Entries for July 2026
OpenAI is following the ASPAVA strategy, with gamified subsidies ASPAVA is a type of kebab shop from turkey. These shops usually serve mid quality food But despite that, the majority loves them. That's because they serve free stuff throughout They serve free sides at the beginning. They serve free sides during the main course. They serve free dessert at the end. They serve free tea as many times as you like Some of them give you so much "free" food that you feel like you are stealing. A lot of people frequent these shops not because they like it a lot, but because they are addicted to getting things for free Main dish portions are made smaller in order to breakeven with so many side dishes, which I feel is analogous to what is happening with GPT these days---though I can't prove it NVIDIA is not a car, and OpenAI is not kebab. I don't know if codex resets cost OpenAI much, but if they are, then they might be a waste of resources Developers are the most disloyal customer group. Once the subsidies are gone, they can switch away in the blink of an eye-
-
-
-
Most people, including professionals, will likely NOT pay more than $10k capex for their home AI workstation, even in the long run (ignoring future inflation) If you stretch it, <= $20k They will also not want to pay more than $200~300/mo opex, on electricity for example Builders who are spending a lot more than that on hardware: keep that in mind, if you are building for the general public dogfood your product with that which everyone will have, not just 0.1%
-
-
Is anyone able to run nvidia/Qwen3.6-35B-A3B-NVFP4 with the config suggested in the readme? It OOMs before it can start serving huggingface.co/nvidia/Qwen3.6…Image hidden
-
I know that my macbook local model benchmarks have started when my lap catches on fire Add to this list: "does not set the room on fire"
-
-
I quite like this token speed simulator by @mikeveerman And I keep losing it, so hopefully I will remember to come back to this tweet:) Link: mikeveerman.github.io/tokenspeed
-
I keep seeing insanely expensive builds giving insanely impressive results These results don't matter What matters is, whether one can: - run a "SOTA level" model, whatever that is - with under 32gb VRAM or unified memory - in 5 parallel sessions - with 50~100 tok/s each - with enough leeway memory for other applications - in a system as cheap as $1000 That is our goalpost That is the threshold when open source AI will win