Stargate AI can cure cancer

AIs, personalized targeted medicine and a little RAGtime

Perfecting Equilibrium Volume Three, Issue 28

Come on and hear, come on and hear Alexander’s ragtime band.
Come on and hear, come on and hear, it’s the best band in the land.
They can play a bugle call like you never heard before,
So natural that you want to go to war;

The Sunday Reader, Jan 26, 2025

Tuesday’s half-trillion dollar Stargate AI investment announcement at the White House was surreal. Then things got really weird. Larry Ellison, the Baba Yaga of Silicon Valley, got up to announce AIs are going to cure cancer.

One of the most exciting things we're working on ... is our cancer vaccine," Ellison said. "You can do early cancer detection with a blood test, and using AI to look at the blood test, you can find the cancers that are actually seriously threatening the person. You can make that vaccine, that mRNA vaccine, you can make that robotically, again using AI, in about 48 hours.

The fact that Ellison was speaking publicly at all was the first weirdness. He’s Silicon Valley’s unknown billionaire; while Bill Gates and Steve Jobs, Jeff Bezos and Mark Zuckerberg et al were commanding magazine covers and society pages – and Elon Musk is everywhere – the Oracle co-founder was apparently in a bunker plotting successful takeovers of tech company after tech company and subsuming them into his platform. The list of tech titans that have disappeared into Oracle’s maw is almost endless: Sun Microsystems. PeopleSoft. Siebel, which built the market for CRM – Customer Relationship Management systems. JD Edwards. NetSuite. Thinking Machines.

But even weirder than Ellison talking was what he said:

Stargate AI will cure cancer.

Is he serious? Is this some Monty Python skit? Or is such a thing really possible?

Yes it is. In fact tailoring medical treatments is a problem well suited to the strengths of Large Language Models, the actual name of the software commonly called AIs.

I’ve long said modern medicine is as much voodoo as science, which annoys my medical researcher son. All three of my children will tell anyone who will listen that I am annoyingly dramatic and given to telling jokes that no one gets or finds amusing except me.

To which I reply: What’s your point?

My son will also tell you I’m not wrong about medicine. Here’s why: Medicine is faced with the Microsoft problem.

For decades everyone has known that all Apple devices “just work,” and just work together, while smart users always waited to install new Microsoft products until the first three or four rounds of patching were done. And this is true, but not because of the software.

Look at the hardware. All Apple devices are made by Apple. And there aren’t that many of them. And Apple routinely changes a port or a dongle or some such to force older items to be trashed and replaced with current iDevices.

Microsoft, meanwhile, has more vendors building devices than Apple has devices. And they range from standard desktops and laptops, to laptops with multiple screens such as the Asus Zenbook Duo 2024 – which switches screens depending on where you place a magnetic keyboard – to flexi screens that change size at the press of a button. So there are a zillion combinations no one predicted, and it takes a while to support them all.

Now do humans.

The current world population is 8.2 billion. That means if someone is one in a million…there are 7,999 just like them.

And yet how often have you seen identical twins?

And yet look at the differences even between fraternal twins.

So why would anyone think that a single formulation of a drug would work the same way in billions of humans with such wildly disparate DNA and gut microbiota?

Obviously one-size-fits-all drugs are severely limited. So doctors basically experiment on their patients – let’s stick a hypodermic needle in them and see what happens!

Sounds like voodoo, doesn’t it? Just listen to those disclaimers on every drug ad ever: (Very LOUD voice) CURES THE HEARTBREAK OF psoriasis!!! (Tiny voice speaking very very very fast) May cause spontaneous combustion, your left pinky toe swelling to the size of a watermelon, and …

Death.

What’s needed is a way to break those 8 billion humans down into their one-in-a-million groups of 8,000.

And sifting through tons of data, finding the key points and summarizing them is exactly what LLMs are best at doing, as we saw when an army of nerds last month in 24 hours defenestrated a bipartisan spending bill laden with pork after congresscritters had introduced it with the words “we’ll have to pass it to see what’s in it.”

Unfortunately for our esteemed congresscritters, we have now entered the third great storm of information technology innovation, when new hardware and software converge to unleash legions of nerds creating things that are entirely new. Large Language Models are not intelligent, artificial or otherwise. But nerds are busy chaining LLMs together into entirely new types of information tools. And digesting reams of bureaucratic speak into neat summaries accompanied by bulleted high points is right up their alley.

Now imagine those same tools turned loose on billions of patients’ medical records and DNA profiles. Suddenly its very possible to find your group of 7,999 replicants and tailor treatment for you.

Will that work? No way to know except to do it and find out. That’s how science works: come up with a falsifiable hypothesis that fits the facts, then test it to see if it works.

Ellison’s hypothesis here is plausible, and falsifiable. That’s progress. As we say out here in startup land, Fail Faster: Hypothesis; Test; Fail; Learn.

Then use that learning to refine and fix your hypothesis as necessary, and do it again. Rinse and repeat until you succeed!

We’ll come back to this subject. First, let’s dive into how you would build such a thing, how LLMs, RAGs and CAGs, agents and workflows all fit together, how training and tuning can make LLMs “hallucinate,” and how to prevent such crazy responses by controlling the data set used to provide answers.

After all, it’s stupid and funny when an AI recommends adding a quarter cup of Elmer’s glue to your pizza sauce to keep the cheese from sliding off because the AI was trained on Reddit and cannot tell meme lord trolling from actual posts from chefs, and you hope no one ended up with a tummy ache.

But’s not funny at all if we’re talking cancer treatments. Indeed, it is entirely unacceptable.

So we’ll look at how training and tuning works, and how to create useful AIs by cleaning and controlling a local data set to be used for answers. And we’ll look at how we can assemble data sets that will form the backbone of the AI powered Virtual Newsroom we’ve been discussing.

Let’s start with workflow, which is simply the order in which work is done. Everything has a workflow. If you’re making breakfast, you heat the pan, cook the bacon and eggs, then eat them. You can’t eat them first; it’s disgusting, and you’ll get trichinosis.

Seriously. Don’t do that.

For computers, workflow is the famous If A, then B; else C. Let’s look at how that works by laying out a simple budget spreadsheet. Here’s a spreadsheet that compiles a company’s monthly expenses. Each month is simple: A cell to enter that month’s expense for each category, and then a monthly total. Note at the top that the monthly total – cell F16 – is actually the formula =SUM(F10:F15), which totals the values entered into cell F10-F15.

Here's the same spreadsheet expanded out to cover the entire year. There’s now an annual total, which is a formula that adds up the 12 monthly formulas. And note that the formulas still work even though there’s no entry for rent in November, when the company got a free month for renewing their lease. GoTo Cell AD22; IF there’s a value add it to the running total; ELSE GoTo Cell AD23.

Yes, this is simple logic and simple math. But these simple building blocks can be snapped together to form the most sophisticated of budget tools. Add another sheet like this for revenue, then a third of formulas calculating profit and loss. Add a fourth sheet with an equipment depreciation schedule, and you’re on the path to a spreadsheet that calculates taxable income, EBITDA, and all the other accounting arcana so beloved by investors.

That’s workflow: a series of simple steps done in a logical order to achieve a reliable outcome. In our next installment we’ll look at how AI agents can choose between multiple workflows to create truly sophisticated applications such as medical diagnostics and virtual newsrooms.