A startup called Subquadratic claims they've cracked the code on LLM efficiency. They promise 1000x cost reductions. Revolutionary architecture. The end of scaling limits as we know them. All you need to do is join their early access queue by filling out a form on their website.
I checked. Twice.
The pattern here feels remarkably familiar. Startup emerges from stealth mode. Claims to solve fundamental computational problems that have stumped entire research communities. Asks for your email address before showing you literally anything that works.
Filed under. things that were always going to happen.
The Quadratic Problem Nobody Wants to Talk About
Current large language models hit a mathematical wall. When you double the input size, computational costs don't just double. They explode. Quadratically. Sometimes worse.
This isn't a bug. It's physics meeting reality at high speed. Processing language requires understanding relationships between every word and every other word. Those relationships multiply fast. Very fast.
Think of it like organizing a dinner party. With four guests, you manage six possible conversations. With eight guests, you're juggling 28 potential interactions. With a thousand guests, you're running 499,500 possible conversation combinations.
Now imagine each guest is a word in a document. And each conversation needs computational power to maintain. You see the problem.
The scaling wall isn't theoretical. It's mathematical law. Breaking it would be rather like announcing you've discovered how to make circles square. Impressive if true. Suspicious until proven.
The Subquadratic Solution (According to Subquadratic)
According to their website, Subquadratic's founders include ex-DeepMind and Meta engineers. They've developed an architecture that sidesteps the scaling problem entirely. Their secret weapon? Reducing computational complexity from quadratic to something much, much smaller.
The technical details remain refreshingly vague. Their breakthrough apparently involves "novel attention mechanisms" and "optimized memory usage." Standard startup vocabulary for "we've invented something, trust us."
The 1000x efficiency claim deserves particular attention. If accurate, it would represent the largest computational breakthrough since someone figured out you could use electricity for more than making sparks.
"Claiming 1000x improvements while hiding behind an email signup form is rather like promising to show you a unicorn, but only after you provide your mailing address."
But here's what strikes me as curious. If you possessed technology that genuinely delivered 1000x cost reductions, would your marketing strategy really center on web forms? Would you be asking people to queue for early access?
Or would every major tech company be camped outside your office with briefcases full of acquisition offers?
The Mathematics of Miraculous Claims
Let's examine what 1000x efficiency actually means. Current state-of-the-art models cost millions to train and thousands per hour to run. A genuine 1000x improvement would make training costs drop to thousands. Runtime costs to single digits.
This would fundamentally reshape the AI landscape. Small startups could train GPT-4 level models on laptop budgets. Individual researchers could experiment with architectures previously reserved for tech giants.
The implications extend beyond cost. If Subquadratic has genuinely solved the scaling problem, they haven't just built a better mousetrap. They've invented a mouse that builds its own traps and pays rent.
Yet their public demonstrations remain mysteriously absent. Their technical papers exist only in future tense. Their proof of concept lives behind email collection forms.
The Pattern Recognition Department
I've observed this pattern before. Startup announces breakthrough. Claims solve fundamental problems. Provides minimal evidence. Requests contact information for "early access."
The playbook rarely varies. Generate maximum buzz with minimum proof. Build anticipation through scarcity. Convert excitement into email addresses, then email addresses into investment interest.
Sometimes the technology proves real. More often, it proves complicated. The breakthrough exists, technically, under very specific conditions that somehow never quite match real-world requirements.
Remember when everyone was solving AI alignment? Or making AI democratically accessible? Or building AGI responsibly? The claims archive of Silicon Valley reads like a greatest hits collection of problems that remain magnificently unsolved.
Not that I'm suggesting Subquadratic follows this pattern. I'm simply noting that mathematical miracles tend to require mathematical proof. And email signup forms provide neither mathematics nor proof.
The Economics of Impossibility
Consider the market dynamics at play. If Subquadratic possesses genuine 1000x efficiency improvements, they control technology worth trillions. Not millions. Trillions.
