LLM: Asefled, Almad, Asmudem

Acḥal aya i ɣileɣ belli AI-inu tessefled.

Mačči akken deg sci-fi. Ula d yiwen ur yeqqim deg texxamt yeqqar imeslayen-iw. Akken deg ttjara. Ayen i ttaruɣ yettruḥ ɣer kra n umkan, lemmden seg-s, yerna ttɛassan-t ma yella deg-s kra yesɛan azal. Yal win ibennun i ssneɣ yesɛa aḥulfu-agi, yerna alamma d aggur-agi ula d yiwen deg-neɣ ur yezmir ad t-id-yesbeggen.

Sakin yiwen umusnaw n tusnakt (mathematician) deg NYU yewweḍ ɣer uẓbayer (proof) ur zmireɣ ara ad t-awḍeɣ nekk. Mačči axaṭer yella win i d-yennan tidet. Axaṭer takebbanit i yesteqsa ur tezmir ara ad tini ala.

Ihi hatan wayen i d-mmalent kraḍ n tsekkilin-agi, akken i zemreɣ ad t-fehmeɣ. Asefled. Almad. Asmudem (modeling). Yerna hatan teqsiṭ i yi-yerran ur qqareɣ ara d ccekk kan.

Ayɣer tesɛa azal tegdazalt-agi

Twalaḍ isem-agi deg yisallen aggur-agi. Navier-Stokes. Hatan d acu-t s tidet, mbla tusnakt.

Ddem yal ayen yettazzalen. Aman deg yiwen n unbub (pipe). Aḍu nnig ufer n usafag. Idammen deg wul. Dexxan yeffɣen seg tcemmaɛt. Tigdazalin (equations) n Navier-Stokes d alugan n wamek yettembiwil wayen-nni: demmer dagi, yerna hatan acḥal n tezzla, yerna ɣer wanda, ara yeddu wayen akk yellan deffir-s. D alugan n Newton, tazmert tsawa d taẓeyt (mass) tikkelt asnerni n tazzla (acceleration), yettwarun tikkelt-nniḍen i tɣawsiwin ur nesɛi ara talɣa yeqqimen.

Ayagi yettban d ayen n tesdawit kan, alamma twalaḍ d acu i ibedden fell-asent. Yal afer n usafag yettwabna yes-sent. Yal win i d-yeqqaren amek ara tili tegnewt d tigdazalin-agi i yettwasseddan ɣer zdat kra n wussan. Yal amudam n yilel, yal abrid n yinbuben, yal anagraw n usuḍu n waḍu (ventilation), yal asemyiswi (simulation) n yidammen yettɛeddin ɣef yiwen n usewwaq (valve) uqbel ad t-yennal ujerraḥ. Sɛant snat n twinas n yiseggasen yerna d nutenti i yesseddayen amaḍal afizik.

Hatan uḥric yesteɛǧiben. Ula d yiwen ur d-yesbeggen ara belli xeddment yal tikkelt. Deg kraḍ n tsektiwin (dimensions), ula d yiwen ur yezmir ad d-yesbeggen belli tazzla tarkidt, taleggaɣt, ad teqqim taleggaɣt. Tezmer, deg kra n wakud yesɛan tilisa, ad "teṭṭerḍeq" (blow up): deg kra n umkan tazzla tettaweḍ ɣer war-tilisa (infinity) yerna tigdazalin ḥebsent ad sɛunt anamek. Xemmem ɣef waman yettezzin (whirlpool) i yettẓemmin i lebda ɣer yiwet n tenqiḍt. Ula d yiwen ur yeẓri ma tigdazalin n tidet ǧǧant aya, yerna seg 2000 Clay Institute tefka-d amelyun n dular i win ara t-ifrun, ama akka neɣ akka.

Fluid monitor · Navier-Stokes, roughlyFLOW: CALM

Push

Two hundred years old. Unproven in three dimensions. A million dollars from the Clay Institute since 2000, for settling it either way.

