I guess it's not gambling, or it'd be covered by the UKs existing laws around gambling that set the minimum age to 18.

edit: I'm pointing out the UK has apparently decided lootboxes are not gambling, because if they did classify it as gambling it'd be covered by existing gambling laws that restrict it to 18+.

Not that I personally hold that opinion, though I can see how I could have phrased my original message better.

It's a stupid decision by the government, they should be 18+ and recognised for being gambling.

> I'm pointing out the UK has apparently decided lootboxes are not gambling

Wow, considering how the UK has been going full Taliban on everything why stop at lootboxes? Guess the politicians are getting some money/bribes from the lootbox companies.

There's a nuance here. Consider these two positions that both mean that the UK Government does not currently recognise lootboxes as a form of gambling:

A: The UK Government has decided that lootboxes are not gambling. B: The UK government has not decided whether lootboxes are gambling or not.

The current situation is a lot closer to B than A. The UK Government has decided not to extend the existing Gambling laws to cover video game loot boxes. The existing laws did not automatically cover them, so a decision has to be made whether to amend the laws or not. Instead it has asked the industry to self-regulate. In 2022 there was a review that asked for improvements, restricting some access and mandatory spending controls. There have been other reviews and studies.

It seems quite clear that the industry is not doing enough (no surprise, it's a cash cow for them) so it's likely to mean the Government will crank up the legislation to force their hand, but the wheels of Government move slowly; way too slowly for many people who have been affected by this.

S3 was well aware of the pain when I was there ~10 years ago, just considered themselves handcuffed by the decisions made before the idea of a cloud was barely a twinkle in a few people's eyes, and even the idea of this kind of scale of operation wasn't seen as even remotely probable. The namespace issue is one of a whole long list of things S3 engineers wish they could change, including things like HTTP status code behaviour etc.

I've never really understood S3's determination not to have a v2 API. Yes, the V1 would need to stick around for a long time, but there's ways to encourage a migration, such as having all future value-add on the V2, and maybe eventually doing marginal increases in v1 API costs to cover the dev work involved in maintaining the legacy API. Instead they've just let themselves, and their customers, deal with avoidable pain.

V1 never dies. You support it forever, including for customers who desperately want v2-only features but would rather escalate than migrate.

AWS has a privileged position compared to other deprecation struggles in the industry, though. They can price the v2 version aggressively/at a loss to incentivize migration without major bottom line impact.

And sure, v1 is forever, but between getting to the point where new accounts can’t use it without a special request (or grandfathered in sweetheart rates, though that might be a PR disaster) and incentivizing migration off for existing users could absolutely get s3v1 to the point where it could be staffed for maintenance mode rather than staffed as a flagship feature.

It’d take years, but is totally possible. Amazon knows this. If they’re not doing it, it’s because the costs don’t make sense for them.

I've been finding it very practical to run the 35B-A3B model on an 8GB RTX 3050, it's pretty responsive and doing a good job of the coding tasks I've thrown at it. I need to grab the freshly updated models, the older one seems to occasionally get stuck in a loop with tool use, which they suggest they've fixed.

I guess you are doing offloading to system RAM? What tokens per second do you get? I've got an old gaming laptop with a RTX 3060, sounds like it could work well as a local inference server.

I'm getting about 15-20 tok/s with a 128k context window using the Q3_K_S version.

For running the server:

    $ ./llama.cpp/build/bin/llama-server --host 0.0.0.0 \
      --port 8001 \
      -hf unsloth/Qwen3.5-35B-A3B-GGUF:Q3_K_S \
      --ctx-size 131072 \
      --temp 0.6 \
      --top-p 0.95 \
      --top-k 20 \
      --min-p 0.00

In the article, they claim up to 25t/s for the LARGEST model with a 24GB VRAM card. Need a lot of RAM obviously

Can you give an example of some coding tasks? I had no idea local was that good.

Changed into a directory recently and fired up the qwen code CLI and gave it two prompts: "so what's this then?" - to which it had a good summary across stack and product, and then "think you can find something todo in the TODO?" - and while I was busy in Claude Code on another project, it neatly finished three HTML & CSS tasks - that I had been procrastinating on for weeks.

This was a qwen3-coder-next 35B model on M4 Max with 64GB which seems to be 51GB size according to ollama. Have not yet tried the variants from the TFA.

3.5 seems to be better at coding than 3-coder-next, I’d check it out.

I personally have used Qwen2.5-coder:14B for "live, talking rubber duck" sorts of things.

"I am learning Elixir, can you explain this code to me?" (And then I can also ask follow-up questions.)

"Here is a bunch of logs. Given that the symptom is that the system fails to process a message, what log messages jump out as suspicious for dropping a message?"

"Here is the code I want to test. <code> Here are the existing tests. <test code> What is one additional test you would add?"

"I am learning Elixir. Here is some code that fails to compile, here is the error message, can you walk me through what I did wrong?"

I haven't gotten much value out of "review this code", but maybe I'll have to try prompting for "persona: brief rude senior" as mentioned elsewhere.

