1. Gabatarwa & Babban Jigo

Wannan binciken, wanda ya dogara ne akan aikin Herbert L. Roitblat, yana gabatar da ra'ayi mai saba wa juna da kuma mai tsauri game da labarin da ke faruwa game da zuwan Hankali Gabaɗaya na Wucin Gadi (AGI). Babban jigon ya nuna cewa samfuran Hankalin Wucin Gadi na Halitta (GenAI) na yanzu da na gaba, ciki har da Manyan Samfuran Harshe (LLMs), ba su da ikon samun AGI ta asali saboda wani ƙuntatawa na asali da ake kira "bashi na dan adam." Wannan bashi yana nufin dogaro mai nauyi, wanda ba za a iya gujewa ba ga shigarwar dan adam don tsara matsala, ƙirar gine-gine, da kuma bayanan horo da aka tsara. Takardar ta bayar da hujjar cewa haɗarin gaske daga AI ba ya fitowa daga hankali mai girma, amma daga amfani da gazawar sa ta asali tare da amincin dan adam.

2. Ma'anar Bashi na Dan Adam

Bashi na dan adam shine tsarin ra'ayi na asali wanda ke bayyana dalilin da yasa Hankalin Wucin Gadi na zamani ba ya kan hanyar zuwa hankali gabaɗaya.

2.1 Ma'ana da Abubuwan Da Suka Haɗa Shi

Bashi na dan adam ya ƙunshi dogaro guda uku masu mahimmanci:

  • Matsaloli Masu Tsari Da Kyau: Dole ne mutane su tsara ayyuka ta hanyar da AI za ta iya sarrafawa.
  • Ƙirar Gine-gine: Tsarin hanyar sadarwar jijiyoyi (misali, Transformer) ƙirƙira ce ta dan adam.
  • Bayanan Horo Da Aka Tsara: Manyan tarin bayanai ana tattara su, tace su, da kuma lakafta su ta hanyar mutane.

Wannan bashi yana nufin tsarin AI ba sa ƙirƙirar sabbin tsare-tsare na warware matsala amma suna inganta cikin iyakokin da mutane suka ayyana.

2.2 Shigarwar Dan Adam A Matsayin Dogaro

Ana yawan fassara nasarar samfura kamar GPT-4 cikin kuskure. Roitblat yana jayayya cewa sun yi nasara ne saboda mutane sun riga sun warware matsalolin hankali na asali, suna barin samfurin ya yi "ƙididdiga masu sauƙi" kamar raguwar gradient. Samfurin mai ƙarfi ne mai amfani da tsari, ba mai ayyana matsala ko mai warwarewa a ma'ana gabaɗaya ba.

3. Matsaloli Na Asali Zuwa AGI

3.1 Iyakancewar Koyon Tsarin Harshe

GenAI na yanzu yana jefa kowace matsala a matsayin matsalar koyon tsarin harshe. Ko shi ne coding, ƙirar hoto, ko tunani, hanyar da ke ƙarƙashinta ita ce hasashen token na gaba (kalma, facin pixel) bisa ga alaƙar ƙididdiga a cikin bayanan horo. Wannan hanya tana da iyaka ta asali ga matsalolin da ke buƙatar tunani wanda ba na harshe ba, m, ko sabon abu wanda ba a ƙunshe shi a cikin bayanin dan adam na baya ba.

3.2 Rashin 'Yancin Kai Na Gaskiya

AGI yana buƙatar 'yancin kai—ikon saita manufofinsa, ayyana sababbin matsaloli, da kuma samun ƙwarewa ba tare da takamaiman umarni ba. Kamar yadda Lu da sauransu (2024) suka lura, LLMs kawai suna bin umarni. Ba su da ƙwaƙƙwaran buri ko iyawa don ƙwarewar ƙwarewa ta kai, wanda shine ginshiƙin hankali gabaɗaya.

3.3 Matsalar Nau'in Matsala

Wani babban cikas shi ne gazawar gane nau'ikan matsaloli da yawa. Wasu matsaloli, kamar "matsalolin fahimta" (misali, matsalar Dots Tara), ba za a iya warware su ta hanyar haɓakawa ko daidaita tsari daga bayanai ba. Suna buƙatar sake tsara sararin matsala—wata iyawa da ba ta nan a cikin tsarin koyo na tushen gradient na yanzu.

4. Tsarin Kimantawa Mai Kura

4.1 Rashin Isasshen Ma'auni

Ma'auni kamar ARC-AGI bai isa ya auna gabaɗaya ba. Wucewa gwaji baya bayyana yadda aka wuce shi. Samfuri na iya amfani da dabara mai kunkuntar gwaji (misali, haddacewa) ko ƙa'idar tunani gabaɗaya. Ma'auni suna auna aiki, ba asalin gabaɗaya na iyawa ba.

