smcnally

groq did an ASIC for llama and now for nvidia. Their cloud service is fast.

> NVIDIA Groq 3 LPU Inference Accelerator > The NVIDIA Groq 3 LPU is the next generation of Groq’s innovative language processing unit. Each LPX rack features 256 interconnected LPU accelerators that, together with the NVIDIA Vera Rubin platform, supercharge inference. Each LPU accelerator delivers 500 megabytes (MB) of SRAM, 150 terabytes per second (TB/s) of SRAM bandwidth, and 2.5 TB/s scale-up bandwidth.

https://www.nvidia.com/en-us/data-center/lpx/

Does this work on early Velop mesh nodes, by chance?

Some of unsloth studio’s code is Apache? Or some of lmstudio is?

Or more-realistic flawed looks as the case is here.

That’s beautiful. But you get aurora even sans G4, true? Was this more intense and spectacular?

Indirect Nuclear Fusion

“Mac-only” was disappointing to read, but OBS’ render performance has been fine on macos and linux even with older hardware. James Webb calls anything heavier than helium “metal.”

It’s also a Wonder Dog, a Canadian power trio not featuring Neal Peart, and a moment when we shouldn’t evacuate the Death Star.

> The technical term is enteral ventilation via anus (EVA).

Anecdatally, I have encountered multiple people with congenital capabilities re enteral locution via anus.

Wishfully, training astronauts for enteral ventilation via anus during extravehicular activities that involve writing an ongoing Prince song would be called “EVA EVA 4EVA.”

> After the Shuttle program ended in failure, work on reusable launch systems stopped for decades.

“ended in failure” sounds harsh in this context. Or is that just me whistling past the sunk cost fallacies?

`model.safetensors` for Qwen3-0.6B is a single 1.5GB file.

Qwen3-235B-A22B has 118 `.safetensors` files at 4GB each.

There are a bunch of models and quants between those.

Beyond solid benchmarks, Alibaba's power move was dropping a bunch of models available to use and run locally today. That's disruptive already and the slew of fine tunes to come will be good for all users and builders.

https://huggingface.co/collections/Qwen/qwen3-67dd247413f0e2...

The ROT13 cipher for API is NVK. NVidia Knows

> The DESI collaboration is honored to be permitted to conduct scientific research on I’oligam Du’ag (Kitt Peak), a mountain with particular significance to the Tohono O’odham Nation.

Anyone here know how a request like this was made or the permission given? I haven’t seen this previously.

A recent financial report on the media industry noted The Onion is on the verge of collapse due to, quote, “not being able to able to make sh*t up that is more idiotic than current reality.”

darktable does all of this. It’s a complex application like Aperture or Light Table. You run it on your own macos, Windows or Linux computer. You can write your own software to extend or change it. Photos.app does most of this sans the Windows, Linux or “write your own” parts.

“Your Honor, why should I bother obeying a law my elected representatives could not be bothered to write?” seems like it should be a reasonable defense, but you’re right that the onus is on us as The Governed to know and understand every bit of slop and hallucination on the books.

OpenAI and Anthropic are definitely among the leaders. Playing catch-up to these leaders' mind-share and technology is some of the motivation for others. Calling the progress being made in the space by Google (Gemini), MSFT (Phi), Meta (llama), Alibaba (Qwen) "nice and all" is a position you might be pleasantly surprised to reconsider if this technology interests you. And don't sleep on Apple and AMZ -

In the space covered by Tabby, Copilot, aider, Continue and others, capabilities continue to improve considerably month-over-month.

In the segments of the industry I care most about, I agree 100% with what the commenter said w/r/t expecting major improvements every few months. Pay even passing attention to huggingface and github and see work being done by indies as well as corporate behemoths happening at breakneck pace. Some work is pushing the SOTA. Some is making the SOTA more widely available. Lots of it is different approaches to solving similar challenges. Most of it benefits consumers and creators looking use and learn from all of this.

A compiler in the mix is very helpful. That and other sanity checks wielded by a skilled engineer doing code reviews can provide valuable feedback to other developers and to LLMs. The knowledgeable human in the loop makes the coding process and final products so much better. Two LLMs with tool usage capabilities reviewing the code isn't as good today but is available today.

The LLMs overconfidence is based on it spitting out the most-probable tokens based on its training data and your prompt. When LLMs learn real hubris from actual anonymous internet jackholes, we will have made significant progress toward AGI.

That is definitely an issue with many LLMs. I've had limited success including instructions like "Don't invent facts" in the system prompt and more success saying "that was not correct. Please answer again and check to ensure your code works before giving it to me" within the context of chats. More success still comes from requesting second opinions from a different model -- e.g. asking Claude's opinion of Qwen's solution.

To the other point, not admitting to gaps in knowledge or experience is also something that people do all the time. "I copied & pasted that from the top answer in Stack Overflow so it must be correct!" is a direct analog.

> LLMs also love to double down on solutions that don't work.

“Often wrong but never in doubt” is not proprietary to LLMs. It’s off-putting and we want them to be correct and to have humility when they’re wrong. But we should remember LLMs are trained on work created by people, and many of those people have built successful careers being exceedingly confident in solutions that don’t work.

> I'm not sure this is grounded in reality. We've already seen articles related to how OpenAI is behind schedule with GPT-5.

Progress by Google, meta, Microsoft, Qwen and Deepseek is unhampered by OpenAI’s schedule. Their latest — including Gemini 2.0, Llama 3.3, Phi 4 — and the coding fine tunes that follow are all pretty good.

A deeper version of the same idea is to ask a second model to check the first model’s answers. aider’s “architect” is an automated version of this approach.

https://aider.chat/docs/usage/modes.html#architect-mode-and-...

for x86 or PowerPC?

From a vision PoV, QvQ has done very impressive work analyzing a photo of 0) celebrities, 1) a dog, and 2) a cartoon from The New Yorker. Other models have issues with one or more of these.

https://github.com/smcnally/smcnally.github.io/blob/main/IMG...

Restating the question to start the response is good practice for Actual and Artificial Intelligence.

Hoarder has a chrome plugin, Firefox addon, and apps for Android and iOS for which the app store says something I’ve never seen before:

“Data Not Collected — The developer does not collect any data from this app.”

Among the frightening parts of that is the time and human lives it has taken between Boeing destroying its engineering culture and starting to pay the price for that destruction.