Umhlahlandlela Wobunjiniyela Be-Loop: I-'autoresearch' kanye ne-'Bilevel Autoresearch' Ziwaguqula Kanjani Ama-AI Agents Awenze Umshini Ozimele Wokufunda Izihibe Zokucwaninga zeML

0 0 5 minutes read

Umhlahlandlela Wobunjiniyela Be-Loop: I-'autoresearch' kanye ne-'Bilevel Autoresearch' Ziwaguqula Kanjani Ama-AI Agents Awenze Umshini Ozimele Wokufunda Izihibe Zokucwaninga zeML

Iningi labantu lisasebenzisa i-AI njengebhokisi lokusesha lango-2015. Uyabhala, uyafunda, uyathayipha futhi. Iphethini entsha imiselela leyo manuwali emuva naphambili ngeluphu. Lo mhlahlandlela uyachaza ubunjiniyela be-loop usebenzisa ama-artifact amabili aqinisekisiwe. Imithombo ngeka-Andrej Karpathy autoresearch i-repository kanye ne- Bilevel Autoresearch iphepha. Uhlaka lulandela ukubhalwa ngu-@0xCodila.

Yini iLoop Engineering?

Ukuze uqale, qhathanisa izindlela ezimbili. Ukwaziswa kuwumyalelo owodwa, ngemva kwalokho unquma isinyathelo esilandelayo. Iluphu, ngokuphambene, inhloso imodeli ephishekelayo ize ifike. Imodeli ihlela, yenza, ihlole umphumela wayo, bese iphinda. Uchaza inhloso kanye, futhi iluphu iphatha ukuphindaphinda. Okubaluleke kakhulu, iluphu ithola izindleko zayo kuphela lapho umsebenzi ulinganiseka.

Izingxenye Ezintathu Ezenza Iluphu Isebenze

Ngakho-ke yini ehlukanisa iluphu yangempela ku-chatbot ngokuphinda? Yonke iluphu ethembekile inezingxenye ezintathu.

A isiqinisekisi amamaki umzamo ngamunye. Lelo sheke lingaba ukuhlolwa okuphasayo, imethrikhi ehambayo, noma ibhilidi. Ngaphandle kwesiqinisekisi, i-ejenti ivele ivumelane nayo ngokwayo ngokuphinda.
Izwe ibhala lokho okwazanywa, okwahluleka, nalokho okusele. Ifayela elincane eliseceleni livumela ukuqalisa okulandelayo kuqalise esikhundleni sokuqala kabusha.
A isimo sokumisa ivimbela izindleko zokubaleka. Iluphu iyama lapho umgomo ufinyelelwa, noma ngemva kwemizamo ka-N.

I-Karpathy Loop: Ngaphakathi kwe-'autoresearch'

Lezi zingxenye ezintathu azizona ithiyori. NgoMashi 7, 2026, uKarpathy wakhululwa autoresearchinqolobane yomthombo ovulekile ngaphansi kwelayisensi ye-MIT. Ithumela amafayela amathathu ayisisekelo kanye nemigqa yekhodi engaba ngu-630. Le phrojekthi yanda ezinsukwini ezimbalwa futhi manje isiseduze nezinkanyezi ze-GitHub ezingu-90,000. Yabuye yethulwa njengephethini “i-Karpathy Loop.”

Umklamo mncane ngamabomu, kodwa uqinile. I-ejenti ihlela kuphela train.pyephethe imodeli ye-GPT, i-optimizer (i-Muon ne-AdamW), ne-loop yokuqeqesha. Ayikwazi ukuthinta izinsiza zokuhlola ku prepare.py. Lokho kuhlukaniswa kumisa i-ejenti ekwenzeni ukuhlolwa kube lula esikhundleni semodeli ibe ngcono. Phakathi naleso sikhathi, umuntu uyabhala program.mdimiyalelo umenzeli okufanele ayihloniphe.