4.2 Kuskuren Tabbatar da Sakamakon

Takardar ta nuna babban kuskuren ma'ana a cikin kimanta AI: tabbatar da sakamakon. Tsarin shine: Idan wani abu yana da AGI, zai wuce gwaji T. Abun ya wuce gwaji T. Don haka, yana da AGI. Wannan kuskure ne. Nasarar aiki ba ta nuna amfani da hankali gabaɗaya a zahiri ba, kamar yadda za a iya samar da fitarwa iri ɗaya ta hanyoyi daban-daban (kuma marasa iyawa).

5. Takaddamar AGI Da Gaskiya

Ma'auni Masu Muhimmanci A Muhawarar AGI

  • 88% – Kiyasin kaso na iyawar AGI da ake buƙata wanda aka riga aka cimma (Thompson, 2025).
  • 33,000+ – Sa hannu akan buɗaɗɗen wasiƙar Cibiyar Rayuwar Gaba don dakatar da haɓaka LLM (2023).
  • 2025 – Shekarar Taron Aikin Hankalin Wucin Gadi a Paris.

5.1 Hasashe Da Iƙirari

Yanayin yana da alamar hasashe masu ƙarfi daga shugabannin masana'antu (Altman, 2025; Leike & Sutskever, 2023) na AGI na ɗan gajeren lokaci, galibi ana ƙididdige su (misali, "88% na iyawa"). Ana bambanta waɗannan da gargaɗin alama kamar "agogon amincin AI."

5.2 Damuwa Mai Ɗagawa Da Amsar Ka'idoji

Hasashe sun haifar da damuwa mai mahimmanci. Sanarwar Cibiyar Amincin AI (2023) ta daidaita haɗarin AI da annoba da yaƙin nukiliya. Rahoton Gladstone (Harris da sauransu, 2024) wanda Ma'aikatar Harkokin Wajen Amurka ta ba da izini ya yi gargadin game da haɗarin "kamar WMD" wanda gasar dakin gwaje-gwaje ke haifarwa. Wannan ya haifar da ƙoƙarin tsari, kamar shawarar SB-1047 na California tare da umarnin "kashe wuta," ko da yake an ki amincewa da shi.

6. Binciken Fasaha & Tsarin Lissafi

Ana iya fahimtar iyakancewar samfuran na yanzu a wani ɓangare ta hanyar manufar ingantawa. Ana horar da LLM na yau da kullun don haɓaka yuwuwar token na gaba $x_t$ idan aka ba da mahallin $x_{

$$\mathcal{L}_{LLM} = -\sum_{t} \log P(x_t | x_{

inda $\theta$ su ne sigogin samfurin. Wannan manufar ta tilasta wa samfurin ya zama ƙwararre a tsaka-tsaki a cikin yawan bayanan horo. AGI, duk da haka, yana buƙatar tsawaitawa da taƙaitawa—warware matsalolin da ke waje da ƙullin misalan horo. Ana iya ƙirƙira shingen "matsalar fahimta" a matsayin neman mafita $s^*$ a cikin sarari $S$, inda hanyar daga matsala $p$ zuwa $s^*$ ke buƙatar canji $T$ wanda ba a iya bambanta shi ba wanda ba a koya daga bayanai ba:

$$s^* = T(p), \quad \text{inda } \nabla_\theta T \text{ ba a bayyana shi ba ko sifili.}$$

Koyo na tushen gradient ($\theta \leftarrow \theta - \eta \nabla_\theta \mathcal{L}$) ba zai iya gano irin wannan $T$ ba. Wannan ya yi daidai da hujjoji daga AI na gargajiya, kamar "Matsalar Kafa Alama" (Harnad, 1990), wanda ke tambayar yadda ma'anoni za su iya tasowa daga sarrafa tsarin magana kawai.

Hoto: Tazara Tsakanin Tsaka-tsaki Da Tsawaitawa

Zane na Ra'ayi: Jirgin sama mai girma 2D yana wakiltar sararin yuwuwar matsaloli da mafita. Gajimare mai yawa na maki yana wakiltar bayanan horo (matsalolin da mutane suka bayar da mafita). Samfuran GenAI na yanzu sun fi dacewa da neman mafita a cikin wannan gajimaren (tsaka-tsaki). Ja "X" yana nuna alamar "matsalar fahimta"—mafitarsa tana waje da gajimaren. Babu wata hanyar gradient mai santsi da ke kaiwa daga gajimaren zuwa "X"; isa gare shi yana buƙatar tsalle a cikin tunani, wanda raguwar gradient ba zai iya cimma shi ba. Wannan yana wakiltar bashi na dan adam a zahiri: samfurin yana ƙuntata cikin gajimaren bayanai da mutane suka bayar.

