@deredleritt3r covered k1.5 here!
moonshot's research taste & adherence to it has always been good
they were also one first chinese labs to achieve multi-modality too
https://t.co/r3chtFSzk6
other types of data is sparser, but by subsidizing routers (中专站) & purchasing data from routers
they've gotten a lot of data + subsidizing & developing their own harnesses now (kimi-code, Zcode)
also moonshot is just brilliant, wrote some ways they approached solving the data problem in the K1.5 here
https://t.co/CIHo0z86j2
@IbrahimDagher20 so expert data is striving (lots of highly educated & specialized phds)
there's a big surplus of peking u, tsinghua u students to help provide & label data
phds in china generally dont graduate early, they do 4 years+ and msotly can't take on jobs outside the lab
@badlogicgames i wrote this lookback on the k1.5!
they intentionally went for simpler and scalable training methods on top of their k-series transformer base
https://t.co/r3chtFSzk6
@mvanhorn the strategist prompt is really interesting because it's evolutionary context engineering
the strategist is loaded fresh to avoid context pollution but given traces and transcripts
the prompt uses that to try new directions
https://t.co/9evH8SV12d
what's interesting is the approach to goal building
when you set a goal, an acceptance criteria is set
the goal ends when a quorum agrees the goal is done
if the work keeps failing
a strategist is invoked that can change the approach to achieve the goal
what's interesting is the approach to goal building
when you set a goal, an acceptance criteria is set
the goal ends when a quorum agrees the goal is done
if the work keeps failing
a strategist is invoked that can change the approach to achieve the goal
what's interesting is the approach to goal building
when you set a goal, an acceptance criteria is set
the goal ends when a quorum agrees the goal is done
if the work keeps failing
a strategist is invoked that can change the approach to achieve the goal
Or… what if we gave you $100 in Codex credits if you tell us what you love about GPT-5.6 Sol or why you switched?
Tweet it, claim your gift, enjoy more usage. First 10k get the free tokens!
https://t.co/8mU93eA13i