I spent a few weeks building a Neuro-Symbolic Manufacturing Engine. I proved that AI can design drones that obey physics. I also proved that asking AI to pivot that code to robotics is a one-way ticket to a circular drain. Kwiintsuku ezidlulileyo, ndiyabhalisa ukwakhiwa kwizithuba zayo , inkqubo ye-AI enokuthintela i-intent ye-user ebonakalayo kwi-hardware ebonakalayo. OpenForge Indawo yokuVavanyelela iinkqubo ze-Google Ndingathanda ukuphendula ingxaki elifanelekileyo: Ngaba i-LLM ingangena ngaphezu kokubhala i-Python scripts kwaye ngokwenene iinkqubo ze-physical apho i-tolerance, i-voltage, kunye ne-compatibility kubalulekile? Gemini 3.0 Umxholo, kuboniswa, i-complicated "Yes, kodwa..." Ndingathanda le nophuhliso namhlanje. Nazi post-mortem malunga neengxaki, yaye iingxaki, kunye neengxaki ebalulekileyo phakathi ikhowudi kunye iinkqubo Generating Refactoring I-The Win: I-Drone_4 Works Okokuqala, iindaba elungileyo. I-drone_4 ye-repository ye-drone_4 iye yimpumelelo. Ukuba uklonywa i-repo kunye ne-"Long Range Cinema Drone", i-system isebenza ukusuka kwi-seed ukuya kwi-simulation. I-Cinema ibonelela ukuba "Cinema" ibonelela umlenze kwaye "Long Range" ibonelela kwi-GPS kunye ne-Crossfire protocols. Izixhobo zePhysics: I-Compatibility Engine ihamba ngempumelelo iingxaki zeMoto / i-battery ezinokufutshane okanye i-explode. I-simulation ye-reality: Iifayile ze-USD ezisekelwe kwi-NVIDIA Isaac Sim zibonakalisa ngokwenene. Ndingathanda, ndingathanda ukuba ndingathanda. Kwi-make_fleet.py, ndingathanda nangaphantsi. Ndingathanda ngaphantsi kwi-LLM ukufumana i-flot logic ngokugqithisileyo kwaye ngaphezulu kwi-hard-coded Python orchestration. Ndingathanda ukuba oku yinto yokubuyiselwa ukuxhaswa kwe-Gemini 3.0, yaye akuyona ukuba ndingathanda ukuyifaka ukuba ndingathanda ukuyifaka i-line elinye ye-code. Njengomdla we-concept ye- -Ukuba i-LLM ibhagamshelane ne-creative translation, kwaye i-Python ibhagamshelane neengxaki ze-physics - i-OpenForge inokufumana. Neuro-Symbolic AI I-Failure: I-Quadruped Pivot I-Half II yeengxaki yaba ukuthatha le injini esebenzayo kwaye i-pivot. Ndingathanda ukuguqulwa i-Drone Designer kwi-Robot Dog Designer (i-Ranch Dog). Ndinezisa i-Gemini 3.0 yonke i-codebase (i-88k i-tokens) kwaye ndithanda kwi-refactor. I-refactor yenza i-physics ezintsha, i-agent ezintsha ye-sourcing, kunye ne-solvers ezintsha ze-kinematics. I am officially shelving the Quadruped branch. Kwakhona kubonakala ukuba indlela ndivula le pivot iye yenza ngathi i-circular drain rabbit hole of troubleshooting. Ngaba ndiza ngexesha apho ukuguqulwa kwe-torque isilinganiso yaye ukuguqulwa kwe-inventory sourcing yaye ukuguqulwa kwe-sourcing yaye ukuguqulwa kwe-simulation. Ukuba ndingathanda ukwakha i-Ranch Dog, ndingathanda ukuvelisa kwakhona ukusuka kwakhona, usebenzisa injini ye-Drone kuphela njenge-model ye-reference, ayikho isiseko yokubhalisa. I-Lesson: I-Flattening Effect Yintoni iinjini ye-Drone iye yenza imiphumo kwaye i-Refactor