HPC News Bytes – 20260921

SPEAKER_1 (0:05)

Welcome to HPC News Bytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies.

Doug Black (0:14)

Hi, everyone, welcome to HPC News Bytes. I'm Doug Black, and with me is Shahin Khan. In the lead up to Chinese President Xi's visit to the White House later this week, US Treasury Secretary Besant and Chinese Vice Premier, He LeFang, opened talks in New York yesterday, with an economic agenda that includes extending the trade truce, tariffs, Chinese purchases of US goods, and US concerns about the flow of critical minerals, especially rare earths.

Besant said the two sides would also discuss AI safeguards, including open and closed weight models, as well as ways to avoid a deeper split between the two AI ecosystems. So, mines, trade, chips, and AI policy are now sitting in the same negotiating room, Shahin.

Shahin Khan (1:06)

Yeah, there you have it. When technology dominates geopolitics is maybe when you want to call it technopolitics. Rare earths are critical for technology components, but there are other important minerals, so it makes sense to refer to the superset as critical minerals. We are also seeing growing attention given to both the materials supply chain, with minerals that can cause a pause in the manufacturing process, and the data supply chain that drives advances in AI and access rights. So it's tariffs, aircrafts, agriculture, and critical minerals in the same bilateral negotiation, but higher level, it's about the supply chain for bits and atoms and how they impact commerce and global power dynamics.

The open versus closed discussion, meanwhile, is gaining in importance. And more so with the proposed acquisition of Hugging Face by NVIDIA, which we discussed a few weeks ago. Because governments are now using those terms in negotiations, even though the industry still does not agree on what exactly quote unquote open should mean for AI systems.

Doug Black (2:16)

So the issue is what exactly is open, as in accessible via a traditional open source download model? And an effort called Open Waldo, spearheaded by Greg Kurtzer of CIQ and Rocky Linux, also a guest on our podcast previously, is trying to turn a stronger version of open source into something operational. Open Waldo's argument is essentially that if we are going to call AI open source, then the source should actually be open.

That is a much stronger proposition than downloading a finished model. Open weights are extremely useful. You can run the model yourself, fine tune it, keep your data local, and reduce dependence on an API provider, but you still may not know exactly how the model was produced. Under Open Waldo's stricter definition, open weights is considered only partially open. Fully open source opens all the digital pieces and documents as much of the human process as possible.

Shahin Khan (3:21)

So think of a trained LLM as the black box people ultimately use. Behind it are four major pieces. First is the code, second is the training data, third is the neural network itself, its size, its topology, and the mathematical functions that push information through it.

Fourth are the weights, the billions or trillions of numerical values on the network's connections that are adjusted during the training process until the model produces useful results. There can also be a fifth piece, as you indicated, human intervention during training, people ranking answers, steering behavior, correcting problems, and making development decisions. The effects of such human judgment can be captured in data and weights, but the historical human process itself is harder to save and reproduce.

Open Waldo is putting training data, licenses, provenance, tools, training recipes, and resulting weights into a public, inspectable chain. Since the inputs and processes are also open, another group can inspect the provenance, change the training mix, and given enough compute, attempt to reproduce or rebuild the model. I think it's a really important initiative from Greg.

Doug Black (4:38)

Okay, Reuters reports that SK Hynex is in exploratory talks with Intel about manufacturing memory chips in the US for the first time. The options reportedly include leasing part of Intel's delayed Ohio manufacturing complex or forming a joint venture involving Intel and large Cloud customers. SK Hynex says nothing has been decided and even the type of memory has not yet been specified. The company is already building a $4 billion HBM facility in Indiana, but that plant is for advanced packaging and testing. The memory wafers will still be fabricated in South Korea and shipped to Indiana. Meanwhile, China's CXMT says its fifth generation DRAM platform has entered mass production using more aggressive patterning to increase density and lower costs.

So memory localization is moving on both sides of the US-China technology divide.