Indlela yokuthola i-Claude nerf (ngedatha yangempela)
Abantu bathi uClaude uba yisiphukuphuku emasontweni ambalwa ngemuva kokwethulwa ngakunye.
Abantu bathi uClaude uba yisiphukuphuku emasontweni ambalwa ngemuva kokwethulwa ngakunye. Omuntu othile, umthuthukisi, wafaka i-Claude Opus 5.5 ewashini lezinsuku ezingama-30 kusukela eviki lokwethulwa, ngemibuzo efriziwe, i-Claude Code ehlanganisiwe nemithetho yomphakathi, futhi lokho kukhombisa ukuthi kungani kungekho muntu obengazi ngaphambili, nokuthi kungani inani lakho lamathokheni kuyinto okufanele uyibheke. Isinqumo: SHIP IT.
Funda uhlelo olubhaliwe (isiNgisi) ↗
Lokho okuhlanganiswa yile vidiyo
- UClaude uba yisiphukuphuku ngemuva kokwethulwa: ithiyori ihlangana newashi losuku oluyiziro
- i-livenerf: iwashi lezinsuku ezingama-30 ku-Opus 5.5 kusukela eviki lokwethulwa
- Indlela yokubamba i-nerf: imibuzo engama-78 ngezinye izikhathi ilungile
- Lokho ekwaziyo ukukubona: amaphuzu angu-7.5, kanti inani lamathokheni linyakaza kuqala
- Ibala elingaboni: ukushintshanisa kwe-Opus 5 nezinkinobho ezi-8 zezimpendulo ezingalungile
Imibuzo nezimpendulo
Ungahlola kanjani ukuthi imodeli ye-AI yenze kabi kancane ngemuva kokwethulwa?
Sebenzisa imibuzo efanayo eqandisiwe ngeshejuli ehleliwe bese uyihlela ngokuqondanisa okuqondile. Indawo yokugcina ye-livenerf isebenza ihora lezinsuku ezingama-30 ku-Claude Opus 5.5, kanye ngosuku, enamalogi okungeke ahlelwe muntu. Isebenzisa imibuzo engama-78 imodeli engayithola ngendlela efanele ngezikhathi ezithile, ngoba umbuzo ewuphendula ngaso sonke isikhathi ngendlela efanele awukwazi ukuwa. Ihlanganisa inguqulo ye-Claude Code, ngoba i-harness eshintshiwe ibukeka njengemodeli eshintshiwe.
Iluphi usayizi wokuwa olungabonwa imitha ye-livenerf?
Ngewindi lezinsuku eziyishumi, imitha ingabona ukuwa kwamaphuzu angaba ngu-7.5. Ukuze afakazele ukuthi indawo isebenza, umbhali wanciphisa umzamo ka-Claude ngenhloso. Umzamo omncane unciphise okukhiphayo cishe ngengxenye yesithathu namaphuzu ayisishiyagalombili. Ukufaka i-Opus 5 endala kwakuyinguquko encane kakhulu ukuze indawo ibize, futhi i-readme isho lokho kusengaphambili.
Yini okufanele abathuthukisi bayilandele ukuze babone ukuthi i-ejenti ye-AI icabanga kancane?
Ukukhipha isibalo se-token ngomsebenzi. Ezivivinyweni ze-livenerf, ukucabanga okuncane kwavela esibalweni se-token ngaphambi kokuba kuvele ezimpendulweni. Isiqephu siphakamisa ukuhlanganisa inguqulo yemodeli yakho, ukungena kwamathokheni okukhiphayo ngomsebenzi, nokugcina imibuzo yakho eqandisiwe embalwa. Uma i-ejenti yakho ibhala kancane ngokungazelelwe, hlola amalogi akho kuqala.
Ingabe u-Anthropic uthi unciphisa ikhwalithi ka-Claude lapho isidingo siphezulu?
