Let's make a concrete plan to 'cure every disease in 10y' and see how far we could get.
Our assumptions will be ~infinite IQ 200+ 'thinkers', automated labs by ~2032, $1-2T to spend (~10y Google profits), can recruit any living person, significant influence over US Gov but not omnipotence.
There are three problems to solve:
1) Knowing how to cure a disease.
2) Making the medicine to do it.
3) Showing that you've done it (in humans).
Difficulty on this timeline is 3>1>>2. We will need to identify longest critical path timelines for each problem, and do things in parallel.
We'll use siRNA for loss of function and LNP mRNA for gain of function, because this gives access to every target in the genome and makes transition from knowing through drug design very fast (if you assume AI has enough training data zero shot a small molecule for any target, could use that as well).
We will need manufacturing capacity, which can be built while we sort out what to do, COVID-style. Number of disease x person cases to treat in US is ~1B, which would mean ~1000x current capacity, at cost of ~$100B and requiring ~5M workers. Let's assume the workers can be robots by our deadline date. Let's assume AI makes the molecules 20x more potent/gram, then this expansion could cover global demand.
Main limitation of these drug modalities is not reaching all cell types. To solve this we will have AI nominate conjugated ligands for all cell types based on existing scSeq + proteomic Atlases, create libraries of these and use pooled in vivo screening of a whole organism to thoroughly map biodistribution for each. We'll prototype in primates to get methodology down, then will confirm using 3-5 human decedents (has been done for AAV serotype screens). 5-100B dollars depending on how deep you go on cell count. These conjugates will ~modularly map to any oligo sequence we come up with later. From experience this could be done in <3y.
Showing effects in humans will be the slowest part, particularly for progressive and hard-to-measure disease like Alzheimer's. To make our timeline we will have to develop at minimum prognostic and pharmacodynamic biomarkers for every disease where readouts take >2y. Let's assume doing multiple -omics on blood contains patterns that AI can infer as disease activity.
You might pair this with video records of people with/without disease (either as new study in parallel, or get CCTV from China maybe), assuming AI can track behavioral & functional capacity as added data.
We want access to samples with causal information on disease, ideally timecourse with incidence of lots of diseases. The US Veteran's Admin has blood samples over many years from many many veterans. We'll spend year 1 working with USG to produce multi-omic data to pair with these health records, as has been done at mid scale with UK Biobank. By year 3 this could give us potential prognostic markers for all diseases. We'll come back to these.
For question 1, any sound thinker will tell you that some interventions will treat some cells while messing up others. So we will need two sets of data: First, understand every treatment's effect in every cell. We (today, only
@GordianBio) can do this with pooled in vivo screening, in animals that have already developed the diseases to avoid waiting, to understand what does good/bad things in each cell type.
In parallel, we will start testing safety of the treatments in healthy human volunteers, doing dose escalation of ~40 people x 20k x up/down at 50K per patient (just blood draws) would be like 80B dollars, and within order of mag of how many patients are recruited for (all) trials today (if you think this is daunting, spend 1y doing in mice first). From this we will 1) get toxicity for each target, 2) draw blood and get pharmacodynamic markers for each target, 3) measure blood changes for AI to compare to the VA data, as well as to the cellular changes from the pooled screen, to deconstruct the physiological changes from cellular effects.
So year 4ish you have treatments in animals that benefit each disease/cell type, you put together the data on which interventions cause which types of toxicity systemically with what effects occur in each cell types to learn what cell types to avoid for each target, and use biodistribution data to design around that.
This tells you what interventions where, so you use and/or combine those into trials in diseased patients (~600B for 25K diseases, 4 treatments per), initially with multiple treatment arms for each of your hypotheses and biomarker readouts based on the blood omics for efficacy potential. Get more blood, calibrate, pick best options. Say 3 rounds of 1y trials. All paperwork and analysis ~instant because AI.
We'll try combinations too based on the causal in vivo map, using AI to identify synergies based on effects and inferring regulatory/interactome networks and cell-cell interactions.
If you're good at this you now have a strong treatment for each disease, could imagine running a pivotal with just one or two hundred patients, say 20M and 2y, 500B if 25K diseases (in reality less because most of those 25K diseases are rare+genetic).
Key thing is that the manufacturing scaleup, the delivery enablement, the target discovery, and the biomarker development happen in parallel in the first ~3-4 years, leaving time for a few rounds of biomarker-based trials and then one pivotal per disease.
We'll fall short of 'all diseases', missing: Ones not present in VA data, ones with no natural model system (although we should try ex vivo human organs), a few where neither KD or overexpression solves.
FDA approval may have to wait a few more years to wait for pivotal trial hard endpoints if biomarkers not considered validated surrogates yet.
And of course everything has to go right, etc. etc.
But there's at least directions we can start today that make amazing outcomes happen in our lifetimes.
(h/t discussions with
@SGRodriques)
2/2 Second, on the messaging around AI. I do not agree that my messaging has been disproportionately negative. In fact it has been about equally balanced between risks and benefits: I’ve written one major essay about each, and even in interviews where I discuss the risks, I make sure to frequently mention the incredible benefits as well as proposing possible solutions to the risks (short clips from my interviews that end up on social media tend to be disproportionately negative, as that gets clicks). In fact, I wrote Machines of Loving Grace because I didn’t feel the AI industry was painting an inspiring enough picture of how the technology could radically transform the world for the better. The bulk of the essay is devoted to refuting skepticism of AI’s potential in health and biology, and showing why I think it will actually be possible to cure most human disease in ~5-10 years, as crazy as it may sound to ordinary people and frankly to biologists as well (I used to be one!). And, if you read my most recent essay (Policy on the AI Exponential), I discuss concrete proposals for how to streamline the FDA process to make sure the deluge of AI-accelerated drugs isn’t slowed down by the regulatory process. I feel the urgency here: I lost my father to Hepatitis C only a few years before the development of direct-acting antivirals (sofosbuvir), which cure 95% of patients and probably would have cured him.
I do agree that the public has a negative view of AI (and that this is a big problem), but I don’t think it is primarily caused by me or any other AI leader warning about AI’s risks. I think it is fundamentally a crisis of trust. I think that ordinary people don’t trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over. The causes of this go back decades and AI is just the latest iteration of it. I don’t think that a glitzy marketing campaign with a positive spin (which some have advocated that Anthropic do) is the way to win back that trust — at this point, saying that AI will cure cancer is more a cliche than it is inspiring, and most people think it is deceptive. The thing that will work is *actually curing cancer*. I think by far the most accurate criticism of AI companies including Anthropic is that we haven’t yet delivered on our big promises to benefit the world. That is totally on us, and I think it’s the criticism you should be making, instead of all this stuff about messaging and marketing.
We are however doing our best to fix this: Anthropic is ramping up its efforts very quickly in biology and medicine, and we hope to have incredible results in the coming years and some early glimmers in the coming months. When we’ve actually accomplished something real, the whole world will hear about it, as loudly as possible, you have my word on that. But until then I don’t want to make empty promises, and in the meantime I feel compelled to speak honestly about the very real risks of AI and how to address them. Honesty is the right thing on the merits, and in terms of public credibility and trust it is no worse than, and may in fact be better than, an approach that ignores or distracts from risks which people instinctively understand are real.