ilc

Steve, I remember our talks at XP, other conferences and plugfests fondly to this day.

From my mortal spot, I'm sad I'll never get to have another.

Rice made a good one in 'ya, though I suspect anywhere could have forged you well. You were made of the right stuff.

God speed Steve.

Watch the video carefully. DFlash2's tool call fails on python syntax.

Usually models in this class nail things like that 1 shot, which the other side did.

I don't know the cause. It may be nothing. But I'd like to see the model doing something where its path is a bit more constrained, to help out rule out such oddities.

A quick question: Where can I read your TOS, and other legal documents?

Knowing things like: Will my code end up in Grok? What ownership do I have of my content, etc... matter.

Any one weight, but all of them. And also crushing the architecture itself?

I wouldn't pick up 400gb of hardware to run in that mode. I might try it for fun, but even then you are looking at handling a 95GB active parameter set.

This is NOT a model for most home labs. I'm sure some can and will use it. But most, should steer clear.

To compare a 1 bit quant to the full fat model is misleading.

Honestly this model people at home can tinker with, if you have a big enough Mac. Maybe 4 Strix Halo/DGX Spark, and then at 1 bit quant? Nah.

Use the right sized model, for your hardware. You'll get better results.

Having audited one code harness (Qwen Code) in the past on telemetry. Let's say I'm sus of the Zuck offering free candy.

Sandbox and use Local AI. This is the real answer.

As soon as it said "non-commercial" I noped.

That is so vague with some of what I do, that it isn't worth even looking further.

I'd made the shift to HTTP/Stateless from MCP a few months ago. It's the right thing to do IMHO. Reliability up, problems down. TOON support is natural, if desired, etc.

My only question is how do you handle channels in the architecture now. From what I saw in Claude Code, shifting to a totally http world has some timeout issues if a server drops out and comes back. Because of that I'm stuck writing stubs for my internal use MCP, this is fine for me, but if you are cleaning up semantics: Understanding how we expect clients to act around failure would really help, the story.

This is definitely one where wait 2-3 weeks and look again, feels like the right strategy. It could be good, it could not. It is too volatile to say.

What harness/quant did you use for testing?

Not the parts which matter to trust. Which is my point.

You can state the math, but not why it won't discuss various topics, etc. Once you see the models waffling on subject with objective truths. You wonder what else is wrong.

I do not exempt US models from this. They do it too, ask anything about politics, elections etc. And they can get... weird.

It doesn't take much to create a systemic error class in a model at these scales. And history has shown nation states are willing to do these things.

Just be wary.

Training data, training methodology. All NOT OPEN.

Until we know what a model is trained on, and how it is trained in high detail, I hesitate to call them "Open Source" in any way. They are free. But, we don't know what their priorities are etc. Witness the censorship we see in all models in one form or another. I'm not absolving any side of this.

Just saying: Don't be blind.

Opus: Yeah, it can be a little fuzzy at 600-700k, I prefer it 400K and under in general. But, I've had usable sessions to 850k.

It really depends on the session, and what you are doing how far you can push it.