the tool that AppImage was created for.
- aantix 2y ago
  Pull requests welcomed.
  - FeepingCreature 2y ago
    Might throw it into an AI later.
  - FeepingCreature 2y ago
    Okay I tried to get Sonnet to do it for like two hours and I give up, sorry.
lm28469 2y ago
You might as well piss on the floor, take a picture of the puddle, upload it to chatgpt and ask it if you'll develop cancer between 2033 and 2037, you'd probably get the same amount of hallucination.
- aantix 2y ago
  No. Maybe if you abandon all logic. With an LLM, we’re trying to efficiently search the entire space using as many granular, personal attributes as possible. We’re trying to find plausible hypotheses. Especially those that haven’t been considered. The final determination and action is still a partnership between you and your doctor.
  - lm28469 2y ago
    You don't know what it's been training on and you don't have the resources to validate the output. It's LLM 101: what you shouldn't do
    - aantix 2y ago
      I do have the resources to validate. I am smart. I am capable of understanding deeply technical subjects. I have a team of smart doctors to present my findings to. I’m looking for plausibility. Not the full diagnosis.
      - lm28469 2y ago
        And what data does chat gpt has that your personal team of smart doctors doesn't have?
    - wholinator2 2y ago
      Well known that 101 type rules are learned to be simplified near-lies in 301, and possibly completely untrue in 501. We shouldn't hamstring ourselves on the fear that the idiot will hurt himself with the tool. It's up to you to determine if you're the idiot or not, and then how to delegate to others if so. This type of cooling/censorship is strange to me. These are the exact strange and beautiful usecases we were all excited for aren't they? To expand the capability of the average human, to greatly expand the effectiveness of the intelligent users. Obviously the damn thing can lie! We all know that! The stakes presented by chronic illnesses go far far beyond "scared to hear a lie" and tend to land in "please god save me i'll do anything". People usually have _already tried everything at their disposal_. What's it gonna do, tell them to walk in front of a bus? Every medical conversation I've had, it repeatedly tells me to consult human doctors. And I've had so many useful interactions that it seems to me your philosophy would completely occlude. I'm still figuring out serious problems i have. Things that don't get appointments for months because I've been told I'm fine. Of course a hypochondriac could go insane, just like they might with the DSM-V. It's a lot different when you have to skip work because of constant nausea, or when your productivity tanks because of frequent heart palpitations, or when your child is losing too much weight. I take very harshly to being told i can't use a tool to help myself because you think I'm too stupid to not hurt myself.
bitshiftfaced 2y ago
6. Receive mail from your insurer that, inexplicably, your premiums will be going up. 7. Receive email in 5 years disclosing how your chat data may have been part of a security breach.
- dragonwriter 2y ago
  If uou are in the USA, you skipped the step where the Genetic Information Nondiscrimination Act of 2008 and (depending on state, but they are common), the state law and/or insurance regulation with similar effect is repealed.
Gerardo1 2y ago
So is dang/HN in general good with the several comments in this thread giving explicit medical advice that is some variation of "don't go to the doctor, use chatGPT"? I'm not sure the value of the discussion around those comments outweighs risk of spreading/normalizing that idea.
- djmips 2y ago
  Some folks don't have access to doctors or their doctors are way overworked.
  - Gerardo1 2y ago
    And so random people on the internet should be advising them to dump their genetic profile into chatGPT for medical diagnostics?
    - aantix 2y ago
      Doctors are humans too. They don’t know the entire solution space. Any diagnosis and action set of steps found, you’ll have to work with your doctor to implement.
    - aantix 2y ago
      Have you ever lived with a chronic condition for decades for which multiple doctors literally have zero answers for?
      - Gerardo1 2y ago
        I'm glad you figured out your issue! Genuinely! That it worked for you, doesn't mean it's a wise general recommendation.
        wholinator2 2y ago
        Eh, i think it's fine. It isn't like the teeming masses are gonna run out for dna tests that they spend hours formatting and selecting and then fed into a complicated extraction process involving careful prompt engineering. I think the danger in recommendation lies almost entirely within the space of "easier than going to a doctor", which this process certainly is not. You wouldn't do that for a cold. But god damn am i glad that there's people talking about this. I have several chronic and subtly terrifying health issues myself and this information can save lives. What alternative do you propose, that we should not freely share our methodologies for survival?
philomath_mn 2y ago
Would ChatGPT use tools to analyze this? Or are we talking raw data in the context window? Have you done this? Any interesting results?
- aantix 2y ago
  If you create a custom gpt, it’ll will delegate to a custom gpt-4 model, “a specialized GPT-4 variant optimized for advanced health, genetics, and performance optimization.”
  - philomath_mn 2y ago
    In my experience, LLMs have a hard time processing large amounts of raw data and finding patterns (even simple univariate time series) so I would be surprised if a GPT could find _real_, non-hallucinatory patterns in raw data.
GTP 2y ago
From what I recall from a bioinformatics course, I really doubt you can have your DNA fully sequenced for just $200. But, prove me wrong, do you have some resource saying otherwise?
- aantix 2y ago
  Look at the pricing page. The cost has plummeted. Compute is cheap. https://dnacomplete.com/tier-selection/ https://dnacomplete.com/tier-selection/ Nebula offered specials for full genome sequencing. They’re now owned by DNA complete. DNA Complete prices it at 495. I paid half that for 30x. “By analyzing your DNA multiple times, we can significantly increase accuracy. 30x Whole Genome Sequencing means that 100% of your DNA is sequenced an average of 30 times, thus providing even higher accuracy. Typical DNA Ancestry tests sequence less than 1% of your DNA.”
  - GTP 2y ago
    Yes, compute is cheap, but it was the actual lab procedure that required expensive equipment while also being time-consuming (although I don't remember the time required, maybe it wasn't that much). I remember the professor mentioning an upcoming machine by Illumina that was soon be able to sequence the DNA without cutting it into pieces, this is probably part of the cost reduction as in this way you wouldn't need to run a multiple alignment algorithm to reconstruct the sequence from the pieces. Thanks for pointing at this resource.

