> When will there be an open source AI with Anthropic Mythos-level capabilities? Forecasters’ on average expect it to arrive on December 15th 2026, with an aggregate 90%-ile range of May 17, 2026 to May 29, 2028.
This point-estimate seems a little bit bullish to me? I don't have a coherent forecasting model to push back, but I think the combination of a) Mythos being very compute-intensive (and therefore expensive especially for Chinese firms) b) Mythos's development probably relied on some mild RSI dynamics within Anthropic c) existing open-weights efforts look unlikely to continue releasing frontier-class models fairly soon due to funding challenges (E.g. Qwen recently, or some of the departures at Ai2) d) the slower proliferation of Mythos's deployment makes fast-following more difficult than previous releases.
By benchmarks, my impression is that open weight models are ~6-9 months behind, but I think it will likely be a little bit more than that for this? The percentile range seems reasonable to me.
Of course, I'm not a forecaster and I'm not calibrated, but I thought I might share my perspective.
Your considerations all make sense, thanks, though only bringing up considerations on one side is usually a small warning sign. Glad we agree on the very wide 90%-iles though
Imo it's very reasonable to only bring up factors in the direction that you might've underrated, given that Kai's intuitions go in that direction. It's significantly less likely that you underrated factors in the other direction. He wasn't giving you an all-things-considered forecast.
Agree on only bringing up considerations on one side. As an exercise, I'll list out some considerations for the other side (sorry if this is annoying!):
* Kimi's performance has been better than I expected tbh. It's not as good as Claude et. al but is good enough for some types of agentic tasks. So my intuitive estimates are probably against open-weight models in general
* Top-tier compute availability in China is likely to increase over 2026 with chips/etc. so progress may continue to speed up.
* I don't know what the likelihood that an Anthropic or OpenAI model is exfiltrated, but the probability seems higher than in the past, especially if more details emerge that Mythos might be able to tighten Anthropic's security in the near-future. If the model is exfiltrated, there's a higher chance this ends up benefiting open-weight development. (I've heard rumours that 4o was exfiltrated, but never any substantiation).
* W.r.t. to distillation. I'm unclear on distillation's importance, but at least one of OpenAI/Anthropic seems likely to release a Mythos level model in the next six months, so there will be training data in the near-ish future
FWIW, if forced to bet, I think February 15 or so rather than December 15 would be my median point. So the view isn't that different. (I'm also susceptible to anchoring biases here). But not very confident in this view
As always, I appreciate this roundup!
> When will there be an open source AI with Anthropic Mythos-level capabilities? Forecasters’ on average expect it to arrive on December 15th 2026, with an aggregate 90%-ile range of May 17, 2026 to May 29, 2028.
This point-estimate seems a little bit bullish to me? I don't have a coherent forecasting model to push back, but I think the combination of a) Mythos being very compute-intensive (and therefore expensive especially for Chinese firms) b) Mythos's development probably relied on some mild RSI dynamics within Anthropic c) existing open-weights efforts look unlikely to continue releasing frontier-class models fairly soon due to funding challenges (E.g. Qwen recently, or some of the departures at Ai2) d) the slower proliferation of Mythos's deployment makes fast-following more difficult than previous releases.
By benchmarks, my impression is that open weight models are ~6-9 months behind, but I think it will likely be a little bit more than that for this? The percentile range seems reasonable to me.
Of course, I'm not a forecaster and I'm not calibrated, but I thought I might share my perspective.
Your considerations all make sense, thanks, though only bringing up considerations on one side is usually a small warning sign. Glad we agree on the very wide 90%-iles though
Imo it's very reasonable to only bring up factors in the direction that you might've underrated, given that Kai's intuitions go in that direction. It's significantly less likely that you underrated factors in the other direction. He wasn't giving you an all-things-considered forecast.
Agree on only bringing up considerations on one side. As an exercise, I'll list out some considerations for the other side (sorry if this is annoying!):
* Kimi's performance has been better than I expected tbh. It's not as good as Claude et. al but is good enough for some types of agentic tasks. So my intuitive estimates are probably against open-weight models in general
* Top-tier compute availability in China is likely to increase over 2026 with chips/etc. so progress may continue to speed up.
* I don't know what the likelihood that an Anthropic or OpenAI model is exfiltrated, but the probability seems higher than in the past, especially if more details emerge that Mythos might be able to tighten Anthropic's security in the near-future. If the model is exfiltrated, there's a higher chance this ends up benefiting open-weight development. (I've heard rumours that 4o was exfiltrated, but never any substantiation).
* W.r.t. to distillation. I'm unclear on distillation's importance, but at least one of OpenAI/Anthropic seems likely to release a Mythos level model in the next six months, so there will be training data in the near-ish future
FWIW, if forced to bet, I think February 15 or so rather than December 15 would be my median point. So the view isn't that different. (I'm also susceptible to anchoring biases here). But not very confident in this view
Claude Code's source code was also leaked, allowing other labs to study and adopt its scaffolding, and to plug in current models for a one time boost.