dofm

This is so they can get an hour’s jump on Rubio’s invasion force.

> Alice User goes to Bob's Website on the wifi at Eve's Coffeeshop. Can Eve read or inject sensitive information into Alice's browsing session?

Or more likely can Zac, the random customer who is actually really responsible for running the wifi, having volunteered, since Eve, like all indie coffee shop proprietors, doesn’t know shit about wifi.

> Similarly to “one language is best,” saying “one harness is best” is shortsighted in the best possible case

You are arguing against something that was not said?

I ended up building the patched PHP 7.4 source they made available for one of my machines; I wish I did not have to do that but there we are.

Your regular printer, like mine, presumably puts down a line of melted plastic material with an extruder, bit by bit, so the print time is guided by complexity as well as size.

Resin printers, basically, are stereolithographic. They:

- dip the build plate (upside down) into a thin tray of resin

- push it all the way onto onto a screen at the bottom of the tray

- blast that screen with UV through a translucent LCD display that contains a mask image of the current "slice" of the whole plate, enough that the thin layer of resin partly cures

- lift the plate out of the resin so it peels off the screen and the resin replenishes below it

- repeat, with the next slice.

- then the whole model is taken out by hand, washed off, blasted with UV light to completely cure it, and the (more or less inevitable) supports snipped off.

So instead of printing line by line, the entire plate prints layer by layer — a small print takes (approximately) the same time as a full plate.

There are several big differences, notably that it is single-colour, much higher resolution, the various peel/adhesion forces are managed differently, larger parts need to be hollow with a vent so uncured resin drains out, most resin prints are quite brittle, and the whole process is traditionally really rather cytotoxic.

There are eco resins, tough resins, and printers that do a better job of not being chemically evil; this is one.

I've stayed away from it, because it's not the kind of printing I need (for all except one part that I might end up subcontracting) and because it is chemically nastier than b/w darkroom printing, which until now has been about my limit of risk.

The above is a bit of an oversimplification but it covers the main differences I think (happy to have corrections/additional comments)

https://formlabs.com/uk/blog/ultimate-guide-to-stereolithogr...

This is absolutely true.

He is an ignorant, foolish, undereducated person who does not care to learn detail of things and believes himself an expert in things on the shallowest read possible — he doesn't like to be called stupid so he will believe anyone who seems authoritative, hence Peter Navarro/Ron Vara.

But he is a truly world-class bully.

He knows how it works, he was schooled in how it works (by Roy Cohn, who represented all five Cosa Nostra crime families at one time or another) because he had to deal with the legit mafia (concrete etc.) in New York in the 1970s. He perceives himself to be like a mafia don.

Josh Marshall of talkingpointsmemo once coined the term "dignity wraith" for what happens to people in Trump's orbit; the process of breaking them etc.

Only one person can make Trump himself lose dignity: Putin.

If poor nutrition is the cause (not obviously the case) then it's fair to say our diet in the UK is not so far behind the USA in terms of the availability of unhealthy foods. We maybe eat a bit less of it, but we have rising obesity levels. We don't have anywhere near the same number of morbidly obese or super-obese people, but towards the merely very overweight end of the scale we've probably caught up.

Probably catching more cases in screening, though.

There is an obvious factor for cases in 2023, which is that a lot of backlogged cases from 2021 and 2022 will be turning up, and the mere presence of a backlog means delayed diagnosis of newer cases. Even if COVID itself doesn't cause cancers or make them worse (and that's being studied), the pandemic did.

> I wonder why AI almost never gives me links to Wikipedia

Users on Reddit say…

Co-ordinated disclosure may not be dead just yet, but it has started ringing around to see if anyone knows a memorial stonemason.

If the Zero 3W has USB-C gadget mode support, a gigabyte of RAM and the Pi 4 chip, it’s going to sell very well indeed.

“Can’t bill a minute of it” is the crux of the problem.

There are all sorts of ways you can improve the management of customer WP sites. You can put each site in its own small VM, with caching, with FrankenPHP, set that up with cloud-init, manage the entire fleet remotely with WP-CLI, etc.