Umjikelezo ngamunye uqhuba isilingo esisodwa. I-ejenti ifunda ikhodi, iphakamise ushintsho, iqeqeshe imizuzu emihlanu, bese igcina noma ihlehle. Imethrikhi yamaphuzu ithi val_bpbizingcezu zokuqinisekisa ibhayithi ngayinye, lapho okuphansi kungcono khona. Leso sabelomali sikhiqiza cishe izivivinyo eziyi-12 ngehora, ngakho-ke cishe eziyi-100 zisebenza ngobusuku obubodwa.

Imiphumela ebikiwe iqinile. U-Karpathy ukhombe kokwakhe osekulungiselelwe nanochat Ikhodi yokuqeqesha ye-GPT-2. Ithathe izinsuku ezimbili futhi yaqeda ukuhlola okungaba ngu-700, igcina ukuthuthuka kwangempela okungu-20. Kuhlanganiswe ndawonye, lokho kulungiswa kunqamula isikhathi sokuqeqeshwa sekhwalithi ye-GPT-2 ngo-11%, sisuka ku-2.02 siye emahoreni angu-1.80. Okunye okutholakele kwaba a QK-Norm ukuqaliswa kushoda isiphindaphindi se-scalar, ebesishiye ukunaka kusabalele kakhulu kuwo wonke amakhanda.

Ngokuphawulekayo, abantu bayakhathala ngemva kokuhlolwa cishe kweshumi nambili, kuyilapho iluphu ayikhathali. Ngokuhlukana, isikhulu esiphezulu sakwaShopify uTobi Lütke wagijima autoresearch ubusuku bonke kumodeli yangaphakathi. Ubike ukuthuthuka okungu-19% ngemuva kokuhlolwa okungu-37. I-Karpathy's takeaway: uma unemethrikhi ephokophelwe, uyibhodlela.

I-Prompt vs Loop vs Bilevel Loop

Umehluko uya ucaca ngokuhlanganyela.

Isici	Umyalo wesibhamu esisodwa	I-Karpathy loop (`autoresearch`)	I-Bilevel Autoresearch
Uyachaza	Isinyathelo ngasinye	Igoli, kanye	Igoli, kanye
Ophindaphindayo	Wena	I-ejenti yangaphakathi	I-ejenti yangaphakathi + yangaphandle
Isiqinisekisi	Wena, ngokwenza	`prepare.py` (`val_bpb`)	Imethrikhi efanayo, amaleveli amabili
Izwe	Xoxa kuphela	Ilogi yokuhlola	Ilogi kanye nekhodi ejovwe
Iqhaza lomuntu	Injini	Umbhali we `program.md`	Umbhali we `program.md`
Umphumela obikiwe	Iyahlukahluka	700 ugijima → 20 ukulungiswa, 11% speedup	5x ukwehla okukhulu kwe-val_bpb

Amabhulokhi Wokwakha Amahlanu

Ngenxa yalokho, amaqembu onjiniyela be-AI manje ahlanganisa izihibe zokusebenza kusuka izingcezu ezinhlanu ezingasetshenziswa kabusha:

Okuzenzakalelayo ivula iluphu kushejuli, umcimbi, noma i-trigger.
A ikhono igcina ulwazi lwephrojekthi efayeleni lokumaka, elifundwa kukho konke ukugijima.
Ama-sub-ejenti hlukanisa umbhali kumbuyekezi, njengoba imodeli eyodwa ifaka amamaki ngokuphana kakhulu.
Izixhumi vumela iluphu isebenze ngaphakathi kwamathuluzi wangempela, njenge-tracker yenkinga noma i-Slack.
Ekugcineni, a isiqinisekisi kusasele isango elenqaba umsebenzi omubi. UClaude Code kanye ne-Codex manje sezithumela zonke ezinhlanu.

I-Bilevel Autoresearch: Iluphu Phezulu Kweluphu

Ngokulandelayo, abacwaningi babuza umbuzo obukhali. Uma i-autoresearch ingucwaningo, ungacwaninga ngokuzenzakalelayo ucwaningo? Iphepha locwaningo Bilevel Autoresearch: Meta-Autoresearching Itself uphendula yebo.

