Let's Talk About Jev

Videogan waxaan kaga hadlayaa Jev, model cusub oo ay soo saartay shirkadda TypeSafe AI. Model-kani ma aha large language models-ka aan naqaanno sida GPT iyo Claude, waxaana loogu talagalay classification. Wuxuu u shaqeeyaa sida System One: qaybta maskaxda ee jawaabta isla markiiba soo tuurta adigoon ka fikirin. Qoraal token token uma soo saaro, balse su'aashaada iyo jawaabaha aad sii diyaarisay ayuu hal mar ku process-gareeyaa, kadibna wuxuu kuu soo celiyaa go'aan leh probability iyo confidence. LLM-yadu waa undeterministic, Jev-na jawaabtiisu waa la sii ogyahay, hallucination-na ma sameeyo.

Waxaan sharxayaa saddexda nooc ee aad jawaabta u diyaarin karto: choice, sida customer billing-ka loo diro; score, sida 1 ilaa 10; iyo null oo ah true ama false. Tijaabo 27 su'aalood ah, Jev wuxuu ku dhammeeyay 0.114 ilbiriqsi, GPT-na wuxuu qaatay ku dhawaad siddeed ilbiriqsi iyo badh. TypeSafe waxay sheegtay in input-ku yahay $0.042 halkii million token, output-kuna bilaash yahay, halka ay Fable input-ka ka qaaddo $10, output-kana $50. Kadib waxaan tijaabiyay demo home automation ah ee codka lagu maamulo oo uu Claude end-to-end u dhisay. Markaan dhaho "turn off all the lights" ama "open bedroom one door", isla markiiba wuxuu gartaa nalka ama albaabka loo baahan yahay.

Aniga waxaan u arkaa wax amazing ah. Jev wuxuu ku fiican yahay go'aannada degdegga ah sida spam-ka, comments-ka YouTube iyo qalabka IoT, xisaabtase wuu ku xun yahay. Ii sheega sida aad u aragteen, application-kana GitHub ayaan soo gelinayaa, link-iguna wuxuu ku jiri doonaa description-ka.

A thirteen-minute Somali-language (Af-Soomaali) AI news and demo video from Dugsiiye on Jev, the new model from TypeSafe AI. Mohamed explains that Jev is not a large language model like GPT or Claude but a System One classification model: instead of generating text token by token, it returns a typed, deterministic decision with probability and confidence scores, so it does not hallucinate. He walks through the three answer types you prepare in advance (choice, score and true/false), a speed test where Jev sorts 27 questions in about a tenth of a second while GPT takes over eight seconds, and TypeSafe's pricing of $0.042 per million input tokens with free output. He then demos a voice-controlled home-automation app that switches lights and opens doors instantly, and closes on where classification fits and where it does not, such as math. Free on YouTube.