You can have your own themes, lock down your plugin choices, only use your own code in functionality plugins, all that.

But if your customers can administer the sites, something is going to burn you nonetheless, and when it happens nobody is paying.

Part of it is just that WordPress exists in the middle of the continuum between content and functionality, part of it is the odd mix of users: you can have technically unaware users who you inadvertently empower to make bad decisions. It’s not their fault: it’s an over-permissive architecture.

WordPress, in your own careful hands, can run brilliantly for years. In all the time I have run it (18 years now) my own sites have been fast, busy, stable, reliable: I have even run WP brochure sites for finance industry clients, and I have had full confidence in them.

But if you’re not going to be in charge of it, don’t run it. Because really no amount of monitoring and security mechanisms can stop this eventually happening.

Ultimately I have also decided I am not going to roll out any new WP sites that I have to maintain. But then (despite my admiration for Laravel) I am going to get out of using scripting languages for server side code at all. AI is going to make all internet-facing interpreters dangerous. It’s too large a surface and I don’t have the energy for php2shell Groundhog Day.

Yeah. But he has openly mused about renaming a river, a state and an ocean. Seriously apparently he wants to rename New Mexico.

Granite is not great at any size. Wonder why they picked that.

No worries. And yes, absolutely. Even the tiniest opportunity to poke around with stuff changes my perspective dramatically.

I've picked up lots of stuff this way — just being able to watch the server logs is great!

P.S. I prodded myself into playing with Osaurus; it's actually rather good.

This is because someone's told him he probably can't rename New Mexico?

They really have radically improved the situation for prefill speed! But the generation speed is behind enough that you'd notice. An M4 Max is a great choice I think.

I am not — I'm using a 64GB M1 Max. Mostly with Muse Glimmer, sometimes Qwen 3.8 27B, Gemma 4 12B QAT, currently playing with LFM 2.5 24B-2B a bit.

I'm just saying, the base M6 is not very competitive in terms of memory bandwidth (the M5 Pro has about 80% faster memory bandwidth).

Compared with my old M1 Max, the M5 Pro is radically better at prefill, only about 15% slower at most at generation (and probably performs better at DFlash and MLX than the M1, so it might close the gap or even outperform the M1 Max)

So the models I am using/testing should all be quite fun, assuming the 48GB RAM machine. The Gemma 4 12B and LFM should run well on the 24GB machine. Neither of those are great coding models but they should be pretty good assistants. There are harsher quantisations of those other models that may fit.

Whether playing with these things locally is worth the extra spend, I don't know. It's probably more cost-effective to buy a cheaper Mac, experiment with the smaller models, and use Openrouter for the bigger ones.

Thanks :-)

Mostly I think my own approach has been to try to combat my anxiety and exasperation about all of this by taking this measure of control.

While no really small local model comes super-close to a frontier model, the shape of the technology is settling and I think one can learn so much by approaching it the conventional nerdy way: what's a practical setup I can run on kit I already have? What can I build that helps me explore the problems?

I am quite cynical about AI still, including for generative coding, but I am not cynical about what I can run locally. There is no doubt that they are already helpful to me even in a minimal form.

And the really cool thing about being able to run them locally is to understand the tradeoffs.

For LLMs (apart from Apple's foundation models) the M1 Max still holds its own for the most part.

The base M6 is faster (much much faster) than the M1 Max at prefill (compute bound).

The M1 Max outperforms the base M6 at token generation; this is memory-bandwidth-bound, and while the M1 Max can't actually saturate its memory bandwidth, it gets a lot closer with the GPU.

So I suspect the M6 is better if your AI tasks are decision-based or video/image based, and the M1 Max is better if you're running a coding LLM. Though prefill is a bit of an issue.

If you want to run LLMs: get the M5 Pro, not the base M6.

I mean, it's not the size of the download, it's the unauthorised scraping of community assets strictly against terms and ignoring robots.txt by an American corporation. Resource of a kind that used to get kids threatened with spending most of their life in prison.

