December 19, 2004

December 19, 2004 by yorksranter 0 comments

Admin: TYR User Consultation

I have had the long-term project of redesigning the Ranter for some time. In the last few weeks, work on a new Ranter has proceeded apace and a nearly complete template has been in use on a test blog for several days. The test blog’s purpose is to check the new template’s performance, accessibility and…

HBM memory supply to scale in 2025
each HBM3E chip has a capacity of 24Gb, which is a 50% increase over the 16Gb of HBM3. With greater single-chip capacity of HBM and increasing GPU integration of HBM, large-scale mass production of HBM3E will require more overall DRAM capacity. According to industry analysts, HBM3E employs approximately two to three times the wafer volume […]
Towards Cross-Cultural Machine Translation with Retrieval-Augmented Generation from Multilingual Knowledge Graphs
Finally, we feed the hidden states 𝐡+kg+e to the decoder of the MT model, which is now able to also attend to the entity embeddings from the retriever and fuse them with its hidden states. Our intuition is that the embeddings 𝐞i from the knowledge retriever can contain useful fine-grained, latent information about the retrieved […]
Divide-or-Conquer? Which Part Should You Distill Your LLM? - Apple Machine Learning Research
These results indicate that by using smaller, distilled problem decomposition models in combination with problem solving LLMs we can achieve reasoning with cost-efficient inference and local adaptation // Apple is doing a lot of these! this one has a model that makes a plan, and then the plan is executed by other task specific models
4M-21: An Any-to-Any Vision Model for Tens of Tasks and Modalities
We are able to train a single unified model on diverse modalities by encoding them with modality-specific discrete tokenizers (see fig. 3). For image-like modalities, e.g. RGB or edges, we train ViT-based [24] VQ-VAE [63] tokenizers to map the inputs into a small grid of discrete tokens. For modalities like 3D human poses or image […]
The Fraudulent Science of Success - The Atlantic
Among the four research papers in which Gino was accused of cheating is one about the human tendency to misreport facts and figures for personal gain. Which is to say: She was accused of faking data for a study of when and how people might fake data. Amazingly, a different set of data from the […]
Generalizable Autoregressive Modeling of Time Series Through Functional Narratives - Apple Machine Learning Research
Analogous to the next word prediction task in languages that learns narratives by connecting different words, our autoregressive transformer aims to learn the Narratives of Time Series (NoTS) by connecting different functions in time // transformers are good at predicting the next word, if we look at a time series as a sentence of functions, […]
2410.09407v1.pdf
great paper! doing complicated agent stuff on a phone without RAGging. also MS Phi shines again. why why would Apple do that...
A Hackney Story - The Fence
Another drug dealer who sold crack and heroin as a young man in Hackney over the 1990s and 2000s, and who does not want to be identified because he still sells cocaine [well, that's a good reason!], remembers that LOM was a bigger threat than the police. // also contains the controversial assertion that Hackney […]
Nvidia outpaces revenue estimates with strong data center growth amid mixed analyst reactions
q-o-q now 17%
Japan expected to announce additional US$1.3 billion investment in Rapidus
The government's increased stake in Rapidus through direct investment allows it greater control and governance over the company, marking the first time it will hold equity in Rapidus. By October 2027, when Rapidus is set to begin mass production of 2nm chips, the government plans to exchange factory assets for shares in the company.