X

NVIDIA just trained a 14-billion-parameter AI using evolution, not calculus. Every AI today learns through backpropagation. It computes gradients, adjusts weights, repeats. It works, but it demands precision hardware and enormous GPU clusters. Evolution Strategies offered an alternative. Mutate the model, test it, keep what works. Like biological evolution. The problem was speed. Random mutations on GPUs were painfully slow. EGGROLL fixes this with one trick. It splits huge random matrices into two small ones per mutation. The model mutates, tests, and keeps what works. Hundreds of thousands of mutations run at once. > 100x faster training throughput > 91% speed of pure inference > Pretrains models using only integers > Competitive with backprop on reasoning > Works on non-differentiable systems It pretrained a language model from scratch using zero gradients. It also matched reinforcement learning methods on math reasoning tasks. Everyone kept scaling the calculus to train massive AIs. It turns out, we just needed to evolve.

🚨Breaking: Princeton researchers just ran the numbers on where AI is actually heading. The results should make every founder, investor, and policymaker stop what they are doing. Training OpenAI's next-gen model consumes an estimated 11 billion kWh of electricity. That is enough to power every home in New York City for a full year. More than the annual output of a nuclear reactor. For one model. One training run. And that is before a single user asks a single question. Every time someone uses a reasoning model like o1 or DeepSeek-R1, it costs 33 Wh of energy per query. A standard GPT-4 query costs 0.42 Wh. That is a 79x energy multiplier. Per query. At billions of queries per day. Now here is what nobody is saying out loud. The industry's answer to this is Stargate. A $500 billion compute campus. 5 gigawatts of power. Enough to run 5 million homes. Owned by the same four companies that already control the technology. They are building a new kind of utility. Except you do not elect its board. Meanwhile the models consuming all that energy still cannot reliably reason outside of math and code. Everywhere else they pattern-match. They hallucinate. They confabulate confidence. Princeton's argument is that this is not a scaling problem. It is a structural one. More parameters have not fixed it. More data has not fixed it. The architecture itself is the ceiling. Their alternative: stop chasing one god-model and build thousands of small specialists instead. Each one trained on curated domain data. Each one grounded in verified knowledge. Each one small enough to run on your phone. The energy comparison is not close. A cloud query to a reasoning model uses 33 Wh and 20 milliliters of water. The same query on a local specialist model uses 0.001 Wh. Zero water. That is 10,000 times more efficient. AlphaFold did not beat biologists by knowing everything. It won by going impossibly deep in one domain. A 14 billion parameter model trained on medical knowledge graphs just outperformed GPT-5.2 on complex clinical reasoning. Depth beats breadth when the domain is defined. The question nobody building these systems wants to answer: If the only path to general AI requires the energy output of a small nation, controlled by a handful of companies, running on hardware most of the world cannot access — is that actually intelligence? Or is it just the most expensive pattern matcher ever built?