This is time to first token, which is related to pre filling the cache and weights, not so much thinking.

It depends on the model. For newer stuff, and Qwen3.8 27b specifically, its Qwen Sparse Attention/Gated DeltaNet architecture does much better than a lot of previous models.

The old rule of thumb was that the middle of the context window can get lost in old models, but Qwen3.8 27b and Qwen 3.8 Flash Next work fine, even when using most of the 262k native tokens of context… but this long context requires 8GiB of VRAM just for the context, and this is with FP8 KV quantization.

I still don’t trust much more than 128k though, and most harnesses automatically compact if a request would exceed a certain threshold (e.g., OpenCode is the higher of 15% of max or 16k tokens for its buffer).

For smaller models 16-32k is probably fine unless you’re using a model that is extremely old, like llama3 or phi 2.5. But you’re going to struggle to get anything other than a simple chatbot with very short conversation on a 4GB card. Maybe light translation or document work, but don’t expect anything useful in coding or agentic work.

Aren’t these inherently new or forthcoming releases? One can’t expect community tags to be in place for newly released content.

Digital signature as a means of non repudiation is exactly the way this should be done. Any official docs or releases should be signed and easily verifiable by any public official.

When using an external TPM. Which next to no one does.

Shouldn’t be this hard to find out the attack vector.

Buried deep, deep in their writeup:

RocketMQ servers

• CVE-2021-4043 (Polkit)
• CVE-2023-33246

I’m sure if you’re running other insecure, public facing web servers with bad configs, the actor could exploit that too, but they didn’t provide any evidence of this happening in the wild (no threat group TTPs for initial access), so pure FUD to try to sell their security product.

Unfortunately, Ars mostly just restated verbatim what was provided by the security vendor Aqua Nautilus.

Frankly, surprised it was only that much of a drop.

Earlier this week, Reddit disclosed in a corporate filing that CEO Steve Hoffman sold 500,000 shares, and Reddit COO Jennifer Wong also disclosed that she sold 514,000 shares.

No, not a pump and dump at all! Totally confident in our ability to execute our fiduciary duties!

So the article repeats, several times, “waymo relies on remote operators”. I don’t think the author knows what “self-driving” means.

They’ve been around since 2018, and with only two weeks notice, now are saying they need $150k?! Who the fuck is their CFO, and how did their burn rate get so out of control? Shouldn’t they have started the campaign months ago if this was an expected problem?

This just fucking reeks of embezzlement; grossly mismanaging the funds charitable individuals had entrusted them with.

Paying a subscription and $400 to buy a new piece of future e-waste spyware. Who is dumb enough to buy this crap?

The so-called architect put those hairs there, you moron.

A motorcycle. You can’t outrun the radio.

On Sunday Trump announced an “energy ceasefire” between Kyiv and Moscow [..] Russia then launched a “targeted attack” on a Kharkiv power station just hours later [..] “It’s clear that Putin does not read Truth Social,” Zelenskyy quipped.

Brutal.

The crux of that article is buying into Heidegger‘a definition of a tool, specifically, “thing that lets you do something you wouldn't otherwise have been able to.” If you disagree with that premise, the article doesn’t hold up.

Tools aren’t about letting you do something you would otherwise not have been able to do, they are simply a force multiplier for capability or efficiency. You still need knowledge and skills to do the thing.

I am not a carpenter. My knowledge and skills are near zero. This multiplied by any number is still damn near zero, and I have no idea about the culture, safety, or history of carpentry. I’m not stupid enough to jump right into using an lathe without a lot of research or someone to guide me, because I’m not a fucking moron.

People with programming knowledge and skills using AI will eventually figure out a workflow that works. Laypeople using AI to vibe code? They’re probably fucking morons. Sure, maybe even most of the time, it will be fine. But eventually, someone is going to be degloved.

Explain your thought process here, how did you arrive at the larger bottle being 90% more detergent? It’s EXPLICITLY clear that the concentration is higher in the smaller bottle.

You could complain about the form factor or lack of precision in dosing loads using the higher concentration, but “detergent” is mostly water, which they clearly said they reduced by 75% (same solute, with less water/solvent = higher concentration).

Quick search and going by what it says on the label, the cost per load has not significantly changed, a little more than half a penny’s difference:

Ultra Concentrated (left) $15/60 loads = $0.25/load https://mrsmeyers.com/collections/laundry/products/ultra-concentrated-laundry-detergent-rain-water?variant=50673207640338

Standard (right) $18/74 loads = $0.2432/load https://mrsmeyers.com/collections/laundry/products/ultra-concentrated-laundry-detergent-rain-water?variant=50673207640338

Clickbait garbage

In short, I want Linux to work for me and other technically-minded enthusiasts […] Is this an elitist view? I don’t think so.

Is it elitist to gatekeep Linux for technically minded enthusiasts? Yes.

Especially when they end the article with “I for one hope it never does”. Definition of gatekeeping and toxic elitism.

That’s the benefit of a custom domain, I suppose; you can always change he provider without changing your email.

Not curious, Canonical is widely seen as antithetical to open source ethos. But it is stable and has put in a lot of work for vendor support, which is why so many distros (including Mint) are downstream derivatives from Ubuntu.