- Microsoft imakhazikitsa Phi-4-multimodal, mtundu wa AI womwe umasinthira mawu, zithunzi ndi zolemba nthawi imodzi.
- Ndi magawo mabiliyoni a 5.600, imachita bwino kuposa mitundu yayikulu pamawu komanso kuzindikira masomphenya.
- Mulinso Phi-4-mini, mtundu womwe umayang'ana kwambiri ntchito zosinthira mawu.
- Imapezeka pa Azure AI Foundry, Hugging Face, ndi NVIDIA, yokhala ndi mapulogalamu osiyanasiyana mubizinesi ndi maphunziro.
Microsoft yapita patsogolo m'dziko lamitundu yamitundu yokhala ndi ma multimodal Phi-4, nzeru zake zaposachedwa kwambiri komanso zapamwamba kwambiri zomwe zimatha kupanga nthawi imodzi zolemba, zithunzi ndi mawu. Chitsanzochi, pamodzi ndi Phi-4-mini, chikuyimira a Chisinthiko mu mphamvu ya zitsanzo zazing'ono (SLM), yopereka mphamvu komanso yolondola popanda kufunikira kwa magawo ambiri.
Kufika kwa Phi-4-multimodal sikungoyimira kusintha kwaukadaulo kwa Microsoft, komanso Imapikisana mwachindunji ndi mitundu yayikulu monga ya Google ndi Anthropic. Zomangamanga zake zokongoletsedwa ndi luso lapamwamba la kulingalira zimapangitsa njira yokongola yamapulogalamu angapo, kuchokera pamakina omasulira kupita ku chithunzi ndi kuzindikira mawu.
Kodi Phi-4-multimodal ndi chiyani ndipo imagwira ntchito bwanji?
Phi-4-multimodal ndi mtundu wa AI wopangidwa ndi Microsoft womwe ungasinthe nthawi imodzi mawu, zithunzi ndi mawu. Mosiyana ndi zitsanzo zachikhalidwe zomwe zimagwira ntchito ndi njira imodzi, nzeru zopangira izi zimagwirizanitsa magwero osiyanasiyana a chidziwitso mu malo amodzi oyimira, chifukwa chogwiritsa ntchito njira zophunzirira.
Chitsanzocho chimamangidwa pa zomangamanga za 5.600 biliyoni parameters, pogwiritsa ntchito njira yotchedwa LoRAs (Low-Rank Adaptations) kuti aphatikize mitundu yosiyanasiyana ya deta. Izi zimalola kulondola kokulirapo pakukonza chilankhulo komanso kumasulira mozama nkhani.
Mphamvu zazikulu ndi zopindulitsa
Phi-4-multimodal imakhala yothandiza kwambiri pazinthu zingapo zofunika zomwe zimafuna luntha lapamwamba kwambiri:
- Kuzindikira Kulankhula: Imachita bwino kuposa zitsanzo zapadera monga WhisperV3 pamayesero omasulira ndi makina omasulira.
- Kukonza zithunzi: Imatha kutanthauzira zikalata, zithunzi ndikuchita OCR molondola kwambiri.
- Low Latency Inference: Izi zimalola kuti igwiritse ntchito pazida zam'manja komanso zotsika mphamvu popanda kuchita zambiri.
- Kuphatikizana kosagwirizana pakati pa njira: Kukhoza kwawo kumvetsetsa zolemba, zolankhula ndi zithunzi pamodzi kumawongolera malingaliro awo.
Poyerekeza ndi zitsanzo zina
Pankhani ya magwiridwe antchito, Phi-4-multimodal yatsimikizira kuti ikugwirizana ndi mitundu yayikulu. Poyerekeza ndi Gemini-2-Flash-lite ndi Claude-3.5-Sonnet, imakwaniritsa zotsatira zofanana ndi ntchito za multimodal, ndikusunga bwino kwambiri chifukwa cha kapangidwe kake kakang'ono.
Komabe, amapereka malire ena pa mafunso ozikidwa pa mawu ndi mayankho, kumene zitsanzo monga GPT-4o ndi Gemini-2.0-Flash zili ndi mwayi. Izi ndichifukwa chakuchepa kwake kwachitsanzo, zomwe zimakhudza kusunga chidziwitso chenicheni. Microsoft yawonetsa kuti ikuyesetsa kukonza izi m'mitundu yamtsogolo.
Phi-4-mini: mchimwene wake wa Phi-4-multimodal
Pamodzi ndi Phi-4-multimodal, Microsoft yakhazikitsanso Phi-4 mini, chosinthika chokongoletsedwa ndi ntchito zozikidwa pamawu. Chitsanzochi chapangidwa kuti chipereke Kuchita bwino kwambiri pakukonza chilankhulo chachilengedwe, kupangitsa kuti ikhale yabwino kwa ma chatbots, othandizira enieni, ndi mapulogalamu ena omwe amafunikira kumvetsetsa kolondola komanso kupanga zolemba.
Kupezeka ndi kugwiritsa ntchito
Microsoft yapangitsa Phi-4-multimodal ndi Phi-4-mini kupezeka kwa opanga kudzera Azure AI Foundry, Nkhope Yokumbatira, ndi Catalog ya NVIDIA API. Izi zikutanthauza kuti kampani iliyonse kapena wogwiritsa ntchito omwe ali ndi mwayi wopeza nsanjazi akhoza kuyamba kuyesa chitsanzocho ndikuchigwiritsa ntchito pazochitika zosiyanasiyana.
Potengera njira yake yophatikizira, Phi-4 ndi Cholinga cha magawo monga:
- Kumasulira kwamakina ndi mawu ang'onoang'ono munthawi yeniyeni.
- Kuzindikira zolemba ndi kusanthula kwamabizinesi.
- Mapulogalamu am'manja okhala ndi othandizira anzeru.
- Mitundu yamaphunziro yopititsa patsogolo maphunziro a AI.
Microsoft yapereka a kupotoza kosangalatsa ndi zitsanzo izi poyang'ana pakuchita bwino komanso scalability. Ndi mpikisano wochulukirachulukira pamitundu yaying'ono yazilankhulo (SLM), Phi-4-multimodal imaperekedwa ngati njira yotheka kumitundu yayikulu, kupereka malire pakati pa magwiridwe antchito ndi mphamvu yopangira kupezeka ngakhale pazida zopanda mphamvu.
Ndine wokonda zaukadaulo yemwe wasandutsa zokonda zake za "geek" kukhala ntchito. Ndakhala zaka zoposa 10 za moyo wanga ndikugwiritsa ntchito ukadaulo wotsogola komanso kusewera ndi mitundu yonse yamapulogalamu chifukwa cha chidwi chenicheni. Panopa ndaphunzira luso la pakompyuta komanso masewera a pakompyuta. Izi zili choncho chifukwa kwa zaka zoposa 5 ndakhala ndikulembera mawebusaiti osiyanasiyana pa teknoloji ndi masewera a pakompyuta, ndikupanga nkhani zomwe zimafuna kukupatsani zambiri zomwe mukufunikira m'chinenero chomwe chimamveka kwa aliyense.
Ngati muli ndi mafunso, chidziwitso changa chimachokera ku chilichonse chokhudzana ndi makina opangira Windows komanso Android pama foni am'manja. Ndipo kudzipereka kwanga ndi kwa inu, nthawi zonse ndimakhala wokonzeka kuthera mphindi zochepa ndikukuthandizani kuthetsa mafunso aliwonse omwe mungakhale nawo pa intaneti.