This breakdown of the subsidy model versus pay-per-use is sharp. The point about subscriptions being VC-subsidized loss leaders is something I dunno many folks realize when they're picking services. In my own workflow I've noticed how much the "unlimited" models get quietly capped when usage gets real. The cleanest insight here is recognizing that heavy usrs will eventually hit similar costs regardless of the pricing structure.
Like the Bitcoin and crypto world, we start first years with expensive, home made Graphic Cards, and evolve to ASIC a hardware that is much more efficient. I think end 2026 or 2027 on AI we will face the same principle. Some chinese with big brain will rollout AI ASIC that worth the price, and can run very advanced open source models on home-small offices. And KiloCode will be "THE right tool" to catch that horde of coders that develop with on premise. Great job
People are falling all over themselves today for Opus 4.5. By Q2 or sooner of this year we’ll be running an open weights drop-in replacement for it at home. Build better workflows, task/agentic harnesses and get comfortable with possibly swapping out the engine for a new one every 3-6 months. And as a bonus, just as with x86-64, the hardware stayed the same but software improved; so we can afford to sweat these hardware assets even longer.
Nothing lasts forever, especially free stuff. xAI got either all the data they need or they burned out their (marketing/training) budget.
This breakdown of the subsidy model versus pay-per-use is sharp. The point about subscriptions being VC-subsidized loss leaders is something I dunno many folks realize when they're picking services. In my own workflow I've noticed how much the "unlimited" models get quietly capped when usage gets real. The cleanest insight here is recognizing that heavy usrs will eventually hit similar costs regardless of the pricing structure.
Like the Bitcoin and crypto world, we start first years with expensive, home made Graphic Cards, and evolve to ASIC a hardware that is much more efficient. I think end 2026 or 2027 on AI we will face the same principle. Some chinese with big brain will rollout AI ASIC that worth the price, and can run very advanced open source models on home-small offices. And KiloCode will be "THE right tool" to catch that horde of coders that develop with on premise. Great job
People are falling all over themselves today for Opus 4.5. By Q2 or sooner of this year we’ll be running an open weights drop-in replacement for it at home. Build better workflows, task/agentic harnesses and get comfortable with possibly swapping out the engine for a new one every 3-6 months. And as a bonus, just as with x86-64, the hardware stayed the same but software improved; so we can afford to sweat these hardware assets even longer.