trollbridge

“Distillation” is just a term of art these days. Usage of competitors’ models these days is mostly around RLHF (minus the H, I guess).

GLM 5.2 is light years ahead of anything called “Gemini”.

I was surprised to learn Ethiopia is nearly all EV now.

They are a petroleum products importer, so it’s a big win for them.

Chinese IP law is actually quite solid. You do have to register your trademarks and copyrights properly in China, and then lawsuits have to be filed appropriately according to Chinese law. Is that a problem?

It turns out U.S. law doesn’t have jurisdiction across the entire world, nor does Anthropic and OAI’s rather blatant attempt to buy government influence.

Great! Neither is distillation then.

Are you sure about that? I can't think of any bread bakeries pulling 2.2 gigawatts, which is us-east-1 pulls.

Be cautious with this. My firstborn's first complete sentence was "King Graham fell" (from the sledding scene in King's Quest V).

I'd say modern AI tools could probably do this pretty effectively. They're very effective at describing anything else about an image. I have a workflow that churns through large amounts of images, describes them, and then looks for things I specifically want (in my case, auction listings that are not described accurately on the auction website).

I thought the client side scanning was to protect children? If it suspects an image is bad, it blurs it and pops up a warning including a link to resources to go to for help.

Very different than trying to narc out users to the authorities.

The cost of computer components going up 4X could naively be believed to eliminate 80% of demand.

The great news is you can use DS-V-Pro, MiMo-V2.5-Pro, GLM-5.2, or K3 and see everything.

I tend to use 5.6-Sol now more for one shot type of tasks where all I want is the answer and I’m not going to read the reasoning.

Yes. I taught the non-programmer to ask the harness to set up things like MCP connectors.

Thanks for that. My first question was “What does this do that Unsloth doesn’t?”

Aren’t there all chat interfaces? Same goes for Claude Work, ChatGPT Work, etc

I make sure projects are split into much smaller units so we don’t need some gargantuan bazel based build.

Having a bloated, barely-maintainable codebase and a build that takes hours and needs 100GB of storage is not a badge of honour.

I actually found a 1999 iMac set out for a special rubbish pick up day. The owner of the house was there so I chatted a bit, asked her if she minded if I took it.

It had last been booted on it, complete with working hard drive an; MacOS X 10.3. So $1.299 -> $0.

Or get the orange driving glasses.

DeepSeek has a phenomenal cacheing architecture that specifically can work with less-than-ideal clients like this and get cache hits anyway.

This seems to be quite common. You start with OpenCode (assuming you're someone who skips all the Claude Code hype); then you start thirsting for more, and switch to oh-my-pi.

I generally view the kind of code in projects like OpenCode to be the same as binary machine code, and the harness that works on it is the compiler. The source code is the prompts (I guess), except we live in a weird state now we use the existing machine code + more prompts to make even more machine code.

Isn't it standard journalism? Instead of just giving us the conclusion we're supposed to think, they interview various people in the field and quote what those people think, and then let us draw our own conclusions.

In the U.S., the DOT sets standards, but they only apply to new vehicles sold. Aftermarket lights can be just about anything. (Some are labelled "not for highway use", as if that prevents them from being used on public roadways.)

Certain types of people seem to think they should install really, really bright LED lights in their car and then aim them right at other driver's eyes. I believe this comes from a mindset where one is unable to put one's self in another person's aspect, or where one simply doesn't care about others.

The "lights=safety" nonsense shows no sign of coming to an end. Many private residences also install eye-searing levels of lighting outside that they just leave on all night, which is much cheaper to do now thanks to LED technology. There's really no reason for doing it. The idea it deters crime is a long-standing myth (someone analogous to the idea CCTVs deter crime).

Wouldn't it be more reasonable to just expect people who want to walk places in the dead of night to wear a headlamp? They're cheap, easy to recharge, and can be adjusted for various brightness levels.

For greenfield work, we are using mostly Rust, although small utilities built with Go or Python are also acceptable provided they have a comprehensive set of unit tests, an integration test suite, well-defined specifications and acceptance criteria, and thorough documentation. Once you have that, it's much easier for the AI to work on it. (Our main motivation for doing this is that it takes far fewer tokens now to get stuff done since we have all that, to the point I can leave Qwen 3.6 running in 24/7 loops accomplishing things.)

If your build is bloated and slow, it would seem prudent to go and fix that, possibly converting a codebase from some other language or build system to something that can be built and tested very quickly. Slow builds and slow unit tests are a choice.

"People paid $5,000 for a Macintosh computer when they were new. I found one at a yard sale for $25."

What are you talking about? A properly prompted agent generates a project with a fast, reproducible, and isolatable build. One of the best things about agentic AI is that I spend far less time on devops and on waiting on builds.

$5 with a referral code for the first month.

You get $60 worth of usage. It’s too good to be true, yet it is.

Or a used tower form factor PowerEdge if you want a little bigger, but still designed to be quiet in an office environment.

Considering the results from Kimi K3, it appears most the accusations of them “stealing” via distillation are unfounded accusations.