+− THE DAILY DIFFdev & AI news
REVERT

U-Armin Ronacher ushiye i-GPT-6 Astra yodwa amahora angu-35.

U-Armin Ronacher (Flask, Jinja) unikeze i-GPT-6 Astra umyalo owodwa wayishiya yodwa amahora angu-35: cishe izimpawu eziyibhiliyoni, cishe u-$1,200, ama-commit angu-79, imigqa engu-75,000 — futhi "akukho lutho olubalulekile".

U-Armin Ronacher (Flask, Jinja) unikeze i-GPT-6 Astra umyalo owodwa wayishiya yodwa amahora angu-35: cishe izimpawu eziyibhiliyoni, cishe u-$1,200, ama-commit angu-79, imigqa engu-75,000 — futhi "akukho lutho olubalulekile". Ikhodi eliyitholile yiyona ndaba: amafayela e-C alungiswe yi-Python string-splicing heredocs, i-Python izala i-Node izala i-PowerShell, izivivinyo zeyunithi ezisuswe isikhala esingenalutho ngoba kushibhile ngo-10% ezimpawini. Izilinganiso ezimbili ziyamsekela: izinombolo ze-Earendil's SlopCodeBench (ikhodi ye-agent iphindwe kabili ngokweqile futhi yonakele kunezindawo zokugcina zabantu, 0% strict pass rate) kanye nesilinganiso sika-Quesma sika-$1,500 se-RTK (izinkanyezi ezingu-79k, "89% zezimpawu zigciniwe" ku-counter yayo, +17% ngomsebenzi ngamunye nge-DeepSeek). Futhi: i-Cognition's SWE-2 (i-Kimi K3 egqoke i-trench coat, 92.8 ku-Terminal-Bench 2.1 vs 27.3 ku-Terminal-Bench 4), i-OpenAI's Agents API kanye nokumiswa kokubhalisa kwe-Pro ka-$200, kanye ne-Friday Hacker News ecele izindaba ezimbalwa ze-AI. Isinqumo: REVERT.

Funda uhlelo olubhaliwe (isiNgisi) ↗

Lokho okuhlanganiswa yile vidiyo

  • U-Armin Ronacher: amahora angu-35 e-GPT-6 Astra engagadiwe, ~$1,200, ama-commit angu-79, imigqa engu-+75k, akukho okuthunyelwe
  • I-Earendil / SlopCodeBench: ikhodi ye-agent cishe iphindwe ka-2 ngezwi nangokuguguleka kunezindawo zokugcina zabantu; izinga lokupasa eliqinile lingu-0%
  • I-Quesma: I-RTK ibika u-89% wezimpawu zigciniwe, kodwa izindleko ze-Terminal-Bench zikhuphuka ngo-17% nge-DeepSeek; i-rewrite loop yehluleke izikhathi ezingu-339 zilandelana
  • I-Cognition SWE-2: iqeqeshwe ngemuva kwe-Kimi K3, ifana ne-Fable 5.1 ku-FrontierCode ngengxenye yesithathu yentengo; i-TB 2.1 92.8 vs TB 4 27.3; i-Devin Voice
  • I-OpenAI Agents API (i-Codex harness njengesevisi, e-US kuphela, akukho ZDR); ukubhalisa kwe-Pro ka-$200 kumisiwe ngenxa yesidingo se-Astra

Umbhalo obhaliwe ohunyushwe

Kuhunyushwe kusuka ekulandiseni kokuqala kwesiNgisi. Umsindo namazwibela atholakalayo alawulwa yi-YouTube.

0:00 U-Armin Ronacher, indoda eyabhala i-Flask, wanikeza i-GPT-6 Astra umyalo owodwa futhi wayishiya yodwa amahora angamashumi amathathu nanhlanu. Yabuya nemigqa eyizinkulungwane ezingamashumi ayisikhombisa nanhlanu yekhodi futhi, ngamazwi akhe, akukho lutho olubalulekile, okwenza kube yifekthri yesofthiwe engokoqobo kakhulu eyakhethwa. Utholele ukubhala ngoLwesithathu. NgoLwesine, i-Cognition yathumela i-SWE-2, kanti i-OpenAI yathumela i-Agents API futhi yamisa ukubhalisa kwesikimu sayo samadola angamakhulu amabili,

0:24 ngoba i-Astra yadla amaseva. Futhi ngoLwesihlanu ukubhala kwafinyelela ekhasini lokuqala le-Hacker News, imigqa embalwa ngezansi kokuthunyelwe okucela i-Hacker News ukuthi iyeke ukuthumela ngokuphathelene ne-AI. Kule vidiyo: okukhiqizwa usuku olulodwa nengxenye lwe-Astra engagadiwe, izilinganiso ezimbili eziguqula udaka lube inombolo, ukuthi kungani ithuluzi elinezinkanyezi ezingamashumi ayisikhombisa nanhlanu izinkulungwane zenza isikweletu sakho sibe sikhulu, nokuthi i-Hacker News yazama kanjani ukuhlunga i-AI, isebenzisa i-AI. NgoLwesihlanu, Septhemba 11, futhi lena yi-The Daily Diff.