I iluphu yangaphakathi ihambisana nokwasekuqaleni kukaKarpathy: phakamisa, qeqesha, hlaziya, gcina noma ulahle. I iluphu yangaphandle ubuka iluphu yangaphakathi futhi ifunde ikhodi yayo kanye nemikhondo. Ikhomba lapho ukusesha ngokwako kuhlala kumile. Bese ibhala izindlela ezintsha zePython, izijove ngesikhathi sokusebenza, bese iphinda isebenzise iluphu yangaphakathi.

Umphumela ubambe ibhentshimakhi ye-Karpathy's GPT pretraining. Iluphu yangaphandle isikiwe val_bpb 5x ngaphezu kweluphu eyodwa (-0.045 vs -0.009). Ngokuphawulekayo, amalophu womabili asebenzise i-LLM efanayo, ngakho-ke inzuzo iqhamuke ekwakhiweni kwezakhiwo, hhayi imodeli ehlakaniphile. Ngokwenza umklamo uhlukana ube amazinga amathathu. Ileveli 1 isebenzisa iluphu eyisisekelo. Izinga 1.5 zamapharamitha wokusesha njalo ekuphindaphindweni okuhlanu. Ileveli 2 ikhiqiza izindlela ngeseshini eyimizuliswano emine. Ukuhlolwa okubikiwe kusebenzise i-RTX 5090 32GB kanye nesabelomali semizuzwana engama-300.

Isizathu kufanele siphawulwe. I-loop yangaphakathi ibilokhu ibuyela kokubalulekile okufanayo, ngisho nangemva kokuyeka ukusebenza. Iluphu yangaphandle yephule lawo maphethini ngokuphoqa ukuhlola okungajwayelekile.

Sebenzisa Amacala anezibonelo

Le mibono idlulisa kahle ngaphezu kokuqeqeshwa kwangaphambili. Ngomsebenzi wemodeli, iluphu isesha ama-hyperparameter kuze kube val_bpb amaconsi. Okwesofthiwe, yenza kabusha kuze kube yilapho izivivinyo, izinhlobo, kanye nokudlula kokwakha. Kokuqukethwe, ibhala kabusha kuze kube yilapho wonke amaphuzu erubrikhi esusa umkhawulo. Ukuze uthole idatha, ishuna ipayipi kuze kube yilapho i-schema sibheka ukubanjwa. Icala ngalinye linesici esisodwa: isango elizenzakalelayo elingahluleka umsebenzi.

Zizame ngokwakho: I-Loop In One Prompt

Ithiyori eceleni, ungamuzwa umakhenikha ngaphandle kwe-Claude Code noma i-Codex. Namathisela lokhu kunoma iyiphi imodeli enekhono futhi uyibuke izilungisa yona.

You will work in a loop until the task meets the bar.

TASK:
[describe exactly what you want produced]

SUCCESS CRITERIA (be strict):
- [criterion 1]
- [criterion 2]
- [criterion 3]

LOOP PROTOCOL, repeat every turn:
1. PLAN   - state the single next step.
2. DO     - produce or improve the work.
3. VERIFY - score the result 1-10 on each criterion. Be honest.
4. DECIDE - if every criterion is 8+, print FINAL and stop.
            Otherwise print ITERATING and fix the weakest point first.

RULES:
- Never call it done until every criterion is 8 or higher.
- Each pass must fix the weakest score from the last VERIFY.
- Do not ask questions. Make a sensible assumption and continue.
Begin.

Ngaphansi, ukugeleza kokulawula kuncane. Uhlaka lwamathambo olungezansi lubonisa lezo zingxenye ezintathu ku-Python: isiqinisekisi, isinqumo, nezimo ezimbili zokuma.

current = baseline
best = evaluate(current)                 # verifier: lower val_bpb is better
for step in range(MAX_STEPS):            # stop condition 1: experiment budget
    candidate = propose_change(current)  # agent edits train.py
    score = train_and_eval(candidate)    # train 5 min, then verify
    if score < best:                     # keep only real improvements
        current, best = candidate, score # commit
    # else: discard candidate, restore baseline
    if best <= TARGET:                   # stop condition 2: goal met
        break

Zombili izinguqulo zilinganiselwe. Useyicupha, futhi ukuvala ithebhu kusula isimo. Ukwengeza i-automation, ifayela lesimo, nesango langempela liphendulela lokhu ku-loop ezimele.