7. Tsarin Bincike: Matrix na Iyawar AGI

Don matsawa fiye da ma'auni na kuskure, muna ba da shawarar matrix na kimantawa mai inganci. Maimakon tambaya "Shin ya wuce gwajin?", muna tambaya "Menene yanayin iyawarsa?" Ga kowane aiki T, kimanta tare da ginshiƙai biyu:

  1. Gabaɗaya Hanyar (G): Shin hanyar warwarewa ta musamman ga T (G=0), ta shafi nau'in ayyuka (G=1), ko ba ta da alaƙa da yanki (G=2)?
  2. 'Yancin Kai A Tsara Matsala (A): Shin mutane ne suka ayyana matsala gaba ɗaya (A=0), tsarin ya inganta shi a wani ɓangare (A=1), ko tsarin ya gano/kafa shi da kansa (A=2)?

Misalin Shari'a (Ma'auni na ARC-AGI): Samfurin da ke haddace mafita ga takamaiman tsarin wasanin gwada ilimi na ARC yana ci (G=0, A=0). Samfurin da ya koyi dabarar tunani na gani gabaɗaya wanda ya shafi wasanin gwada ilimin ARC da ba a gani ba yana ci (G=1, A=0). Tsarin da ba kawai ya warware wasanin gwada ilimin ARC ba har ma ya gano sabon nau'in wasanin gwada ilimi na tunani a kansa zai kusanci (G=2, A=2). Samfuran SOTA na yanzu suna aiki a cikin kusurwar (G=0/1, A=0). AGI na gaskiya yana buƙatar aiki akai-akai a (G=2, A=2). Wannan tsarin ya sa kuskuren tabbatar da sakamakon ya bayyana a fili: babban makin gwaji kawai yana tabbatar da aiki, ba babban maki G ko A ba.

8. Hanyoyin Gaba & Bincike

Samun AGI zai buƙaci sauye-sauyen tsari, ba kawai girman gine-ginen na yanzu ba.

  • Samfuran Duniya da Fahimtar Jiki: Dole ne bincike ya wuce hasashen rubutu mara aiki zuwa wakilai masu aiki waɗanda ke gina samfuran ciki na duniya ta hanyar hulɗa, kamar yadda aka gani a ci gaban mutum-mutumi da kwaikwayo (misali, SIMA na DeepMind). Wannan yana rage dogaro ga bayanan harshe da aka tsara.
  • Haɗin Kai na Neuro-Symbolic: Haɗa ƙarfin gane tsari na hanyoyin sadarwar jijiyoyi tare da bayyanannen tunani na AI na alama (kamar yadda MIT-IBM Watson Lab suka bincika) zai iya magance shingen "matsalar fahimta."
  • Manufofin Koyo Na Kai: Haɓaka algorithms na ƙwaƙƙwaran buri waɗanda ke ba da damar tsarin samar da manufofin koyo nasu, matsawa fiye da ayyukan asara da mutane suka ayyana. Wannan fage ne sabo a cikin binciken AI.
  • Sabon Kimiyyar Kimantawa: Ƙirƙirar ma'auni waɗanda ke gwada gabaɗaya (G) da 'yancin kai (A) a fili, watakila ta hanyar buɗaɗɗen jerin ƙalubale da aka ƙirƙira ta atomatik waɗanda ke bincika ƙwarewar meta-koyo da tsara matsala.

"Aikace-aikacen" na farko na wannan binciken shine a cikin manufofi da saka hannun jari: ka'idoji yakamata su mayar da hankali kan haɗari na zahiri, na ɗan gajeren lokaci daga tsarin da ba su da adalci ko aminci, ba hasashen AGI ya karɓe ba. Yakamata a ba da saka hannun jari zuwa ga bincike na asali wanda ke rage bashi na dan adam, ba kawai zuwa girman bayanai da sigogi ba.

9. Ra'ayin Mai Bincike Mai Tsauri

Fahimta Ta Asali: Masana'antar AI tana fama da mummunan lamarin "myopia na fitarwa." Muna sha'awar rubutu mai santsi da hotuna masu ban mamaki, muna kuskuren ƙwararrun ƙididdiga don fahimta. "Bashi na dan adam" na Roitblat shine kalmar da ta dace don wannan dogaro na ɓoye. Shi ne giwa a cikin ɗakin uwar garken. Kowane "ɓarkewar" shine, idan aka duba, shaida ga hazakar dan adam a cikin tsara bayanai da tsara matsala, ba hankalin da aka haifa da inji ba. Labarin gaskiya ba ƙarfin AI bane; shine babban aikin ɗan adam, sau da yawa a ɓoye, wanda ke sa ya yi kama da ƙarfi.

