dannyw

Unless you're clever enough to only poison user agents known to be macOS and CLI. You'd probably get by.

xz-tools was only caught because it had a bug causing slow startups and someone was micro-benchmarking Postgres and decided to dive in further.

I really don't understand your justification.

One of the certificates literally has a private key of 4; it's as compromised as it could quite possibly be. This is not a "one of the intermediates have a weak and possibly crackable key" or "a bunch of incorrect certs were issued but the issuance was patched".

To a MITM attacker, git clones can be easily distinguished from visiting in Safari. You could choose to only poison git clones; without poisoning anything else.

Heck, you can probably build an exploit kit in a hour; camp out around a Starbucks or McDonalds with open wifi, and start compromising git clones and deliver malicious payloads.

Reputable hosting providers, e.g. Fireworks, Cloudflare, Baseten, whatever absolutely _do_ have a reputation to lose.

Even and especially for frontend UI, GLM 5.3 is a whole world better than Astra.

I can see (and believe) Astra beating GLM 5.3 on verifiable hard domains like maths, etc; but we’re at this stage where intelligence scores and even ‘frontier real world unsaturated benchmarks’ really doesn’t reflect real world use anymore.

From my rough recollection, all 3 of the past updates all had the effects of decreasing the score of open weight models.

One was rushed and same-day updated as GLM 5.3 (iirc) got added; of course its main change to rankings was lower GLM scores.

Umm yes, the M6 chip only has a certain amount of memory controllers and channels, which means you can only have so many LPDDR5X modules.

Yes, high-capacity LPDDR5X are made and do exist, but for all practical intents and purposes, you can't buy enough of them to deliver to customers at Apple scale.

It's a shame that only the Mac Studios come with USB-A ports. I know, won't fit on a MBA, but similar size and class laptops to the MBP often come with at least one USB-A.

It's just handy when you have a USB-A device and don't want a dongle, traveling, etc.

While not Asahi/M1, I can tell you my desktop has been Linux Mint for about 5 years, including through distro upgrades, etc and I have had to troubleshoot it like once.

Claude models refuses to perform work on kernel optimisation or ML/AI training work on Chinese inference cards (e.g. Huawei Ascent cards)

World knowledge, factual recall without grounding, more complex instruction following, etc.

Benchmarks don’t really tell you much. For API best will always be a model+effort sweep on your internal evals (and if you don’t have internal evals, do that first).

For individual coding harnesses, honestly I just go with vibes. The best workflow is the one that works for you.

One of my beloved devices was the Nexus 7, it was an excellent affordable template that didn't feel cheap, nor fancy or pro. Just a good media consumption / general purpose small tablet.

I wish Apple would make a similar device. I'm fine with recycled chips and getting a bit gimped on RAM size.

Like the MacBook Neo of iPads; I just really want a media consumption device. A20 Pro means nothing to me.

I’m sure it’ll go as well as banning piracy or cryptocurrency.

For emails over 7 rolling hours, unless you're e.g. Google or Microsoft, you kinda have to, otherwise those systems will automatically shadowban (send-to-spam) you for excessive/abnormal email sending rates. It's not great.

It also doesn't help that a lot of scammers use this, or fake cyber incidents, for phishing; so it's not like they have an easy solution for exempting "security emails" either.

That’s not true in my opinion.

Maybe it’s 1% or 10% today, but a decade ago, having/assembling the technical chops to know how to build something was a huge part of the meat / barrier to entry.

NVIDIA won't be too happy about that; alongside Micron, etc.

US commercial interest =/= OpenAI and Anthropic.

You can even buy a small solar panel for ~$10 with positive and negative leads you solder; and power something like a Pi Pico^ directly. It kinda feels like magic.

^Not recommended literally due to brownouts/etc, but a cheap power management IC on a breadboard can solve that. Alternatively, add a battery/etc.

Home solar panels and batteries are just like this except on a bigger scale; and with (hopefully) more QC.

If you use CLI, all your transcripts/metadata/etc should be in ~/.claude/ and ~/.codex/ or ~/.pi/, etc.

Ask your agent to make you a visualisation, heat map, whatever. I bet it's just that you're using these models more, and for more complex tasks.

AlexNet (2012) can do the first part :)

Pretty much. See “Scaling Laws for Neural Language Models”, a must read IMO https://arxiv.org/html/2001.08361v1

Compute, data, and parameter count are the three logarithmic values that correspond to model performance.

So if you find ways of improving training efficiency (the simplest conceptually is probably local/global sliding window attention like Gemma; the latest DeepSeek and Kimis have quite fancy ones), you will be able to have more effective compute, and train smarter models while keeping data and param count the same.

If you find this fascinating and want to dive more, I encourage checking out nanochat (solid GPT-2 repro; minimal LoC); and trying to make architectural tweaks with a coding agent.

“Learning through experimentation” has never been easier or more fun in my opinion. If you have a GPU, even an older 8GB card (or a few bucks for Runpod), you can pretrain your own small language model and start experimenting / tweaking it.

You get 614GB/s of memory bandwidth on the M5 Max (full chip), more than double the N1X / GB10 (DGX Spark). For decode, that'll directly translate into ~2x faster token generation.

He also got his OpenAI account banned after sharing the findings for 'violating cyber policies'. https://x.com/i/trending/2113715638950912393

As far as I understand, they have more demand and commitments than their production and scaling capacity. Maybe they're doing something right.

“Language Models are Unsupervised Multitask Learners”, the paper behind GPT-2.

Maybe limited compute to handle launch traffic spikes?

Still get the model out there.

It’s highly likely that everyone has software or firmware made by Chinese engineer running in their house.

These defenders aren’t supposed to have access to the “defender models”.

The ruling doesn't ask Apple to get rid of app sandboxing. The ruling explicitly notes this:

> Moreover, because iOS apps are sandboxed, as implemented in other jurisdictions, apps available through alternative distribution will be sandboxed, which further strengthens security and reduces the likelihood that any app will compromise other apps [...]

It’s also great news for app developers.

First, Apple does not follow the same rules they set for developers.

Second, Apple does not treat all developers equally. We have functionality we’re unable to offer to iOS users.

The same feature, including technical implementation mechanism, is available in Claude and Gemini. It’s unfair.

I feel incredibly proud do to be Australian right now.

I like how the court’s declarations are punchy, factual, and dare I say obvious. I’d suggest giving section 1 a read.

There’s going to be a million “what ifs”, from walled gardens to security to ‘you don’t have to use iOS’, but it’s blatantly obvious this is all about competition and money.

It’s anticompetitive, illegal, and the remedy is to cease these anticompetitive actions.

A security vendor entrusted with advanced patches breaking embargo, is much worse than a security researcher breaking embargo.

It sounds like there was no "co-ordination" behind any of this co-ordinated disclosure. You'd have expected there to be an IRC channel, a mailing list, a Slack chat, or something for something of this scale.

Was there even an incident co-ordinator here? Was any of the parties talking to each other?