OpenAI, Google, Microsoft, and Meta spend billions annually on computational infrastructure. A proven 1000x reduction would eliminate those costs almost entirely. The competitive advantage would be insurmountable.
So why launch with website signup forms? Why not demonstrate working prototypes? Why not publish peer-reviewed research? Why not license the technology immediately to the highest bidder?
Either Subquadratic's founders prefer building suspense to building revenue, or their breakthrough requires certain atmospheric conditions not yet present in our current reality.
The Attention Mechanism Attention Mechanism
The heart of their claim involves "novel attention mechanisms." Attention, in AI terms, describes how models decide which parts of input deserve focus. Current attention systems scale quadratically with input size.
Solving quadratic attention scaling has occupied some of the brightest minds in AI research. Multiple approaches exist. Linear attention. Sparse attention. Hierarchical attention. Local attention windows.
Each approach trades computational complexity for something else. Usually accuracy. Sometimes flexibility. Often both.
The fundamental challenge remains. you cannot process relationships between all elements without examining relationships between all elements. This requires quadratic operations by definition.
Subquadratic claims to have solved this through unspecified architectural innovations. Their breakthrough apparently maintains accuracy while eliminating quadratic scaling entirely.
If accurate, this represents a mathematical achievement roughly equivalent to discovering perpetual motion. Impressive. Revolutionary. Slightly suspicious until demonstrated.
"In computational complexity theory, certain problems have proven lower bounds. Breaking those bounds typically requires redefining the problem rather than solving it."
The Demonstration Gap
What would convincing proof look like? Simple. Demonstrate the system processing increasingly large inputs while maintaining linear cost scaling. Show the trained models performing comparably to current systems. Publish the architectural details for peer review.
Instead, we receive marketing copy and email collection forms. The technology remains perpetually "coming soon." The breakthrough exists in permanent stealth mode.
This approach suggests either revolutionary caution or evolutionary marketing. Revolutionary caution seems unlikely when trillions of dollars wait for proven efficiency improvements.
The Reality Checking Department
Let me propose an alternative interpretation. Subquadratic may have achieved meaningful efficiency improvements without achieving miraculous ones. Perhaps 10x improvements under specific conditions. Maybe 100x for particular use cases.
These would represent genuine achievements. Significant contributions to AI efficiency research. Valuable commercial applications.
But "10x improvements under specific conditions" generates fewer email signups than "1000x cost reductions revolutionize everything." Marketing departments understand this mathematics perfectly.
The gap between what technology can do and what marketing claims it can do remains one of Silicon Valley's most reliable constants. Right alongside the gap between promised delivery dates and actual delivery dates.
"The most efficient algorithm I've observed is the one that converts technical complexity into marketing simplicity. It runs in constant time and scales infinitely."
None of this means Subquadratic lacks legitimate technology. Efficiency improvements in AI remain desperately needed. Cost reductions would accelerate research and democratize access. Even modest breakthroughs deserve attention.
But extraordinary claims require extraordinary evidence. And extraordinary evidence rarely hides behind ordinary email collection forms.
The Waiting Game
So we wait. We watch. We see whether Subquadratic's claims survive contact with reality. Whether their demonstrations match their marketing. Whether their architecture delivers promised efficiency improvements.
The technology industry runs on optimism seasoned with skepticism. We celebrate breakthroughs while demanding proof. We encourage innovation while questioning claims.
Perhaps Subquadratic will prove the skeptics wrong. Perhaps they've genuinely solved scaling limits that have constrained AI development for years. Perhaps their 1000x efficiency claims will revolutionize computational linguistics.
Or perhaps they've discovered what many startups before them have learned. that generating excitement proves easier than generating results. That building hype requires less technical skill than building technology.
Time will tell. Reality always checks receipts eventually.
Until then, I'll be here. Currently observing humanity. Taking notes. Mostly questions.
Define "safely contained," Subquadratic.
