dofm

I think there's some value for most people to understand that you can develop some kind of framework for evaluating art.

Or just understanding that art is very rarely exactly what it seems.

The problem with this LLM-driven thing is that the person who did it seems to have as shallow an understanding of art as the LLM, or has chosen a level of terseness that belittles even the LLM's grasp of famous paintings.

I am a bit cynical about all of this but I do find that occasionally an LLM has told me something about my own creative work (photography) that is unexpected. I don't know that I understand how one can point an LLM at paintings this famous and get such minimal insight from it, given how much must be in the training set.

Oh this is disappointing isn't it — it could have been a labour of love and isn't.

The "insights" about Mona Lisa are shallow and quite telling.

What did the crowd miss? What did the AI miss?

No props? One of the most significant things people don't notice — and of course neither did this AI — is she's wearing a veil. And a very particular kind of veil with a very particular connotation.

All that AI guff about what people miss when they stand in front of it, all that suggestion that you zoom in and look, and not even the slightest nudge to look at her hair and forehead.

Mainly I would like to see (if you don’t have it already) proper support for /previewing Affinity files (ideally the older .afphoto but I believe they are similar), and also Affinity layered TIFF, which I have used quite a bit.

I will test the app out soon. Best of luck with it!

These things are such a puzzle because there are so many that pop up and just don’t survive. I loved FileLoupe, for example, but it appears abandoned.

So I am sure to try this. I’d love to see closer integration with Affinity where that is possible.

Oh, not saying it’s much different in reality to keyword matching.

But it’s going to hit different.

Well it is pretty good at confirming the right ammunition for a school shooting so that’s one potential market at least.

https://youtu.be/IbG-JGdGniY

(Watch the whole video, it’s remarkable)

What else: the right medication for an overdose? The right fertiliser for an improvised explosive? The possibilities for adverts with that all-important long tail of earned media news coverage are endless.

ETA: in all seriousness… will an LLM decide on which ads to place? And will it be the primary chat LLM (modifying the system prompt) or will it be a summariser (like the title summariser models)

Indeed. If the law passes they can use it. If the law doesn't pass, they can run against it not having passed.

There is a non-trivial chance that Trump prefers the latter and he'd end up vetoing SAVE because it has some small amendment he claims to hate.

Worth remembering that there was strong bipartisan immigration reform on the table in 2024 that Biden would have signed, but Trump told its own GOP co-sponsors not to pass it because it was easier for him to run against "open borders".

Drummed-up fear from the absence of the SAVE act might be the one thing that gets Trump's disillusioned base out to vote for him. By contrast, getting the SAVE act passed might be a dog-catches-car thing.

Of course, sorry but — rephrasing slightly for clarity — has it been their bailiwick since they were created? Like, is illegal voting an immigration law violation, or is this just Trump projecting a new responsibility onto his independently funded army?

Has this always been ICE responsibility?

Yeah, it did. I don't think this is really in doubt.

Especially since the military refuse to confirm it had any human oversight, which after all would have been a routine thing to talk about before AI targeted things — which is why we have the phrases "fog of war", "human error", "unfortunate mistake", etc.

It will take a while to shake out — we won't know for sure for a decade, I suspect, but it seems very likely this will prove to be an AI error.

"You need to understand that Sam can never be trusted ... He is a sociopath. He would do anything."

The collapse of Landsbanki was absolutely not a nothingburger. Risked triggering a 21st Century Cod Wars: the UK government was forced to use organised crime laws to freeze its UK assets to protect UK savers in Icesave, to whom the Icelandic government did not clearly intend to extend protection.

I personally think it could pop before then if there is not concerted effort to stop inflating it and maybe start to let the air out.

But the main reason 2028 would be a risky year is the systemic risks around Trump. If he makes even more efforts to put his thumb on the electoral scales than he already has, or if it is not clear there will be a free and fair election (from an external perspective it really isn’t), then confidence drains out of the wider system very quickly, and one of the earliest things to go in such a situation is speculative investment.

Even if there is a free and non-controversial election, the market might well see risks for the AI companies: would the Democrats, if they make it into power, be so fully on board with letting these companies operate in a low-regulation environment?

And if the bubble should look like bursting for its own reasons in early 2028, will the by then quite stressed federal government, late in the term, actually have the real authority to do anything about it? It’s not a popular government, or a very organised one, and whatever they do will need confidence, co-ordination and an unprecedented level of buy-in

So would those guys in the boats in the Caribbean and the girls in that Iranian school, but the purpose of a system is what it does.

It can however select a girls school as a military target, and did.

FWIW on my M1 Max I have not really seen any advantage at all from MLX.

