Tilmaamo cusub oo AI ah oo ka soo baxday qofka ka dhigay ChatGPT, waxay ku faraxsan yihiin baraha dhaqaalaha
Diogo Almeida, oo hore u ahaa macallin ka tirsan OpenAI oo ka qayb galay dhismaha ChatGPT, wuxuu maanta soo saaray nidaam cusub oo loo yaqaan Jev, kaas oo aan ahayn model aad u weyn oo hadal soo saara, balse wuxuu bixiyaa xaqiiqooyin iyo go'aanno fudud oo ku habboon shaqooyinka kombiyuutarka.
Jev: Model aan hadal soo saarin, balse bixiya go'aanno fudud
TypeSafe AI, shirkad uu sameeyay Diogo Almeida, waxay soo saartay model cusub oo loo yaqaan Jev, kaas oo aan ahayn model hadal soo saara (LLM), balse wuxuu bixiyaa xaqiiqooyin iyo go'aanno fudud oo ku habboon shaqooyinka kombiyuutarka. Modelkan wuxuu isticmaala qalabka transformer, kaas oo ka duwan kuwa kale ee AI ah.
Jev wuxuu bixiyaa xaqiiqooyin iyo go'aanno fudud oo ku habboon shaqooyinka kombiyuutarka, waxaana loo yaqaannaa 'calibrated decisions.' Wuxuu sidoo kale bixiyaa xaqiiqooyin iyo go'aanno fudud oo ku habboon shaqooyinka kombiyuutarka, taas oo ka dhigaysa mid aad u fudud in la isticmaalo.
Faa'iidooyinka Jev: Sare u socodka, la'aanta khaladaadka, iyo qiimo jaban
Jev wuxuu leeyahay faa'iidooyin badan, kaas oo ka mid ah inuu bixiyo xaqiiqooyin iyo go'aanno fudud oo ku habboon shaqooyinka kombiyuutarka, waxaana loo yaqaannaa 'calibrated decisions.' Wuxuu sidoo kale bixiyaa xaqiiqooyin iyo go'aanno fudud oo ku habboon shaqooyinka kombiyuutarka, taas oo ka dhigaysa mid aad u fudud in la isticmaalo.
Shirkadda Vercel waxay ku beddeshay modelka ChatGPT Luna 5.6 modelka Jev, waxayna ka heshay natiijooyin 5 ilaa 18 jeer oo ku dhawaaqay, iyadoo sidoo kale si fiican uga wanaagsan. Sidoo kale, shirkadda Bryo AI waxay tijaabisay Jev iyadoo ku xiran Gemini, waxayna ka heshay natiijooyin 10 ilaa 20 jeer oo ku dhawaaqay.
Jev wuxuu kordhin karaa modelada kale, sidoo kale wuxuu u baahan yahay in la isticmaalo
Jev wuxuu kordhin karaa modelada kale, sidoo kale wuxuu u baahan yahay in la isticmaalo, waxaana loo isticmaali karaa in lagu xakameeyo modelada kale, sida in lagu xakameeyo modelada kale, iyo in lagu xakameeyo modelada kale.
Modelkan wuxuu sidoo kale u isticmaali karaa in lagu xakameeyo modelada kale, sida in lagu xakameeyo modelada kale, iyo in lagu xakameeyo modelada kale. Wuxuu sidoo kale u isticmaali karaa in lagu xakameeyo modelada kale, sida in lagu xakameeyo modelada kale, iyo in lagu xakameeyo modelada kale.
Diogo Almeida, the researcher who co-invented reinforcement learning from human feedback at OpenAI, has launched TypeSafe AI and released Jev, a transformer-based model that outputs calibrated probabilities rather than text, making it cheaper and faster for automation tasks.
A New Kind of AI Model
Almeida left OpenAI two years ago to found TypeSafe AI, aiming to solve what he described as the problem of optimizing for human language when computers speak a different one. The company released Jev, a transformer-based model that does not generate text but instead produces probabilities or calibrated decisions. Output tokens are free, while input tokens are metered by the billion rather than the million, making the model significantly cheaper to run.
Jev cannot hallucinate because users define the outputs in advance, and it provides confidence scores that are real probabilities, making it ideal for automating workflows. Almeida stated that the model is trained exclusively on synthetic data using a technique called reinforcement learning from calibrated decisions, and he called it one of the best bets he has ever made in his life.
High Demand and Real-World Adoption
Developers are taking a great interest in the product, and TypeSafe AI briefly lost the ability to serve users from its API because demand was so high. Vercel replaced OpenAI's ChatGPT Luna 5.6 with Jev for a classifier to review commands for safety, getting results 5 to 18 times more quickly and with greater accuracy. Bryo AI CTO Nikhil Mudholkar tested Jev against Gemini for classifying business emails, finding Gemini slightly more accurate but 10 to 20 times more expensive.
Mudholkar noted that Jev is the only model that hands back a real probability, which makes it ideal for automating workflows. Armin Ronacher, CTO of Earendil, explained that Jev delegates the hallucination problem to the user, who can set thresholds like 50% or 95% probability to decide whether to act on the output.
Future Potential and Broader Vision
Jev can augment LLMs by acting as a smart check on misbehavior, such as tracking LLM agent traces and preventing jailbreaks. It can also be used for model routing, predicting whether a given workload requires a specific model, which would be expensive if done with an LLM. The model is named after William Stanley Jevons, the 19th-century economist whose paradox describes how the falling cost of a commodity can lead to it being used more and more.
Almeida hopes that Jev will lead to smart software all over the place in a way that is emergent and distributed, much more like the early internet than the mega apps people are trying to build. He stated that TypeSafe AI is not a frontier lab because the main product of frontier labs is fear or hype, and he would like their main product to be intelligence.
Ilaha iyo xuquuqda sawirka
Sawir: TechCrunch Xigasho
