I-Google isanda kukhipha imodeli yokugenca. Nansi umehluko.

Lokho okuhlanganiswa yile vidiyo

• I-Google ikhipha i-Gemini 3.8 Flash + Flash Cyber (86.2% CyberGym, Fairwind Program)
• I-Meta ikhipha i-Muse Spark 1.3: $1.25/$4.25, noma $0.10/$0.20 uma i-Meta ingaqeqesha kuyo
• Izingosi ezintathu zenze amakhasi angu-215,128 wamanga "esofthiwe engcono kakhulu"; I-Perplexity iwanamathisela

Umbhalo obhaliwe ohunyushwe

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

0:00 I-Google ikhiphe imodeli yayo yesithathu ye-Flash emavikini ayisithupha namuhla, kanye nenguqulo eqeqeshwe ukugenca, futhi ngemva kwamahora amane i-Meta yakhipha imodeli ethi ibiza amasenti ayishumi esigidini samathokheni uma uvumela uZuckerberg ukuthi afunde ikhodi yakho, ngakho impi yamanani e-AI manje isinesigaba se-cybersecurity. NgoLwesithathu, futhi umehluko mude. I-Gemini 3.8 Flash yashaya amaphuzu ayinkulungwane eyodwa namakhulu ayikhulu ku-Hacker News, I-Muse Spark 1.3 yathatha cishe angamakhulu ayisikhombisa ngaphezulu, futhi phakathi nalokho, indawo yokucwaninga yathola amawebhusayithi amathathu ashicilele amakhulu amabili neshumi nanhlanu ezinkulungwane

0:28 wamakhasi amanga esofthiwe engcono kakhulu ngenhloso yokuthi i-Perplexity iwanamathisele. Futhi namuhla: okuthunyelwe okukuncenga ukuthi ubambelele ku-Firefox yakho kwathola amakhulu ayisishiyagalolunye namashumi ayisishiyagalombili nesishiyagalombili amaphuzu, futhi ikhasi losizo le-Mistral lokuyeka ukuqeqeshwa lathola cishe amakhulu amahlanu, iningi livela kubantu baseYurophu abathola inketho yobukhosi iqeqesha ngemiyalo yakho ngokuzenzakalelayo. Kule vidiyo: lokho i-Gemini 3.8 Flash ebiza ngempela, imodeli yokugenca i-Google izothengisa kuphela kubalingani abangamakhulu ayisithupha namashumi amahlanu, izinga le-Meta lamasenti ayishumi nokuthi yini ebiza ngempela, kanye neyodwa yezinga elingabhalwanga yinoma iyiphi inkampani.

0:59 NgoLwesithathu, Septemba 2, futhi lokhu The Daily Diff. I-Gemini 3.8 Flash ingamasenti angamashumi ayisikhombisa nanhlanu esigidini samathokheni ngaphakathi namathathu namashumi ayisikhombisa nanhlanu ngaphandle, okufanayo ne-3.7 Flash yamaviki amathathu adlule, nenothi elincane elisho ukuthi intengo iphindeka kabili ngoJanuwari 1, okuyisivivinyo samahhala esinezinyathelo ezengeziwe. Ezilinganisweni ze-slide-deck, ezingahlulwanga, ithola amaphesenti angu-54.9 ku- HLE-Verified, i-X ithi ihlula i-Sol ne-Opus 5 ku-Terminal-Bench, futhi uLogan Kilpatrick washicilela i-73.7 ku-DeepSWE, okuyisilinganiso esisodwa eside

1:29 sokufaka amakhodi esingakakhunjulwa okwamanje, ngokucatshangwa. USimon Willison wayicela ukuthi yenze into epholile nge-HTML futhi wathola ukulingisa kwamachashazi emasekhondini ayishumi nantathu ngamasenti angu-1.8, isebenza ngamaphirimitha angamashumi ayisithupha ngomzuzwana, ngoba amashumi ayisithupha ayefakwe ngokuqinile ku- ikhasi. Okuthunyelwe kwe-Google uqobo kuchaza izinzuzo ngomusho engingawukhawula: I-3.8 Flash isebenza kanzima. Yenza izinyathelo ezengeziwe zokucabanga, ibiza amathuluzi ngokuphindaphindiwe, futhi, ngokuphindaphinda, ingasebenzisa amathokheni amaningi, okuyindlela yebhizinisi yokuthi imitha igijima ngokushesha. Ibhodi elizimele le-DeepSWE liyavumelana kuzo zombili izindaba: I-Flash ithola amaphesenti angamashumi ayisikhombisa nane,

2:02 iboshwe ne-Opus 5 ngekhono lesihlanu lezindleko ngomsebenzi, kodwa idinga izinyathelo eziyikhulu namashumi ayisithupha nesithupha namakhulu ayikhulu namashumi amane nantathu ezinkulungwane amathokheni okuphumayo ukuze ifike lapho, ngaphezu kokuphindwe kabili kwalokho okuchithwe i-Sol. I-Artificial Analysis yashisa amathokheni ayikhulu namashumi amane ezigidi namakhulu ayinkulungwane namashumi ayisikhombisa nesishiyagalombili amadola ukuze nje iyilinganise. Kushibhile ngamathokheni, ikhuluma kakhulu ngomsebenzi, futhi ithokheni yokuqala ifika ngemuva kwamasekhondi ayishumi nambili okucabanga. Bese kuba ingxenye ethakazelisayo.

