Sabitlenmiş Tweet
Jonathan Chang
2.9K posts
Jonathan Chang
@ChangJonathanC
inference @poolsideai, building https://t.co/uEbfxzEztO
Zurich, Switzerland Katılım Mayıs 2020
982 Takip Edilen1.8K Takipçiler
interestingly, with 5.6 i feel i can orchestrate more agents at the same time.
because the model is much more reliable and goal driven. i just need to think about what i want and less about what i want it to do.
English
It is a truth universally acknowledged, that this
was a young man in want of a father. His father
was a very proud man, and his mother had been
dead many years. His father was in town. He was
sent for. He was told that he was to fight. It was
very dreadful. It had the name of an angel of
light. He wished to run away. He did not run
away. He was in great danger. He was made
unhappy. His father did not love him. There
were two young ladies. One was a very strange
creature. The other was a lively young lady. But
his father had a design against the world. He
meant to be the means of uniting them all. It
would be the end of the world. The young man
would not. He wished to live. Everybody was
very happy.
English
@sasuke___420 just ask model to give you reasoning summaries lol
English
What are the things you do as a coder or developer every day that have little to do with actually writing code?
English
I have been saying we should preserve thinking blocks since the first gen reasoning models.
now with the addition of openai, basically every lab does it.
x.com/ChangJonathanC…
Jonathan Chang@ChangJonathanC
I recently learned that @AnthropicAI discards thinking tokens in a multi-turn conversion. at the first glance it seems weird to remove part of the "context" that was causal to the final response. maybe they are using this as some form of context-distillation to improve the model output when no thinking is used, or simply to conserve the context window but I don't like it. sometimes it thinks a lot and output a short answer, and when I ask another question that it already thought about in the previous thinking chunk, it doesn't have access to it and can't answer directly. I have to ask it to include the thought process in the response. however, I found adding a fake reasoning budget tag at the end of response would trigger it to think longer, without being identified as "extended thinking", and I get to preserve the whole cot.
English
in hindsight, this codex cli pr over a month ago already hinted at the persisted reasoning support
github.com/openai/codex/p…
Jonathan Chang@ChangJonathanC
GPT 5.6 supports persisted reasoning this should solve the reasoning sink issue
English
sol None is more expensive than sol low and get slightly better result on browserComp
English
> Computer Use is not allowed to use the app 'com.openai.codex' for safety reasons.
but if you use codex cli to do it, gpt 5.6 is smart enough to bypass it
English
Deepmind folded all the proteins in the world, why don't OpenAI solve all Erdos problems? @thsottiaux
English
vibe proofing with gpt 5.6
Graffiti.pc Conjecture 143:
gist.github.com/changjonathanc…
English
@xeophon i think o3 and gpt models use generic code exec tool in chatgpt, not some special slicing / zoom tools.
i don't see any hidden tool from my results
English
@ChangJonathanC fun, smol test
but dunno whether the models behind apis still have their slicing / zoom tools, openai had them in o3 days iirc
English