3.5 is doing a good job of reviewing code, even without prompting it to be brief and/or rude.

I've been using opencode pointing to the local model running llama.cpp.

The last thing I was having it build is a rust based app that essentially pulls data from a set of APIs every 2 minutes, processes it and stores the data in a local database, with a half hourly task that does further analysis. It has done a decent job.

It's definitely not as fast or as good as large online models, but it's fast enough and good enough, and using hardware I already had spare.

Which models would that be?

unsloth's quantized ones. They mention on the site that this links to that a couple of days ago they released updated freshly quantized versions of Qwen3.5-35B, 27B, 122B and 397B, with various improvements.

I had a similar speed jump, went from a Celeron 233 to the Athlon 1Ghz. Such a huge uplift.

That must've been an even bigger jump since Celerons were notably slower than PIIs.

The old mental model doesn't fit how any OS manages RAM. Every OS plays all sorts of fun guessing games about caching, predicting what resources your program will actually need etc. The OS does a lot of work to ensure that everything just hums along as best as possible.

When I was at AWS over a decade ago, there was endless complaints about the elevator algorithms by engineers, with the usual egotistical tech-bro insistence that they could do better. Things used to really suck at lunchtime when folks would flood to the elevators and be stuck waiting for ages. Those same geniuses could never figure out the benefit of staggering lunch times.

Someone got really tired of it, and somehow organised a hackathon weekend, with the elevator company, and let teams of engineers have at it.

Every single team failed to come up with better algorithms. All the complaints stopped dead.

Is AWS an environment where attempts to discover improvements are commonly mocked?

I don't think the disdain was for seeking improvements. It was for the techbros thinking they can solve any problem, even in domains they have never worked in, better than anyone actually working in said domain.

The disdain on my part was very much towards the egotistical tech bros who were convinced they could do better in fields they had no background in.

It used to get really tiring seeing the rhetoric about every single field of expertise. So many tech bros that were simultaneously experts in law, geopolitics, elevators, building codes, architecture, sports, transportation, finance, global logistics, and beyond. Literally from one second to the next. Any time anything wasn't 100% perfect, it was because the people working in the field were idiots and they could have done it so much better.

I've been testing the same with some rust, and it's has spent a fair bit of time going through an infinite seeming loop before finally unjamming itself. It seems a little more likely to jam up than some other models I've experimented with.

It's also driving itself crazy with deadpool & deadpool-r2d2 that it chose during planning phase.

That said, it does seem to be doing a very good job in general, the code it has created is mostly sane other than this fuss over the database layer, which I suspect I'll have to intervene on. It's certainly doing a better job than other models I'm able to self-host so far.

> it's has spent a fair bit of time going through an infinite seeming loop before finally unjamming itself.

I think this is part of the model’s success. It’s cheap enough that we’re all willing to let it run for extremely long times. It takes advantage of that by being tenacious. In my experience it will just keep trying things relentlessly until eventually something works.

The downside is that it’s more likely to arrive at a solution that solves the problem I asked but does it in a terribly hacky way. It reminds me of some of the junior devs I’ve worked with who trial and error their way into tests passing.

I frequently have to reset it and start it over with extra guidance. It’s not going to be touching any of my serious projects for these reasons but it’s fun to play with on the side.

Some of the early quants had issues with tool calling and looping. So you might want to check that you're running the latest version / recommended settings.

> and it's has spent a fair bit of time going through an infinite seeming loop before finally unjamming itself

I can live with this on my own hardware. Where Opus4.6 has developed this tendency to where it will happily chew through the entire 5-hour allowance on the first instruction going in endless circles. I’ve stopped using it for anything except the extreme planning now.

I don't know much about how these models are trained, but is this behavior intentional (ie, the people pulling the levers knew that this is how it would end up), or is it emergent (ie, pulling the levers to see what happens)?

I haven't seen a page on HF that'll show me "what models will fit", it's always model by model. The shared tool gives a list of a whole bunch of models, their respective scores, and an estimated tok/s, so you can compare and contrast.

I wish it didn't require to run on the machine though. Just let me define my spec on a web page and spit out the results.

I took a quick look, the dependency on php 8.5 is mildly irritating, even Ubuntu 26.04 isn't lined up to ship with that version, it's on 8.4.11.

You mention in the README that the goal is to run things in a standard environment, but then you're using a near bleeding edge PHP version that people are unlikely to be using?

I thought I'd just quickly spin up a container and take a look out of interest, but now it looks like I'll have to go dig into building my own PHP packages, or compiling my own version from scratch to even begin to look at things?

Those are quite good:

https://launchpad.net/~ondrej/+archive/ubuntu/php

Anyway, whatever you write in an earlier PHP version is likely to work on future versions. PHP has remarkable BC.

If you're just experimenting, might as well start in the browser:

https://alganet.github.io/phasm/

Not all extensions available there, but it has the essentials.

Hell, the number of times I've used a lot of the data structures that come up in leetcode exercises without at least looking at some reference material is pretty small. I usually assume I'm going to misremember it, and go double check before I write it so I don't waste ages debugging later.