kadoban

Probably depends if you're trying to treat it as "subagent(s)" or as two different models working on the same stream of thought/context.

Both are, I think, valid options. Depends on the workflow you're looking for.

The open models catch up, the _US_ open models catch up much more slowly. Different incentives and cultures.

> Why does that seem unlikely? The difference between frontier and open weights has been much less than 12 months for a while now.

Frontier and open weights? Yes.

Frontier and _US_ open weights? No.

I'd expect a US open weights provider to suddenly get bought or go non-open if they get near Opus 5.5 anyway.

Wow, that's pretty good.

> Open weight US based Opus 5.5 within 12 months.

Seems unlikely. Maybe open weight at all near there? That'd be great.

The list of shit that Intel was within the ballpark of the right tech or 5 years mistimed from having a huge hit is ridiculous. Fuckers don't have good management, but also must have terrible luck to keep _just_ missing the mark?

> And, back then, this was the normal or is ice covering the normal?

"Normal" is probably not a very useful framing. We're currently in an interglacial period in an ice age. So, on one scale we'd expect ice there, on one scale we'd not. On another scale we're just waiting for the sun to swallow the Earth and burn it all away. What's "normal" out of that?

If you happen to be asking if the current human-caused warming is an issue, yes it is. The Earth doesn't care, and "normal" doesn't care. We do though, we're fucking up our environment pretty good and it's causing misery and death.

Even if it just ends up being a base to throw and llm at to update, isn't that still useful? Sounds to me like it'll probably skip the whole part where my agent would have to give me 4 different plans, searches the internet, argues that it's impossible, works out subtools to generate, etc.

Can't tell if they plan to release the model weights? Very uninteresting if not. Otherwise seems like potentially some nice post training.

> This has been known for a long long time too (decades+), its not a recent discovery.

We've known what caused the effect for a long time. We did _not_ know how they achieved the nanoparticle mix and size in the glass, last I checked. This story is about someone actually taking the basic idea and making a real object.

You are implying that the law cannot be dehumanizing?

One practical-ish thing you could reasonably do with this (beyond just better understanding how models work so you can make better choices/tradeoffs), is post-train some of the small-to-medium models, fine-tune them on your own data or priorities.

You don't need to change many weights, relatively, to add capabilities or modify behavior. Look up LoRas for instance.

I'm of the opinion that post-training existing models is an area with a lot of benefit and fun in the near future (it's already being done here and there of course).

I doubt you _really_ have to know everything in this course to know how to do that, but I'd want this as background personally for it.

Relatively sure it's not. I'm using pi and I can't find any place it's doing any such thing. My current theory is that it's behavior trained into the model either accidentally or intentionally, probably during post-training in some harness that _does_ give hints of that sort?

Any idea what the ~rule is or how it's meant to work? I assume it'd work better if I can wire up my harness to handle it, but no idea. It didn't seem to know itself, from it's training. I didn't search too much yet though.

DeepSeek v4.1 flash seems to do this thing where it starts saying that we should wrap up the current work and start again "next session" (whatever that means), that the "context budget" is almost up, or some other wording I'm forgetting, as the context gets longer.

I've taken to just manually making it write a handoff file and clearing context completely. I'm pretty much manually compacting instead of my harness doing it. Feels weird, but it does work.

Almost any movie makes heavy use of CGI. Combining them with practical effects, like stop-motion, yields better results than sticking to one technique.

The practical grounds the digital, and the digital enhances the practical.

What has been ruined?

Hm? I see a model download link to hf.

These people building on or tweaking open models is _great_, it should be encouraged. There's a lot of room to spend some effort improving models in various ways that the original creators didn't have time for.

Your favorite agent is really good at organizing media files, if you haven't tried that and it happens to fit your situation. I usually make it write out a list of move/renames as a script so I can double-check, but bet it's fine without most of the time.

You could cloak where it's coming from more easily this way.

Yeah, this would be a much better version of the well-meaning but poorly produced youtube videos I typically use to fix things up around the house. Probably a lot of value there.

That kind of coordination doesn't bother me, personally. That's likely just early release/access to reviewers, which isn't inherently problematic.

The part that looks weird to me is some models get tested all up and down their effort level, but this one seems to just have been tested at the level where it will come out worst (I'd expect). Maybe it'll get more coverage over time?

I've also noticed the numbers (and relative rankings) change right after some model releases, which felt a little like someone didn't like who came out on top so they tweaked it.

I wonder what AA's funding and affiliations are. I get the sense that they have their hand on the scale at least a bit.

You don't actually need the signing if you have the counters (if your ~only goal is to prevent replay attacks of that form).

Put it in a container with a build step and lock down versions. That's what I did. Then you can choose what to update, audit code, etc.

Whatever agent will be more than capable of wiring that up.

A mechanism that works at scale X doesn't necessarily work at scale Y.

Different forces and material properties are all at relatively different levels.

The one I know offhand that will mess with stuff is the square-cube law. If you double the linear dimensions of something, it's twice as big, linearly, but 4 times the surface area and 8 times as heavy. The strength of anything that depends on area (like friction) won't scale at the same rate as anything that depends on mass or volume or anything.

Not to mention I'm sure this thing can only print a few different materials. Unless those are also _all_ of the materials it's made out of, it couldn't even make itself.

It's not a crazy question to ask / theory to posit. The next message down explains why it's unlikely and the whole conversation sounds reasonable to me.

The cryptgraphy community is pretty reasonable folks, on the whole.

> We present Ultramarine 1 Flash, an agentic model for professional knowledge work. Ultramarine applies large-scale reinforcement learning to a strong open base model, GLM-5.3 Flash, and improves on it across all six knowledge-work benchmarks we evaluate [...]

In case anyone was curious what it's based on. (No judgement either way from me, IMO depends how well they did the post-training, which I can't tell).

Qwen 3.8 27b is a great model, but 2 bit quant of it is _not_ Opus 4.8 level.

The model they're running is Qwen 3.8 27b. It's maybe somewhere around Opus 4.6 level, it's quite good. I didn't check the benchmarks, that's from memory/vibes.

The ~issue is though that they're using a pretty low quant, 2 bit. 4 bit are _pretty_ close to full ability, maybe degraded a bit. 2 bit is _noticably_ worse. It's about as small of a quant as you can get while still pretending it works, for this model.

This won't replace claude for you, unless you baby it a lot and are okay with it being kind or derpy.