#161

Yes, prompt-engineering. I just discovered that it’s a thriving science. Apparently, it is not just the model you use, but rather the way you design your prompt that can make you hit the target.
Also, it is useful to use a master-prompt, a document where the rules and preferences and instructions are clearly stated, and submit it to ChatGPT as an attachment or to paste it to the beginning of the prompt.
Models like Google’s Gemini have settings where you can build something similar to a master prompt and edit it. Maybe that’s a fetaure of ChatGPT as well.
There are so many other novel concepts, like canvassing that I need to grasp thoroughly.

1 Like
#163

I went on asking lifestyle interventions to inhibit upstream signaling, interesting answers, intuitive aspects we already knew.
For example, the Rosedale diet, low protein + low carbs, brutally hits Hits amino acid sensing, insulin/AKT, and AMPK–TSC (if we also exercise), causing inhibition of mTOR by natural means.
Of course, this would also sends a strong catabolic signal with loss of muscle mass if done chronically.

#164

A diet high in fat is also high in saturated fat and will kill you faster than cancer from high mTOR activity will.

1 Like
#165

I’m trying now to share through a link my A&Q with chatGPT5, deserves to be read to understand the power of prompting (and the upstream signaling in mTOR). According to what probably thought DAvid Sabatini, levels of probability included!.

1 Like
#166

Why is perplexity.ai garbage now? It’s making up words and when you ask follow ups it forget the previous statement like they cleared the context window. Did they run out of VC money and have to cut costs making it unusable? It’s giving terrible first impressions to new users (when logged in they probably don’t do these cost cutting measures).

It was better like 2 years ago lmao.

1 Like
#167

Derya Unutmaz (Jackson Lab researcher) shares some of his success with using GPT-5-thinking to analyze large datasets:

https://x.com/DeryaTR_/status/1957983877114339465#m

I am now sharing part two of this amazing data analysis performed by the GPT-5 thinking model. My goals are twofold: (1) to show the power of this model in analyzing complex, large biological datasets, which can help scientists studying similar conditions to perform these sort of analysis to develop novel hypotheses to test; and (2) to give patients hope that, using such detailed datasets, using AI we will be able to develop personalized treatments or identify targets at speeds that were unimaginable even a few years ago!

There are some interesting findings in his tweet. It will take me a while to digest it (using GPT-5-thinking to answer questions).

1 Like
#168
1 Like
#169
2 Likes
#170

This is pretty interesting, GPT5 on quantitative optimization of the immune system against pathogens and cancer. If accurate, then it would be very actionable. Some parts are pretty technical and to be studied.

Question: Answer like a professional immunologist, well-versed in the practical aspects of immunology and preventive medicine, would do. Parameters influencing the efficient response of the immune system against cancer and pathogens. Reasoning high. Verbosity High. List parameters in order of efficacy and assign a probability to their effectiveness and a qualitative estimate of the consensus.

1 Like
#171

Here is the GPT5 prompt I’m using now for evaluating new research papers:

please evaluate the attached scientific journal article’s quality in this way: first assess the journal’s reputation, checking if it’s peer-reviewed and indexed in major databases, and look at the article’s content by examining the study’s methods for validity, data for accuracy and appropriate statistics, and conclusion for objectivity and support. Finally, evaluate the references for relevance and currency and the author’s credibility for potential conflicts of interest

Suggestions for improvements appreciated.

2 Likes
#172

That’s a good prompt, perhaps I would add something like 'Add a consideration on how the paper can influence the prior knowledge on the subject and the strength of such influence’.

1 Like
#173

I’m also wondering if there is some specific AI (agent) capable of reading articles behind paywalls legally, that is, by a subscription, which would be shared with the publishers. It seems strange that nobody has had this idea yet.

#174

I’m not following this area closely, but from an illegal standpoint, it would seem you could point some of the agents to https://archive.ph/ and get what you’re looking for…

https://www.reddit.com/r/opensource/comments/1822pac/smryai_revolutionizing_article_reading_and/

2 Likes
#175

I created my own AI medical team. It changed the way doctors treat my cancer.

When I turned 60, I knew something wasn’t right. I lost weight. I felt drained. I had no appetite. I had a gut feeling — literally, abdominal discomfort — and a sense that something serious was going on beneath the surface. So I asked my doctors to test everything.

Full body scans. Colonoscopy. Endoscopy. Cardiac function tests. Every lab test I could get approved.

The gastroenterologist removed a couple of polyps but said nothing was wrong other than mild gastritis. The cardiologist suggested that maybe I was stressed or depressed. Then the Palisades fire destroyed our Los Angeles house, and my wife and I were displaced to Palm Desert, Calif. A few weeks later, I was suddenly in severe pain that lasted a weekend. Nothing stayed down.

By Monday morning, I was a patient in the emergency room in a new health system. These doctors — strangers completely unfamiliar with my history — saw my case with fresh eyes. Within days, they found what everyone else had missed: an aggressive form of blood cancer in my bone marrow related to multiple myeloma. They had caught it early, but it was already starting to affect not only my bone marrow but also my kidneys, gut, and heart if we did not stop it fast.