But I agree that this kind of theft by US corporations is no longer news.

(Yes, it's theft: they were aware they may not do it and according to the article attempted to get around further restrictions. I mean it's literally even theft of the $3 you mention — you don't have to be in possession the $3 to illegally deprive its owner of it under common theft statutes)

Shouldn’t be embarrassing to admit.

It’s a huge new subject area and it’s completely exhausting to learn. I still feel like a total novice.

I have an M1 Max (more memory bandwidth) so I am not sure how much advice I can give about models, but I would suggest maybe trying the new LFM2.5-24B-A2B. It’s surprisingly good. If you want to mess about with multimodal work, Gemma 4 12B is really impressive. (I mostly use Muse Glimmer but it might be rather slow on an M1 Pro)

If you want to have the space to learn the terminology, tinker with what you can control etc., then you have three easy options: LM Studio, Unsloth Desktop or the llama.cpp native Mac app. But don’t get put off by the llama.cpp command line; most of these tools are running it behind the scenes, and most entries on Hugging Face have recommended parameters.

MLX is an option but so far I haven’t found it to have an edge over simple GGUF on Metal on the M1 at least; it seems it might have more advantages on the later M series chips, but it is usually slower for me.

I have explored running OpenCode inside a VM (you can “remote into it” from the OpenCode GUi desktop app). You might want to look at Paseo, which can do a similar thing but with many more coding agents.

I am quite intrigued by Osaurus; on macOS 26 and up it looks like it will manage a container for you.

You might also want to look at Orchard. Not done this myself yet but I can see the appeal.

https://orchard.andon.dev/ai.html

I am an experienced web dev in my fifties. Six months ago everything I have written above would have been total gibberish to me. Learning some of the open weights stuff has made me feel much less helpless than I think I might if I just developed a Claude dependency. Nothing matters more to me than grasping the fundamentals.

I still don’t really use LLMs for hands-off coding. But I have found that a local LLM can be a useful sidekick; even at the most basic level it’s a better Google that remembers the context of queries.

“Garth is the most significant artist that I've worked with and I've worked with Lulu and four other people, so we’re talking crème de la crème.” — Dean Learner

So why are we doing this, if it is not better?

:-) Nice.

The oldest web app code of my own that I still maintain/evolve turned 20 this year, so yours is older than mine. I think it's largely a Ship of Theseus at this point, though I did just remove some code the other day that was absolutely untouched since 2008.

This is true, but up until about two and a half years ago, we built IT systems to be more reliable, consistent and auditable than humans, and to provide some security guardrails on human malfeasance.

Now we're OK with building systems that are an order of magnitude faster but behave no better, because what?

> Altman already has a couple I think.

If they aren't patrolled by giant, iris-scanning eyeballs he's missed out on a vertical.

> At the time of this post, that’s a 33 years and 8 months old application,

Pointing this out without a trigger warning is just rude ;-)

Some of us are older than we think, yo.

> By that logic you can't trust any model, neither trained by Cohere nor trained by yourself from scratch, because it's pretrained on public data that might easily be poisoned on purpose

Yes. Even if it's poisoned training data, RLHF fuckups, or something more malignant: the model, wherever it comes from, is in your supply chain and it makes you more vulnerable.

As much as I doubt that Alibaba have any great interest in global espionage, and this is much more likely to just be a consequence of being trained on Alibaba's own products, the fact is there is a geopolitical dynamic here.

I dunno — they have made a claim and backed it up?

This is a risk with any model (even in situations like this where it's still more likely to be a training error than malignant).

It just tells you that you have to think as carefully about open weights models as you would third party libraries. It's just another supply chain risk, only a bit more probabilistic.

I have enjoyed using Qwen but I do think it's worth bearing in mind that this is at least possible and it might not always be a screwup.

And that's putting aside the risk of all of this happening for more generalised reasons like the training set or RAG results being poisoned.

In practice it's surely most likely that the issue is that this pattern is in many of their own apps and they are doing RL training on their own code.