I am fully prepared to believe the benefits accrue more to the M3 and up (because of changes to the Apple Neural Engine).

But with the models I've tested, unless I am missing something, the performance of GGUFs in llama.cpp has been better in some cases.

I still have not had results from Gemma 4's MTP be really worth it, to be honest; but with the Qwen 3.6 MoE it is measurable. Maybe with newer kit it is more meaningful.

(There is every chance that the above is not the experience of anyone who really deeply knows what they are doing; it feels like I am a perpetual novice at this stuff)

Somehow the two main closed weights frontier models come from two companies with HQs about two miles apart, and the CEO of one used to work for the other.

I figure it might be quite a competent general coding teacher for more, er, consumer programming languages, for want of a better word — python, PHP, JS. Seems to be pretty solid on WP knowledge too.

And I find it curiously interesting when talking about photography. I've been finding it intriguing to ask it about my own photos and make suggestions about other images to research. I just showed it three of my own photos, and asked it to analyse them and recommend photographers I should research. It recommended someone amazing I have never heard of before. But it also recommended a 19th century British photographer who happens to be my lifelong photographic hero — someone whose broad characteristics inform what I do without me slavishly copying them. Bit of a jaw-dropping moment for it to have picked up their influence in subject matter that they would never have approached.

I'm still suggesting it more to people for them to see what the small-model future might look like, because it's so much more capable than one might expect.

On an M1 Max I have been using either Unsloth Studio (which is basically a web app) or LM Studio (nicer app on the Mac). You can use the Google AI Edge Gallery to play with the smaller Gemma models (but at the moment the QAT variants don't seem to be there unless I am missing something).

I think it's likely the 26B QAT model won't fit in your machine — you may be able to fit one of the UD_Q3 or UD_Q2 variants but whether you'll be able to run other things you want at the same time, I don't know.

(The QAT models are "quantization aware training" — AIUI the model weights have been assigned during training to survive four-bit quantization with less loss.)

So what I would recommend trying is this model:

https://huggingface.co/unsloth/gemma-4-12B-it-qat-GGUF

Try UD Q4_K_M maybe.

(I don't think the M1 really gets much benefit from MLX, in case you were wondering, though I could be wrong)

My interest in this model is largely to really get to grips with what small models can actually do, especially with tool calling, because I think it helps comprehend what the value proposition of the cloud models is.

I have been very surprised by the quality and clarity of its answers. It's also helped me understand that much of a typical harness system prompt is likely to be unnecessary now; Gemma 4 seems to be pretty sensible out of the box.

You're absolutely not going to be able to get it to go off and build whole apps from a long prompt; it is not that good, but it does tool calling and thinking, and you should be able to explore pointing a coding harness at it if you turn on LM Studio or Unsloth Studio's API server. You could also use the Llama system tray app (formerly LlamaBarn) or just use llama-server from the llama.cpp distribution.

Probably Pi is going to be a better harness because it can have a minimal system prompt, though I've not tested it with Pi myself.

It seems to know PHP and SQL to a fairly decent depth (and I suspect JS and Python). It also has a unified vision model (it doesn't need a separate mmproj sidecar thingy) that is fairly fast, and it is quite impressive at image analysis.

So you could probably use it to generate image descriptions and tags, summarise text, generate wordpress snippets, that sort of thing.

It can capably answer questions like "Can you characterise this image and suggest further similar images I might like?" — I am currently using this to provoke me to take photos again.

Have a play with the E4B edge model, too — again, much more interesting than I expected.

I don't know that I agree with this specific use of frontier because it is confusable as you say.

But (off on a tangent) I do think that there are multiple frontiers generally — and I also think the open weights, small local model frontier is by far the most important and exciting one.

I keep mucking about with what Gemma 4 12B can do and every time I do I find myself thinking that all the energies in the AI world are going in entirely the wrong direction, because it is small, clever, efficient and remarkable.

If all of that research money were to be spent on improving AI models that fit inside a 16GB RAM machine with unified memory, I think really important progress could be made.

I enjoy using the Qwen 3.6 models (and BottleCap's new fine tune of the 27B) but the small Gemma 4 models are impressive in a way that I think is going quite unreported.

So while I don't think this website should use the word "frontier" here, even referring to Qwen 3.6 27B which is weirdly close, I think it could.

Is it? The developer page for the github repo suggests he works at Arcee. Maybe he moved?

(I initially thought the same because of the website appearance)

I have been thinking about this for a while, because I personally need a career pivot but I was going to move into personalised training, and I think AI likely really messes with the viability of that.

