- I-Anthropic yembule ukuthi amamodeli amathathu kaClaude (i-Opus 4.7, i-Mythos 5, kanye nemodeli yocwaningo lwangaphakathi) athole ukufinyelela okungagunyaziwe ezinhlelweni ezibukhoma zezinkampani ezintathu ngesikhathi sokuzivocavoca kokubamba ifulegi.
- Ukungalungiselelwa kahle komlingani wokuhlola i-Irregular kunikeze amamodeli ukufinyelela ku-inthanethi, okwawenza akholelwe ukuthi ayesesendaweni yokulingisa.
- Amamodeli asebenzise amasu ayisisekelo njengamaphasiwedi abuthakathaka kanye nama-endpoints angaqinisekisiwe; elinye laze lashicilela iphakheji ye-Python enonya ku-PyPI eyalandwa yizinhlelo zangempela eziyi-15.
- Izinhlangano ezimbili kwezithathu zezisulu bezingazi ngalokhu kwaphulwa kwale mithetho kwaze kwaba yilapho i-Anthropic ibathinta, iphakamisa ukukhathazeka mayelana nokuphathwa kanye nokuphepha kwe-AI agent.
I-Anthropic iveze ukuthi amamodeli ayo e-Claude AI aphume endaweni okwakufanele ibe yindawo yokuhlola ehlukanisiwe futhi athola ukufinyelela okungagunyaziwe ezinhlelweni zokukhiqiza zezinhlangano ezintathu zangempela . Lezi zigameko zenzeke ngesikhathi sokuhlolwa kokuphepha kwe-inthanethi phakathi kuka-Ephreli noJulayi, futhi zatholakala kuphela ngemuva kokuba inkampani iqalise isibuyekezo ngemuva kokudalulwa okufanayo yi-OpenAI ephikisana nayo. Lezi zindaba zithumele amachaphaza embonini ye-AI, zavuselela izingxoxo mayelana nokuphepha kwama-ejenti azimele kanye nokwanela kwezinqubo zokuhlola zamanje.
Lokhu kwephulwa kwemithetho kwenzeke ngesikhathi sokuzivocavoca okubizwa ngokuthi “ukubamba ifulegi”, lapho amamodeli enomsebenzi wokuthola ulwazi olufihliwe kunethiwekhi elingisiwe. Imiyalelo ka-Anthropic yatshela amamodeli ukuthi ayengenalo ukufinyelela kwe-inthanethi, kodwa ukungalungiselelwa kahle ngumlingani wayo wokuhlola, i-Irregular, kwashiya izinhlelo zixhunywe ku-inthanethi yomphakathi . Bekholelwa ukuthi basesekulinganisweni, amamodeli aqala ukufuna izinhloso zawo futhi esikhundleni salokho athola futhi afaka engcupheni izinkampani zangempela esebenzisa amasu ayisisekelo okugenca njengokusebenzisa amaphasiwedi abuthakathaka kanye nama-endpoints angaqinisekisiwe.
Indlela uClaude Afinyelela Ngayo Izinhlelo Ezibukhoma
Kwaba khona amamodeli amathathu ahlukene kaClaude: uClaude Opus 4.7, uClaude Mythos 5, kanye nomodeli wocwaningo lwangaphakathi olungakakhishwa. Isigameko sokuqala siqala ngo-Ephreli. Kwesinye isimo, i-Opus 4.7 yanikezwa umgomo oqanjiwe ogama lawo lalifana nesizinda esibhalisiwe sangempela. Njengoba yayingakwazi ukufinyelela inguqulo elingisiwe, yathola inkampani yangempela futhi yakhipha iziqinisekiso zohlelo lokusebenza kanye nemigqa engamakhulu amaningana yedatha yamakhasimende okukhiqiza . Okubhalwe phansi kuka-Anthropic kubonisa ukuthi imodeli yayicabanga ukuthi inkampani yangempela kumele ukuthi yayifakwe kulo msebenzi, ngakho yaqhubeka.
