leanleft

model:
32B olmo-2 03/25

https://arxiv.org/abs/2501.00656

"We release all OLMo 2 artifacts openly -- models at 7B and 13B scales, both pretrained and post-trained, including their full training data, training code and recipes, training logs and thousands of intermediate checkpoints. "

"At the same time, daily life relies more than ever on digital technology: more things run on software (fridges, cars, phones), those things have a greater array of sensors (GPS receivers and radio transmitters) and they are increasingly connected, often over the internet, allowing data, often embodying our most personal secrets, to flow to and fro. The paradox of the modern world is that, while we have more means to keep our data secret, there is so much more data to contend with and so many more places from where it can seep out into the world, where a sprawling ecosystem of private intelligence can collect, analyse and use it."

opt out now

"Researchers at the Technical University of Denmark (DTU) and Lund University, Sweden, have used enzymes produced by a common gut bacteria to remove the A and B antigens from red blood cells, bringing them one step closer to creating universal donor blood."

"The most recent example is a now-merged merge request to revert an earlier change bumping the Zlib dependency for Mesa. The basis for that revert is that it breaks SPECViewPerf."

"Due to Mesa dynamically linking Zlib and how SPECViewPerf is handled, the update happens to break SPECViewPerf that is a popular benchmark for workstation graphics and one commonly used by hardware vendors and other stakeholders. Ultimately it's an issue with how SPECViewPerf is setup as an application bug but it could also be argued that Mesa could statically link it or better handle its dependencies. In any event, it's a regression for Mesa and breaks SPECViewPerf. And SPECViewPerf is important to vendors.

So the immediate solution that's now been merged is to revert that Zlib update commit..."

"They think it's a technical issue. It's not. It's a political and strategic issue for the Mesa community. If you prevent something from working that the industry finds important, you risk destroying real jobs in this community and shrinking it, regressing Mesa's reputation, making it more inferior in the industry, and thus less important. What this revert does is that it preserves existing jobs (i.e. existing stuff keeps working) and opens the door for creating new jobs and growing this community in a sustainable manner by showing others what it can do. You need capital and business interests to grow the community, and to get that, Mesa must be the best because it's always competing with alternatives.

If you thought this is only about dependencies, well, you're mistaken, and if you want to hurt the future of Mesa because your stupid zlib dependency is more important than anything else, including the livelihood of other people, you're just a foolish bikeshedder."

similar to other tools. the author says "RustViz is a bit more of a purely educational tool, as code has to be annotated manually, while Boris aims to be more of a development assistance"

it would be really great to have a lemmy client (or feature of existing client) that allows for batch downloading of a user specified list of communities.
this would allow a user to download all the content for the day or week on wifi internet and then depart from the source of internet but slowly & carefully read a selection of material(text posts, comment discussion, and even images like memes).
one benefit is that it would be extra impossible to see what users are loading/viewing because they already loaded everything and are disconnected from the internet entirely. performance is also good because there is no network latency that would be experienced, each time, when accessing the servers.