0:53 Ifekthri. Umyalo owodwa: yakha i-Python enezintambo ezibonakalayo kanye ne-lexical ukucacisa. I-Astra yazigcina amanothi ayo, yavula ama-subagent ayo, futhi yagijima waze u-Armin wamisa, usuku olulodwa nengxenye kanye namadola angamakhulu ayishumi nambili ngemuva. Ama-commit angamashumi ayisikhombisa nane, ngakho amadola ayishumi nanhlanu nengxenye nge-commit ngayinye, okulingana nalokho umuntu akukhokhisayo, ngaphandle kokuthi umuntu uyagcina ngokuqeda. Ikhodi iyona ndaba. Esikhundleni sethuluzi lokuhlela, i-Astra ilungisa amafayela e-C ngokudlulisa ama-heredoc e-Python afunda ifayela, afake umsebenzi, futhi awubhale kabusha.

1:20 Ukuze iqhube isivivinyo esisodwa se-Windows yabha i-Python eyazala i-Node eyazala I-PowerShell, i-call stack enepasipoti. Izivivinyo zeyunithi ezizifaka azinazo izikhala ezimhlophe nhlobo, ngoba ngaphandle kwe-indentation zishibhile ngo-10% ezimpawini. Futhi uhlu lwemisebenzi luvela ku-one, two, three kuya kumsebenzi 8b2c2b2b checkpoint one, okuyindlela owazi ngayo ukuthi i-intern ibe yodwa isikhathi eside kakhulu. Ithiyori ka-Armin: imodeli iklonyeliswa ngokuqeda imisebenzi emide futhi cishe ayi ijeziswa ngekhodi embi, ngakho izihambisi zethuluzi ezifakwe izimpawu zivuza ku-codebase,

1:48 futhi abantu abambalwa ababukayo, akubalulekile kangako. Uleso sigaba usibiza ngokuthi I-AGI Uma Ungabheki. I-Hacker News yengeze ezomnotho: ikhodi eningi isho izimpawu eziningi zokuyinakekela, okwenza udaka kungabi iphutha kodwa ukubhalisa. Ungakulinganisa udaka? Inkampani ka-Armin yazama kuleli sonto. U-Earendil wagijima izinombolo ze-SlopCodeBench: ikhodi ye-agent cishe iphindwe kabili ngezwi futhi iphindwe kabili yonakele kunezindawo zokugcina zabantu eqhathaniswa nazo,

2:10 futhi lapho isilinganiso sisula inkumbulo ye-agent phakathi kwamaphoyinti okuhlola, izinga lokupasa eliqinile layo yonke imodeli abayihlolile lingu-zero. Isilinganiso sodaka esithembeke kakhulu abasitholile bekuyinombolo yomugqa eluhlaza, eyeka ukusebenza uma umuntu eyilungiselela, ngakho sicela ungatsheli amamodeli. Bese kuba yi-wallet. I-RTK inezinkulungwane ezingamashumi ayisikhombisa nane zezinkanyezi ze-GitHub nezithombe ze-YouTube ezithembisa uku nciphisa isikweletu sakho se-Claude Code; icindezela ukuphuma kwe-terminal ngaphambi kokuba i-agent

2:34 iyifunde. I-Quesma yayisebenzisa ku-Terminal-Bench ngamakhulu ayishumi nanhlanu amadola ezimpawu. Nge-Fable isikweletu sehla ngamaphesenti amahlanu, cishe konke okuvela kumsebenzi owodwa; nge-DeepSeek umsebenzi ojwayelekile waba expensive ngamaphesenti ayishumi nesikhombisa. Ngaleso sikhathi i-counter ye-RTK yabika u-89% ogciniwe, ngoba ibala amabhayithi asusiwe, hhayi ukujika okwengeziwe okubangelwe: imitha ehlakaniphile e ikhokhisa umakhelwane. Futhi iphutha elilodwa lenguqulo libhale kabusha i-find ibe umyalo owahluleka, izikhathi ezingamakhulu amathathu namashumi amathathu nane zilandelana, ngentengo ephindwe kayisishiyagalolunye.

