Why the Future of AI is Dumb & Dirty - Salar
SubscribeSign in
Why the Future of AI is Dumb & Dirty<br>Why the next paradigm shift is horizontal, open, and cheap
Salar<br>Jun 13, 2026
Share
During the Russia-Ukraine conflict, asymmetric warfare was proven to be effective by Operation Spiderweb. Ukraine's sub-$1000 smuggled FPV drones caused roughly $7 billion in damage to Russian strategic bombers. Iran's primary weapon during the Iran-US conflict has been the Shahed-136 drone and its newer jet powered variant, the Shahed-238. The domestic production cost estimates of the Shahed runs around $10,000-50,000 per unit. These drones have caused the loss of F-15Es and a Chinook for Kuwait. The Pentagon's response was to reverse-engineer the same airframe into its own cheap drone (LUCAS) rather than rely purely on expensive interceptors - even the side with trillion dollar defense budget ended up building the $10k drone. A state of the art drone that cost a million dollars to make seems impressive but, when it's facing a 100 dumb drones that cost 10 thousand dollars each to make, it stands no chance. This was the key takeaway from the lecture our professor gave in a PDC class last semester.<br>The Pattern
If you look into the past, it seems like all new technology goes through the same cycle: A new technology is accomplished, it's made available to the general public, it's improved to the highest degree a publicly available tech could reasonably be, improving it further runs into the bottleneck of financial feasibility, and cheaper, dumber versions of the tech are commoditized that are good enough for the general public. In the end, the tech that's used the most widely is the one that's very cheap to manufacture and distribute. The first cars, in the 1880s, were effectively luxury novelties that almost nobody could afford. It took roughly two decades of pushing the technology forward before the Ford Model T arrived in 1908 and forced the industry to build something the average person could actually buy. That same logic still plays out today -- the largest share of the car market belongs to Toyota, Volkswagen, and Hyundai, all of which focus on building cars that are affordable and good enough rather than maximally advanced. Each year there are minor improvements to a Civic, but it's nowhere near a Rolls-Royce that most people will never own.<br>This pattern can also be seen in computing Infrastructure too. As Drepper said in his "What Every Programmer Should Know About Memory", scaling these days is most often achieved horizontally instead of vertically, meaning today it is more cost-effective to use many smaller, connected commodity computers instead of a few really large and exceptionally fast (and expensive) systems. The same logic that pushed cars toward the Civic pushed computing toward racks of cheap, interchangeable hardware instead of a few exotic supercomputers.<br>This is the common pattern throughout all tech, whether it's consumer tech, military tech, or business infrastructure. The final stage of the life cycle is to create versions of the tech that's cheap to manufacture and good enough for most people. Not to say that the efforts to higher ceiling of the said tech stop -- it just means the majority of the market ends up being consumed by the dumb and dirty version of the tech, not the cutting-edge one.<br>AI's turn
This pattern is starting to emerge in AI. When OpenAI's ChatGPT (utilizing the GPT-3.5 model) came out in 2022, all major tech companies started racing towards the common goal of creating the best model the world has seen. by 2025-26, with releases like Anthropic's Sonnet and Opus, OpenAI's GPT-4 and GPT-5, it seems like the race has produced AI capable of solving most people's day-to-day problems. As much as people like to criticize the new models, as working software engineer, Sonnet and Opus seem to get the job done for what I need, and I suspect that's true for most people.<br>Two forces are now pushing AI toward the same end state as cars and computing. The first being economic: Anthropic and OpenAI are reportedly losing money on the compute required to serve their SOTA models to the public, and the direction ther're both signaling is exactly the commoditization endpoint this essay describes -- Sam Altman, at BlackRock's Infrastructure Summit in March 2026, said the future of intelligence is a metered utility like electricity or water, explicitly reaching for the 'too cheap to meter' line from the nuclear industry. That's the SOTA-loses-money, commodity-tier-wins pattern, stated outright by the person running one of the two labs racing to build the SOTA. NVDIA's push the commoditize datacenter hardware is the same horizontal scaling story Drepper described, just one layer down the stack. The second is instability at the frontier itself.<br>Just this week, Anthropic was forced to disable public access to its newest model, Fable 5, along with Mythos 5, after the US government issued an export control...