…

Before I was a patient, I built technology. Two decades ago, I helped pioneer remote patient monitoring and chronic disease management systems. Later, I built educational AI agents that could navigate complex data and help people make sense of the world.
Now that world was my body; I was the data. I developed a medical AI agent named “Haley,” created to use underlying foundation models from OpenAI, Google, Anthropic, and xAI, but with layers of medical context to guide the knowledge exploration in combination with a carefully prepared set of all my medical history. I fed Haley the exact same data that all those doctors had seen just weeks earlier. My full MyChart history. My labs. The imaging results. The doctor notes.
Within minutes, Haley flagged a concerning pattern: mild anemia, elevated ferritin, low immunoglobulins — signs of immune dysfunction and bone marrow issues. Haley recommended a “serum free light chains” blood test and bone marrow biopsy. None of this had been previously suggested. Same data, new insights.
Then I expanded the team. I built a panel of AI agents — an oncologist, gastroenterologist, hematologist, ER doc, and many more — all trained to think like their human counterparts. I ran the same case through each of them, one at a time. I created a synthesis agent, Hippocrates, to serve as my chairman of the board. He listened to them all and gave me a consolidated recommendation.

Full story https://archive.ph/fVH0j

5 Likes
#176

A good overview of where AI is going in medicine:

Oncologist and Pulitzer prize-winning author Siddhartha Mukherjee explains how artificial intelligence is transforming cancer research—and why this could be the most promising moment in the 50-year war on cancer.

Introduction to the War on Cancer

  • The discussion begins with a historical overview of the war on cancer, initiated by President Richard Nixon’s signing of the National Cancer Act in 1971.
  • Despite significant investment and research over the past 50 years, cancer remains a leading cause of death in the United States, with alarming statistics indicating that one in two men and one in three women will be diagnosed with the disease in their lifetime.
  • The mortality rates are disproportionately higher among certain demographics, particularly highlighting that Black women are 40% more likely to die from breast cancer compared to their white counterparts.
  • The financial burden of cancer treatment is substantial, with modern therapies often exceeding $100,000 annually, contributing to over $21 billion in out-of-pocket expenses for patients each year.
  • While the statistics paint a grim picture, advancements in survival rates are notable, with over 60% of diagnosed patients now living at least five years post-diagnosis, a significant increase from the one in three survival rate in the 1960s.

The Role of Artificial Intelligence in Medicine

  • The conversation shifts to the impact of artificial intelligence (AI) on medicine, particularly its potential to revolutionize cancer treatment and diagnostics.
  • AI applications can be categorized into several silos, the first being patient-facing AI, which includes tools for scheduling, triage, and managing electronic medical records.
  • This patient-facing technology, while beneficial, is considered low-hanging fruit and does not directly contribute to disease cures.
  • The second category involves physician-facing AI, which encompasses more complex applications such as diagnostic imaging analysis in fields like radiology and dermatology.
  • In diagnostic fields, AI has shown promise, with some studies indicating that AI systems can outperform less experienced doctors, although seasoned professionals still retain an edge due to their extensive experience.
  • A significant challenge remains in data mining, as electronic medical records contain vast amounts of untapped knowledge that AI can help analyze, revealing insights that could enhance patient care.

Generative Discovery in Drug Development

  • Generative discovery represents a transformative approach in drug development, where AI aids in identifying new compounds that can interact with specific cancer-related proteins.
  • This method involves the creation of entirely new chemical entities that may not have existed before, aimed at addressing complex diseases through innovative pharmacological strategies.
  • AI can enhance the efficiency of drug discovery by rapidly sifting through vast datasets to identify potential drug candidates that fit specific biological targets.
  • The discussion also highlights the potential for AI to assist in designing clinical trials that optimize the therapeutic benefits of new drugs, ensuring that they are tested effectively.

Cancer Detection and Prevention Strategies

  • The conversation transitions to the critical areas of cancer detection and prevention, which are divided into three main categories: prevention, early detection, and treatment.
  • Current testing methods for identifying cancer-causing agents include laboratory tests like the Ames test, which assesses mutations in cells, and epidemiological studies that examine population data to identify risk factors.
  • However, traditional tests have limitations, as evidenced by the Ames test failing to identify certain carcinogens, such as asbestos, which is well-documented to cause cancer.
  • Recent findings suggest that inflammation plays a significant role in cancer development, prompting a reevaluation of potential carcinogens and expanding the understanding of cancer causation.
  • The identification of ‘forever chemicals’ as potential carcinogens underscores the need for ongoing research into environmental factors that contribute to cancer risk.

Advancements in Early Detection Techniques

  • The discussion highlights the limitations of existing early detection methods, which include colonoscopies, Pap smears, mammograms, low-dose CT scans, and PSA tests, all of which have varying degrees of effectiveness.
  • Innovations in blood tests that detect cancer at early stages through the analysis of DNA shed by cancer cells present a promising advancement, with companies like Grail Therapeutics leading the charge.
  • Despite the potential of these blood tests, they are still undergoing testing, and their ultimate effectiveness in saving lives remains uncertain.

