at most 18 months

if they don't have a mythos by then something went seriously wrong

@scaling01 Recall when 5.3 was considered a national security issue ? just another time happening that's all

@scaling01 To be fair, we are still not yet at Opus 4.5 level with open models, which I think was an even bigger step change that Mythos is (even if not by much).

That will be a huge milestone by itself, and clearly did not happen yet (or does anyone claim otherwise?).

at least 8 months

@scaling01 in Mr. al Gaib we trust

@scaling01 Nobody cares about mythos.

We care about our stuffs to be done.

@scaling01 If the Huawei Ascend 950 clusters come online in Q4 as planned, Chinese models will catch up sooner. The question then becomes will they still be open-source...

@scaling01 depends on how fast they steal 😅

I strongly agree, and we will also have to see if whatever company attains this capability (when it does) decides to release it or go the safe route like Anthropic.

My own intuition is that Mythos is also a real break in capabilities from the trendline, a bit of a "freak model" in a sense, and so even OpenAI will not be train such a model until ~next year.

But we shall see what happens with OpenAI's next release. It will certainly be an important data point.

(And of course, OpenAI is not to be underestimate as GPT 5.5 has clearly shown in the realm of mathematics)

@scaling01 kinda bold to assume they close the gap at all without copying the messy details

@scaling01 Do I get to cross off June because it’s already started?

OSS benched at Fable 5 level by November?

I believe the over/under is maybe somewhere in November but I’d rather bet first of the year

I just think AI development is getting steeper in general, only reason

@scaling01 It may take a year for open source models to be this capable, but it will still be a huge model way over 1t parameters. Probably much longer for an open source model as capable as mythos to be as small as 1t parameter.

@scaling01 I’d say for general capability in all fields as a full model it will take open models around 18 months

I dont think oss is as behind as we think. Kimi k2.6 is a 1T model and is on par with gpt 5.2 (which is probobally around 2-3T) and beats Opus 4.5 (3-5T)... Theoretically this only means China is compute limited, not talent limited. If Mythos pushes the CCP even harder. I would assume they may fund the AI race more.

@scaling01 K3.1 suffices

@scaling01 you are way too anthropic piled

@scaling01 I put it at 16-18 months. They've made distillation harder + more expensive and one can only do so much with limited compute.

@scaling01 Pliny said we will have one before end of the year idk but I believe him

@scaling01 The 12 month estimate feels more grounded than most timelines

Usually these gaps take longer to close than people expect

@scaling01 Especially if Anthropic's distillation protections are effective.