A public back-and-forth between Wharton professor Ethan Mollick and investor Ramez Naam centers on how far behind open-weight AI models trail closed frontier systems, with Mollick pegging the coding lag at 8-12 months while Naam cites evals showing roughly four months on average.
Assuming open models continue to lag about 8-12 months behind closed source (at least in coding), the countdown to hardening IT systems against Mythos-class models is now at 4-8 months
Having publicly available and relatively safe defensive Mythos-class models today is important
Epoch AI data since January 2026 puts the average open-weight lag at four months on an aggregate capabilities index, narrowing from earlier one-year estimates and directly supporting Naam's shorter window.
Mollick flags a closing 4-8 month period before advanced capabilities reach wide open availability, urging firms to shore up IT systems ahead of that shift.
Users are optimistic that open source AI models from Chinese labs like DeepSeek are rapidly closing the gap to closed frontier models thanks to fast iterations, while others find the pace scary or dismiss the short timelines as unrealistic.
No Digg Deeper questions have been answered for this story yet.
😭 bro, if you're sad about this, I'd hate to be you when DeepSeek's Inner Mongolia Division gets the clusters spinning. With $7.4B in Wenfeng's war chest, you'll see how fast this bad boy can go
so it seems like unfortunately we're staying pretty much on-trend and open models are pushing their way towards fable-performance with the usual 4-5 month lag.
opus and 5.5 were already very good at cyber especially with scale + the right scaffolding. these models are cheaper, have no safeguards or kyc, and can likely be fine-tuned + rl'd for even further improvements without significant effort
i think it's very obvious at this point that mythos access has *not* been enough for radical changes and hardening of global security
we're in for a very rough time very soon
@tenobrus Wait, what part of this is unfortunate? You were expecting chinese OS to catch up faster and this disappointed you?
I think we are in a much more optimistic timeline than we thought, the chances of US closed models having a monopoly have never seemed lower
@tenobrus oh no, it's only starting!
@teortaxesTex it's so fuckin over
@emollick Curious that you express the lag as 8-12 months. The evals seem to say 4 months, 8 at the longest, right?
Assuming open models continue to lag about 8-12 months behind closed source (at least in coding), the countdown to hardening IT systems against Mythos-class models is now at 4-8 months
Having publicly available and relatively safe defensive Mythos-class models today is important
@tenobrus This is false. There’s no reason to think Fable tier open sources models will be available that soon. In 2027 yes perhaps
@teortaxesTex it's so fuckin over
@tenobrus *fortunately
We need an intelligence explosion. Machine intelligence should never be held hostage to the whims of petty tyrants. China and IS has every right to take over the frontier if US chooses to cede it.
@teortaxesTex China needs to crack lithography ASAP. can't wait to see what the future looks like
@wolfefriend @tenobrus There is no RSI ramp
@tenobrus and If mythos is already good enough to start the RSI ramp then ???? what do we do
@emollick I haven't seen any compelling evidence that Mythos was actually better at this. It seems to have just been a case of people hearing that it was good at it and then setting it loose on a task they hadn't given to other models.
@woke8yearold @tenobrus In the grand scheme of things, 2027 is soon
@teortaxesTex Huawei is holding them back
@woke8yearold @tenobrus yeah man that's in 7 months
@tenobrus i agree on the whole but i’m also somewhat doubtful that proliferation will continue at the same rate if frontier models get locked down hard enough that “open source” models can no longer be distilled from them
@woke8yearold @tenobrus 3-4 months was Deepseek v1, it was peak opensource. Opensource is like 10 months behind now idk where he got the 4 months estimate from
@emollick Mythos 5 was a marketing genius in the product terms rarely ahead of others
I have used all frontier models extensively I would rate
ChatGP5.5, Gemini Pro then Claude
@tenobrus This is normal and good. As models gradually get better cybersecurity professionals around the world are using them to find and fix bugs. Locking a new model behind glass does nobody any good. Nothing is getting fixed, might as well release it.
@ramez I think the evals overestimate ability in creative or out-of-distribution tasks.
You can’t really measure how far behind open source is in terms of time, but there’s every reason to think the gap is increasing. Compute constraints, distillation being played out, etc. Anthropic arguably even has special sauce no one else has, including OpenAI. people just repeat this as a talking point