upvote
Interesting. You're claiming that the Opus/Sol/Neon class and Fable/Astra/Carbon class models are similar sizes to each other? That would be surprising to me.

I've found that while Opus 5.5 is often pretty nice for practical work, Fable (even 5.0) is still significantly smarter on hard tasks and makes far fewer "dumb mistakes"

reply
A Microsoft Azure page briefly claimed Sol runs with less looped transformer passes; and so you could speculate that Sol and Astra might share weights; but have different “loop” factors.

That would make a lot of sense from an inference optimisation perspective. The same prefill and decode machines can hold one set of weights, and serve both Astra and Sol requests.

At (multi-)terabyte-scale model weights, and when you have to offer not just the latest like Fable 5.2 but still serve 5.1, 5.0, etc; heck I think Opus 4.6 is still around; you really care about overall compute utilisation.

Idle GPUs are wasted GPUs; and even over the fastest interlinks you’re not hotswapping multi terabytes of weights.

With MoEs, there’s probably stuff you can play around with varying the number of active experts; or other clever optimisations.

Maybe Sonnet 5.5 could be just Opus 5.5 with a different router and less active params? Who knows; I’m just theorycrafting now :)

reply
> A Microsoft Azure page briefly claimed Sol runs with less looped transformer passes;

This doesn't really make sense given how the models released.

OpenAI released Astra long before latent-space looping was a thing at the frontier space.

Then they released Sol 6, which was fast, clearly not looped. Then they release Sol 6.1 which was nearly as good as Astra, but way cheaper, and WAY slower -> this would follow it just being a looped version of Sol 6.

It doesn't seem likely that Astra 6.2 is just a looped Astra 6... It's too fast, and they'd go bankrupt if they're charging the same price... It may be a further looped version of Sol than Sol 6.2 - which would explain why it's not as much better as we would've hoped.

It may also explain why their release is not going as well as they hoped. It's hard to predict how your first looped models are going to fare.

If you look at how long it took Google to recover from their failed 3.5 Pro, it would be surprising if GPT recovered from their rumored failed run on Astra in under a month. That would give them reason to release a looped Sol and call it Astra.

reply
Looped transformers have been known for years now. I don't think it's something new.
reply
They are JUST arriving to frontier models.

Even in the open weight Chinese models as small as DeepSeek V4.1 Flash - nothing is looped.

reply
I think they're responding to the claim Astra can't be looped, there's clear and convincing research, as well as public comms, that it is. It's unclear why Astra can't be looped because looping is new, but, 6.1 Sol released a week later can be.
reply
Astra was the first shipping looper. Unconvincing evidence would be reporting x OpenAI's quick reaction to it via a blog post from their chief research (it broke what's considered a safety taboo). Convincing evidence would be research from ex. Neel Nanda on serial reasoning, recent update at https://x.com/EpochAIResearch/status/2122762411287748696?a=J...
reply
I heard rumours that Sol 6.1 was cut down Astra because Sol 6.0 was received badly. But I thought that was the exception (and a result of Open AI's actual Sol-class model being uncompetitive) rather than the rule.
reply
That would line up - on release and for the few following weeks Sol 6.1 felt like Astra but maybe more task oriented, recently it’s been a noticeable downgrade in the level of specification needed to successfully set off.
reply
deleted
reply
Fable & Opus, Astra & Sol, and Carbon are all thought to be >1T param models - AFAIK, most people think Opus & Sol are ~5T param models and Astra, Fable are 10T+.

I think the rumor-mill consensus is that Google's models are significantly smaller while attempting to swing at the same level - but the ratios AFAIK are thought to be around the same - i.e. Neon is somewhere around half as big as Carbon, big enough that a newer checkpoint can make it temporarily a little bit better.

reply
So many figures and ratios pulled from a single ass
reply
There's only a limited amount of hardware for training and serving models at scale, and which form an upper bound for what model size/inference can support.

A GB300 rack has 72x288GB=20TB VRAM. Do the math and 10T parameter models are just about the max size you can easily train and serve (at 1M context) without doing a lot of custom hardware development or limiting your scale.

reply
The Bitter Lesson - Richard Sutton (2019)

https://news.ycombinator.com/item?id=19393432

reply
deleted
reply