I first got into deep learning in 2012, when AlexNet came out. I was CTO of Jetpac, a startup that aimed to provide information about bars, hotels, and restaurants by analyzing public photos, for e…
Nice read but he answers the question himself at the end:
"Nvidia these days feels like Sun did then, and so I bet over the next few years there will be a lot of chatbot startups based on cheap PCs with open source models running on CPUs. "
In 1998 Linux was already available and able to be deployed. His company isn’t getting attention because he’s selling a promise that if he’s given $100m, he might be able to pull a rabbit out of the hat and deliver improved efficiency.
If you had $100m and could buy servers that would give you revenue tomorrow or spend $100m on R&D that might not ever pay off, which would you do?
He’s also making some false equivalencies. For instance, his startup (Moonshine) seems to be a plain STT dev… STT is dirt cheap and dime-a-dozen these days. Not to speak of more advanced speech ingestion coming out daily, like https://huggingface.co/stepfun-ai/Step-Audio-R1
He’s also assuming the same level ‘unpicked’ optimization exists as we had in 2012, with Fermi GPUs and barebones software support. It does not. Look at frontier stuff, like bitnet models with sparsity and CPU LUT kernels, and there is just not much theoretical room for dramatic optimization there.
I don’t mean to refute his general point; his sentiment on what investors are (stupidly) doing is spot on. But also, it’s not 2012.
Nice read but he answers the question himself at the end:
"Nvidia these days feels like Sun did then, and so I bet over the next few years there will be a lot of chatbot startups based on cheap PCs with open source models running on CPUs. "
In 1998 Linux was already available and able to be deployed. His company isn’t getting attention because he’s selling a promise that if he’s given $100m, he might be able to pull a rabbit out of the hat and deliver improved efficiency.
If you had $100m and could buy servers that would give you revenue tomorrow or spend $100m on R&D that might not ever pay off, which would you do?
+1
He’s also making some false equivalencies. For instance, his startup (Moonshine) seems to be a plain STT dev… STT is dirt cheap and dime-a-dozen these days. Not to speak of more advanced speech ingestion coming out daily, like https://huggingface.co/stepfun-ai/Step-Audio-R1
He’s also assuming the same level ‘unpicked’ optimization exists as we had in 2012, with Fermi GPUs and barebones software support. It does not. Look at frontier stuff, like bitnet models with sparsity and CPU LUT kernels, and there is just not much theoretical room for dramatic optimization there.
I don’t mean to refute his general point; his sentiment on what investors are (stupidly) doing is spot on. But also, it’s not 2012.