MallocVoidstar

IIRC VLC used the wrong primaries for converting to RGB for a long time (years) even after it being reported to them as wrong

'Predicts the most likely tokens' is not the same as 'pulls people towards normality'.

I think a therapist telling teenagers to kill themselves is always bad and should lead to revocation of license and prosecution, even if they helped other people.

If a therapist tells 10 people not to kill themselves, and convinces 5 patients to kill themselves, would you say "this therapist is doing good on the whole"?

Why should "some people have been helped by AI" outweigh "ChatGPT has convinced multiple teenagers to commit suicide"?

I'm going to assume this is a predominantly AI-written article since for some reason it's talking about GroqCloud serving Llama 2, which they don't.

It claims they serve Llama 2 7B @ 750 tokens/s with 2K context, but over on OpenRouter Groq is listed as serving Llama 3.1 8B @ 1300 tokens/s with 128K context. (And the official GroqCloud site says 840 tokens/s.)

Where are you seeing 79% accuracy? 79% only occurs on the page as a win rate, not an accuracy

They can't train on the API.

I used a VPN that had a virtual location of China for a while, which avoided ads on some websites; China blocks those sites, so those sites don't have any ads in China, but the VPN exit wasn't actually in China so it could reach the sites fine.

Related: https://www.reuters.com/business/disney-sends-cease-and-desi...

> Walt Disney has sent a cease-and-desist letter to Alphabet's Google, CNBC reported on Thursday.

Pirates are generally slow to transition formats, but AV1 is also not better than H.265 (in practice) at the high-bitrate encodes.

Scene rules say to start with --crf 17 at 1080p, which is a pretty low CRF (i.e. it results in high bitrates): https://scenerules.org/html/2020_X265.html

AV1 would most likely result in slower encodes that look worse.

AV1 is still worse in practice than H.265 for high-fidelity (high bitrate) encoding. It's being improved, but even at high bitrates it has a tendency to blur.

According to the RFC:

>The messages in classic UDP-based DNS [RFC1035] are inherently unordered and have low overhead. A competitive HTTP transport needs to support reordering, parallelism, priority, and header compression to achieve similar performance. Those features were introduced to HTTP in HTTP/2 [RFC7540]. Earlier versions of HTTP are capable of conveying the semantic requirements of DoH but may result in very poor performance.

I'd bet basically all their clients are using HTTP/2 and they don't see the point in maintaining a worse version just for compatibility with clients that barely exist.

you don't "appreciate" anything, you're just posting LLM comments

I started using Firefox with version 1.5, as did many of my friends, and we were doing it because it was flat out better. We did not care about 'stagnating' or standards.

You told an LLM to generate three possible responses to HN articles and then just started pasting all three?

> I'm not too worried about someone DDOSing my personal site. Yeah, they could do it. And then what? Who cares?

Your host, assuming you're hosting your site on a VPS. Many of them have a policy of terminating clients who get DDoSed.

> As they say in security, "no one will burn a zero day on you!". For your small blog with one hundred visitors per month, it's probably the same: "no one will burn their DDoS capabilities on you!"

The last I saw you can hire DDoS as a service for like $5 for a short DDoS, and many hosts will terminate clients who get DDoSed.

I don't think there are any up-to-date leaderboards, but models absolutely degrade in performance the more context they're dealing with.

https://wandb.ai/byyoung3/ruler_eval/reports/How-to-evaluate...

>Gpt-5-mini records 0.87 overall judge accuracy at 4k [context] and falls to 0.59 at 128k.

And Llama 4 Scout claimed a 10 million token context window but in practice its performance on query tasks drops below 20% accuracy by 32k tokens.

At the time I'm quite certain I was using uBlock Origin pre-MV3. I don't think I also had my DNS-based adblocker yet, though.

Much cheaper and no blackouts. HeheStreams was $100/year for NBA/NHL/NFL/MLB, the NBA's equivalent was $200/year in 2021.

A lot of the pirate stream sites I've run into break entirely if you have an adblocker enabled. I'd guess it's a combination of filter lists not being tested on them along with much more aggressive ads (from sketchier ad networks).

The vast majority of pirate stream sites are monetized in some way. If I was going to use one I'd probably prefer to pay some small amount rather than deal with the hellish ads the 'free' ones use.

Previously: https://torrentfreak.com/hehestreams-iptv-admin-sentenced-to...

They ban VPNs unless you're logged in (and randomly shadowban users using them), and probably other datacenter IPs.

> What we’re going to see is as energy becomes a problem

This is much more likely to be an issue in the US than in China. https://fortune.com/2025/08/14/data-centers-china-grid-us-in...

> There are plenty of external citations, so rather than relying on a model to recall information internally, it’s likely effectively just summarizing external references.

And according to Tim Bray, it's doing that badly.

> All the references are just URLs and at least some of them entirely fail to support the text.

The existence of trans people isn't sexist.

The highest post in the chain we're in:

>PSA: author is Alejandra Caraballo [...]

The reply:

>Yep, this whole article is basically just -->his<-- very biased opinions [...]

My reply:

>Alejandra Caraballo is a woman.

I'm correcting them.

The Grokipedia article on Gamergate claims that Gjoni "revealed" that Quinn was sleeping with reviewers for good reviews, but he himself later admitted he had zero evidence for that and claimed it was a typo that made people believe he was accusing Quinn of it. Why is being outright wrong a good thing?