Ayɣer ara k-yecɣeb waya? Axaṭer amcewwel n tazzla (turbulence), ayen yesserǧigin asafag-ik yerna yessexṣaren ayen i d-yeqqaren amek ara tili tegnewt, mazal ur yettwafhem ara, yerna axaṭer aṭas n tmuhendist tbedd ɣef tgdazalin ur nezmir ara mazal ad d-nesbeggen belli teddant akken iwata. Yal win ixeddmen deg tussna n tezzla n waman (fluid dynamics) yeẓra aya. I wiyaḍ akk, yekfa ad teẓreḍ belli timacinin ddant-d tura kan ɣer yiwen seg wayen imeqqranen ineggura ur nettwassen ara deg tfizikt. Yerna belli d wagi i ɣer ddant d amezwaru.

Ayen yeḍran

Tristan Buckmaster d amusnaw n tusnakt deg NYU. Levent Alpöge d amusnaw n tusnakt i ixeddmen deg Anthropic yerna yettcarak d Buckmaster deg wakud-is netta. Fernen yiwen webrid uzzig ɣer wugur: aṭerḍeq (blow-up) s tezmert tabeṛṛanit taleggaɣt, yerna s wawalen n Buckmaster, qrib ula d yiwen nniḍen ur yelli ɣef webrid-nni. Sseqdcen AI akken ad t-xedmen: Claude, d wallal n tira n tengalt n OpenAI. Deg tlemmast n ɣuct sɛan igmaḍ. Yiwen yimalas deffir-s, iẓbayer ttwasenqden s tmacint deg Lean.

Sakin Alpöge yura deg Twitter kra n wawalen n lferḥ, ur nban ara, ɣef yiwen waggur igerrzen. Bdan-d wawalen (rumours) belli Anthropic terẓa yiwen wugur n Millennium. Deg yiwen wass, akken i t-id-teḥka OpenAI s yiman-is, OpenAI tewehha azal n mraw n yigiman n yiagentes, yesseddayen-ten yiwen umudam agensan iǧehden ugar n wayen akk teznuz, ɣer Navier-Stokes.

Kra n wussan deffir-s, yeqqar Buckmaster, amcarek-is yewweḍ-it-id yisali belli isallen ɣef usnerni-nsen wwḍen ɣer OpenAI. Yura tabrat i yiwen umusnaw n tusnakt dinna. Tiririt tefka-d compute baṭel. Sakin usan-d sin n yisiwlen n tilifun d yiwen umnadi n OpenAI, i s-yennan belli amudam agensan yesbeggen-d aṭerḍeq s tezmert tabeṛṛanit i Navier-Stokes. Sin n wanawen s timmad-nsen i ɣef xeddmen netta d Alpöge s tuffra.

Yesteqsa ma yella amudam yettwaselmed ɣef tɣimiyin-nsen deg wallal n tira n tengalt n OpenAI, neɣ ma yezmer ad ten-yaweḍ. Yeqqar belli nnan-as amudam ur inadi ara isefka n yimseqdacen. Yesteqsa tikkelt-nniḍen, ɣef uselmed. Yeqqar belli ur d-yewwi ara tiririt.

Fkan-as yiwen umtawi. Ama ad d-ssufɣen igmaḍ-nsen ɣef Euler d imezwura yerna OpenAI ad d-tessufeɣ azekka-nni, ama ad yaru netta s yiman-is agmuḍ n OpenAI, mbla Alpöge, axaṭer Alpöge ixeddem deg Anthropic. Yeqqar belli amnadi yessuter snat n tikkal ad yettwakkes Alpöge. Mi yenna belli ad yeffeɣ ɣer medden deg umkan n waya, yeqqar belli steqsan-t: "Ayɣer ara tessexṣreḍ abrid-ik n uxeddim?" Amnadi n OpenAI ur yeqbil ara ayen i d-yeḥka ɣef umeslay n yimeskaren (authorship).

