Mafunde Awiri a AI: Kupindula Bwino, Ma Agents Odziyimira Pawokha, ndi Ngongole Yakuyandikira ya Infrastructure | B4N1

Mafunde Awiri a AI: Kupindula Bwino, Ma Agents Odziyimira Pawokha, ndi Ngongole Yakuyandikira ya Infrastructure

6 days ago · b4n1 Read time: 8 min

Mafunde Awiri a AI: Kupindula Bwino, Ma Agents Odziyimira Pawokha, ndi Ngongole Yachitukuko Ikubwera

Mawonekedwe a AI ali mu chisinthiko champhamvu, koma zomwe zikuchitika masiku ano zikuwonetsa kusinthana kofunikira kwa mphamvu. Tikuwona kufunikira kochita bwino kwambiri m'mamodeli akumalire limodzi ndi kuwonekera kwa mabungwe odziyimira pawokha, pomwe misika yazachuma ikuwonetsa chenjezo lokhudza kuchuluka kwa ndalama zomwe zikufunika kuti zithandizire kusinthaku. Izi sizongowonjezera zowonjezera; ikuyimira kutanthauzira kofunikira kwa kuthekera kwa AI, kamangidwe kake, ndi zenizeni zachuma zomwe zikuthandizira kukula kwake.

OpenAI's GPT-6.1 Terra: The Efficiency Imperative

Kutulutsidwa kwa OpenAI kwa GPT-6.1 Terra kukuwonetsa kusintha kwabwino. Zodziwika kuti "zotsika mtengo komanso zachangu" kuposa zomwe zidalipo kale, izi ndizoposa kupindula kowonjezera; ndichofunika kamangidwe. Pachitsanzo chakumalire, kukwaniritsa zonse ziwiri zochepetsera mtengo komanso kuwongolera liwiro nthawi imodzi kumapereka kukonzanso kwakuya kwa stack yowerengera.

Kuchokera kumalingaliro aukadaulo, "zotsika mtengo komanso zachangu" mumitundu yayikulu yachilankhulo (LLM) nthawi zambiri zimalozera kuzinthu zingapo zofunika:

  1. ** Quantization ndi Sparsity: ** Kuchulukitsa kwaukali pambuyo pa maphunziro (mwachitsanzo, 4-bit kapena 2-bit inference) kuphatikizapo njira zamakono zowonongeka zowonongeka zingathe kuchepetsa kwambiri kukumbukira kukumbukira ndi kuwerengera katundu popanda kutaya kwakukulu mu khalidwe. Izi zimatanthawuza ma GPU ochepa ozungulira komanso VRAM yochepa pa tokeni.
  2. Mainjini Ololera Okhazikika: Zingwe zolozera mwamakonda, zogwiritsa ntchito zida zapadera (monga ma ASIC amtundu wina kapena ma GPU amtundu wotsatira okhala ndi ma tensor cores), amatha kupititsa patsogolo kutulutsa komanso kutsika kocheperako. Njira zofananira zongopeka, pomwe mtundu wocheperako, wothamanga umapereka zizindikiro zomwe mtundu wokulirapo umatsimikizira, zimathandiziranso kufulumira.
  3. Zomangamanga Zapamwamba Zachitsanzo: Ngakhale kuti core transformer idakalipo, zatsopano m'magawo ang'onoang'ono, monga Mixture-of-Experts (MoE) omwe ali ndi njira zoyendetsera bwino kapena zolemba zatsopano, zingayambitse ntchito yabwino pa parameter.

Zotsatira zamagulu a uinjiniya ndizozama. Njira yotsika mtengo kwambiri yam'malire imapangitsa kuti anthu azitha kupeza mwayi, zomwe zimapangitsa kuti kuphatikiza kwa AI kwapamwamba kutheke pamitundu yambiri yamapulogalamu omwe kale anali otsika mtengo. Ganizirani za chithandizo chamakasitomala munthawi yeniyeni, kupanga zinthu zamunthu payekhapayekha, kapena mapaipi ovuta kusanthula deta pomwe kuyimba kwa API kunali kovutirapo. Izi zimasintha mtengo wa Total Cost of Ownership (TCO) kwambiri.

