This preprint led to this enthusiastic X post that doctors better start packing it up because AI is going to take their jobs away. Taking a look at that post, it's clear that if doctors follow the advice of o1-preview, they will bankrupt insurance companies, or be met by firm denials for several of the suggested tests. In fact, the justifcation for some of the recommended tests is quite weak. Others have critiqued OpenAI's Strawberry model, and their findings are very interesting. OpenAI has never published on what data their models have been trained, but it's likely that they were trained on New England Journal of Medicine's CPC cases. As such, their uncanny ability to predict may be nothing more than data leakage – performing inference on data that you've already seen and have been trained on. I, too, get the sense that this makes o1 overestimate the likelihood of very rare diseases it sees, making them seem more common than they actually are. Plus, there is this observation that if you query it again, it may give you a different answer. We don't know how the prompts were constructed, or whether this was a zero-shot or one-shot query, but still. But the paper does reveal that there are cases where o1 gets the diagnosis that GPT-4 missed, so what if you relied just on one LLM? Some have suggested that doctors consult two LLMs in tandem, but has time for that?
––– 凄い –––
Scientists have discovered a protein that helps cancer cells become resistant to CAR_T cell therapy. This protein is YTHDF2, and is a N6-methyladenosine reader, and regulates antitumor functions of macrophages. The precise mechanism of this is unclear, but blocking can lead to tumor death, and these scientists have identified an agent that can achieve this, called CCI-38. Blocking YTHDF2 enhances CAR-T cell activity. Original article here.
––– 凄い –––
MIT scientists may have found a way to magnetize antiferromagnetic materials with terrahertz light (high end of infrared, just before microwave). Could have applications in memory storage technology.
––– 凄い –––
Eating brain mushrooms (Gyromitra esculenta) may be the cause of amyotropic lateral sclerosis in a small French village.
––– 良くない–––
I remember there were high hopes that metformin was going to enhance breast cancer therapies. That hasn't panned out. There's still hope that it has antiaging properties, and many in the field are already taking it. We'll see.
––– 良くない–––