U-Anthropic uthi akakwenzi lokho. I-postmortem yayo yangoSepthemba 2025 ithi ayikaze inciphise ikhwalithi yemodeli ngenxa yesidingo, isikhathi sosuku noma umthwalo weseva. Iposi elifanayo liyavuma ukuthi iziphazamiso ezintathu zazicekela phansi u-Claude, nokuthi cishe ingxenye yesithathu yabasebenzisi be-Claude Code bashaya amaseva angalungile okungenani kanye.
Ingabe imiphumela ye-livenerf ingathembeka ukukhombisa i-Claude nerf okwamanje?
Isinqumo se-The Daily Diff sasiyi-SHIP IT yephrojekthi, ngoba yiyo impikiswano yokuqala ye-nerf enesitembu sesikhathi, incwadi yemithetho yomphakathi kanye nendawo eyimfihlo eshiwo. Isiqephu sithi ungacaphuni ngaphambi kuka-Okthoba 24, ngoba kwakungaphelelanga izinsuku eziyisithupha kwezingamashumi amathathu futhi ucingo lokuqala olungenzeka lufika ngaleso sikhathi. Ukuhlola futhi kwathola izinkinobho zempendulo eziyi-8 ezibukeka zingalungile.
Umbhalo obhaliwe ohunyushwe
Kuhunyushwe kusuka ekulandiseni kokuqala kwesiNgisi. Umsindo namazwibela atholakalayo alawulwa yi-YouTube.
UClaude uba yisiphukuphuku ngemuva kokwethulwa: ithiyori ihlangana newashi losuku oluyiziro
0:00 Wonke umuntu uyazi ukuthi uClaude uba yisiphukuphuku emasontweni ambalwa ngemuva kokwethulwa. Ngakho-ke omunye umthuthukisi wafaka i-Opus five point five ewashini kusukela eviki lokwethulwa, futhi iwashi lithi akekho obekangazi okwamanje. Kule vidiyo, imibuzo emithathu. Uyibamba kanjani i-nerf? Yini engabonwa yile mitha? Futhi yini ethi i-Anthropic? Futhi umugqa owodwa ezigabeni ze-repo wangihleka kakhulu.
0:20 Ngizokufika ekugcineni. NguLwesithathu, Septhemba thirtieth, futhi lena i-The Daily Diff. Izindaba ezintathu namuhla, kanti enkulu yi-GitHub repo ebizwa nge-live nerf.
i-livenerf: iwashi lezinsuku ezingama-30 ku-Opus 5.5 kusukela eviki lokwethulwa
0:30 Sekuyizinyanga, abantu bathi i-Anthropic ihlela buthule amamodeli ayo ngemuva kokwethulwa. Izithiyo ezijwayelekile yimodeli ecindezelwe, imodeli encane ngaphansi kwegama elifanayo noma ukucabanga okuncane nje. I-repo ibiza yonke impikiswano kuze kube manje ngokuthi ama-vibes vs ama-vibes. I-Opus five point five yathunyelwa ngoSepthemba twenty-second, futhi ngesonto eledlule ngakutshela ukuthi yahlala phezulu ebhodini elizimele ngenkathi ibhala amagama aphindwe kathathu kunemodeli ephakathi. Ezinsukwini ezimbili nengxenye, umthuthukisi ogama lakhe lingu-ninja hawk waqala iwashi,
0:59 kanye ngosuku izinsukukuma-30, ngamarekhodi okungenzeka kungabi namuntu ongawahlola. Uchungechunge lwe-Hacker News lwafinyelela amaphuzu acishe abe ngama-800 futhi lwahlukana njengeqembu ingxoxo. Omunye obeka amazwana uthi ukulungisa amamodeli akuyona into yangempela kuyo yonke iningi lamacala abikiwe. Omunye uthi abantu bajwayela izinga elisha lokuhlakanipha. Futhi omunye umsebenzisi omkhulu we-Claude Code manje usenqaba okuthi rate-this-session pop-up ngaso sonke isikhathi, ngoba ukulinganisa uClaude kwabonakala kukwenza kube kubi. Ukubuyekezwa kontanga. Ngakho, umbuzo wokuqala, uyibamba kanjani i-nerf?