Tiyeni tilingalire kufananitsa kongoyerekeza, kusonyeza mtundu wa zopindulitsa zomwe tingakhale tikuwona:

| | Metric | GPT-6.0 (Zongopeka) | GPT-6.1 Terra (Yoyerekeza) | Impact Engineering | | | :-------------------- | :-------------------- | :----------------------- | :-------------------------------------------------- | | | Kuchedwa (P95) | 500ms | 250ms | Imayatsa kuyanjana kwa ogwiritsa ntchito munthawi yeniyeni, kuchedwa kwapang'onopang'ono. | | | | Mtengo pa 1M Zizindikiro | $15.00 | $7.50 | Kuchepetsa ndalama zoyendetsera ntchito zama voliyumu ambiri. | | | | Kudzera (Zizindikiro/s) | 1,000 | 1,800 | Imathandizira ogwiritsa ntchito nthawi imodzi kapena kukonza ma batch akuluakulu. | | | | Memory Footprint | Mkulu | Wapakati | Itha kuloleza zida zazing'ono, zotsika mtengo. | | | | Kugwiritsa Ntchito Mphamvu | Mkulu | Wapakati | Amachepetsa ntchito ya carbon footprint ndi magetsi. | |

💡 Key Takeaway: GPT-6.1 Cholinga cha Terra pakuchita bwino sikungokhudza liwiro; ndi zakusintha kuthekera kwachuma kwa AI yapamwamba. Magulu a mainjiniya tsopano akuyenera kuwunikanso njira zawo zophatikizira AI, poganizira zakugwiritsa ntchito zomwe poyamba zinkawoneka zodula kapena zochedwa. Izi zimatsegula zitseko zakuya, kufalikira kwa AI kumakampani onse.

Sewero Lapawiri la DeepSeek: Ma Models Olemera Otsegula ndi Ma Autonomous Coding Agents

Nthawi yomweyo, DeepSeek yapanga cholowa chodziwika bwino ndi V4.1-Flash-1017 yotseguka yolemetsa komanso choyimira chodziyimira pawokha. Njira yamitundu iwiriyi imayang'ana zonse zoyambira zoyambira komanso gawo lachitukuko cha AI.

Kutulutsidwa kwa chitsanzo chotseguka cholemera ngati V4.1-Flash-1017 ndi mwayi waukulu kwa anthu otseguka. Amapereka njira ina yamphamvu ku mayankho a eni ake, kulimbikitsa zatsopano kudzera powonekera komanso makonda. Kwa mabungwe omwe ali ndi zofunikira pazinsinsi za data kapena zosowa zapadera za domeni, cholozera chowoneka bwino chimalola kukonza bwino nkhokwe zawo popanda kutumiza zidziwitso zachinsinsi ku ma API ena. Izi zitha kupangitsa kuti pakhale mitundu yapadera kwambiri, yochita bwino yomwe imaphatikizidwa mozama mumagulu omwe alipo kale.

Komabe, chomwe chikusokonekera kwambiri ndi DeepSeek's prototype autonomous coding agent, yopangidwira kulimbana ndi omwe akupikisana nawo monga a Claude C, monga tafotokozera mu AI mwachidule kuyambira Novembara 1, 2026. Ichi si chida chomaliza chomaliza kapena cholembera chapamwamba. Wothandizira wodziyimira pawokha amatanthauza dongosolo lotha:

  1. Kukonzekera Mogwirizana ndi Zolinga: Kugawa ntchito zapamwamba kukhala ntchito zazing'ono zomwe zingatheke.
  2. Kugwiritsa Ntchito Zida: Kuyanjana ndi machitidwe akunja (ophatikiza, ochotsa zolakwika, kuwongolera mtundu, ma API, ma database) kuti akwaniritse mapulani.
  3. Kudzikonza tokha ndi Kusinkhasinkha: Kusanthula zotuluka, kuzindikira zolakwika, ndikukonzanso mobwerezabwereza njira yake.
  4. Boma Losalekeza: Kusunga nkhani pazochitika zingapo ndi chitukuko.

Izi zimapitilira magwiridwe antchito a "copilot" kupita ku "madivelopa-mmodzi" kapenanso "wotsogolera wotsogolera" m'malo odziwika bwino. Tangoganizirani wothandizira amene angalandire pempho la mawonekedwe, scaffold polojekiti, kulemba code yofunikira, kupanga mayesero, kuwayendetsa, kulephera kuthetsa, komanso kutumiza pempho la kukoka - zonsezi ndi kulowererapo kochepa kwaumunthu.

Zomangamanga, wothandizira wotere amadalira malupu apamwamba kwambiri, omwe nthawi zambiri amaphatikiza LLM "yokonzekera", LLM "yogwiritsa ntchito zida", ndi "reflector" LLM, yoyendetsedwa ndi ndege yowongolera.

