/Tech2h ago

Open Source AI Tops ErdosBench Ahead of GPT-5 as China Leads

2702635

#595

Original post

⿻ Andrew Trask@iamtrask#595inTech

Everyone is *missing* what actually happened here... this is the craziest weekend I've ever seen in AI

1. #opensource AI just beat closed source AI

Let that sink in... open source is no longer 12 months... 6 months.. or 4 months behind. Because Fable/Mythos is shut down, the most capable AI which *anyone* can use right now... is open source 😱.

2. For the first time (ever?) China is offering the world's most capable general AI system

Up until ~2 days ago... the leading AI reasoning system has *always* been from a US company. If ErdosBench (a reasoning benchmark) is right... that's no longer the case. 🤯

3. Anthropic released Mythos *BARELY* in time for the US to maintain its lead... for China to NOT be the leading reasoning model in the world 🌎.

4. The USG handed 1st place *back* to China by banning Fable/Mythos... 🤦

What an *insane* weekend. Truly a crazy weekend. I can't remember a more insane time in AI. This is crazy.

Przemek Chojecki | PC@prz_chojecki

Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh

We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs.

Kimi 2.7 is amazingly good. More below.

8:52 AM · Jun 15, 2026 · 283 Views

Sentiment

Users express excitement that open source AI like Kimi is topping benchmarks ahead of GPT-5 because the surge feels inevitable yet faster than expected with many untapped workflow opportunities.

Pos

100.0%

Neg

0.0%

3 comments with sentiment.

Cluster Engagement

Posts from X

Most Activity

VIEWS352BOOKMARKS1LIKES3

⿻ Andrew Trask@iamtrask

folks are *missing* the bigger story here:

1) open source just caught closed source (!) 2) china just caught the US (!!) 3) fable was *barely* released in time to maintain the lead (!!!) 4) by banning fable... the USG handed china the lead. china is the world's most capable AI model now (!!!!) 5) ensembles of models are beating individual models... OpenRouter is the world's most capable frontier AI system atm (as opposed to traditional "frontier AI Labs")

tldr: centralized AI is in trouble... and it's IPO time...

Przemek Chojecki | PC@prz_chojecki

Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh

We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs.

Kimi 2.7 is amazingly good. More below.

1h35231

REPLIES2

⿻ Andrew Trask@iamtrask

@amir_harati not new to science - but is new to products

1h132

⿻ Andrew Trask@iamtrask

@amir_harati 100%

1h122

⿻ Andrew Trask@iamtrask

@amir_harati like... "test time compute" also wasn't new... it's fancy beam search.

1h102

Amir Harati@amir_harati

@iamtrask is this actually new? i see these type of fusion models for a while so is the new thing publishing the evals or they are doing something really different?

1h181

tsunami_crypto@ls_brd

@iamtrask open source winning feels inevitable but i didnt think wed see it THIS fast

whats the trickiest problem it solved?

2h101

Amir Harati@amir_harati

@iamtrask yes so many low hanging fruits . I started working on some workflow automation and was amazed not a lot of solution already exists

1h81

⿻ Andrew Trask@iamtrask

@amir_harati but the engineering wasn't there for a while in the leading products.

1h92

⿻ Andrew Trask@iamtrask

honestly i think there's been an insane amount of groupthink in the space. and the economics has been conflicting with it for quite some time... and folks are now waking up because:

- time: truth will out - open source maturity: is making the cost differential of ensembling deeper - data soverengty: there's big conversations happening here

1h12

⿻ Andrew Trask@iamtrask

@ls_brd i honestly don't know yet. analysis is still coming in.

2h9