The part nobody prices in is the stranded-asset scenario. These underwritings assume a demand curve that keeps rising for a decade. If inference efficiency improves the way training efficiency did — and there’s real pressure in that direction now — a lot of this capacity gets repriced hard.
The bank behavior is the tell: debt that gets packaged and sold on right after underwriting is the classic ‘we have concerns but we’re not holding the bag’ move.
I would say there are definite advances for inference. Custom backends built by AI for specific models, such as Strata, makes efficient leveraging of hardware much easier. Engrams, such as those in Qwen Next-Flash, allows for using more parameters via SSD, rather than memory alone.
These and other things over the past few years have allowed me to use bigger models, faster. Given another year of such progress, I think the value of subscribing to big American companies would start drying up.
Sorry, ignorant question from someone who didn’t understand much of your comment - do you mean that AI models are writing code that improves them? Cos that would be interesting.
Not improvements to the AI themselves, but rather the software that runs them. Backends like Unsloth are typically generalists - they are compatible with all sorts of hardware, but in exchange, they don’t incorporate the latest improvements or ideas. Projects run by pure human implementation tend to require a lot of oversight to ensure that additions don’t break functionality.
AI is really good at collecting and aggregating information. This makes them ideal for creating tailor-made products that have a narrow focus, since they raid many Githubs and papers, then attempt to make them into a product. Strata is one such backend, which is specially designed for Qwen Next-Flash.
Strata runs much faster than Unsloth and other generalist backends, but lacks flexibility in what it is suited for.
At least, that is my admittedly basic understanding of the situation…
The part nobody prices in is the stranded-asset scenario. These underwritings assume a demand curve that keeps rising for a decade. If inference efficiency improves the way training efficiency did — and there’s real pressure in that direction now — a lot of this capacity gets repriced hard.
The bank behavior is the tell: debt that gets packaged and sold on right after underwriting is the classic ‘we have concerns but we’re not holding the bag’ move.
I would say there are definite advances for inference. Custom backends built by AI for specific models, such as Strata, makes efficient leveraging of hardware much easier. Engrams, such as those in Qwen Next-Flash, allows for using more parameters via SSD, rather than memory alone.
These and other things over the past few years have allowed me to use bigger models, faster. Given another year of such progress, I think the value of subscribing to big American companies would start drying up.
Sorry, ignorant question from someone who didn’t understand much of your comment - do you mean that AI models are writing code that improves them? Cos that would be interesting.
Not improvements to the AI themselves, but rather the software that runs them. Backends like Unsloth are typically generalists - they are compatible with all sorts of hardware, but in exchange, they don’t incorporate the latest improvements or ideas. Projects run by pure human implementation tend to require a lot of oversight to ensure that additions don’t break functionality.
AI is really good at collecting and aggregating information. This makes them ideal for creating tailor-made products that have a narrow focus, since they raid many Githubs and papers, then attempt to make them into a product. Strata is one such backend, which is specially designed for Qwen Next-Flash.
Strata runs much faster than Unsloth and other generalist backends, but lacks flexibility in what it is suited for.
At least, that is my admittedly basic understanding of the situation…