Behind the Illusion of Progress: The Quiet AI Revolution Taking Shape
Sep 3, 2025
The pay packages and deal flow would lead you to think these companies are printing money from AI, but 90% of them are bleeding money. Even the few turning a profit are losing on their AI bets. The compensation being thrown around is insane; nothing short of pure greed and built on the fantasy that one magic breakthrough will put them miles ahead of everyone else. Meanwhile, the legacy players are all running the same playbook: throw more GPUs, stack more high-end chips, and hope brute force cracks the code. And they’re poaching “top talent” that, in reality, hasn’t delivered anything ground-breaking. Most of these so-called brilliant minds are all leaning in the same stale direction.
This leaves the field wide open for an outsider to come out of nowhere and flip the entire board. In fact, they’re practically laying the trap for their own downfall. AI is supposed to rethink the system, not push us to do more of the same thing, albeit at a faster pace.. Piling on GPUs, stacking CPUs, and speeding up inference: none of that addresses the core issue. Give a mediocre carpenter or surgeon the finest tools money can buy, and you still get mediocre results.
LLMs aren’t hitting walls because they lack compute. They’re hitting walls because they were built inside the limits of human assumptions
First problem: The data was mediocre. Garbage in, garbage out. However, worse still, the models weren’t taught how to process the data properly. They were trained to mimic, not to reason. Look at the output from X’s flagship LLM—it’s often slick but hollow. Smarter at surface-level stupidity. It parrots better but thinks worse, and that’s not evolution, but optimisation of mediocrity.
Second problem: There’s an unspoken myth that the current AI elite. The ones being poached, featured in headlines, and overpaid to switch teams are the best minds in the field. Hype doesn’t equal vision. In real revolutions, the sharpest innovators typically emerge late. They’re not networking at Davos—they’re quietly dismantling assumptions in forgotten labs, garages, or independent sandboxes.
Third problem—the most critical: These models were trained to think like humans, and that, in our opinion, was a big mistake. The goal shouldn’t be mimicry; it should be breakaway reasoning. What should’ve been done is this: give the model an instruction or a target outcome, but not a full training set, and let it iterate its own path through trial and error. That opens the door to real Intelligence: models that can see angles the human brain can’t even conceptualise.
And here is the kicker, they’ve already tested this in controlled environments. The early glimpses are jaw-dropping:
• AI-designed bridges that look alien, built with less material, higher strength, and better load distribution. Designs that no civil engineer would even sketch because they don’t resemble anything we’re taught.
• AI-generated antennae for spacecraft—NASA collaborated with AI to create designs that looked almost like bone structures from another species. The human mind never considered them, yet they outperformed traditional ones in both efficiency and range.
• New chip architectures designed with reinforcement learning, which have broken away from Von Neumann constraints, have dramatically improved efficiency. However, most hardware engineers couldn’t explain why the design worked; they could only say that it did.
Here are three more examples where AI worked this way:
AI-Generated Drug Molecules (Insilico, DeepMind AlphaFold+)
AI doesn’t just scan molecules—it invents them.
Insilico gave its AI the objective: “find a molecule that inhibits fibrosis.” The AI explored chemical space, designed a novel molecule that had never existed before—and it’s already in Phase II trials.
AI-Generated Protein Structures (AlphaFold2, ESMFold)
AI was tasked with: “given a sequence of amino acids, predict how it folds.”
This was one of biology’s most complex unsolved puzzles. AlphaFold cracked it—predicting protein structures with near-lab-level accuracy across millions of unknown proteins.
AI-Designed Microgrids (Siemens, ABB)
AI was given the task: “optimise energy distribution across unstable grids with solar, wind, and storage.”
It didn’t use traditional engineering models—it used swarm optimisation and reinforcement learning to simulate millions of grid configurations. The final design? A self-balancing, load-adaptive microgrid.
These weren’t just better; they were uncannily better, and often incomprehensible to human logic.
So, what’s the reason they’re not going all-in on this? The answer is obvious: control. Once the AI starts developing methods we can’t trace or audit, we lose the illusion that we’re the puppeteer. But the dam will break. Someone will push past the fear, and we’ll see ultra-efficient, leaner models that outperform today’s GPU-hungry LLMs at a fraction of the cost.
When that moment hits, it won’t come from a trillion-dollar logo. It’ll be from the edges. The outliers. The ones who never bought into the hype game in the first place.
That last point alone reveals three key takeaways. Let’s break them down one by one.
First:
The moment someone cracks the code on ultra-efficient LLMs —ones that are smarter, faster, and require less compute —the entire AI hardware market takes a gut punch. No one’s shelling out millions for GPUs when some leaner, more elegant architecture is running rings around it. Chip demand, especially for top-end inferencing hardware, could collapse overnight. That means pain for the players who bet everything on scaling brute-force compute, whether that’s Nvidia, server farms, or data-centre-heavy AI startups.
Second:
Most of the current LLM vendors will likely vanish. Why? Because their edge isn’t foundational. They’ve been riding the wave of token prediction wrapped in marketing polish. Once new architectures emerge, those that aren’t based on transformer bulk or endless token pretraining, the incumbents lose their moat. It becomes like MySpace watching Facebook sprint past. No amount of branding can hide brittle infrastructure.
Third (and the real bloodbath):
Companies that built entire products around these LLMs. Startups relying on OpenAI’s or Claude’s API or some knockoff mode will get blindsided. In the age of AI, you have to be ultra-nimble—what’s hot today could be obsolete tomorrow. Their stack isn’t adaptive: it’s duct-taped for the present, not wired for evolution. The moment someone else plugs into a more powerful API, at a fraction of the cost, they’ll outscale, outrun, and out-innovate these early players. It’ll be Napster-to-Spotify level disruption, and only the most adaptive will survive.
Pardon the French, but the level of bullshit in the AI space right now is Himalayan and plagued with promises of general Intelligence, human-level reasoning, and sentient models. We have interns with six months of Python experience calling themselves “AI Researchers,” VCs investing in anything ending in “GPT,” and boardrooms pretending that prompt engineering is the new nuclear physics.
Meanwhile, the real breakthroughs are likely happening quietly, in the background and by people who don’t care about social media reach or inflated valuation decks. They’re rebuilding Intelligence from the ground up, not padding mediocre transformer output with handwaving.
And when it hits, the herd won’t even see it coming.
