big goose

Noob,NLP

[big goose @Goose_Doro]

@willcb infra is all u need, What about post-training fine-tuning then? It seems like that is also needed. I can only say it is slightly easier

[big goose @Goose_Doro]

@chetaslua I have encountered this before, but after launch, the speed is at the standard rate, 50+ tokens/s. It seems there is a slight leak on the frontend, but the OpenAI backend has not opened it up

[big goose @Goose_Doro]

Today's GPT-5.6 Sol is shockingly stupid. Even my grandmother is smarter than it is.

[big goose @Goose_Doro]

When looking at publicly released benchmark scores for this kind of model, especially when there are very few public benchmark scores, do not focus on which benchmarks they published scores for. Instead, look at which benchmarks they did not publish scores for. For example, with terminal bench, why only publish 2.1 and not 4.0? And with Longcontext, why is there only MRCR and nothing else? The only explanation is that the scores are bad.

[big goose @Goose_Doro]

@latepostnews 我猜应该是哪个大说哪个。应该是后者，

[big goose @Goose_Doro]

@HJain13 @Lentils80 hack issue

[big goose @Goose_Doro]

@datacurve that's mean your Bech need a big update

[big goose @Goose_Doro]

Nah bro, that’s not how you evaluate new models.
Don’t look at the benchmark scores they show off. Look at the major, recent benchmarks they conveniently leave out.
Take Terminal-Bench for example: why test on 2.1 instead of the latest 4.0? All I can say is that the latest, unhacked Terminal-Bench 4.0 results are just too embarrassing to look at.

[big goose @Goose_Doro]

@sheriyuo @liulicheng10 来了吗，如来

[big goose @Goose_Doro]

@synthwavedd Why not GPT-5.6 Astra? After all, they are both stars.

[big goose @Goose_Doro]

@eliebakouch so maybe now GPT model is close to 200？

[big goose @Goose_Doro]

@merettm “The depth of the computation graph”, does this refer to the number of model layers?

[big goose @Goose_Doro]

@omouamoua just update rightnow

[big goose @Goose_Doro]

@theo But as an ordinary subscriber, the cache read price cut has little to do with us. Still, a price cut is always a good thing.

[big goose @Goose_Doro]

@xxzengyibuke @VersunPan 实际提供的推理成本远小于API定价， ollama有自部署的

[big goose @Goose_Doro]

@nickbaumann_ The magical remote compression trick of Codex

[big goose @Goose_Doro]

@feitong_yang @vicapow 还有个小问题，后续GPT的模型更新了，prism中提供的模型也会更新吗？

[big goose @Goose_Doro]

@feitong_yang @vicapow 能否来点更加明确的提示？ 现在确实都不用overleaf了，在猛猛用Prism。

[big goose @Goose_Doro]

@nateberkopec A wide variety of needs. Some are enterprises, some are concerned with data compliance, and others want to use the models somewhere other than Claude Web or Claude Code.

[熊师傅 weight decay 了吗 @bigeagle_xd]

they could have published a nice report on architecture modification during post/continual training, but they chose to mislead the community with a clickbait title.

I saw it yesterday and thought they were unprofessional, but found out today it was all intentional.

> That would have cut our visibility by 90%. The fact is, you saw the title, you engaged with it. It worked its purpose.

i doubt if they really want to make contributions, but i'm sure they want to be known.

let's check the author list, @jm_alexia @RheaSukthanker @CameronPashmina @Emy_Aze, they wanted to be famous, but instead, they made themselves infamous.