Ibone Igijima

Idemo esebenzisanayo engezansi igqwayiza iluphu eyodwa egcwele: phakamisa, qeqesha, qinisekisa, bese ugcine noma ubuyele emuva. Lungisa ithagethi kanye nomkhawulo wesinyathelo, futhi ubuke val_bpb ukuwa kuze kube yilapho isimo sokumisa sivutha.

1){ cx.strokeStyle=”#76B900″; cx.lineWidth=2.2; cx.beginPath(); state.hist.forEach((v,i)=>{ i?cx.lineTo(X(i),Y(v)):cx.moveTo(X(i),Y(v)); }); cx.stroke(); // gcwalisa cx.lineTo(X(state.hist.length-1),H-18); cx.lineKu(X(0),H-18); cx.closePath(); cx.fillStyle=”rgba(118,185,0,.10)”; cx.gcwalisa(); // iphuzu lokugcina const li=state.hist.length-1; cx.fillStyle=”#5c9000″; cx.beginPath(); cx.arc(X(li),Y(state.hist[li]),3.4,0,7); cx.gcwalisa(); } } // ———- ui ———- umsebenzi paintStats(){ $('s-exp').innerHTML=state.exp+'/'+state.maxExp+''; $('s-best').textContent=state.best.toFixed(4); $('s-kept').textContent=state.kept; $('s-rb').textContent=state.rb; } umsebenzi clearStages(){stages.forEach(s=>s.className=”stage”);} umsebenzi okhanyayo(i){clearStages(); izigaba[i].classList.add('on');} umsebenzi welogi(html,cls){ const d=document.createElement('div'); d.className=”ln”; d.innerHTML=html; const L=$('log'); L.appendChild(d); L.scrollTop=L.scrollHeight; } const sleep=ms=>Isithembiso esisha(r=>setTimeout(r,ms)); // ———- iluphu ———- umsebenzi we-async runLoop(){ uma(ugijima) uyabuya; ukugijima=iqiniso; $('btn-run').disabled=iqiniso; $('log').innerHTML=''; log('Umenzeli we-$: funda uhlelo.md → qala ukuhlola'); // setha kabusha isimo sokusebenza, gcina izilawuli state.best=1.20; state.exp=0; state.kept=0; state.rb=0; state.hist=[1.20]; state.mechs=0; paintStats(); drawChart(); let improveScale=0.020; // ubukhulu bokuthuthukiswa obujwayelekile lapho ushintsho lusiza ukunamathela=0; // imigijimo elandelanayo yokungazuzi (i-loop yangaphakathi ithafa) ngenkathi(state.exp<state.maxexp state.best="">state.target){ state.exp++; // 01 phakamisa lit(0); $('isimo').textContent=”Isivivinyo “+state.exp+' · iphakamisa ushintsho ku-train.py…'; lindela ukulala(220); // iluphu yangaphandle (ye-bilevel): yonke imigijimo embalwa ebambekile, faka indlela entsha yokusesha vumela ijovwe=amanga; uma(state.bilevel && stuck>=3){ state.mechs++; ibanjiwe=0; ngconoIsikali*=1.6; // umfanekiso: uphula ithafa elijovwe=iqiniso; log('↻ iluphu yangaphandle: ukusesha kumisiwe → indlela ejovwe #'+state.mechs+' (hlola izikhombisi-ndlela ezintsha)'); lindela ukulala(340); } // 02 isitimela (philisa ibhajethi yamaminithi angu-5) sikhanyisiwe(1); $('isimo').textContent=”Isivivinyo “+state.exp+' · ukuqeqeshwa (isabelomali samaminithi angu-5)…'; const tb=$('tbar'); tb.style.width=”0%”; kokuthi(vumela p=0;p<=100;p+=8){ tb.style.width=p+'%'; ulinde ubuthongo(16); } // 03 qinisekisa — khiqiza ikhandidethi val_bpb lit(2); $('isimo').textContent="Isivivinyo "+state.exp+' · iqinisekisa i-val_bpb…'; lindela