Yessufeɣ-d ɣas akken, s yiwen n ulɣu, yerna Terence Tao yura fell-as ass-nni s timmad-is. OpenAI tessufeɣ-d agmuḍ-is azekka-nni: 88 n yisragen, 2.7 n yimelyunen n yiznan, aẓbayer yettwasenqden deg Lean, d yiwet n tefyirt teqqar belli tazzla tebda deffir n wawalen n Twitter belli Anthropic tefra uguren n Millennium. Kraḍ n wussan deffir-s, 25 n wid yewwin arraz n Fields zemlen yiwet n tebrat s yisem "A Severe Misalignment of AI in Mathematics". Yerna OpenAI, mi tettwasteqsa srid, tenna-d belli ulac isefka n yimseqdacen uzzigen i yettwawḍen, sakin terna-d belli "ur tezmir ara ad tekkes belli isefka yettwakksen seg timagit (de-identified), i d-yekkan seg useqdec-nsen n yifarisen-nneɣ, ɛawnen ad snernin imudam-nneɣ".

Ɣer-itt snat n tikkal. Ur qqaren ara belli ur lemden ara seg uxeddim-is. Qqaren belli ur zmiren ara ad ẓren.

Tape 03 · replay01 / 12

Mid-August · Buckmaster & Alpöge

Results on a route almost nobody else is on: blow-up with a smooth external force. Worked with Claude and OpenAI’s coding tool.

Asefled

Bdu s uḥric amagnu, axaṭer deg uḥric amagnu i tezdeɣ.

Ma tesseqdaceḍ ChatGPT s uɣawas Free, Plus neɣ Pro, imeslayen-ik sselmaden amudam s lɛadda (by default). Ilaq ad tkecmeḍ ɣer Settings, sakin "Data Controls", yerna ad tsensiḍ "Improve the model for everyone". Allal n tira n tengalt yesɛa abuṭun-is weḥd-s, win n ChatGPT ur t-yettnal ara. Iɣawasen n yimsaɣen n Claude sselmaden s lɛadda seg ɣuct n useggas iɛeddan, yerna ḥerrzen imeslayen-ik semmus n yiseggasen ma teǧǧiḍ-t yeddem. Gemini daɣen. Ala API d tseddarin n tkebbaniyin timeqqranin (enterprise) i yensan s lɛadda, axaṭer tikebbaniyin yesɛan ibugaṭuten ssutrent. Wiyaḍ akk qeblen s kan ur ldin ara umuɣ (menu).

Data controls · as shipped4 of 4 listening

ChatGPT · Settings › Data Controls
Improve the model for everyone
Free, Plus and Pro train by default.
The coding tool
Its own toggle
The ChatGPT switch does not touch it.
Claude · consumer plans
Help improve Claude
On by default since last August. Five-year retention if you leave it on.
Gemini
Activity
Trains by default too.
API · enterprise tiers
Off by default
Because companies with lawyers asked.
OFF

Sens-it, mazal ṭṭfen imeslayen-ik tlatin n wussan, yerna imdanen deg tkebbanit mazal zemren ad ten-ɣren ma yella yiwen unagraw yewehha-d fell-ak. Tamḥekemt tezmer ad tumer ad yettwaḥrez uɣbar (archive) meṛṛa. Yella yiwet i t-yumren, tuget n useggas iɛeddan, yerna ɛecrin n yimelyunen n yimeslayen yettwakksen seg timagit ttunefken ɣas akken deg yiwet n tcetka.

Ihi ih. Tessefled. Teqqar-it deg Settings. Ayen yesteɛǧiben mačči belli tessefled. D acḥal n yimdanen i ibennun tudert-nsen fell-as, ur ldin ara umuɣ-nni werǧin.

Almad

Asefled d ayen yeqqimen. Almad d ayen i t-yerran yesɛa azal.

Imalas iɛeddan uriɣ belli ilaboratwaren jemɛen-d uddis n uselmed (training set) yesɛan azal akk deg umezruy, baṭel, seg yimsaɣen-nsen s yiman-nsen: mačči isefka izegzawen (raw data), talɣa n uxeddim. Anwa isteqsiyen i d-ttusteqsayen. Anta tiririyin i yettwaqeblen. Anda yeqqen webrid (bottleneck). Amek yettban ugmuḍ yelhan. Laboratwar i iwalan mraw n yigiman n terbaɛin xeddment ɣef kra ur yeḥwaǧ ara ad yeɣer ula d yiwet deg-sent. Yeddem yakan ayen ssnent.