# Chifaniziro chosavuta cha Pythonic cha loop yayikulu ya wothandizira wodziyimira pawokha
kalasi AutonomousCodingAgent:
    def __init__(self, planner_llm, coder_llm, debugger_llm, zida):
        self.planner = planner_llm
        self.coder = coder_llm
        self.debugger = debugger_llm
        self.tools = zida # mwachitsanzo, {'git': GitTool(), 'compiler': CompilerTool()}
        self.context = [] # Kusunga mbiri yamakambirano, zidule zama code, zolakwika

    def execute_task(yekha, high_level_goli):
        plan = self.planner.generate_plan(high_level_goal, self.context)
        self.context.append(f"Cholinga: {high_level_goli}\nPulani: {pulani}")

        kwa sitepe mu plan.steps:
            sindikizani(f"Kuchita sitepe: {step.descript}")
            code_snippet = self.coder.generate_code(step.description, self.context)
            self.context.append(f"Generated Code:\n{code_snippet}")

            yesani:
                # Gwiritsani ntchito chida kuti mupereke nambala (mwachitsanzo, yesani kuyesa, phatikizani)
                zotsatira = self.tools['compiler'].run(code_snippet)
                ngati zotsatira.has_errors:
                    error_report = self.debugger.analyze_error(result.errors, code_snippet, self.context)
                    sindikiza (f"Zolakwika zapezeka: {error_report}")
                    # Agent amadziwongolera: amawonjezera zolakwika pamutu, konzekeraninso kapena ma codenso
                    self.context.append(f"Vuto: {error_report}")
                    # Mutha kubwereranso ku coder_llm ndi zolakwika
                zina:
                    print("Njira yopambana.")
                    self.context.append(f"Kuyenda bwino: {step.descript}")
            Kupatulapo monga e:
                sindikiza (f"Kuletsa kwalephera: {e}")
                self.context.append(f"Kuletsa kwalephera: {e}")
                # Yang'anani zolephera zosayembekezereka, konzekeraninso

        final_review = self.planner.review_overall_goal(self.context)
        print(f"Ntchito yatha. Ndemanga yomaliza: {final_review}")
        kubwerera komaliza_review

# Kugwiritsa ntchito zitsanzo (malingaliro)
# wothandizira = AutonomousCodingAgent(GPT61Terra, DeepSeekV41, DeepSeekV41, {'git': GitTool(), 'compiler': PythonCompiler()})
# agent.execute_task("Tsimikizirani mawu omaliza otsimikizira ogwiritsa ntchito ndi JWT.")

Kusinthaku kumafuna kuunikanso kwa kayendetsedwe ka ntchito, kamangidwe ka mapulogalamu, ndi kutsimikizika kwabwino. Madivelopa a anthu adzasintha kuchoka pa kulemba boilerplate kupita kukufotokozera zolinga zapamwamba, kuwunikanso ma code opangidwa ndi othandizira, ndikumanga makina omwe othandizira amagwira ntchito mkati.

The AI ​​Build-out Mantha: Kuwerengera Zachuma

Ngakhale kupita patsogolo kwaukadaulo kukusangalatsa, nkhani yotsutsa ikuwonekera kuchokera kumisika yazachuma. Lipoti laposachedwa la Wall Street Journal, losonyezedwa ndi AI News Today, limasonyeza kuti Credit Default Swaps (CDS) pangongole yamakampani akuluakulu aukadaulo—Amazon, Alphabet, Microsoft, Meta, and Oracle—afika pamlingo wapamwamba kwambiri kuyambira 2021. Wolakwa? "AI Kupanga Mantha A kunja."

Credit Default Swaps kwenikweni ndi mapangano a inshuwaransi motsutsana ndi kampani yomwe ikulephera kubweza ngongole yake. Mitengo ya CDS ikakwera, zikuwonetsa kuti osunga ndalama amawona kuti ali pachiwopsezo chachikulu chokhazikika. Mfundo yakuti izi zili pamtunda wa 2021, nthawi yakusakhazikika kwa msika, ndi chizindikiro chokulirapo cha mantha omwe amagulitsa ndalama.

Mantha a "AI Build-Out" sizongopeka. Amayimira ndalama zazikuluzikulu zomwe zimafunikira kuti apikisane nawo mpikisano wa zida za AI: Ma Data Center: Kupeza malo, kumanga malo okulirapo, ndikuwonetsetsa kuti magetsi ndi zida zoziziritsa zikuyenda bwino.

  • Chip Mwapadera: Kugula ma GPU apamwamba kwambiri (monga ma NVIDIA H100, B200s, kapena ma ASIC okhazikika), omwe amabwera ndi ma tag okwera mtengo komanso nthawi yayitali yotsogola.
  • Kugwiritsa Ntchito Mphamvu: Ndalama zoyendetsera malo opangira datawa ndi zakuthambo komanso zikukulirakulira. Kuphunzitsa LLM imodzi yayikulu kumatha kugwiritsa ntchito magetsi ochulukirapo ngati nyumba mazana ambiri pachaka.
  • Kupeza Talente: Mpikisano wowopsa wa ofufuza apamwamba a AI ndi mainjiniya umakweza mtengo wamalipiro.