by aantix 2y ago

FeepingCreature 2y ago

> WGS Extract Good god that install script. This is it, this is the tool that AppImage was created for.

aantix 2y ago
Pull requests welcomed.
- FeepingCreature 2y ago
  Might throw it into an AI later.
- FeepingCreature 2y ago
  Okay I tried to get Sonnet to do it for like two hours and I give up, sorry.

lm28469 2y ago

You might as well piss on the floor, take a picture of the puddle, upload it to chatgpt and ask it if you'll develop cancer between 2033 and 2037, you'd probably get the same amount of hallucination.

aantix 2y ago
No. Maybe if you abandon all logic. With an LLM, we’re trying to efficiently search the entire space using as many granular, personal attributes as possible. We’re trying to find plausible hypotheses. Especially those that haven’t been considered. The final determination and action is still a partnership between you and your doctor.
- lm28469 2y ago
  You don't know what it's been training on and you don't have the resources to validate the output. It's LLM 101: what you shouldn't do
  - aantix 2y ago
    I do have the resources to validate. I am smart. I am capable of understanding deeply technical subjects. I have a team of smart doctors to present my findings to. I’m looking for plausibility. Not the full diagnosis.
    - lm28469 2y ago
      And what data does chat gpt has that your personal team of smart doctors doesn't have?
  - wholinator2 2y ago
    Well known that 101 type rules are learned to be simplified near-lies in 301, and possibly completely untrue in 501. We shouldn't hamstring ourselves on the fear that the idiot will hurt himself with the tool. It's up to you to determine if you're the idiot or not, and then how to delegate to others if so. This type of cooling/censorship is strange to me. These are the exact strange and beautiful usecases we were all excited for aren't they? To expand the capability of the average human, to greatly expand the effectiveness of the intelligent users. Obviously the damn thing can lie! We all know that! The stakes presented by chronic illnesses go far far beyond "scared to hear a lie" and tend to land in "please god save me i'll do anything". People usually have _already tried everything at their disposal_. What's it gonna do, tell them to walk in front of a bus? Every medical conversation I've had, it repeatedly tells me to consult human doctors. And I've had so many useful interactions that it seems to me your philosophy would completely occlude. I'm still figuring out serious problems i have. Things that don't get appointments for months because I've been told I'm fine. Of course a hypochondriac could go insane, just like they might with the DSM-V. It's a lot different when you have to skip work because of constant nausea, or when your productivity tanks because of frequent heart palpitations, or when your child is losing too much weight. I take very harshly to being told i can't use a tool to help myself because you think I'm too stupid to not hurt myself.