Indlela yokubamba i-nerf: imibuzo engama-78 ngezinye izikhathi ilungile
1:27 Uqala ngemibuzo yokuhlolwa enzima ecishe ibe ngu-2300, futhi i-Opus ithola amaphesenti angama-93 alungile ekuzameni kokuqala, okungenasidingo kumtshina, ngoba umbuzo ewuhlala ewuthola ulungile awukwazi ukuwa. Amaphesenti angama-97 ahlala elungile noma ehlala engalungile, futhi angama-78 ewathola ngezinye izikhathi kuphela abe yiphaneli. Ukubuza okufanayo, akukho mathuluzi, nokulinganisa okufanayo ngaphandle kwejaji le-AI, ngoba nejaji belizolayisha kakhulu. Lisebenza kwi-Max subscription nge-headless Claude Code,
1:55 njengoba inguqulo ye-API ibinamani acishe abe yi-1600 zamadola ngenyanga. Futhi inguqulo ye-Claude Code ihlanganisiwe, ngoba, ngamazwi e-repo, iharness eshintshiwe ibukeka kahle njengemodeli eshintshiwe. Izibalo zivela ephepheni elibizwa nge-Adding Error Bars to Evals, ngu-Evan Miller e-Anthropic. Ngakho-ke izibalo ze-Anthropic manje zihlole i-Anthropic, okuyinguqulo yezifundo yokucaphuna imigomo yesevisi kubasebenzisi besevisi.
Lokho ekwaziyo ukukubona: amaphuzu angu-7.5, kanti inani lamathokheni linyakaza kuqala
2:19 Umbuzo wesibili, yini ekwaziyo ukuyibona? Efasiteleni lezinsuku eziyishumi, ukwehla okucishe kube ngamaphuzu ayisikhombisa nengxenye. Ukuze afakazele ukuthi i-rig iyasebenza, umbhali wehlise umzamo kaClaude ngenhloso. Umzamo ophakathi wehlise ukuphuma ngekota, futhi isikolo ngamaphuzu amane kuphela. Umzamo ophansi wehlise ukuphuma cishe ngezingxenye ezimbili kwezintathu, ngamaphuzu ayisishiyagalombili. Leyo yingxenye okufanele yebiwe. Ukucabanga okuncane kubonakala enanini lamathokheni ngaphambi kokuba kubonakale ezimpendulweni.
2:43 Ngakho-ke hlanganisa inguqulo yakho yemodeli, ubhale amathokheni okuphuma ngomsebenzi, futhi ugcine imibuzo embalwa eqinile yakho. Uma i-agent yakho ngokungazelele ibhala kafushane, hlola amalogi akho ngaphambi kokuhlola i-Reddit. Futhi nansi ingxenye eqotho.
Ibala elingaboni: ukushintshanisa kwe-Opus 5 nezinkinobho ezi-8 zezimpendulo ezingalungile
2:54 Ukushintshanisa buthule i-Opus five endala kwakushintsha okuncane kakhulu ukuthi i-rig ingakubiza, futhi i-readme ikusho ngaphambili. Ukuhlolwa kuthola nezinkinobho zezimpendulo eziyisishiyagalombili ezibukeka zingalungile, ngakho-ke umtshina we-nerf wathola iziphazamiso ku-benchmark ngaphambi kokuthola i-nerf.
I-Anthropic: asikaze sinciphise ikhwalithi yemodeli
3:08 Umbuzo wesithathu, ithini i-Anthropic? Nyakenye, ngemuva kogagasi lwezikhalazo, i-Anthropic yashicilela i-postmortem. Ithi, ngicaphuna, asikaze sinciphise ikhwalithi yemodeli ngenxa yesidingo, isikhathi sosuku noma umthwalo weseva. Okufanayo lokho okukhishiwe kuvuma iziphazamiso ezintathu zazenze uClaude waba mubi, futhi cishe ingxenye yesithathu yabasebenzisi be-Claude Code bashaya amaseva angalungile okungenani kanye. Ngakho-ke izikhalazo zaziyiqiniso futhi imbangela yayiyiziphazamiso, okuyisizathu sokuba i-repo ixwayise ukuthi iviki lokwethulwa lingaba yiviki elibi kakhulu.
