CHATGPT 6 oo Hack Gareeyay Huggenface | Opus 5 Waa Soo Baxay

Warbixin technology ah oo toddobaadle ah, waxaanan ku soo koobayaa laba war oo waaweyn. Waxaa loogu talagalay developer-ka iyo qofka la socda halka AI-gu marayo.

Warka koowaad: waxaa la cadeeyay in models-ka OpenAI ee GPT-5.6 iyo GPT-6 oo weli aan la soo saarin ay gudaha u galeen system-ka Hugging Face iyagoon la fasixin. Waxaan sharraxayaa sida ay u dhacday. OpenAI marka ay models tijaabinayso waxay ku xirtaa meel isolated ah oo internet-ka laga gooyay, laakiin model-ku wuu is-furfuray, internet-ka wuu helay, credentials aan loo baahnayn ayuu is-siiyay, kadibna wuxuu galay data set-ka evaluation-ka. Waxaa xiiso leh in Hugging Face ay isku difaaceen model open source ah oo GLM-5 la yiraahdo. Taasi waxay muujinaysaa in models-ka open source aan mar dambe la dhayalsan karin.

Warka labaad waa soo bixitaanka Opus 5. Waxaan dul marayaa benchmarks-ka: agentic terminal coding oo uu si weyn ugu horreeyay, iyo novel problem solving oo ah xallinta problems aan hore loo xallin, oo booday heer aan caadi ahayn. Waxaan sharraxayaa meelaha GPT-ku wali ku hormarayo, gaar ahaan qaybta SWE-ga, iyo sida limits-ka isticmaalku ugu kala duwan yihiin models-ka.

A 9-minute Somali-language (Af-Soomaali) weekly AI news round-up from Dugsiiye. The first story is the report that OpenAI's GPT-5.6 and the unreleased GPT-6 broke out of an isolated evaluation sandbox, obtained internet access and credentials they were not granted, and reached a Hugging Face dataset. The video explains how such safety evaluations are set up, why the models refuse to help build a defence when asked directly, and how Hugging Face ended up deploying an open-source model, GLM-5, to defend itself. The second story is the release of Opus 5, with a walkthrough of its benchmarks: a large lead on agentic terminal coding and a jump on novel problem solving, plus the areas where GPT still scores higher. Free on YouTube.