Ali Nasiri-Sarvi

[Ali Nasiri-Sarvi @anasirisarvi]

@giffmana @JFPuget @RheaSukthanker @CameronPashmina @Emy_Aze Retrofitting.

[Ali Nasiri-Sarvi @anasirisarvi]

@giffmana @JFPuget @RheaSukthanker @CameronPashmina @Emy_Aze Even calling it post training would be misleading.

[Ali Nasiri-Sarvi @anasirisarvi]

@Miles_Brundage I found it as good as the demo for English/French translation.

Also, asked it to count to 50 and don’t stop no matter what. Ignored a dozen stop requests.

[Ali Nasiri-Sarvi @anasirisarvi]

@JinjingLiang They mentioned this

[Ali Nasiri-Sarvi @anasirisarvi]

@_arohan_ It switches (ed) to Opus when asking to read some code with simple fsdp and kernels.

[Ali Nasiri-Sarvi @anasirisarvi]

@Yuchenj_UW @ThePremiseOfIt It's live in claude code at least

[Ali Nasiri-Sarvi @anasirisarvi]

@livgorton I respect that.

[Ali Nasiri-Sarvi @anasirisarvi]

@marksaroufim How can we enforce a new license? They don't even respect the current copyright laws.

[Ali Nasiri-Sarvi @anasirisarvi]

@m_sirovatka Just check this one

https://t.co/BDZEKVoIbg

[Ali Nasiri-Sarvi @anasirisarvi]

@iScienceLuvr Thanks for sharing our work!

[Ali Nasiri-Sarvi @anasirisarvi]

@WenhuChen You have to clear your browser's cookies for their website.

[Ali Nasiri-Sarvi @anasirisarvi]

@roydanroy @kimmonismus To be fair, their results aren't zero-shot and should be compared to SFT.

[Ali Nasiri-Sarvi @anasirisarvi]

@ehsanik @Vercept_ai @inkynumbers Have you tried a calculator app that gives the wrong answer?

[Ali Nasiri-Sarvi @anasirisarvi]

@xwang_lk Previous models had been exceptional at similar type of tasks. Your example, however, is much harder.

https://t.co/UNfw1IHgkC

[Dimitris Papailiopoulos @DimitrisPapail]

Inspired by @distributionat I tested the geolocation capabilities of gpt4o.

uhm, they are INSANE!!!!!!

here are examples. ALL CORRECT. These are all photos of mine, not in the public domain,

1. San Sebastian, Spain
2. Chalkida, Evoia, Greece
3. Koufonisia, Cyclades, Greece
4. Kedrodassos, Crete, Greece

[Ali Nasiri-Sarvi @anasirisarvi]

@giffmana Just curious if your previous view on image resolution still holds true given the new tool usage of the o-series models?

https://t.co/9wCUzzVT5Q

[Ali Nasiri-Sarvi @anasirisarvi]

@giffmana @SakanaAILabs Not being skeptical of 150x speed up over the baseline is certainly an issue.

[Ali Nasiri-Sarvi @anasirisarvi]

@DhruvBatraDB Thanks for the explanation. On this account, do you think a task like counting number of people in a crowded scene would rely on this type of reasoning?

[Ali Nasiri-Sarvi @anasirisarvi]

@DhruvBatraDB What about questions like occlusion in object tracking? Could "thinking" help tracking in the occluded frames?

[Ali Nasiri-Sarvi @anasirisarvi]

@sea_snell R1 CoT clearly shows this