3:01 Ithuluzi elibulala izimpawu line-loop. Isikhukhula ngokwaso. I-Cognition's SWE-2 iyi-Kimi K3, imodeli yaseShayina evulekile, eqeqeshwe ngemuva yaze yafana ne-Fable 5.1 kusilinganiso se-Cognition ngesigaba esithathu sentengo; I-Hacker News: kungani wonke amamodeli aseMelika e-AI empeleni angama-Kimi egqoke i-trench coat. Ku-Terminal-Bench 2.1 ihamba phambili etafuleni lonke; ku-Terminal-Bench 4, leyo akekho oyikhumbule okwamanje, ithola amashumi amabili nesikhombisa kuya kwamashumi amahlanu nesithupha e-Fable,

3:28 ngakho kumasilinganiso e-slide-deck ikholamu engcono kakhulu isivivinyo wonke umuntu asevele pasile. I-Cognition futhi yamemezela i-Devin Voice, unjiniyela wakho wesofthiwe we-AI osanda kuthola i-landline, ngoba into eyodwa eyayishoda ekubhaleni ikhodi nge-agentic bekuyimculo wokubamba. I-OpenAI, ngalo lolo suku: i-Agents API, okuyi-Codex harness ethengiswa njengesevisi. I-OpenAI iqhuba i-sandbox, indawo yokuhlala yedatha yase-US kuphela, akukho ukugcina idatha nge-zero, ngakho i-agent ikhumbula i-codebase yakho kahle njengoba i-ChatGPT yenza. Bese i-OpenAI yamisa ukubhalisa kwesikimu se-Pro samadola angamakhulu amabili,

3:57 ngoba isidingo se-Astra, quote, asikaze sibe khona: umkhiqizo wokuqala onoluhlu lokulinda ukukhokha okwengeziwe. U-Armin washisa i-reset ephelele efekthri yakhe. Yena uyisidingo. Okuholela ku-Ask HN kaLwesihlanu: singakwazi yini ukunciphisa isikhukhula sezindaba ze-AI, amakhulu ayisithupha namashumi amahlanu nesithupha amaphuzu, ngoba, quote, njalo uma umuntu e-OpenAI noma I-Anthropic iqhuqha, kuba khona okuthunyelwe okuyishumi phezulu.

4:17 Izimpendulo bekungamahlelo: i-hcker dot news isusa izindaba ze-AI nge-BERT i-classifier eqeqeshwe ezinkulungwaneni eziyisishiyagalombili zezibonelo, i-AI yokususa i-AI, eyondliwe utshani; futhi i-uBlock regex efanisa i-A-I ngokunganaki icala futhi isusa i-imeyili, nsuku zonke, main, domain, kanye ne-train, ngakho i-regex, njengenjwayelo, ingumlimi. Ngalobo ntambama, amakhulu amane neshumi nanhlanu amazwana baphikisana ngekhasi lokusekela le-Claude lithatha ukuthi u-Claude uneminyaka engu-ishumi nesishiyagalombili nangaphezulu.

4:38 Ikhasi libhalwe ngoMeyi. Akukho okushintshile. Ngisho nezindaba ze-AI ezingezona izindaba ziyizindaba, okuyinto, ukulunga, nayo imodeli yami yebhizinisi. Uma ungathanda ukufunda lokhu kunokungizwa ngikushoyo, i-diff ifika ebhokisini lakho leposi njalo ekuseni — mahhala ku-the daily diff dot dev, isixhumanisi ngezansi. Kuyizindaba ze-AI, ngokungenakugwenywa. Ngakho — isinqumo sanamuhla ngalefekthri yesofthiwe yamahora angamashumi amathathu nanhlanu: revert. Ama-commit angamashumi ayisikhombisa nane akekho owawafunda akuyona i-codebase, iyisikweletu esine-git

5:04 log; i-Astra iyamangalisa, futhi ifekthri yingxenye okumele ibuyiselwe emuva. Futhi yilokho i-diff yanamuhla. Ngingu-Niko wase-Axrisi. Hlanganisa ngokunomthwalo wemfanelo.

Imithombo

  1. Armin Ronacher — Astra for Coding: Why Are We Doing This Again?lucumr.pocoo.org
  2. HN: Astra for Codingnews.ycombinator.com
  3. Earendil — Measuring the sloppiness of codeearendil.com
  4. HN: Measuring the sloppiness of codenews.ycombinator.com
  5. Quesma — RTK reports huge token savings, but our cost benchmarks disagreequesma.com
  6. HN: RTKnews.ycombinator.com
  7. RTK (Rust Token Killer)github.com
  8. Cognition — Introducing SWE-2cognition.com
  9. HN: Cognition SWE-2news.ycombinator.com
  10. OpenAI — Agents APIdevelopers.openai.com
  11. HN: OpenAI Agents APInews.ycombinator.com
  12. TechCrunch — OpenAI puts Pro subscriptions on hold due to Astra demandtechcrunch.com
  13. Ask HN: Can we please limit the AI news flood?news.ycombinator.com
  14. HN: Claude is only available to people over 18 yearsnews.ycombinator.com

Amavidiyo ahlobene