ukulala(200); const helps=Math.random() < (state.bilevel?0.55:0.42); // ingxenye yezinguquko ezisiza vumela i-cand; if(helps){ const gain=improveScale*(0.4+Math.random()); cand=Math.max(state.target-0.02, state.best-gain); }okunye{ cand=state.best + Math.random()*0.03; // ukuhlehla } // 04 nquma ukuthi(ikhandi < state.best-1e-6){ state.best=cand; state.igcinwe++; ibanjiwe=0; izigaba[3].className="iphasi yesiteji"; izigaba[2].className="iphasi yesiteji"; log('exp '+String(state.exp).padStart(2,'0')+' · val_bpb '+cand.toFixed(4)+' ✓ GCINA'+(ijovwe?' (ngemuva kokujova)':')); }okunye{state.rb++; ibanjiwe++; izigaba[3].className=”ukuhluleka kwesiteji”; izigaba[2].className=”ukuhluleka kwesiteji”; log('exp '+String(state.exp).padStart(2,'0')+' · val_bpb '+cand.toFixed(4)+' ↻ BUYELA EMUVA'); } tb.style.width=”0%”; state.hist.push(state.best); paintStats(); drawChart(); lindela ukulala(260); } clearStages(); const hit=state.best<=state.target; $('isimo').textContent = hit ? 'Isimo sokumisa: okuqondiwe kufinyelelwe esivivinyweni '+state.exp+'.' : 'Isimo sokumisa: isabelomali sokuhlola ('+state.maxExp+') sifinyelelwe.'; log('– iluphu imisiwe – best val_bpb '+state.best.toFixed(4)+' · igcinwe '+state.igcinwe+' · ihlehliselwe emuva '+state.rb+ (state.bilevel?(' · izindlela ezijovwe '+state.mechs):'')); ukugijima=amanga; $('btn-run').disabled=false; postHeight(); } function resetAll(){ if(running) return; state.best=1.20; state.exp=0; state.kept=0; state.rb=0; state.hist=[1.20]; state.mechs=0; clearStages(); $('tbar').style.width=”0%”; $('log').innerHTML=''; $('isimo').textContent=”Ilungile. Cindezela okuthi “Run loop” ukuze uqale.”; paintStats(); drawChart(); } // ———- izilawuli zezintambo ———- $('i-target').addEventListener('input',e=>{state.target=+e.target.value; $('v-target').textContent=state.target.toFixed(3); uma(!Iqala);{w} $('izinyathelo').addEventListener('input',e=>{state.maxExp=+e.target.value; $('v-steps').textContent=state.maxExp; uma(!igijima){paintStats();drawChart();}}); $('i-bilevel').addEventListener('shintsha',e=>{state.bilevel=e.target.checked; $('tg').classList.toggle('amberon',e.target.checked); $('c-tag').textContent=e.target.ihloliwe?'okuphansi kungcono; ·} i-bilevel iqhubeke'; $('btn-run').addEventListener('chofoza',runLoop); $('btn-reset').addEventListener('chofoza',resetAll); // ———- shintsha usayizi + ukuphakama okuzenzakalelayo kwe-WordPress ———- umsebenzi postHeight(){ const h=document.documentElement.offsetHeight+40; if(window.parent) window.parent.postMessage({loopDemoHeight:h},'*'); } umsebenzi onResize(){fitCanvas(); drawChart(); postHeight();} window.addEventListener('resize',onResize); // ———- init: amanani wangempela ekulayisheni ———- fitCanvas(); paintStats(); drawChart(); postHeight(); setTimeout(postHeight,300); })(); </state.maxexp>

Source link

nimda 2 hours ago

0 0 5 minutes read