Ngakhale makampani monga Santander ndi Novartis akuwonetsa kale kukwera kwa ndalama zoyendetsedwa ndi AI ndikuchepetsa mtengo, monga momwe [CNBC] (https://www.cnbc.com/2026/11/01/ai-investments-company-earnings.html), ndalama zotsogola zimakhalabe zazikulu. Kupanikizika kwazachuma kumeneku kumakhudza tsogolo lamakampani aukadaulo. Kukwera mtengo kubwereka kwa zimphona izi kungayambitse:

  • Kupanga Kwapang'onopang'ono: Kuchepetsa bajeti ya R&D pomwe makampani amaika patsogolo kubweza ngongole.
  • Kulemba Ntchito Kuyimitsidwa/Kuyimitsa Ntchito: Njira zochepetsera mtengo zomwe zikukhudza kukula kwa ogwira ntchito.
  • Kuphatikizika: Osewera ang'onoang'ono akuvutika kuti azithamanga, zomwe zimapangitsa kuti azigula.
  • Kusintha kwa Mitengo Yamtambo: Opereka mitambo pamtambo angafunike kusintha mitundu yamitengo kuti abwezerenso mtengo wa zomangamanga, zomwe zingakhudze oyambitsa ndi mabizinesi ang'onoang'ono.

⚠️ Kuopsa Kwa Zomangamanga: Kukwera mtengo kwa zomangamanga za AI kumabweretsa chiwopsezo chokhalitsa. Zosankha zaumisiri pazosankha zachitsanzo, kukhathamiritsa kwamalingaliro, komanso magwiridwe antchito apakatikati sadangokhudzanso magwiridwe antchito; ndi zisankho zofunika kwambiri zachuma zomwe zimakhudza mwachindunji tsamba lamakampani ndi malingaliro amsika. Kunyalanyaza TCO ya AI kudzatsogolera ku zitsanzo zosagwira ntchito.

Zowona Zatsopano Zauinjiniya: Kuyenda Mwachangu ndi Ngongole

Kuphatikizika kwa mitundu yotsika mtengo, yothamanga kwambiri komanso kukwera mtengo kwa zomangamanga za AI kumapereka chododometsa. AI ikupezeka mosavuta pakungoyerekeza, koma mtengo wonse womanga ndi kusamalira chilengedwe ukukwera. Izi zimafuna gawo latsopano lazachuma kuchokera ku utsogoleri wa engineering.

Magulu a uinjiniya ayenera kudziwa bwino mtengo wa zosankha zawo za AI. Izi zikutanthauza:

  1. Zomangamanga Zodziwa Mtengo: Kuika patsogolo kuyika kwachitsanzo koyenera, njira zogwiritsira ntchito monga kuchulukitsa ndi kudulira, ndikusankha mosamala pakati pa mafoni a API a eni ake ndi zitsanzo zotseguka zotseguka zochokera ku TCO.
  2. Njira Zachitsanzo Zophatikiza: Kuphatikiza mphamvu ya zitsanzo zam'malire (monga GPT-6.1 Terra) chifukwa cha kulingalira kovuta ndi zitsanzo zomveka bwino zotseguka (monga DeepSeek V4.1-Flash-1017) pa ntchito zapadera, kukhathamiritsa kwa ntchito zonse ndi mtengo.
  3. Embracing Agency Workflows: Kukonzekera zam'tsogolo momwe ma AI amathandizira ntchito zambiri zachitukuko. Izi zimaphatikizapo kupanga ma API olimba, mawonekedwe omveka bwino, ndi machitidwe oyesera omwe othandizira angagwirizane nawo. Madivelopa adzasinthira ku orchestration ya othandizira, kutsimikizira, ndi mapangidwe apamwamba kwambiri.
  4. Kulimbikitsa Zomangamanga Zokhazikika: Kukankhira zida zogwiritsa ntchito mphamvu zamagetsi, kukhathamiritsa kagwiritsidwe ntchito ka malo a data, ndikuyang'ana magwero amphamvu ongowonjezedwanso kuti muchepetse kukhudzidwa kwa chilengedwe ndi ndalama za AI compute.

Mawonekedwe apano ndi malo ofunikira kuti apange zatsopano, kukakamiza mainjiniya kuti asangokankhira malire a kuthekera kwa AI komanso kuti azitha kudziwa bwino zachuma zomwe amatumizidwa. Mpikisano wa odziyimira pawokha ukukulirakulira, ndipo kudziwa lusoli ndikofunikira pantchito zotsimikizira zamtsogolo komanso njira zamabizinesi. Makampani omwe amadziwa bwino zaukadaulo komanso zachuma pakusintha kwa AI ndi omwe azichita bwino.

##Magwero

Have Questions?

Let's Talk

Contact me on LinkedIn