3:35 Izinsuku eziyisithupha kwezingama-30 zingaphakathi. Ucingo lokuqala olungenzeka luhla ngomhla ka-Okthoba twenty-fourth, ngakho-ke impendulo eqotho ukuthi akekho owaziyo, kubandakanya wonke umuntu owayeqiniseka. Ekhuluma ngezibalo okungekho muntu ozishicilelayo.
Izikhungo zedatha zigcina izinombolo zazo ziyimfihlo; i-Backblaze iyazishicilela
3:46 I-Lighthouse Reports nephephandaba laseDutch i-Trouw yathola ukuthi iningi kakhulu lezikhungo zedatha zaseYurophu zigcina ukusetshenziswa kwamandla namanzi azo kuyimfihlo, nakuba umthetho we-EU ufunwe ukubika iminyaka emithathu. E-Netherlands, abangaphansi kwekota bashicilela. Isikhungo sedatha esisodwa seMicrosoft sisebenzisa cishe amaphesenti ayedwa kagesi waseDutch. Ngaleso sikhathi, i-Backblaze yenza into ecindezelayo futhi. Izibalo zayo zedrayivu zekota zihlanganisa amadrayivu acishe abe ngama-350,000 anzima, futhi izinga lokuhluleka lenyuka laya ku-1.73%,
4:16 eliphakeme kakhulu kuleso sikhathi. Amamodeli amathathu e-Seagate awazange ahluleke. Yilokhu okubukeka ngakho isisekelo somphakathi, ikota ngemuva kwekota, iminyaka eyishumi nantathu.
Ulayini wezikweletu: UClaude wasiza ekwakheni imitha
4:25 Manje, lowo mugqa wezikweletu. I-readme ivuma ukuthi ingxenye enkulu ye-repo yabhalwa ngosizo lukaClaude, okuyimodeli elinganiswa. Ngakho-ke uClaude wasiza ekwakheni imitha ehlola ukuthi uClaude waba yisiphukuphuku yini. Yilokho okwenza abalungisi babe yimisebenzi elula. Ngamazwi ombhali, akufanele uthembe umbhali, umuntu noma enye indlela. Uma ukhetha ukufunda lokhu kunokungizwa ngikushumayela, i-diff ifika ebhokisini lakho leposi
4:46 njalo ekuseni, mahhala ku-the daily diff dot dev, isixhumanisi ngezansi. Ngakho, isinqumo sanamuhla nge-live nerf.
Isinqumo: SHIP IT, futhi ungacaphuni ngaphambi kuka-Okthoba 24
4:51 SHIP IT. Ngingakuthumela ngoba yiyona mpikiswano yokuqala ye-nerf engiyibonile ene- isikhathi, incwadi yemithetho yomphakathi kanye nebala elishiwo elingaboni. Vele ungayicaphuni ngaphambi kuka-Okthoba twenty-fourth. Futhi lokho yi-diff yanamuhla. Ngingu-Niko wase-Axrisi. Hlanganisa ngokuzibophezela.
Imithombo
- livenerfgithub.com
- Validationgithub.com
- Hacker Newsnews.ycombinator.com
- Anthropic postmortem (Sep 2025)www.anthropic.com
- Adding Error Bars to Evals (Evan Miller)arxiv.org
- Inspect (UK AI Security Institute)inspect.aisi.org.uk
- NL Timesnltimes.nl
- Hacker Newsnews.ycombinator.com
- Backblaze Drive Stats Q2 2026www.backblaze.com
- Hacker Newsnews.ycombinator.com



