81–90 of 111 posts
The drop in memory stocks seems counterintuitive to me. The demand for memory isn't going to go down, we'll just be able to do more with the same amount of memory.
[flagged]
Earlier quoted context omitted.
It will also help with local inference, making AI without big players possible.
It's already possible. Post-training is vastly more important than model size. (There's bigtime diminishing returns with increasing model size.)
My experience doesn't disagree, at least. I've been using Qwen for coding locally a bit. It is much better than I thought it would be. But also still falls short in some obvious ways compared to the frontiers.
The drop in memory stocks seems counterintuitive to me. The demand for memory isn't going to go down, we'll just be able to do more with the same amount of memory.
I think that either investors were extremely skittish that the stocks might crash and jumped at the first sign of trouble (creating a self-fulfilling prophecy) or they were trading on non-public information and analysts who don't have access to said information are reading too much into the temporal coincidence of the Google Research blog highlighting this paper.
This part sounds especially cool. I did not think about this application when reading the other articles about TurboQuant. It would be cool to have access to this performance optimization for local RAG.
"The TurboQuant paper (ICLR 2026) contains serious issues in how it describes RaBitQ, including incorrect technical claims and misleading theory/experiment comparisons. We flagged these issues to the authors before submission. They acknowledged them, but chose not to fix them. The paper was later accepted and widely promoted by Google, reaching tens of millions of views. We’re speaking up now because once a misleadin…
Earlier quoted context omitted.
The hyperscalers do not want us running models at the edge and they will spend infinite amounts of circular fake money to ensure hardware remains prohibitively expensive forever.
> of circular fake money Oh it gets worse than that, the money which caused all of this by OpenAI was taken from Japanese banks at cheap interest rates (by softbank for the stargate project), and the Japanese Banks are able to do it because of Japanese people/Japanese companies and also the collateral are stocks which are inflated by the value of people who invest their hard earned money into the markets So in a way…
We will not see memory demand decrease because this will simply allow AI companies to run more instances. They still want an infinite amount of memory at the moment, no matter how AI improves.
If models become more efficient we will move more of the work to local devices instead of using SaaS models. We’re still in the mainframe era of LLM.
"The TurboQuant paper (ICLR 2026) contains serious issues in how it describes RaBitQ, including incorrect technical claims and misleading theory/experiment comparisons. We flagged these issues to the authors before submission. They acknowledged them, but chose not to fix them. The paper was later accepted and widely promoted by Google, reaching tens of millions of views. We’re speaking up now because once a misleadin…
I guess I'm trying to understand. I'm hearing this paper has been around for a year -- I would think that many companies would have already implemented and measured its performance in production by now... is that not the case?
The main breakthrough [rotating by an orthogonal matrix to make important outliers averaged acrossed more dimensions] comes from RaBitQ. Sounds like the RaBitQ team was much more involved, and earlier, and the turbo quant paper very deliberately tries to avoid crediting and acknowledging RaBitQ.
My understanding is that the efficacy of these methods isn't in dispute, what turboquant did was adapt the method that was being used in vector databases and adapted it for transformers, and passed it of more as a new invention than an adaptation.
The drop in memory stocks seems counterintuitive to me. The demand for memory isn't going to go down, we'll just be able to do more with the same amount of memory.
Given that increasing model size doesn't yield proportional increases in intelligence, there is a world where these datacenters don't have a positive ROI if we make these models even a fraction as effective as the human brain.