ye-Quadruped iye yenza imiphumo? Oku kubandakanya ubungakanani obungapheliyo ngathi ndingathanda kwi-Gemini 3.0 ( kunye nezinye iimodeli ze-high-context). Xa ukwakha ukusuka kwelanga, wena kunye ne-AI ukwakha i-architecture step-by-step. You lay the foundation, then the framing, then the roof. Nangona kunjalo, xa ufuna i-LLM ukuba isicelo esebenzayo, ayibonakali umdla we-code. Ayibonakali izifundo ze-batch. pivot Ikhowudi yokuqala ye-Drone ilawulwa kwiinyathelo ezahlukeneyo kunye ne-linear. Kukho iingcingo ezizodwa yokusebenza kwengxaki kunye neengxaki zangaphakathi ezivela kwiingxaki ezidlulileyo. I-Gemini 3.0, ngokuchasene neengxaki yokusebenza, . Yenza izinyathelo zonyango ezahlukileyo ezahlukileyo kwiinkqubo ezininzi, ezininzi. Kwi-surface, ikhowudi iye ndingathanda kunye ne-Pythonic. Kodwa kwimeko, iye iye yandiswe iintlobo ze-structural-bearing walls ezininzi zithembisa isicelo. flattened the architecture I-coding yenzelwe yi-guide ye-styling, yaye akuyona i-structural necessity. I-Paradox of Capability: Gemini 2.5 vs. 3.0 I-project yokukhuthaza i-contra-intuitive reality: Gemini 2.5 was safer because the code it confidently spit out was truncated pseudo-code. Kwiimvavanyo ezidlulileyo, i-outputs iye yenzelwe ukubonisa indlela yokwakha. Uya kufuneka ukwakha isicwangciso yokwakha izicwangciso ngaphakathi kwiprogram. Kwiimeko, kunokufunda ifayile epheleleyo. Kwiimeko, kufuneka uqhagamshelane isicwangciso ngokufanelekileyo. I-Gemini 2.5 yandisa ukuba ndiye u-Architect. Ndingathanda ukuba uqhagamshelane kwi-program-by-program, ukucacisa ngokufanelekileyo into efunyenweyo. Ndingathanda ukuxhaswa kwimandla ye-AI. I-Gemini 3.0 inesantya kunye nokuxhomekeka ukwenza yonke into ngexesha elifanayo. Yenza i-illusion enokwenzeka ye-One-Shot Pivot. I-Gemini 3.0 ikhiqiza ikhowudi ebonakalayo ngokuzenzakalelayo kodwa i-structurally rotten ngaphakathi. I-Stafolding phase ifumaneka. Umhla lokugqibela Ukuba ufuna ukwakha I-Generative Manufacturing Engine, okanye nayiphi na inkqubo efanelekileyo kunye ne-LLMs, apha iimpazamo yam lokugqibela ye-OpenForge: I-Greenfield i-Easy, i-Brownfield i-Hard: I-LLM ziyafumaneka kwi-building ukusuka kwangaphambili. Zininzi ezininzi kwi-renovating i-complex, i-architectures ezikhoyo ngaphandle kwe-massive human-hand-holding. Don't Refactor nge Prompts: Ukuba ufuna ukuguqulwa inkqubo ye-app, awufuni i-AI ukuba uqhagamshelane oku kwi-X. Kwimeko, uthathe i-flow ye-logic ye-app entsha, kwaye uthetha i-AI ukwakha i-app entsha usebenzisa le map ye-logic. I-Architecture is Still King: Ungayifumana i-codebase njenge-document ye-fluid enokuthi ifumaneka yi-LLM. Ungayifumana ne-staffolding. I-OpenForge ibonise ukuba sinokufumana i-bridge phakathi kwe-intent ye-user ebonakalayo kunye ne-engineering ye-physical. Thina nje akuyona umntu kwi-architecture chair. Ngoko ke, i-Gemini 3.0 iyona ingxaki enkulu ukusuka kwi-2.5. I-Part of what I’m exploring here is how to get the best out of a brand-new tool.
