@MilksandMatcha: "Most of the real world is actually the long tail. And how do you learn the long tail as cheaply as possible is one of …
Summary
In a tweet, Sarah Hooker argues that GPUs are ill-suited for the long-tail distribution of real-world data, suggesting a need for alternative AI hardware.
View Cached Full Text
Cached at: 06/18/26, 02:06 AM
“Most of the real world is actually the long tail. And how do you learn the long tail as cheaply as possible is one of the most important questions right now.”
@sarahookr on why GPUs — built for dense, predictable matmuls — are the wrong tool for the world AI is actually moving into.
Presented with @Cerebras and @alyciazcary
Similar Articles
@TheAhmadOsman: We’re entering an era where we use the models that run on expensive hardware to make models runnable on cheap hardware …
Ahmad Osman argues we're entering an era where large models running on expensive hardware are used to create models runnable on cheap hardware, with ODS's mission being to make local AI the default on everyday devices like a 2013 laptop.
@LottoLabs: The skills you learn from running local models is more valuable than the cost of the hardware
This tweet argues that the skills gained from running local AI models are worth more than the hardware cost.
@TheAhmadOsman: Don't you think your GPUs will be renting their own GPUs to train a better and more improved version of themselves
A Twitter exchange speculating whether GPUs will eventually rent more GPUs to train improved versions of themselves, while noting that insufficient compute and token generation will remain the bottleneck for AI progress for years.
"Hardware is the only moat" - Should we buy new hardware now or wait?
The article discusses the growing importance of hardware as a competitive advantage in AI, noting that leading labs are prioritizing product competitiveness and compute scale over pure AGI research. It highlights the resulting strain on consumer GPU availability and the increasing costs for hardware upgrades.
@Skaly__Bull: Traditional AI stack is walking dead They just don't know it yet $10K enterprise servers, data-center GPUs, racks and c…
The author argues that the traditional enterprise AI stack is obsolete, claiming a $599 Mac mini running Ollama can handle 80% of AI workloads locally for a fraction of the cost of renting cloud GPUs.