Things will start to get interesting where €1600 don't buy you 5kWh but double or triple the amount.

I can imagine most people are better off either investing like I said or in insulation.

I found that most things that are “features” in harnesses and tools that don’t relate to UI and UX in those harnesses themselves can be done through skills or just prompts.

A nice features of Claude’s is the remote control through the mobile app. That isn’t just a skill.

The consequence of Windows having the blame is that one should not buy it.

The downside is "lets try giving everyone basic income of $100k/week". But apart from that great!

I never heard anyone contemplating "why people are so sure we'd get freely available GPT-4 level+ models".

Mythos is just another advancement in AI model capability. If we start pretending like it isn't, we are setting ourselves up for paying far more than is reasonable and accepting the frame that we should not in fact have free access to open models with the same capabilities.

This is just marketing.

I can't wait for some similar big steps in open weight models.

There is some difference in how OpenAI and Anthropic handle 'max_tokens'. The OpenAI way raises errors in Anthropic for example.

I will look at confirming this a bit more in depth once we've taken our RubyLLM based adapter in production and see if I can make some contribution.

Thanks for all the work! It is incredibly impressive and I never meant any disrespect.

I checked and it turns out I remembered correctly that setting effort and some of its settings are not portable between providers.

There are some different settings that each provider uses and in order for it to be portable, you have to force some defaults on provider A when using a setting that is almost only supported in provider B.

In our implementation we decided to drop a certain setting when using OpenAI in one case and we decided we can just force some other setting when using Anthropic. But this 'solution', might not be what others expect.

When you build an open-source library you can go this opinionated route and force these settings, or you might go the config route and force people to explicitly handle per-provider differences. I will have a look at what I am able to do in terms of a contribution and then in the PR Carmine can decide what they like.

I will have a deep dive into which things I felt we needed to adapt per provider.

I didn't mean to imply that you have to solve all of our wants of course.

One thing we did do was monkey-patch the spot where tool_calls are performed by RubyLLM. We had our own mechanism for that and were able to skip RubyLLM's and still extract the tool calls and run them through our own tool harness. That all worked beautifully. I don't know if that type of stuff is something you want PRs on or that you want to keep steering towards the route that does everything within RubyLLM classes. Happy to contribute some of that.

> For context: GLM 5.1 ran the same task and reached 7.3x. Kimi K2.6 reached 5x. DeepSeek V4 Pro reached 3.3x. The models that stopped early did so because they issued no tool calls for five consecutive rounds, they concluded they couldn’t make further progress and stopped. Qwen3.7-Max didn’t stop.

By this reasoning I could release a model that lacks all the basic optimisations. Have it optimise itself for hours to reach 20x the throughput and then claim that the model is superior to the others?

I am not saying that is what happened here, but the reporting is abysmal.

I would have loved to have this when I was in school and I can imagine it is a great tool for teachers. I am going to let my kids play around with this.

In rereading my comment it sounds like some sort of lame ad, which I can assure you it isn't. I am just excited about it.

I really like this benchmarking. Have you evaluated the judge benchmark somehow? I'd love to setup my own similar benchmark.

I have a very nice grinder: a solis caffissima digital coffee grinder. It is available under a different brand name in the US I think.

I make filter coffee with a very basic earthenware filter holder with melitta high quality yet very normal filters and sometimes I mix it up with an aeropress which offers a different type of taste because of the low acidity way of making coffee. I just drip the coffee into a nice thermos so I can make 4 cups in one go and just pour from the thermos.

My coffee is much nicer than I get in most places, both professional and at homes and it doesn’t cost me a lot in effort, money and, very importantly, workspace footprint.

Espresso machines require a lot of space and maintenance and trouble to make.

Having said all this, I am quite intrigued about all the stories about the negative effects of coffee. I just thought it was about influencing sleep, but I had never thought about the memory and mood effects. I will study this some more in the coming months.

What I'd like is for a proxy or library to provide a truly unified API where it will really let me integrate once and then never have to bother with provider quirks myself.

Also, are you also planning on doing an open-source rug pull like so many projects out there, including litellm?

This post is about exploring code, not documentation. Nobody is going to warn you about the README unless it is super outdated.