The Most Useful Thing AI Has Ever Done For Me
n=1 and unexpected
I struggle with depression. I’m lucky in that it’s on the milder side and I hope to be on the other side of the worst of it. But it’s of course, not good and I’d like to minimize its impact on my life.
One of the primary drivers of depressive symptoms for me is disruption and dysfunction in my gut microbiome1. Because of this I’ve been working with an excellent and nuanced functional nutritionist for the past seven years and conducted dozens of experiments, some of which have been life changingly beneficial. Adding foods, eliminating foods, trying many combinations of supplements — probiotics, prebiotics, postbiotics, and more. One of the persistent challenges is that it’s a moving target. What once might have been a critical pillar of the protocol could become actively detrimental if things change sufficiently.
This sounds like a lot of work and effort: why would anyone do this? Feeling even ok is so much better than being depressed. And feeling actively good is a gift. I’m extremely privileged in many ways here: having the resources to get the help that I need and the time and support to benefit from it. I’ve made tremendous progress in this area.
Recently I’d been working with o3 on trying to improve the digestibility and bioavailability of vegan protein powders. I was curious about things like if pre-soaking or cooking might help (probably not and it’s complicated). And one morning I had the idea to ask a broader question about whether food sensitivities2 are often correlated and gave a few examples. What followed was an extensive, interactive differential diagnosis session. At the end, I had a new hypothesis to test — that I might be allergic to nickel. It’s absolutely critical to emphasize that the value here is a new idea and framing around things and not a definitive answer. But it’s so valuable to me because in years of effort, this particular issue had never been surfaced as a potential cause.
Nickel allergy is not good news for a vegan. Foods that are high in nickel include legumes, soy, oats, nuts, and nut butters as well as my favorite food: chocolate3. And in the medium category are things like grains, many seeds, and tomatoes. And to further complicate things, nickel content in the exact same food can vary dramatically due to the soil, region, season, and other factors. I could probably eat a diet of just those foods for an extended period of time.
But, I had really high curiosity here. So I embarked on a low nickel diet as an experiment.
The initial results are extremely promising. My mood and energy have been higher. Eczema (which is often thought to be a skin issue, but is actually usually food / digestion related) symptoms have been improving.
Does this definitively mean that I have a nickel sensitivity? No. Maybe there is some other explanatory cause. Perhaps it’s just one of the foods that I’m avoiding. Or maybe it’s the addition some of the new foods that I’m trying. It could be the placebo effect of me wanting this to work. Maybe other external factors in my life are driving the changes. The improvements might not be persistent over a much longer time horizon.
Aren’t food experiments so fun and satisfying in their extreme clarity and lack of ambiguity? 🧌
While it would be amazing to really understand the underlying causality and mechanisms in play, I’m also just happy to take the win and enjoy feeling better for this time period, however long it lasts4.
I’ve seen some of the studies suggesting that in some test contexts, AI models can do better than physicians at differential diagnosis. I already found those results plausible5. But it’s so powerful to literally experience something like this firsthand and in a way that qualitatively has improved my life.
If you’re struggling with issues like these and have the resources, I highly recommend seeking professional help — without years of experimentation and guidance I probably wouldn’t have been able to ask the right questions, make a good plan, or have the confidence to try this. But for the challenging edge cases like mine or for the many who do not have the resources to get professional help, it’s amazing to see what’s possible6.
I’ve known this to be true based on extensive personal experience and it’s been interesting how skeptical many people are about this. Haven’t done an extensive lit review, but here’s some recent research.
This is a long conversation with o3 — it probably won’t be of interest to most, but I think it’s an interesting artifact and perhaps a useful example of what’s possible in iterative conversation with the models.
I eat a lot of chocolate. However much you think is unreasonable, I probably eat a little more than that. I am, at the very least, emotionally addicted to chocolate. I would have never considered even taking a break from chocolate on purpose for a even a day without an extremely compelling reason. I’m still a little surprised that I was open to even trying this.
And of course there are many knock-on effects like reducing consumption of many nutrients that I was previously, which could cause long term problems, so it might not be a long term solution. But I have learned about the BraMa-Ni diet, which attempts to address these long considerations (I had worked out a roughly similar plan with o3, but it’s good to know that it’s a thing and has been studied).
Doing better on one task among many in a simulated environment is starkly different from doing the job. This is not a claim that I find it plausible that AI is already a better replacement for physicians. But this is a task that AI seems particularly well suited for.
And the models are the worst they’ll ever be. And imagine this being properly productized into a food and symptom tracker with specialized prompts, access to the appropriate studies, and humans in the loop to verify and assist.