The Broccoli Problem: AI’s Precision Glitch
The Broccoli Problem is the most memorable teaching device in Marc Chaikin’s “The End of Elon?” presentation, and it has earned its nickname honestly: it works. Ask an AI model how many times the letter C appears in the word “broccoli,” and it can fail. On camera, Chaikin demonstrates exactly that, and then delivers the line that turns a spelling stumble into an investment thesis: “That lack of precision can’t be solved with better programming. You can’t solve it by feeding more data into the models. Because the floating point problem isn’t a software problem. It’s a hardware problem.”
Why the demo works
The letter-counting failure looks like a quirk, but it exposes something structural. Today’s AI models process text and numbers through hardware that stores values in low-precision formats, FP16 and below, chosen for speed. A 16-bit float can represent only about 65,000 distinct values, so tiny quantities get rounded away. Chaikin’s worked example makes it concrete: add .00001 five times, and FP64 hardware returns 1.00005, while FP16 hardware “spits back the number one.” In his framing: “FP16 chips can process 65,000 unique values at once… FP64 chips can handle 18 quintillion calculations at once.”
For spelling, rounding is harmless. For science, it is fatal. Simulating a fusion reactor, modeling a drug molecule, computing stress on a bridge wing: these accumulate billions of small operations where each rounding error compounds into meaninglessness. That is why every serious scientific supercomputer for decades has run FP64, double precision, and why the AI industry’s FP16 choice, correct for chatbots, leaves a ceiling on what AI can do in the physical sciences. Chaikin calls this “AI’s darkest secret.”
The fix, and the investment
The fix is hardware: FP64-powered AI machines. The first, called Lux, is being built at Oak Ridge National Laboratory in Tennessee, “the world’s first FP64-powered AI supercomputer,” described by its designers as “the first dedicated U.S. AI factory for science.” Its scale numbers are the presentation’s most vivid: about “the area of a 7-Eleven convenience store… Literally 0.001% the size of Colossus,” drawing “as little as 24 megawatts” versus Elon Musk’s Colossus campus at roughly two gigawatts, yet “more than a trillion times more powerful” for the precision-dependent work that matters. Lux becomes the first node of American Atlas, a network of nine new AI micro clusters plus “more than 50” converted DOE supercomputers, Frontier, El Capitan, Aurora, Perlmutter among them, “all becoming part of one giant AI breakthrough machine.”
The picks follow the buildout. The free one is AMD, whose Instinct accelerators already power the DOE’s Frontier and El Capitan, with “At least $1 billion dollars about to hit this company’s account.” The paid American Atlas Portfolio report covers “17 companies that have signed deals… They’re all working directly on building American Atlas,” led by a networking company providing “advanced networking solutions… The unsexy stuff that will make the next stage of the AI revolution possible.”
The honest limits
Two caveats keep the Broccoli Problem honest. First, the “283 trillion times more powerful” style comparisons count representable values, not real-world throughput; FP16 and FP64 machines are built for different jobs, and the presentation is careful, when read closely, to qualify its claims with “regarding major breakthroughs.” Second, the demo does not mean ChatGPT is worthless or that AI stocks collapse; it means scientific AI needs precision hardware, which is a new layer of the opportunity, not a refutation of the old one. Chaikin’s own framing supports that reading: “What I’m showing you today is a different layer of the same opportunity. I want you to have both.”
What the demo means for investors
The broccoli demo is the best sixty seconds of science education in any AI-era pitch we have reviewed, and it is worth being precise about what it does and does not show. Asking a chatbot to count the C’s in “broccoli” produces a wrong answer because the model does not see letters; it sees tokens. Chaikin’s framing is that this precision failure cannot be fixed with better programming or more data, “Because the floating point problem isn’t a software problem. It’s a hardware problem.”
The hardware fix is FP64: 18 quintillion addressable values against FP16’s 65,000. That is why the Department of Energy is building Lux, the FP64 supercomputer at Oak Ridge, for work where being approximately right is worthless: reactor design, drug discovery, materials science. Generative AI and scientific AI are different jobs on different hardware, which is the entire investment thesis.
Questions readers ask
Can’t software just get better at counting letters? For the letter-counting party trick, yes, and that nuance is the demo’s honest limit. The presentation’s real claim is about numerical precision in scientific simulation, where low-precision arithmetic genuinely accumulates errors that no amount of model scaling fixes. That is textbook computer science, and Chaikin teaches it correctly.
What is the trade? The free pick is AMD, whose Instinct accelerators power the DOE’s existing flagship machines, Frontier and El Capitan. The 17-stock American Atlas Portfolio holds the smaller builders of the FP64 network. Our full review covers the whole thesis.
Where to learn more
Start with our End of Elon? American Atlas review, then the FP64 technical explainer and the Lux and Oak Ridge detail. The full presentation, with the broccoli demo, is at the Chaikin Analytics offer page, discounted from $499 with a 30-day refund guarantee.
NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.