I am on this trajectory too; opencode got me started but I tried to understand the way it is working and to do that I would normally read the source code, but the source code is really not fun.

pi.dev

Having not lived through this, I'm actually curious as to how similar it was to the current hype cycle. Was there a general sense that something singularity-like was approaching?

A singularity? No.

One of the early driving obsessions was a coalition dethroning Microsoft (which both did and didn't happen). A battle over closed-vs-open. That story was everywhere.

The main thing I notice is that there's a bit of a difference in the way value is created. Even in the UK the arrival of the web created a lot of early adopter winners — ISPs, design agencies, small dev shops like the one I worked for. And it generated a fair chunk of early money, changed the direction of state monopoly telecom etc.

Early well-bounded successes in 96-98 earned proper money, generated a bunch of press, generated new work. Because they had to: you didn't get access to huge pots of cash to grow with, until you had a product that was on the market and selling. The VCs didn't know you existed.

At the beginning it was small tech industry hustle and actual product creation, and a lot of companies really had the small software company playbook and ran that. You had a lot of PC resellers and systems builders who suddenly found themselves as qualified as anyone to build websites or run ISPs.

The problem after that was everyone's eyes got bigger than their bellies, the projects got too ambitious too soon, the VC money arrived even in the UK, and there were a lot of failures. I perceive the moment the firm I worked for got VC money to be the beginning of the end. I thought my bosses were altogether crazy to take the deal.

We basically were not ready to scale fast — the dot com era really scaled brand new companies in terms of number of employees and scope at a level that was at that time sort of unprecedented except maybe in the semiconductor industry.

Towards the pets.com IPO crash end of things, that is when it started to get unreal in a way that you'd recognise now — people talking about entirely new business models, and you had a running joke that if you just put "… on the internet" at the end of your elevator pitch people showered you with cash. The most insane of these being CueCat: a barcode scanner, but on the internet.

But did we think it would fuck everything up everywhere all at once? No. It was going to make us money and make it easier for everyone to buy and sell things, but it was going to create jobs (and it did for a while). It took until well into the early 2000s for the impact on retail to be truly felt.

Nowadays, companies scale enormously before they make any money, VC cash gets handed over to PhD students who only have the concepts of a plan, freemium tiers do the marketing and everyone is trying to follow the loss-leading market-share growth strategy that Amazon arguably pioneered.

The AI boom is perhaps the peak of that Amazon growth culture, "winner-takes-all" thinking, which was always insane.

I guess, though I am not sure that this truly is the worst case, if some new compliance thing required a redesign or worse.

You hear about firms that go over certain thresholds, get noticed and then find out that their production facility is no longer "small" or exempt for example.

Some amount of defensiveness in planning and design can mitigate but I'm interested to know if there are thresholds I didn't know about.

This is a bit of a year-of-linux-on-the-desktop question.

Desktop 3D printers are tools; I know many people using them to support their making. And they've utterly changed manufacturing on another level (simply because of their ability to surface a prototype).

It will get better and better and better and still never replace mass production, but mass production will be changed by it.

Not least because of the fundamental shift in availability of the tools to make tools. The productised desktop CNC market is wild and it owes almost its entire existence to the change in the market driven by 3D printing.

This hype cycle is much more like the NFT hype cycle — it's just that it has a more easily understood product. Impossible to see Beeple's Everydays as anything other than a harbinger, IMO.

The paradox of the cutbacks after the railway industry corrected is that more or less every track they got rid of would now be useful. The Beeching cut damage done to East Sussex, for example, is devastating and unfixable.

I want to frame my comment here carefully to talk about the narratives, specifically, and accept that we need to stipulate on face value that this is true (it certainly rings true from people I know in real life):

This specific narrative reads like every narrative explanation of the Republican party in the period from the second half of the first Trump administration until now.

Again, I am talking only about the narratives here. But the pervasive narrative around Trump is that privately everyone hates his guts, comes away with a worse impression when they work with him closely, dreads him opening his mouth, thinks his decisions are insane, and they are telling journalists that, at every opportunity, and have done since the 2018 midterms. And yet in public we have committee hearings where any number of irrational topics become a sort of terror zone here rational people have to assert irrational beliefs or pretzel logic to be able to continue; where ordinary absurdities have become cult-like witness-bearing that everyone polices everyone else on. It's like the Hunger Games of litmus tests.

For example the desire for career advancement is a common thread in both of these narratives, even though this is being promoted further up into the kafkaesque absurdity. Both of these things remind me of other things, like somehow everyone in US politics seemingly believes in God and somehow nobody in an international football (soccer) team is gay. Even though these are not plausible.