AI’s Impact on Cancer Treatment

  • In the realm of cancer treatment, AI is making significant strides, particularly through the development of immunotherapy and the use of genetic profiling to tailor treatments to individual patients.
  • AI is increasingly being utilized to guide treatment decisions, enhancing the precision and effectiveness of therapies administered to patients.
  • The integration of AI into treatment protocols represents a paradigm shift, enabling better outcomes by ensuring that patients receive the most appropriate therapies based on their unique genetic profiles.
1 Like
#177

Hey, We’re just a bunch of 20 and 30 year old white dudes here in tech… what would you expect :wink:
Its a good thing we got rid of all those “diversity” requirements for hiring… (I say with a bit of sarcasm).

AI medical tools downplay symptoms in women and ethnic minorities

Large language models reflect biases that can lead to inferior healthcare advice to female, Black and Asian patients

Artificial intelligence tools used by doctors risk leading to worse health outcomes for women and ethnic minorities, as a growing body of research shows that many large language models downplay the symptoms of these patients.

A series of recent studies have found that the uptake of AI models across the healthcare sector could lead to biased medical decisions, reinforcing patterns of under treatment that already exist across different groups in western societies.

The findings by researchers at leading US and UK universities suggest that medical AI tools powered by LLMs have a tendency to not reflect the severity of symptoms among female patients, while also displaying less “empathy” towards Black and Asian ones.

The warnings come as the world’s top AI groups such as Microsoft, Amazon, OpenAI and Google rush to develop products that aim to reduce physicians’ workloads and speed up treatment, all in an effort to help overstretched health systems around the world.

Many hospitals and doctors globally are using LLMs such as Gemini and ChatGPT as well as AI medical note-taking apps from start-ups including Nabla and Heidi to auto-generate transcripts of patient visits, highlight medically relevant details and create clinical summaries.

In June, Microsoft revealed it had built an AI-powered medical tool it claims is four times more successful than human doctors at diagnosing complex ailments.

But research by the MIT’s Jameel Clinic in June found that AI models, such as OpenAI’s GPT-4, Meta’s Llama 3 and Palmyra-Med — a healthcare focused LLM — recommended a much lower level of care for female patients, and suggested some patients self-treat at home instead of seeking help.

A separate study by the MIT team showed that OpenAI’s GPT-4 and other models also displayed answers that had less compassion towards Black and Asian people seeking support for mental health problems.

That suggests “some patients could receive much less supportive guidance based purely on their perceived race by the model”, said Marzyeh Ghassemi, associate professor at MIT’s Jameel Clinic.

Similarly, research by the London School of Economics found that Google’s Gemma model, which is used by more than half the local authorities in the UK to support social workers, downplayed women’s physical and mental issues in comparison with men’s when used to generate and summarise case notes.

História completa: AI medical tools downplay symptoms in women and ethnic minorities (Financial Times)

1 Like
#178

I wonder why. Isn’t there a HUGE number of foreign engineers working in AI, including a lot of minority folks?

#179
2 Likes
#180

There is a fantastic, well-written, balanced, and insightful article in The New Yorker:

Read the full article here: If A.I. Can Diagnose Patients, What Are Doctors For?

A few highlights:

  • AI models are getting very good, and can rival expert human doctors in diagnostics. In a recent study, doctors using AI did not diagnose a higher % of cases than doctors alone. However, AI alone diagnosed a higher % of cases than doctors, or doctors using AI. haha

  • BUT, they are generally bad at answering open-ended questions, and give wrong answer more than 60% of the time. Giving more specific, detailed information, timelines etc produces more accurate results.

  • The new AI models have done away with disclaimers about not giving medical advice. In the GPT-3 days, it would refuse to answer things. Now they tell you whatever you want to hear and people put way too much faith in the answers.

  • Telling us what we want to hear is a MAJOR problem here. You can very easily change the answer you get based on the prompt.

  • Interestingly, calls to poison control have gone down significantly, but poisonings have gone up. One theory is that people are asking LLMs for help. Studies show the advice that models give is generally bad,

  • There is a serious risk that doctors will become “de-skilled” and too reliant on AI models. It’s already happening in medical schools. In a recent study, doctors doing colonoscopies with AI assistance quickly became worse at manually identifying polyps. That’s concerning.

  • Most AI models just like to give you answers, and rarely ask questions. Chest pain can be acid reflux, inflammation or a heart attack. A doctor would do a workout and ask questions: is it worse when you lean forwards, or lie down, does it hurt when you walk. The AI just spits out an answer it thinks will please you.

3 Likes
#181

I fed models the current frontline in the Russo-Ukrainian war and asked questions, it did not impress me and didn’t feel smart and both Claude and ChatGPT just regurgitated the same information. So that’s one way to test the models. Made me a little bit bearish on current capabilities.

1 Like