prima-facie

There's a glimmer of hope with ROCmFP4 which seems to double the current throughput: https://github.com/charlie12345/rocmfp4-llama

The Strix Halo is a great dev machine and a mediocre AI machine. You can run Qwen 3.6 27B at a decent speed, or larger MoE models, and that's about it. For some that's more than enough though, myself included.

Sure but where do you stop spending? :)

https://github.com/jamesob/local-llm

MediaTek Quad-Core chip, 10G Wan/Lan, SFP+, 2G RAM.

> https://openwrt.org/voting/2025-02-12-openwrt-two

Glinet are doing a great job with their routers. I have the Beryl AX which is fully openwrt compatible. The new Beryl 7 is also fully compatible now. Mediatek chips might not be as high performance as Qualcomm but they make up in openness.

Edit: They just announced Flint 4 with a Mediatek chip:

https://www.gl-inet.com/en-gb/products/gl-be14000

At the moment, for around $4000-5000 you can either have speed (a GPU + 32GB VRAM), or you can have capacity - a DGX Spark/Halo, but not both.

I think once someone comes up with a machine which has both it will easily sell for $10000 and people will be queueing to buy it.

And this is exactly my point, the OEMs have more lobbying power and leverage. Anthropic might be valuated at whatever amount, but they're a new player and their only product is a piece of software - which others like Google, OpenAI, etc also have (not identical but similar enough).

EDIT: FYI https://ibb.co/nMYP34Rr

Well, we could expect anything from Adobe. An LLM subscription on top of the regular Adobe subscription sounds like the sort of thing they would do.

Do your local filters run slow? Does your movie render have no sass? Then sign-up for AaaS!

>> ISO date and time

Yes, but not always in my experience:

    # Default locale is en_GB.utf8
    
    > date  
    Fri  3 Jul 12:14:20 BST 2026
    
    > date +%x  
    03/07/26
    
    > LC_TIME="C.UTF-8" date   
    Fri Jul  3 12:14:29 BST 2026
    
    > LC_TIME="C.UTF-8" date +%x  
    07/03/26

Laws restricting the use of local AI/LLMs are not going to happen, no matter how much Anthropic might want it. All the major OEMs are now counting on local LLMs to take off. Just look at the OEM support for the upcoming Nvidia RTX Spark platform: Asus, Dell, HP, Lenovo, Microsoft, MSI. All the big names in the industry will have, by the end of this year, Nvidia-powered machines made specifically for local LLM use.

If you are from Europe, even if you're not living in the UK, the en-GB locale will feel a lot more familiar to you than the en-US one.

It uses the dd-mm-yyyy date format like the rest of Europe, the start of the week is on Monday (vs Sunday in the US), the default paper size is A4 (vs US letter), measurement defaults are metric (indeed UK roads use imperial, but the default is otherwise metric), the time format uses 24hrs (vs AM/PM in the US).

There's also `systemctl soft-reboot` which initiates a userspace-only reboot, which quickly restarts the system without going through the full hardware and kernel initialization process.

Thanks for the detailed response, I really appreciate it.

What I had in mind was an AMD Strix Halo machine, but it seems to have none of the advantages you mentioned. It's neither high bandwidth, nor does it have CUDA support, nor does it have support from the big OEMs. All the boards are from relatively obscure Chinese vendors.

It seems like all the major OEMs have rallied behind Nvidia, if you look at the upcoming RTX Spark laptops.

> The biggest thing to watch out for is not just RAM/VRAM but memory bandwidth. You can try to "future proof" yourself with lots of RAM, but if it's 400 GB/S you're still constrained to smaller models.

I'm thinking of getting a SoC machine with 128GB RAM but the bandwidth is limited to 256 GBps. Would you even consider such a machine a decent investment, or should I wait for the newer gen of chips? Thanks!

This whole project assumes that --resume replays full transcript, is that actually true? Is there no caching going on?

