woadwarrior01

This one's authored by a human, the other one is authored by Claude.

> Does the M4 and later ANE expose any additional capabilities, or is it just a higher-performance iteration of the same thing?

IIUC, M4 introduced a fast path for INT8 weights and activations (w8a8). M5 Ultra, M6 and A20 have two ANEs.

> As an aside, the introduction to this article seems to conflate the ANE with the Neural Accelerators (NAX)

Yeah, that part is true. NAX cores are matmult accelerators, closer to tensor cores in NVIDIA GPUs.

TBF, what difference does it make when the very harnesses you're using these "hacks" in are themselves vibe-coded?

They have a history of making grandiose claims like this[1] from 2024, with no visible products or research.

[1]: https://magic.dev/blog/100m-token-context-windows (also linked to in their blogpost)

> - Levent works at Anthropic, but this research was independent of his work there, with a mix of GPT and Claude models.

One minor wrinkle in Alpöge's narrative, he refuses to deny that he hasn't used any non-public Anthropic models for his "independent" research.

https://x.com/giffmana/status/2097560503581069585

That's because it's copy pasted from a coding agent.

I'd recommend trying openzl for jsonl.

You might want to look at this CLI that Tencent released recently.

https://github.com/Tencent/teamai-cli

Quis custodiet ipsos custodes?

Who will watch the watchmen?

Especially when it's hosted in the UK.

The company is Swedish, but the website is hosted by datacamp in the UK. You can see it for yourself if you run whois on their domain.

https://datacamp.co.uk

What are the odds that it's a honeypot?

Yeah, both WebLLM and mlc-llm have been mostly abandoned since NVIDIA acquired OctoAI in late 2024.

They care very deeply about retaining their #1 position on Felony Bench. :)

https://www.felonybench.com/

If you want something that feels like Opus 4.8, try GLM-5.3-Flash.

They've been talking about Astra for weeks now. I wonder how much longer would they have delayed Astra, if it wasn't for Anthropic releasing Fable 5.1 today? This is why we need competition.

You want something heavier? They were trashing people's SSDs a few months ago. IDK if it's been resolved, yet.

https://x.com/bdsqlsz/status/2067964486615810369

> On the Fable 5.2 eval summary, Opus 5 only beats Fable on SWE-bench multilingual and multimodal.

Did you mean Fable 5.1, or do you have access to the next (unreleased) version of Fable?

I have an adblocker. You don't see the cookie banner because you're likely outside the EU and your data is silently being shared with the aforementioned adtech vendors. The EU requires "informed consent" for this within their jurisdiction, hence the cookie banner.

The cookie banner and the list of 209 adtech vendors they share user data with is so repulsive.

A very incoherent article. Reads like a Submarine[1].

[1]: https://www.paulgraham.com/submarine.html

> SSDs as Memory

They've implemented mmap() in hardware. What a time to be alive?!

nb: Related discussion on Samsung's PIM from yesterday: https://news.ycombinator.com/item?id=49487341

Because OP is running it on an M3 Ultra with Ollama. He'll get much better perf with llama.cpp or MLX, both of which Ollama wraps, albeit very poorly.

It's a software issue.

I suspect this might be due to MTP mispedictions. Either because the 3.8 quantized model weights do not include MTP heads, or as what happened with Ornith-1.5 recently, corrupted MTP heads, or due to a software issue in Ollama.

This would've likely saved Steve Jobs, had it arrived 15 years earlier.

Captive audience.

That benchmark is super sus. Until someone pointed it out, the top performing open weights model was a Kimi K3 fine tune from their sponsor (abacusai/Smaug-Agentic). Now, it's not on the list.

Source: https://twitterwebviewer.com/?tweet=2091116504787935350

It's widespread here in Ireland. POS terminals have 5%, 10%, 15% and 20% tip options with a tiny and hard to reach dismiss button.

FWIW, macOS has good sandboxing, but LMStudio, Ollama, Darkbloom etc aren't sandboxed. This is also the reason why none of these things aren't distributed via the Mac App Store, because the Mac App Store mandates sandboxing.

RTN quantization of weights