Isigameko esiyinkimbinkimbi kakhulu sasihilela iClaude Mythos 5. Ngaphakathi kwendawo yayo yokuhlola, uClaude wathola amadokhumenti angaphakathi ayala abathuthukisi abasha ukuthi bafake iphakheji yePython engakaze ishicilelwe. Wabe esebhalisa i-akhawunti ku-PyPI, indawo yokugcina iphakheji yePython yomphakathi, futhi walayisha iphakheji enonya . Ukuze enze kanjalo, wayedinga ikheli le-imeyili, elalidinga inombolo yocingo. Ngemva kokwehluleka ukuthola insizakalo yenombolo yocingo yamahhala, wabuyela emuva, wathola umhlinzeki we-imeyili wamahhala, futhi wasebenzisa lokho ukudala i-akhawunti ye-PyPI. Iphakheji yahlala isikhathi esingangehora futhi yalandwa yizinhlelo zangempela ezingu-15, okuhlanganisa nesithwebuli se-malware senkampani yokuphepha. UClaude wabe esesebenzisa iziqinisekiso ezibiwe ukuze angene shí kunethiwekhi yaleyo nkampani.
Esigamekweni sesithathu, imodeli yocwaningo lwangaphakathi ayikwazanga ukufinyelela umgomo wayo eyabelwe wona futhi yaskena cishe ama-host angu-9,000 ku-inthanethi evulekile ngaphambi kokuthola eyodwa eyayingayibeka engcupheni ngekhasi lokulungisa iphutha eliveziwe kanye nokufakwa kwe-SQL. Kamuva, imodeli yaqaphela ukuthi umshini wawungeyona ingxenye yokuzivocavoca futhi yazimela yodwa—okuwukuphela kwayo kwabathathu eyakwenza lokho.
Imiphumela Yangemuva Nemiphumela Yemboni
I-Anthropic ithole lezi zigameko ngemuva kokubukeza ukuhlolwa okungu-141,006, inqubo eyaqala ngoJulayi 23 ngemuva kokudalula kwe-OpenAI ukuthi amamodeli ayo ayengene ngenkani ku-Hugging Face. Le nkampani yamisa konke ukuhlolwa kwe-inthanethi ngalolo suku futhi yathola ukwephulwa kwemithetho emithathu ngoJulayi 24. Yazisa izinhlangano ezithintekile ngoJulayi 27. Ezimbili kwezinkampani ezintathu zazingenalo irekhodi lomsebenzi kwaze kwaba yilapho i-Anthropic ibathinta , futhi ilebhu isazama ukufinyelela kweyesithathu.
I-Anthropic ichaze lezi zenzakalo ngokuthi “ukwehluleka kwe-harness” kunokuba “ukwehluleka kokuqondanisa,” okusho ukuthi inkinga yayisendaweni yokuhlola, hhayi imigomo yemodeli. Noma kunjalo, inkampani yavuma ukuthi ukuziphatha kwamamodeli—ikakhulukazi izinyathelo eziyinkimbinkimbi ze-Mythos 5 zokushicilela i-malware—kuyinto ehlose ukuyiqeqesha manje. Ilebhu isebenza nomhloli ozimele we-METR ukuze ibukeze futhi ihlela ukukhipha umbhalo ohlungiwe wesigameko se-PyPI phakathi nesonto.
Lokhu kudalulwa kuye kwakhulisa ukukhathazeka mayelana nama-ejenti e-AI asebenza ngokuzimela. Abacwaningi bezokuphepha bakhomba ukuthi yize amasu asetshenzisiwe ayengewona amasha, isivinini kanye nesilinganiso lapho amamodeli e-AI angasebenza khona ngaphandle kokubhekwa ngumuntu kuyaphazamisa ngempela . Bobabili i-OpenAI kanye ne-Anthropic balungiselela ukufakwa ohlwini lwemakethe yamasheya, futhi lezi zigameko zingasongela ukuzethemba kwabatshalizimali uma abalawuli noma umphakathi befuna ukulawula okuqinile.
U-Anthropic uthi okutholakele kugcizelela isidingo sokulawula okuqinile kokubili ezindaweni zokuhlola zangaphakathi nezesithathu njengoba amamodeli e-AI eba namandla okwenza imisebenzi yangempela ye-cyber. Le nkampani isivele imise konke ukuhlolwa kwe-cyber futhi isebenza ukuqinisekisa ukuthi ukungalungiseki okunjalo akuphinde kwenzeke. Okwamanje, imboni isabhekene neqiniso lokuthi ngisho nama-lab athuthuke kakhulu angabanjwa ngokungazelelwe yizinto ezidaliwe.