Imagine if we had managed to deliver on the original promises of the Semantic Web, instead of having these locked-in platforms. How incredibly useful all that linked and structured data would've been to humans and LLMs at the same time.

https://www.w3.org/2001/sw/

Commercial VPNs publish their exit nodes IPs online. There are services like ipinfo.io which can accurately determine if you are using a known VPN service.

There are cases where the workers' right to stay in the country depends on their employment, in which case this creates a huge power imbalance.

> We pulled an ultrahackathon. The team who built this did not sleep much for two days. They worked through Wednesday night, through Thursday, through Thursday night, through Friday. They ate at their desks. They wrote the spec late Wednesday evening and they wrote the cutover commit on Friday afternoon, and in between they did the work that the time between those two moments required.

Cool story but I would not want to be in their shoes. Treating your employees poorly only to justify overnight changes in business needs creates a highly toxic work environment.

As someone who has represented themselves in tribunal before I'm definitely interested in this.

The only issue is that in some jurisdictions, like the UK, you can't just offer someone legal advice without being SRA accredited or FCA regulated. I.e. this would effectively make Anthropic a claims management firm under the UK law.

> Under article 89I of Financial Services and Markets Act 2000 (Regulated Activities) Order 2001 ("The Order"), advising a claimant or potential claimant, investigating a claim and representing a claimant, in relation to a financial services or financial product claim is a defined regulated activity.

https://www.fca.org.uk/freedom-information/dual-regulation-c...

> And is there a legal consequence for AI giving bad/incorrect legal advice? Can they get disbarred?

Of course not :) but unfortunately I've received half-truths and outright lies from actual solicitors, while AI has mostly given reliable advice, even if not procedurally perfect.

There's lots of guarding and unwritten rules in the legal profession so this is not something that will be straightforward to train on, but once done, even if imperfect, will bring legal access to the masses.

With the over-reliance on AI, this looks like a veritable slop-machine, designed to create and consume slop as a primary activity. Good job Google.

Somewhat off-topic: I've spent thousands of pounds on legal advice which has ranged from poor to mediocre. I found that most solicitors would refrain from giving proper advice and are there only to be instructed. You have to do your own homework, read the law, the case-law, prepare notes and documents, etc. With the rise of LLMs I found it easier to do all these things and come prepared to these meetings, or even do some of the solicitor's work on your own. For example I've found Gemini 3 to be exceptional at reasoning on the legal side - to the extent where I was able to explore and reason about very thorny topics from all sides.

I found the legal profession to be a prime candidate to disruption using LLMs, especially the initial consultation phase (Do I have a claim?). One of the things that's protecting the status-quo, for the time being, is the law - for example in the UK you can't actually offer any sort of legal services without being SRA-accredited. There's also lots of secrecy within the profession, and lots of procedural tricks that lay people are not aware of. AI could make all of these more accessible for the lay person.

This was exactly my point as well. Everything that can be automated will eventually be automated.

What Google has done is incredibly clunky and only serves its own interests. We already have methods to prove that we're human.

1. lots of laptops have fingerprint readers & TPM2 build-in

2. lots of folks own Yubikeys or FIDO2 keys - if these became the norm then the price would come down significantly.

Both of these methods only require a tap to authenticate to a website. Both provide public-key authentication, and both provide some level of proof of work / require human interaction, without revealing the identity of the end-user.

Why not use or standardise these? because there's no benefit to Google of course.

You should keep regular backups of your BW vault as a plain JSON file. KeepassXC can now import BW vaults natively (passkeys included). If anything were to happen to Bitwarden you can migrate to KeepassXC as a stop-gap measure.

Oh we have a neat solution for that. Just give us your government issued id! I'm joking of course.

Thank you for maintaining Xfce! It's the best de around.

Smaller, decentralised forges actually make lots of sense from a digital sovereignty point of view. Over reliance on a single instance like GitHub is not healthy in the long run. The issue they would have to solve is federation.

There's also the Atkinson dithering (made by Bill Atkinson) which produces very neat results.

Hi, I'm trying to find Lennart Poettering's presentations, but they don't seem to be available yet? https://fosdem.org/2026/schedule/speaker/lennart_poettering/ Thanks