AI vs Cancer in LatAm @ https://xitter.cf/t.co/GWHLFMs42R
Joined July 2010
- Tweets534
- Following1.1K
- Followers242
- Likes1.3K
Data tokenization is about to make a big comeback as validating provenance becomes central to data licensing agreements
We’ve spent the last few years deploying hardware to generate proprietary data across a network of cancer centers, in a region completely out-of-distribution for existing models.
Our next step is connecting frontier labs, pharma and cancer centers into a single market network.
The key to winning as a data marketplace is to simultaneously be the highest paying buyer and the lowest cost seller. This explains why data markets ultimately become duopolies.
The only bio companies with durable data businesses are those that:
At least: control a live stream of longitudinal data,
Ideally: have access to biospecimens to generate incremental modalities,
At best: own the last mile deployment of 1st/3d party models
Data licensing will not be a durable revenue stream for every bio company.
For generic datasets, pricing will stop making sense as demand plateaus and brokers flood the market.
For specialized datasets, licensing will stop making sense as frontier labs start to encroach on their domains.
This will likely lead to a slew of data-centric acquisitions.
In my experience, the recent frenzy for bio data brokerage overestimates the supply of existing datasets, while underestimating costs/challenges of generating new/better data. This will become evident as demand shifts towards deeper, outlier datasets.
Victor Angel retweeted
How did we come to understand cancer in the light of evolutionary theory?
Nothing makes sense if not with population genetics!
In our latest @NatureRevCancer review we unpack how population genetics theory applied to somatic cells allows us to measure and predict cancer evolution from molecular data.
Computational modelling in oncology has gone a long way but we need to go much further.
The frontier is combining evolutionary theory with machine learning.
Link below👇
En México ni se imaginen
What accessing frontier cancer care looks like in 2026:
-Go through 1-2 years of standard of care (poison/slash/burn - original poster forgot radiation and hormone suppression). You’re only a good candidate for a vaccine if you are cancer free or have minimal residual disease.
-Hire PhDs / Lawyers to fight with your hospital for your tumor tissue, which is likely stored in a way that makes sequencing it hard / impossible
-Pay $50K+ to sequence your tumor tissue
-Pay $1M+ for a custom immunotherapy / vaccine with no efficacy or safety info
-Find an oncologist and medical setting to sponsor this plan, the biotechs that made you the drug can’t administer it and you’ve likely alienated your actual oncologist by now
-Hire lawyers: Petition the FDA for an n-1 clinical trial or relocate to Germany (or maybe Montana) to take vaccine series
-Take the series. You won’t actually know if it worked (no data) so you monitor all your markers constantly with ctDNA and other testing to search more therapies and to see if the vaccine activated your immune system and worked.
There’s a lot of promise here and people who can access frontier cancer care are so fortunate. Every cancer patient wants every chance they can get. But it’s bs to say we’ve solved cancer in 2026.
Alignment is an issue of memetics. In the context of AGI, alignment is less of an engineering challenge and more of a profound cultural and psychological endeavor. We cannot steer a “super intelligence” away from mind viruses that we ourselves are unconsciously carrying into its high-dimensional space.
Natural products are the ideal launching pad for exploring the latent biodesign space
Tempus’ genius is their ability to acquire ground truth patient data/biosamples at scale and at a profit.
Tempus trained a foundation model on 1.67M real cancer patients, fusing clinical trajectories with DNA, RNA, and H&E pathology into one patient-state embedding. Overall-survival AUC jumped 0.563 to 0.774, and treatment-benefit ranking improved in 9 of 11 cohorts.
We have argued for years that multimodal models beat curated features in oncology. This is strong evidence at real scale.
arxiv.org/abs/2608.24688
