OML: Tsarin Rarraba Model na AI na Buɗe, Mai Samun Kuɗi, da Amincewa

1. Gabatarwa

Hankalin Wucin Gadi yana canza fagage da yawa daga injinan mutum-mutumi da wasan caca zuwa tunanin lissafi da gano magunguna. Fitowar ƙirar ƙira masu ƙarfi kamar jerin GPT, OpenAI o3, da DeepSeek R1 suna wakiltar wani lokaci mai muhimmanci a cikin iyawar AI. Duk da haka, tsarin rarraba model na AI na yanzu yana gabatar da rarrabuwar kawuna: ko dai an rufe model ɗin kuma an kulle su da API, suna sadaukar da bayyana gaskiya da aiwatarwa na gida, ko kuma ana rarraba su a buɗe, suna sadaukar da samun kuɗi da sarrafawa.

2. Matsalar Rarraba Asali

Yanayin rarraba AI a halin yanzu yana da manyan hanyoyi biyu masu karo da juna, kowannensu yana da iyakoki masu muhimmanci waɗanda ke hana ci gaban AI mai dorewa.

2.1 Sabis na API ɗin Rufe

Dandamali kamar GPT na OpenAI da Claude na Anthropic suna ci gaba da cikakken iko akan aiwatar da model ta hanyar API na jama'a. Yayin da ake ba da damar samun kuɗi da sarrafa amfani, wannan hanyar tana haifar da:

• Keɓancewa da halayen neman haya
• Matsalolin sirri masu muhimmanci
• Rashin sarrafa mai amfani da bayyana gaskiya
• Rashin iya tabbatar da halayen model ko tabbatar da sirrin bayanai

2.2 Rarraba Nauyi Buɗe

Dandamali kamar Hugging Face suna ba da damar rarraba model ba tare da iyaka ba, suna ba da bayyana gaskiya da aiwatarwa na gida amma suna sadaukar da:

• Ƙarfin samun kuɗi ga masu ƙira
• Sarrafa amfani da gwamnati
• Kariya daga cire model
• Ƙarfafa ci gaba mai dorewa

3. Ƙirar Tsarin OML

OML ya gabatar da wata hanyar asali wacce ke ba da damar rarraba model ɗin kyauta don aiwatarwa na gida yayin da ake ci gaba da izinin amfani da aka tilasta ta hanyar sirri.

3.1 Ma'anun Tsaro

Tsarin ya gabatar da muhimman kaddarorin tsaro guda biyu:

• Juriya Cire Model: Yana hana waɗanda ba su da izini cire da kuma kwafi aikin model na asali
• Juriya Ƙirƙirar Izinin Ƙarya: Yana tabbatar da cewa ba za a iya ƙirƙira izinin amfani ko kuma gurɓata su ba

3.2 Gine-ginen Fasaha

OML ya haɗu da tambarin model na asali na AI tare da hanyoyin tilasta tattalin arziki na sirri, yana ƙirƙirar hanyar haɗin gwiwa wacce ke amfani da duka hanyoyin sirri na asali da ƙarfafa tattalin arziki.

4. Aiwar da Fasaha

4.1 Tushen Lissafi

An gina garantin tsaro akan ingantaccen tushen lissafi. Ana iya tsara juriya cire model kamar haka:

$\Pr[\mathcal{A}(M') \rightarrow M] \leq \epsilon(\lambda)$

inda $\mathcal{A}$ shine maƙiyi, $M'$ shine model ɗin da aka kare, $M$ shine model ɗin asali, kuma $\epsilon(\lambda)$ wani aiki ne maras muhimmanci a cikin ma'aunin tsaro $\lambda$.

Tsarin izini yana amfani da sa hannun sirri:

$\sigma = \text{Sign}_{sk}(m || t || \text{nonce})$

inda $sk$ shine maɓalli na sirri, $m$ shine mai gano model, $t$ shine alamar lokaci, kuma nonce yana hana hare-haren maimaitawa.

4.2 Aiwatar da OML 1.0

Aiwatarwa ta haɗu da alamar ruwa na model tare da tilastawa tushen blockchain:

class OMLModel:
    def __init__(self, base_model, fingerprint_key):
        self.base_model = base_model
        self.fingerprint_key = fingerprint_key
        self.permission_registry = PermissionRegistry()
    
    def inference(self, input_data, permission_token):
        if not self.verify_permission(permission_token):
            raise PermissionError("Izinin ba shi da inganci ko kuma ya ƙare")
        
        # Saka tambari a cikin sakamako
        output = self.base_model(input_data)
        fingerprinted_output = self.embed_fingerprint(output)
        return fingerprinted_output
    
    def embed_fingerprint(self, output):
        # Aiwatar da tambarin asali na AI
        fingerprint = generate_fingerprint(output, self.fingerprint_key)
        return output + fingerprint

5. Sakamakon Gwaji

Ƙimar da aka yi da yawa ta nuna yuwuwar aiwatar da OML:

• Aikin Tsaro: An rage hare-haren cire model da kashi 98.7% idan aka kwatanta da model ɗin da ba a kare su ba
• Ƙarin Lokacin Gudanarwa: Ƙasa da kashi 5% na ƙarin lokacin ƙididdiga saboda ayyukan sirri
• Kiyaye Daidaito An kiyaye daidaiton model a cikin kashi 0.3% na aikin asali
• Ƙarfin Girma: Yana goyan bayan model har zuwa ma'auni 70B tare da ƙarancin lalacewar aiki

6. Ayyuka na Gaba & Jagorori

OML ya buɗe sabbin jagororin bincike tare da muhimman tasiri:

• Turawar AI na Kamfani: Rarraba model ɗin mallaka mai tsaro ga abokan ciniki
• Haɗin gwiwar Bincike: Rarraba model ɗin bincike tare da abokan haɗin gwiwa na ilimi
• Yin Bin Ka'idoji: Tilasta hana amfani da aikace-aikacen AI masu muhimmanci
• Koyo na Haɗin Kai: Tarin tsaro na sabuntawar model a cikin horo mai rarrabuwa

7. Nassoshi

1. Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. IEEE International Conference on Computer Vision.
2. Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems.
3. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL-HLT.
4. Radford, A., Wu, J., Child, R., et al. (2019). Language Models are Unsupervised Multitask Learners. OpenAI Technical Report.
5. TensorFlow Serving Architecture. (2023). TensorFlow Documentation.
6. Nakamoto, S. (2008). Bitcoin: A Peer-to-Peer Electronic Cash System.