2:23 I-Gemini 3.8 Flash Cyber iyimodeli efanayo enezindlela zokuphepha ezikhululekile kubo, ngokuphindaphinda, abavikeli abathembekile, futhi uSundar uthi ishaya amaphesenti angu-86.2 ku-CyberGym, isilinganiso sokuzenzakalela sokuthola ubungozi. Iyalungisa futhi: amaphesenti angu-47.2 ku-CWE-Bench ngokumelene no-47.8 kumodeli ehamba phambili yasezingeni eliphambili, ithimba le-Chrome lithi yakhiqiza izilungiso ezifanele eziphindwe ka-2.6 kunamamodeli amakhulu kakhulu okuhweba, futhi abacwaningi befu le-Google bayisebenzisa ukuze bathole ubungozi obukhulu ngaphansi kwamahora amabili, umsebenzi ovame ukuthatha izinyanga, noma intsha eyodwa enomdlandla.

2:54 I-Google igcizelela ukuthi ibeke eqhulwini ukulungisa kunokusebenzisa, okuyindlela ehloniphekile yokusho ukuthi imodeli ingakwazi ukuhamba kalula phakathi kwezikhala, bavele bayicela ukuba ingenzi kanjalo. Ngakho awukwazi ukuyithola. Ukufinyelela kuhamba ngohlelo olusha lwe-Fairwind: ohulumeni, izingqalasizinda ezibalulekile, nabalingani abahloliwe abangamakhulu ayisithupha namashumi amahlanu abanemfuneko izinto ezimbili. Njalo ilabhorethri elisha manje linekhono lokugenca elingeke likudayisele, futhi kuleli sonto lishibhile.

3:16 Amahora amane kamuva, i-Meta yakhipha i-Muse Spark 1.3, futhi uZuck walibiza ngekhono elisha, ukusebenza okushibhile kakhulu ukukala, okuqinisile, ngenkanyezi eyodwa ubukhulu beMenlo Park. I-endpoint ejwayelekile iyidola namasenti angamashumi amabili nanhlanu ngaphakathi namadola amane namasenti angamashumi amabili nanhlanu ngaphandle, ngaphezu kwe-Gemini. I-endpoint yomnikeli iyizisenti eziyishumi ngaphakathi, izisenti ezingamashumi amabili ngaphandle, ukufunda kwe-cache ngekota yesenti, kushibhile ngokuphindwe kayishumi nanhlanu, futhi umehluko kuphela ikholomu ethi 'yasendaweni ukuthuthukisa imikhiqizo yethu'.

3:40 Ngakho isaphulelo yikhophi yeseshini yakho, futhi okokuqala akekho okufanele afunde ilayisense, ngoba i-Meta yababhala ilayisense njengohlu lwamanani. Amazwana aphezulu: manje sekucacile ukuthi ukweba amathokheni ami kubiza malini. Amazwana esibili ayazibuza ukuthi kuzothatha isikhathi esingakanani ngaphambi kokuthi othile akhiphe ukhiye we-AWS kumodeli eqeqeshwe ngamacebo abanikeli. Kumarekhodi e-Meta, i-Spark 1.3 ithola u-75.4 ku-DeepSWE, ngaphezu kwe-Sol ne-Opus 5, futhi i-Meta ithi isebenzisa amaphesenti angamashumi amabili ambalwa amacala amathuluzi

4:06 namaphesenti angamashumi amabili nanhlanu ambalwa amathokheni kune-1.2, okuphambene kakhulu nokubheja kwe-Google, ngakho namuhla izinkampani ezimbili ezinkulu zokukhangisa emhlabeni azivumelani ngokuthi ukukhuluma kakhulu kuhle yini. Ngaleso sikhathi i-Trellner Research yabuza i-Perplexity ngesoftware engcono kakhulu ezigabeni ezingamakhulu amathathu namashumi ayisishiyagalombili futhi yafunda yonke inkomba: amaphesenti angama-59.8 avela kumasayithi ngaphandle kwekhulu eliphezulu eliyizinkulungwane eziyikhulu, i-Wikipedia yacashunwa kathathu kwabayisikhombisa nesigamu sezinkulungwane, namasayithi amathathu odadewabo anezinkulungwane ezingamakhulu amabili neshumi nanhlanu

4:34 abhale imihlahlandlela yokuthenga anezihloko zamakhasi awo asekhaya athi 'Amaqiniso Nekhasi Lokugxila', aqondiswe ku-crawler, hhayi kuwe. Impi ye-AI SEO isiqalile, futhi isikhali sayo sokuqala yi-regex enesabelomali sokumaketha. Lokho kungamanani amaningi ngoLwesithathu olulodwa; uma kunalokho ungafunda lokhu kunokuba ungizwe ngikushoyo, i-diff ifika ebhokisini lakho leposi njalo ekuseni — mahhala ku-daily diff dot dev, isixhumanisi ngezansi. Ngakho, isinqumo sanamuhla: SHIP IT. I-Gemini 3.8 Flash ilingana ne-Opus 5 kubhodi elilodwa i-Google engazange ilibhale,

5:00 ngekota yentengo. Vele ufunde ibhili lethokheni, futhi ufunde izihloko zamakholomu e-Meta. Yilokho okuhlukile namuhla. Ngingu-Niko wase-Axrisi. Hlanganisa ngokuzibopha.

Imithombo

1. Introducing Gemini 3.8 Flash and 3.8 Flash Cyberblog.google
2. Fairwind Programblog.google
3. DeepSWE v1.1 leaderboarddeepswe.datacurve.ai
4. Artificial Analysis: Gemini 3.8 Flashartificialanalysis.ai
5. Muse Spark 1.3 pricingdeveloper.meta.com
6. Introducing Muse Spark 1.3research.meta.ai
7. Hang on to Your Firefoxwww.newsonaut.com
8. Mistral: opting out of traininghelp.mistral.ai
9. HN: Gemini 3.8 Flashnews.ycombinator.com
10. HN: Muse Spark 1.3news.ycombinator.com
11. HN: 215,128 "best software" pagesnews.ycombinator.com