WhitneyLand

The features here don’t seem game changing. The most compelling parts are mostly already available in Claude or Codex or their related apps and services.

The biggest concern is that if you want to use SOTA models I don’t see how they can match what you get with the subscription plans of Anthropic and Open AI, whether your spending $20 or $200 a month.

Even if they could match what you get in terms of token quantity, they are giving their tools away for free for the foreseeable future and Cursor is not.

Saying his name like “girdle”, is the closest English pronunciation I’ve seen.

The actual German ö is hard for me to figure out without having a native speaker around to practice with.

StepFun is an interesting model.

If you haven’t heard of it yet there’s some good discussion here: https://news.ycombinator.com/item?id=47069179

A bit misleading to say they take 14x less memory, no one is doing inference with 16-bit models.

Call me a killjoy I hate April fools jokes.

Great quote from Hilbert, I think it’s also a useful thought for software development.

“The edifice of science is not raised like a dwelling, in which the foundations are first firmly laid and only then one proceeds to construct and to enlarge the rooms,” the great mathematician David Hilbert wrote in 1905 (opens a new tab). Rather, scientists should first find “comfortable spaces to wander around and only subsequently, when signs appear here and there that the loose foundations are not able to sustain the expansion of the rooms, [should they] support and fortify them.”

“The output is no longer 1-bit”

I thought that constraint was the whole idea?

Agentic AI CPU? No.

It’s a CPU designed for an AI cluster. Their last CPU Grace was the same thing and no one called it agentic.

Vera now just has more performance/more bandwidth. It’s cool, I’d like to have one of these clusters, but this is not new.

It’s marketed as agentic AI because that’s fashionable in 2026.

The sad thing is, this is already a paywalled site.

I’m afraid someone who wants to support professional journalism and agrees to pay ~$300/yr for an NYT subscription still gets most (all?) of this nonsense?

If they had a big result like, native 1.58-bit quality clearly matches top peers, they would be saying that prominently in the repo.

The engineering/optimization work is nice, but this is not what people have been waiting for, as much as, can’t the Bitnet idea that seemed promise really deliver in a competitive way.

“Cursor has to pay Anthropic's retail API prices (or close to it) for access to Opus 4.6. So to provide a Claude Code-equivalent experience using Opus 4.6, it would cost Cursor ~$5,000 per power user per month. But it would cost Anthropic perhaps $500 max.”

Cursor seems to be in a tough spot. Just heard the swix podcast on their big new cloud agents thing, and it’s looking like a pretty small moat these days.

Yes. Is there a basic chat app for iOS that prioritizes full intelligence over full duplex?

Agree ChatGpt advanced voice mode is so bad for quality of the actual responses. Old model, no reasoning, little tool use.

I just want hands free conversations with SOTA models and don’t care if I have to wait a couple of seconds for a reply.

Opus 4.6 was getting this wrong only last week.

Nice work.

I think you should remove Claude as a contributor to your repo. It probably weaseled its way in on its own, I think it’s the developers job to talk about the tools they used not the tool company.

There’s a bit of a hidden cost here… the longevity of GPU hardware is going to be longer, it’s extended every time there’s an algorithmic improvement. Whereas any efficiency gains in software that are not compatible with this hardware will tend to accelerate their depreciation.

Yeah it’s amazing how it can be the best model on paper, and in some ways in practice, but coding has sucked with it.

Makes you wonder though how much of the difference is the model itself vs Claude Code being a superior agent.

800 bits is an incomprehensible number of possibilities…yet tiny in comparison to the number of .gifs that could be drawn.

“MRIs…suck at cancer”

Wrong? I understand MRIs are the standard for certain types of cancer like brain and spinal tumors.

With respect to whole body MRI they can be less effective because it’s not optimized, accuracy can be traded for area.

But as a general statement MRIs do not suck at cancer.

One question is should the writer be dismissed from staff. Or can they stay on at Ars if for example, it was explained as an unintentional mistake while using an LLM to restructure his own words and it accidentally inserted the quotes and slipped through. We’re all going through a learning process with this AI stuff right?

I think for some people this could be a redeemable mistake at their job. If someone turns in a status report with a hallucination, that’s not good clearly but the damage might be a one off / teaching moment.

But for journalists, I don’t think so. This is crossing a sacred boundary.

The quote is not suggesting a quantum computer can’t be simulated classically, it can in fact, just slowly, by keeping track of the quantum state where n qubits is 2^n complex amplitudes.

It relates more to the Bell results, that there doesn’t exist a hidden variable system that’s equivalent to QM.

The 2023 paper even if true doesn’t preclude the 2026 paper from being true, it just sets constraints on how a faster attention solution would have to work.

No, people think humans use it a lot less often than AI, because it’s true. Especially for casual writing.

The contrast might become even greater because some humans that did use them have stopped to avoid false accusations.

Actually I thought it was a great example clarity, focus, and economy of words that AI is not capable of at this point in time.

What about the Advanced Voice feature, has this been updated to 5.x models yet?

First off, this is a cool project, look forward to some interesting insights.

I would suggest adding some clarification to note that longer measure like 30 pass rate is raw data only while the statistically significant labels apply only to change.

Maybe something like Includes all trials, significance labels apply only to confidence in change vs baseline.

For Kimi quantization is part of the training also. Specifically they say they use QAT, quantization aware training.

That doesn't mean training with all integer math, but certain tricks are used to specifically plan for the end weight size. I.e. fake quantization nodes are inserted to simulate int4.

Iirc the paper was solid, but it still hasn’t been adopted/proven out at large scale. Harder to adapt hardware and code kernels to something like this compared to int4.

Interesting.

If this were the case however, why would labs go through the trouble of distilling their smaller models rather than releasing quantized versions of the flagships?

Good point thanks for the clarification.

The broader point remains though which is, “you can run this model as home…” when actually the caveats are potentially substantial.

It would be so incredibly slow…

Its often pointed out in the first sentence of a comment how a model can be run at home, then (maybe) towards the end of the comment it’s mentioned how it’s quantized.

Back when 4k movies needed expensive hardware, no one was saying they could play 4k on a home system, then later mentioning they actually scaled down the resolution to make it possible.

The degree of quality loss is not often characterized. Which makes sense because it’s not easy to fully quantify quality loss with a few simple benchmarks.

By the time it’s quantized to 4 bits, 2 bits or whatever, does anyone really have an idea of how much they’ve gained vs just running a model that is sized more appropriately for their hardware, but not lobotomized?