OpenAI Stops Advancing new model
Sharaxaad
Videogan waxaan kaga hadlayaa OpenAI oo ku dhawaaqday inay si ku meel gaar ah u joojisay training-ka models-ka, ugu yaraanna laba isbuuc oo kale. Sam Altman wuxuu qoray inay joojinayaan RL-ka oo ah reinforcement learning, taasoo ah marxaladda model-ku soo maro kaddib pre-training-ka. Reinforcement-ka ayaa model-ka lagu baraa qaabka uu u jawaabo, qaabka uu wax u qoro, dabeecadda iyo role-ka uu yeelanayo, meeshaasna waa meesha shirkaduhu isku kala baxaan. Sababta ay sheegeen waa safety-ga iyo security-ga, maadaama ay soo wadaan model cusub oo la yiraahdo Astra oo qaybtaas aad ugu xooggan.
Waxaan sidoo kale sheegayaa in laba isbuuc aan yarayn, maadaama reinforcement-ka ay ku lug leeyihiin dad badan oo model-ka jiheeya. Shirkado kale ayaa la sheegaa inay isla arrintaas samaynayaan, Anthropic-na waa kuwa ugu buunbuuniya in xoogaa slowdown lagu sameeyo. Dhinaca kalese way adag tahay in shirkadaha Maraykanku speed-ka gaabiyaan, maadaama shirkadaha Shiinaha maalin kasta wax cusub soo saarayaan. Qwen 3.8 ayaa soo baxay isagoo uncensored ah, macnaheeduna waa in model-ku aanu su'aalahaaga kaa diidin. Waxaan ku gunaanadayaa inaan wada arki doonno halka arrintu ku dhammaanayso.
English summary
A four-minute Somali-language (Af-Soomaali) AI news video from Dugsiiye on OpenAI pausing the training of its new models. Sam Altman said the company is stopping reinforcement learning for at least two more weeks. The video explains where RL sits: pre-training comes first, then reinforcement learning shapes how a model answers, how it writes, the character and the role it takes, and that is where the labs differentiate themselves. It covers the stated reason, safety and security ahead of a coming model called Astra, why two weeks is a long pause when so much of RL is human work, and why American labs find it hard to slow down while Chinese labs ship something new every day, including the uncensored Qwen 3.8. Free on YouTube.
Waxaad ku baranaysaa
- Sababta OpenAI ay training-ka models-ka u joojisay
- Meesha RL-ku ka galo: pre-training iyo reinforcement learning
- Waxa reinforcement-ka lagu baro model-ka: jawaab, dabeecad iyo role
- Safety-ga, security-ga iyo model-ka cusub ee Astra
- Anthropic iyo doodda slowdown-ka ee shirkadaha AI-ga
- Tartanka shirkadaha Shiinaha iyo Qwen 3.8 oo uncensored ah
Cutubyada
Su'aalo
- Muxuu OpenAI joojiyay?
- Waxay si ku meel gaar ah u joojisay training-ka models-ka, gaar ahaan RL-ka oo ah reinforcement learning-ka. Sam Altman wuxuu qoray inay ugu yaraan laba isbuuc oo kale sii joojinayaan.
- Waa maxay sababta ay sheegeen?
- Waxay leeyihiin safety-ga iyo security-ga ayaan tixgelinaynaa. Waxay soo wadaan model cusub oo la yiraahdo Astra oo qaybta security-ga aad ugu xooggan, videoga waxaanan ku xusayaa model-kii Hugging Face ee sandbox-ka ka baxsaday.
- Model uncensored ah waa maxay?
- Waa model aan waxba kaa diidin: waxaad u dirto wuu qabanayaa. Models-ka kale waxaa jira su'aalo ay kugu yiraahdaan taas ma jawaabi karo, taasna waxaa lagu baray reinforcement-ka. Qwen 3.8 oo dhinaca Shiinaha ka soo baxay ayaa videoga tusaale loogu soo qaatay.