Tsarin Ma'ana: Hujjar tana da sauƙi kuma tana da ma'ana. 1) Ayyana manufa (AGI a matsayin mai warware matsala mai 'yancin kai, gabaɗaya). 2) Bincika kayan aiki (GenAI a matsayin mai daidaita tsari akan bayanan ɗan adam). 3) Gano rashin daidaituwa (aikin asali na kayan aikin yana dogara ne da pre-processing na ɗan adam). 4) Gano kuskuren (rikiɗar fitarwar kayan aiki tare da buƙatun manufar). 5) Bayyana gazawar tsarin (hanyoyin kimantawa waɗanda ba za su iya bambanta tsakanin haddacewa da fahimta ba). Wannan ba falsafa bane; shine asalin alhakin injiniya.

Ƙarfi & Kurakurai: Ƙarfinsa shine sukar sa na asali. Yana kai hari kan jigon dukan labarin "AGI yana kusa" ta hanyar tambayar gine-ginen bege. Kuskurensa, watakila, shine cewa bai shiga cikin hujjar da ke gaba daga fitowa ba—yuwuwar sabbin iyawa (kamar tunanin sarkar) suna fitowa a cikin girma ta hanyoyin da ba mu fahimta ba tukuna. Duk da haka, takardar ta mayar da martani daidai cewa fitowa ba sihiri bane; har yanzu yana da iyaka ta manufar horo $\mathcal{L}_{LLM}$. Ba za ku iya fitowa 'yancin kai daga aikin asara wanda ba shi da lokaci don haka ba.

Fahimta Mai Aiki: Ga Masu Tsara Manufofi: Ku kau da kai daga takaddamar sci-fi. Ka'idoji abin da ke gaban ku: sirrin bayanai, son zuciya na algorithm, maye gurbin ma'aikata, da farashin muhalli na horo. "Kashe wuta" don samfurin da ba zai iya ɗaure takalmin sa ba wasan kwaikwayo ne na tsaro. Ga Masu Saka Hannun Jari: Ku kasance masu shakku sosai game da kowace kamfani wanda ƙimar sa ta dogara ne akan samun AGI. Ku yi fare akan kamfanonin da ke warware takamaiman matsaloli masu mahimmanci tare da AI mai ƙarfi, ba waɗanda ke sayar da AGI vaporware ba. Ga Masu Bincike: Ku daina bin jagororin ma'auni. Ku fara ƙirƙira gwaje-gwaje waɗanda da gangan suke ƙoƙarin karya ruɗin fahimtar samfurin ku. Ku bi gine-ginen da ke rage bashi na dan adam. Hanyar gaba ba ta hanyar ƙarin bayanai iri ɗaya bane, amma ta hanyoyin koyo daban-daban na asali. Agogon ba yana ƙidaya zuwa AGI ba; yana ƙidaya zuwa lokacin da muka gane cewa muna inganta aikin da ba daidai ba.

10. Nassoshi

  1. Roitblat, H. L. (Source PDF). Wasu abubuwan da za a sani game da samun hankali gabaɗaya na wucin gadi.
  2. Chollet, F. (2019). Akan Ma'aunin Hankali. arXiv preprint arXiv:1911.01547.
  3. Lu, Y., da sauransu. (2024). [Nassoshi akan LLMs suna bin umarni].
  4. Harris, J., Harris, T., & Beal, B. (2024). Rahoton Gladstone. Ma'aikatar Harkokin Wajen Amurka.
  5. Cibiyar Amincin AI. (2023). Sanarwa akan Haɗarin AI. https://www.safe.ai/work/statement-on-ai-risk
  6. Cibiyar Rayuwar Gaba. (2023). Dakatar da Manyan Gwaje-gwajen AI: Buɗaɗɗen Wasiƙa. https://futureoflife.org/open-letter/pause-giant-ai-experiments/
  7. Harnad, S. (1990). Matsalar Kafa Alama. Physica D: Abubuwan da ba su da layi, 42(1-3), 335–346.
  8. Zhu, J., da sauransu. (2017). Fassarar Hoton zuwa Hoton mara Haɗin gwiwa ta amfani da Hanyoyin Sadarwar Adawa na Tsarin Zagaye. Proceedings of the IEEE International Conference on Computer Vision (ICCV). (CycleGAN a matsayin misalin koyo ba tare da bayanan da mutane suka tsara ba—ƙaramin mataki na rage nau'in bashi na dan adam).
  9. DeepMind. (2024). SIMA: Wakilin AI Gabaɗaya don Muhallin Wucin Gadi 3D. https://www.deepmind.com/sima (Misalin binciken da ke tafiya zuwa wakilai masu gina samfuran duniya).