Anthropic oo ka shaqaynaysa nidaam furan oo lagu caddeeyo aqoonsiga moodeelka AI

Nidaamka FlashProve waxa uu higsanayaa in uu bixiyo caddayn cryptographic ah oo muujinaysa in natiijadu ka timid moodeel gaar ah.

Anthropic waxa ay ka shaqaynaysa nidaam open source ah oo la yiraahdo FlashProve, kaas oo loogu talagalay in lagu xaqiijiyo moodeelka AI ee soo saaray natiijada la helay, sida lagu sheegay qoraal blog ah oo la daabacay Talaado.

Caddaynta 'provable inference' iyo ka-hortagga khiyaanada

Nidaamka FlashProve waxa uu bixiyaa waxa cilmi-baarayaashu u yaqaaniin 'provable inference' — caddayn cryptographic ah, oo ah dammaanad xisaabeed, in natiijadu ka timid moodeel gaar ah. Waxa uu u ogolaanayaa isticmaalayaasha in ay xaqiijiyaan in jawaabta ay heleen ay dhab ahaan ka timid moodeelka ay u malaynayaan in ay isticmaalayaan.

Sida saxeexyada iimaylka ee cryptographic, FlashProve waxa uu ka hortagi karaa kuwa isku dayaya in ay moodeel u ekaysiiyaan mid kale, xitaa marka khiyaanada ay ka timaado shaybaadhka dhisay oo maamula moodeelka. Cilmi-baarayaashu waxa ay sheegeen in nidaamku ka hortagi karo in shirkad AI si qarsoodi ah u beddesho moodeel jaban si ay kharashka u yarayso, ama mid qaali ah si ay natiijooyinkeeda waxqabadka u kordhiso.

Fingerprint-ka moodeelka iyo xaqiijinta xogta

Si nidaamku u shaqeeyo, shaybaadhka AI waxa ay tahay in ay daabacaan 'fingerprint' moodeel kasta oo ay soo saaraan, kaas oo markii dambe lagu barbardhigi karo natiijooyin gaar ah oo si cryptographic ah loo xaqiijin karo. Fingerprint-ku ma muujiyo miisaanka moodeelka — tirooyinka gudaha ee qeexaya hab-dhaqankiisa, kuwaas oo shirkadaha moodeelka xiran sida Anthropic iyo OpenAI ay si gaar ah u hayaan.

Taasi waxa ay ka dhigan tahay in shaybaadhku ay daabici karaan fingerprint-ka iyaga oo aan bixin hantidooda ugu qiimaha badan, taas oo u ogolaanaysa nidaamka in uu la shaqeeyo moodeelada ganacsi sidoo kale. Nidaamka waxa loo qaabeeyay in uu fudud yahay, si xaqiijintu uga dhacdo qalab maxalli ah. FlashProve waxa kale oo loo isticmaali karaa in lagu xaqiijiyo sheegashooyinka dhinacyada saddexaad ee ah in hawsha gaar ah ay qabatay moodeel horumarsan.

Koodka furan iyo xaaladda cilmi-baarista

Qayb ka mid ah ku dhawaaqista, cilmi-baarayaashu waxa ay sii daayeen faahfaahin farsamo oo buuxda iyo kood shaqaynaya oo qaybo kala duwan oo nidaamka ah, kaas oo si cad loo heli karo iyada oo hoos timaadda shatiga open source ee GitHub.

Cilmi-baarista nidaamka ayaa socota, mana jiraan qorshayaal lagu geynayo Anthropic ama shaybaadh kale. Cilmi-baarayaashu waxa ay sharraxeen: 'Qoraalkan cilmi-baarista waxa uu ujeeddadiisu tahay in la wadaago horumar dhexdhexaad ah, ma aha ajandaha cilmi-baaris oo dhamaystiran. Waxa aan tan si cad u wadaagnaa iyada oo ujeedadu tahay in dadka kale ay eegaan waxa aan qabanay, qaataan, oo ay horumariyaan.'

The open source research system aims to give users cryptographic proof of a model's identity, without revealing its private weights.

Anthropic is working on an open source system called FlashProve to verify which AI model produced a given output, according to a blog post published Tuesday. The system is built to provide what researchers call "provable inference" — a cryptographic proof that a given output came from a specific model.

How FlashProve Would Confirm a Model's Identity

FlashProve is meant to let users confirm that the answer they received really came from the model they think they are using. It would be robust against impostors attempting to disguise one model as another, even when the impostor is the lab that built and runs the model. The researchers note the system could prevent an AI company from secretly swapping in a cheaper model to cut costs, or a more expensive model to inflate its performance results.

Fingerprints Without Exposing Private Weights

To work, the system requires AI labs to publish a "fingerprint" of each model they release that can later be matched to specific outputs and verified cryptographically. Crucially, the fingerprints do not reveal a model's weights, the internal numbers that define how it behaves and which closed-model makers like Anthropic and OpenAI keep private. The system is also designed to be lightweight, so verification could be performed on a local device.

Open Source Code, With Research Ongoing

Researchers released a detailed technical specification and working code for various elements of the system, which is publicly available under an open source license on GitHub. Research in the system is ongoing, and there are no plans to deploy it at Anthropic or any other lab. "This research post is meant to share intermediate progress, and not a completed research agenda," researchers explain in the announcement post. "We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it."