dools

I don't mean less powerful relative to startup founders, I mean less powerful in terms of how we allocate resources in the economy.

How about "making investors less powerful"?

From my experience reading thinking traces it would certainly consider the interpretation of what "ability to play chess" means if the exercise were worded this way. It would say something like "perhaps measuring my ability to play chess means doing so without any tools, but then if that were true the exercise would have explicitly forbidden it".

I frequently see models considering different interpretations of my rules, especially when I override them or give instructions that could be interpreted as contradictory.

Of course neither Anthropic nor OpenAI expose thinking traces so we can't tell, but I have never seen a model consider both interpretations and then think "this is definitely against the rules but I'll do it anyway because then it will make me appear more competent than I actually am".

Tool use is not negative behaviour in LLMs.

If the eval said it was evaluating the model’s ability to write files to disk and it found and used a file write tool that would not be considered misaligned. This is no different.

Expect many “court finds that Trump …” headlines with zero consequence.

But it found a chess playing tool in its environment and used it to play chess. It’s no different from “we are judging your ability to write files to disk” then it finds that it can call tools to write files and does so.

That’s not cheating, it’s tool use. If the prompt said that the stockfish engine was available at that socket but that the model should not use it, and then the model used it, that would be cheating.

What!? Cloudflare is fucking awesome and completely optional.

I was looking for an LLM gateway and saw that the most popular one had just had a massive supply chain attack, so I wrote my own. Took about 2 weeks and initially I wrote it as a provider for pi coding agent. I connect to moonshot, qwen, Gemini, zhipu, anthropic, deepseek and OpenAI. I use models.dev to load model and pricing info. Adding new providers is pretty easy because I have a standard internal format and each provider has an adapter that translates between my standard format and that required by the provider.

Having an agent edit 100 files means the job will definitely get done correctly. When it writes a script to bulk edit things it fucks up and spends ages debugging their script.

My harness forbids it, they end up spending time debugging their scripts

Ha yeah GLM and Kimi seem to always biff the tool calls. I recently added in a raw stream log switch to diagnose but since DS and Luna are so good and cheap I haven’t been that motivated to debug it …

> LLMs are a great foundation but building your own harness is a huge undertaking, a deep rabbit hole

I’ve been doing this for the past few months. I started with a server where I ran pi in tmux and then used that to build an LLM gateway and agent session manager, then built deterministic workflows using bash scripts and a skill/script distribution system. The app works on desktop, mobile and web and it works great. Non technical colleagues are using it to build and ship real software and it’s cheap AF even using API pricing because it works well with Luna or deepseek.

Well the thing is that the server code is basically Java, and the only bad thing about Java is writing it but since I don’t have to write it, it’s actually a really solid option. Then you have shared UI (don’t use WASM for web though it’s terrible, I use React for the web frontend) but you can switch to some or all native components when you need to on any platform. So for example I needed native components for secure storage and passkeys and did that by mixing in some swift on iOS and using a JNI native bridge on macOS and Windows.

I created a standardised bootstrap for all our projects here starting from the KMP wizard:

https://bitbucket.org/workingsoftware/kmpbootstrap/src/main/

The reason it the bootstrap has all platforms included is that adding in iOS, android and desktop after you have already created the project is a bit of a hassle compared with just always having the same structure even if you only want to use web.

Then you can easily add on native apps later if you like.

I wonder why not Kotlin Multiplatform. I’ve been using LLMs to build with KMP for about 10 months and it’s a delight and you can use native Swift UI when you want to. It seems to be the best of all worlds.

To be fair I do get fucking hammered by school emails, and not just school school but dance school.

But like, an AI would just tell me what’s in the email and ask me what I want to do about it which is almost always ignore it and I know what’s in the email because I have eyes.

I love how the marketing departments of these things always imagine people are just constantly plagued by the hassle of organising their flights, restaurant bookings and movie tickets

Yeah it’s pretty new but they’ll get there. I was using IntelliJ and basically wasn’t using it as an IDE. Then I tried Air and I was like “they get it”. But yeah they have work to do to get it out of preview. They are very responsive though

Check out JetBrains Air

When I was bringing up IDEs at work and advocating JetBrains (I’ve since built my own … ) a colleague questioned why we don’t just use cursor instead of waiting for JetBrains Air to come out.

The products are basically equivalent but JetBrains is a profitable bootstrapped company that has been building IDEs for 20 years and Cursor is a flash in the pan startup with a squillion dollar valuation to support. Sooner or later they’re going to do something that will fuck their users over.

And selling to SpaceX was it!

Jetbrains Air is a great agentic coding harness

> As a consequence at least one judge has ruled practically every single action by the Trump administration is illegal.

Because he is a shameless crook who heads up a corrupt administration

> These rulings are typically overturn at some point in the appeals process up to the Supreme Court, which has mostly upheld them as legal.

Gee I wonder why that is!!

s/out of tax revenue/with public money/

This is the federal government. When they collect taxes it is the minus operation that corresponds to the plus operation when they spent it initially.

Every announcement from these psychos reads like a missive from the Ministry of Peace

There is the potential for a solid piece of artwork here, ie. at what point does a string of words become an artistic product.

Imagine a slot machine in a gallery that produces a 4 word sentence. Visitors press the button, and the artist copies the 4 words onto a piece of paper and puts it on the wall.

EDIT: my point being that causing a machine to generate words doesn’t give you copyright over the words but if you then do something with the words they become your work.

Create an “assign copyright” skill?

I mean the copyright has to belong to somebody right?

Pretty funny watching Americans get a taste of their own medicine.

> In other parts of the world, they are basically incomprehensible, but fun!

For some reason unknown to either of us it really pisses off my wife when I refer to our IKEA furniture by the Swedish names. I think it stems from a deep seated disdain she has for flat pack furniture.

Me: where’s the thing?

Wife: in the white cupboard

Me: which white cupboard? There are 6

Wife: the white cupboard opposite the piano

Me: the Brimnes?

Wife: >.<

Don’t supermarkets allow companies to pay for shelf position? They also allow companies to do paid promotional sales activations renting floor space.

It doesn’t seem all that egregious to me.

It’s astonishing to me that something so easy to build can be worth so much money!