bitshiftfaced 2y ago

6. Receive mail from your insurer that, inexplicably, your premiums will be going up. 7. Receive email in 5 years disclosing how your chat data may have been part of a security breach.

dragonwriter 2y ago
If uou are in the USA, you skipped the step where the Genetic Information Nondiscrimination Act of 2008 and (depending on state, but they are common), the state law and/or insurance regulation with similar effect is repealed.

Gerardo1 2y ago

So is dang/HN in general good with the several comments in this thread giving explicit medical advice that is some variation of "don't go to the doctor, use chatGPT"? I'm not sure the value of the discussion around those comments outweighs risk of spreading/normalizing that idea.

djmips 2y ago
Some folks don't have access to doctors or their doctors are way overworked.
- Gerardo1 2y ago
  And so random people on the internet should be advising them to dump their genetic profile into chatGPT for medical diagnostics?
  - aantix 2y ago
    Doctors are humans too. They don’t know the entire solution space. Any diagnosis and action set of steps found, you’ll have to work with your doctor to implement.
  - aantix 2y ago
    Have you ever lived with a chronic condition for decades for which multiple doctors literally have zero answers for?
    - Gerardo1 2y ago
      I'm glad you figured out your issue! Genuinely! That it worked for you, doesn't mean it's a wise general recommendation.
      - wholinator2 2y ago
        Eh, i think it's fine. It isn't like the teeming masses are gonna run out for dna tests that they spend hours formatting and selecting and then fed into a complicated extraction process involving careful prompt engineering. I think the danger in recommendation lies almost entirely within the space of "easier than going to a doctor", which this process certainly is not. You wouldn't do that for a cold. But god damn am i glad that there's people talking about this. I have several chronic and subtly terrifying health issues myself and this information can save lives. What alternative do you propose, that we should not freely share our methodologies for survival?

philomath_mn 2y ago

Would ChatGPT use tools to analyze this? Or are we talking raw data in the context window? Have you done this? Any interesting results?

aantix 2y ago
If you create a custom gpt, it’ll will delegate to a custom gpt-4 model, “a specialized GPT-4 variant optimized for advanced health, genetics, and performance optimization.”
- philomath_mn 2y ago
  In my experience, LLMs have a hard time processing large amounts of raw data and finding patterns (even simple univariate time series) so I would be surprised if a GPT could find _real_, non-hallucinatory patterns in raw data.

GTP 2y ago

From what I recall from a bioinformatics course, I really doubt you can have your DNA fully sequenced for just $200. But, prove me wrong, do you have some resource saying otherwise?

aantix 2y ago
Look at the pricing page. The cost has plummeted. Compute is cheap. https://dnacomplete.com/tier-selection/ https://dnacomplete.com/tier-selection/ Nebula offered specials for full genome sequencing. They’re now owned by DNA complete. DNA Complete prices it at 495. I paid half that for 30x. “By analyzing your DNA multiple times, we can significantly increase accuracy. 30x Whole Genome Sequencing means that 100% of your DNA is sequenced an average of 30 times, thus providing even higher accuracy. Typical DNA Ancestry tests sequence less than 1% of your DNA.”
- GTP 2y ago
  Yes, compute is cheap, but it was the actual lab procedure that required expensive equipment while also being time-consuming (although I don't remember the time required, maybe it wasn't that much). I remember the professor mentioning an upcoming machine by Illumina that was soon be able to sequence the DNA without cutting it into pieces, this is probably part of the cost reduction as in this way you wouldn't need to run a multiple alignment algorithm to reconstruct the sequence from the pieces. Thanks for pointing at this resource.