Ma jiro hagaajin, kaliya qor code! Shaqada cusub ee kooxda Jeff Clune: Meta Agent si toos ah ayuu u kobciyaa module-ka xusuusta

U socoshada Software 3.0, AI waxay bilaabeysaa inay qorto code-ka Python si ay u kobciso maskaxda.

Qeybta qoto dheer ee horumarinta Agent, xusuusta (Memory) had iyo jeer waa dhib aan laga gudbi karin.

In kasta oo awoodda moodooyinka aasaasiga ah ay sii kordheyso, haddana asal ahaan waa kuwo aan xaalad lahayn (Stateless) inta lagu jiro habka sababaynta, taas oo xaddideysa awoodda Agent si joogto ah u urursato waayo-aragnimo.

Xalka ugu weyn ee warshadaha ee wax ka qabashada xusuusta, ha ahaato RAG ama soo koobidda daaqadaha simbiriirixan, asal ahaan waxay ku sii jiraan marxaladda xeerarka heuristic ee si macmal ah loo naqshadeeyay.

Module-ka xusuusta ee gacanta lagu sameeyay aad buu u jilicsan yahay oo ay adag tahay in la wareejiyo. Prompt-ka iyo macquulka soo celinta ee si taxadar leh loogu hagaajiyay nidaamyada wada hadalka ayaa inta badan si toos ah u fashilma marka la geliyo hawlaha qorsheynta fog (sida ALFWorld) ama ciyaaraha istaraatiijiyadeed ee adag.

Si wax looga qabto dhibaatadan, borofisar UBC ah, cilmi-baare hore ee OpenAI Jeff Clune iyo kooxdiisu waxay bixiyeen xal geek ah.

Maadaama aan la garanayn qaabka xusuusta ee ugu fiican, Agent ha qorto code-ka Python si uu u naqshadeeyo.

Tani waa waxa hadda la sii daayay ALMA (Automated meta-Learning of Memory designs for Agentic systems).

Laga bilaabo ADAS ilaa ALMA: Naqshadeynta otomaatiga ah ee ku saleysan code

ALMA waa sii wadida dariiqa farsamada algorithm-ka ee AI-ga ee kooxdu dhowaan kor u qaadday.

Gudaha ADAS (Automated Design of Agentic Systems), kooxdu waxay caddeeyeen in code-ka uu yahay meel raadin oo hufan marka loo eego miisaanka shabakadaha neerfaha ama Soft Prompts marka la naqshadaynayo qaab dhismeedka Agent. Code-ku wuxuu leeyahay dhammaystirka Turing, wuxuuna leeyahay sharraxaad aad u xooggan.

Kadib gudaha DGM (Darwin Gödel Machine), kooxdu waxay soo bandhigtay fikradda sahaminta furan ee algorithm-ka kobcinta, iyadoo ilaalinaysa kaydka naqshadeynta, dhiirigelinaysa moodelka inuu sahamiyo xalal cusub.

ALMA waxay dhaxashay paradigm-ka code-ka ee ADAS iyo istaraatiijiyadda kobcinta ee DGM, iyadoo diiradda saaraysa goobta codsiga ee qaybta ugu tiirsan khibradda macmalka ah ee nidaamka Agent - xusuusta.

Habka shaqo ee ALMA

Habka hawlgalka ALMA waa wareegga meta-barashada caadiga ah. Meta Agent si toos ah uma qabato hawsha, laakiin waxay mas'uul ka tahay barnaamijyada. Habka waxaa ku jira afar marxaladood:

• Fikirka: Falanqee kaydka naqshadeynta xusuusta ee hadda jira, oo ku saleysan waxqabadka taariikhiga ah si loo abuuro qorshayaal hagaajin
• Qorsheynta: U rog fikradaha macquul pseudo-code ah
• Hirgelinta: Qor code-ka Python ee la fulin karo, qeex hawlaha asaasiga ah
• Qiimeynta: Ku faafi code-ka la soo saaray deegaanka sandbox si loo fuliyo hawsha, jawaab celin tilmaamayaasha waxqabadka

Inta lagu jiro habka kobcinta, ALMA waxay soo saari doontaa geed naqshadeyn oo weyn. Marka tirada tillaabooyinka isdaba joogga ah ay kordho, code-ka xusuusta ee la soo saaray wuxuu si tartiib tartiib ah uga kobcayaa macquul kayd fudud una kobcayaa qaab dhismeed garasho oo adag.

Qaab dhismeedka xusuusta ee kobcay

Naqshadaynta xusuusta ee ALMA soo saartay waxay muujisay kala duwanaansho weyn hawlo kala duwan:

• MiniHack (Sahaminta godka): Waxay naqshadaysay module-ka Khatarta iyo Isdhexgalka, si cad u diiwaangelisa hawlgallada keena dhiigbaxa iyo gardarrada bahalnimada
• Baba Is AI (Xujooyinka macquulka ah): Waxay naqshadaysay Maktabadda Istaraatiijiyadda, iyadoo diiwaangelinaysa isku dhafka xeerarka ee looga baahan yahay in lagu gudbo heerarka

Tani waxay muujineysaa in AI ay aqoonsan karto astaamaha hawsha: ciyaaraha badbaadada waxay u baahan yihiin inay fiiro gaar ah u yeeshaan khatarta, ciyaaraha xujooyinka waxay u baahan yihiin inay fiiro gaar ah u yeeshaan soo koobidda xeerarka.

Natiijooyinka tijaabada

ALMA waxaa lagu barbar dhigay saldhigyada waaweyn ee afarta deegaan ee TextWorld, ALFWorld, MiniHack, iyo Baba Is AI.

Moodelka GPT-5-mini, celceliska heerka guusha ee ALMA wuxuu gaarayaa 53.9%, taas oo ka fiican G-Memory (46.0%) iyo Soo Celinta Trajectory (48.6%).

Marka la eego hufnaanta kharashka, ALMA waxay celcelis ahaan isticmaashaa kaliya 1,319 tokens, halka Soo Celinta Trajectory ay isticmaasho ilaa 9,149 tokens, G-Memory sidoo kale waxay gaartay 6,055 tokens. ALMA waxay ku beddeshay waxqabad wanaagsan oo leh kharash qiyaastii 1/7 ilaa 1/5 ah.

Gabagabo

ALMA waxay muujineysaa suurtagalnimada kala guurka Software 2.0 (Neural Networks) una guurto Software 3.0 (AI-Generating Algorithms).

Horumarinta Agent, naqshadeynta module-ka xusuusta waxay muddo dheer ku tiirsaneyd dareenka injineerada. ALMA waxay caddeysay in, iyada oo loo marayo meta-barashada iyo code-ka, AI ay si toos ah u ogaan karto qaab dhismeedka xusuusta ee ugu fiican iyadoo ku saleysan deegaanka gaarka ah.

Xiriirinta kheyraadka

• Warqad: https://arxiv.org/pdf/2602.07755
• Code: https://github.com/zksha/alma
• Bogga mashruuca: https://yimingxiong.me/alma