Employees at OpenAI and Anthropic might have additional reasons to be spooked, from firsthand observations of post-Mythos (let's call it Mythos+1) and post-Astra models that are probably already trained (targeting hardware of 2027). The capabilities of Mythos+1/Astra+1 (even if they are not internally deployed outright yet) might be notably different from Mythos/Astra. There are already the mathresults, but the qualitative impressions from looking at the models in other kinds of situations could also be telling.
The plans to pretrain Mythos+2 and Astra+2 (targeting hardware of 2028) sometime in 2027 would also be underway (and not in late 2027), so that there's enough time for possible retries and post-training later in 2027. A Mythos+2/Astra+2 model could be about 200T total params, which is by itself alarming.
This isn't being observed in public, and isn't being observed at Meta/Google/SpaceX. Plausibly both Gemini Argon and Muse Wonder are Opus-scale models, created without the benefit of distillation from the more capable Mythos+1 or even Mythos scale models (at least in the case of Meta). They are probably not yet capable enough for their own HuggingFace moment, but at least Meta seems on track to make a Mythos-scale model soon, then a Mythos+1 sometime in 2027, and a Mythos+2 in late 2027, plausibly catching up to Anthropic and OpenAI (which would also have a Mythos+2 by then). A lot of work to do, and the alarm from personal direct observations of a Mythos+1 only starts in 2027, maybe after the actual Mythos+1 and Astra+1 (from Anthropic and OpenAI) are already publicly released.
(For Google, it might be the case that Carbon is Mythos-scale and their biggest model so far, but there aren't necessarily plans to go further as fast as feasible, so they don't necessarily have a Mythos+1 or plan to catch up to Mythos+2 in 2027. And SpaceX hasn't yet demonstrated a model beyond Sonnet scale, though Grok 4.7 is only slightly less convincing as evidence of feasibility of future plans than Muse Spark 1.3 was just a few days ago, before Wonder came out.)
Source? Personal correspondence, vibes from reading twitter, speculation...
(I'm not trying to say "you're wrong!"; I'm trying to calibrate if this is new signal/analysis)
This is analysis/speculation, but it is based on public information I've seen recently. Here are a few examples:
https://x.com/karpathy/status/2109361546505966046
https://x.com/robbensinger/status/2110033571301982214
https://x.com/DKokotajlo/status/2099600298855829616
I think it is not too big of a leap to conclude that the frontier lab employees have seen things internally that terrify them, but they may not be allowed to share exact details or publicly reveal solid evidence. (A somewhat related example: OpenAI has not allowed any independent auditors to investigate the hack of their internal research infrastructure that occurred after the Hugging Face incident.)
If anyone disagrees with this assessment, please let me know your reasoning.