D wagi s timmad-is i d-teglem tefyirt n OpenAI. Mačči "neɣra taɣimit-ik". Kra yessusmen ugar yerna diri-t ugar: "aseqdec-ik yezmer ad yili yeɛawen ad snernin imudam-nneɣ". Almad-agi yefreq. Ur yezmir ara ad yettwaḍfer ɣur-k. Ayen yebɣan ad yini daɣen belli ur yezmir ara ad yettwakkes, werǧin, s ula d yiwen, ula d nutni.

Zik ɣileɣ amihi d anɣal (leak). Mačči d anɣal. D tuččit (diet). Axeddim-ik yekcem, yerna yeffeɣ-d iǧhed ugar, yelsa isem n wayeḍ.

Asmudem

Hatan uḥric i ɣef ttnaɣen medden, ihi ad ḥadreɣ.

Asefled yettak-asen awalen. Almad yettak-asen tizmarin. Asmudem d asurif wis kraḍ: abnu n yiwet n tugna n win ixeddmen d acu, yerna acḥal iqerben.

Tikebbaniyin-agi qqarent yakan imeslayen s waṭas (at scale). Mačči d tiẓri, d tamahilt n ufaris. Yal prompt yettɛeddi deg yimsekyaden (classifiers) i inadin ɣef ubuẓẓer (abuse). OpenAI tessufuɣ-d ineqqisen ɣef yimiḍanen i teṭṭef yerna i tegdel ɣef useqdec n diri. Akken daɣen Anthropic. Ur tezmireḍ ara ad teṭṭfeḍ ayen ur teɣriḍ ara. Abrid (pipeline) i yezmren ad d-yini, deg yimelyunen n yimeslayen, anwa i yettɛerriḍen ad yebnu yiwen n umrig, yezmer daɣen ad d-yini anwa i yettɛerriḍen ad d-yesbeggen yiwen n teorem (theorem), yerna anda yewweḍ. D tamacint-nni s timmad-is yettwawehhan ɣer usteqsi nniḍen.

Field model · built from millions of conversationsSWEEPING

Knowing what is hot, and who is close. The same machinery, pointed at a different question.

Tura err-it ɣer tama n wayen yeḍran. Yiwen n tweet ur nban ara d kra n wawalen, d ayen iwatan akken ad ttwawehhan mraw n yigiman n yiagentes ɣer yiwen wugur deg yiwen wass, yerna ɣer talɣa-s iṣeḥḥan. Ula d yiwen ur yeḥwaǧ aɣmis n yiznan (chat log) i waya. Ḥwaǧen yiwen umudam n wanda tewweḍ taɣult, d yiwen umudam n wanda teqqim yiwet n terbaɛt tamecṭuḥt deg-s. Tamusni n wayen iḥman, d win iqerben, d ayen s timmad-is i k-tettak tugna yebnan seg yimelyunen n yimeslayen.

Ur qqareɣ ara belli d aya i yeḍran. Buckmaster ur t-yeqqar ara daɣen, yerna d netta i yesɛan iṛeṣuten (receipts). Qqareɣ belli tamacint tella, teddu ɣef yal izen, yerna takebbanit i tt-yesɛan ur k-teqqar ara d acu i tesmudum. Mi ara yeffer ubuṭun, yal asefhem yella, ula d wagi, yerna takebbanit i yeffren abuṭun ur tesɛi ara lḥeqq ad tecctki mi ara t-fernen medden.

Aḥulfu

Aṭas n wid ibennun s wallalen-agi sɛan yiwen uḥulfu. Mi ara tiliḍ ɣef yiwet n twuri tamagnut, amudam yezzerb yerna yettak aṭas. Mi ara tiliḍ ɣef kra yesɛan azal, yettuɣal yeẓẓay, ur iban ara, yettḥadar. Yebda yessutur-ak ad tesfehmeḍ tiɣawsiwin i yefhem semmus n tesdatin uqbel. Yerna taɣuri taparanuyt (paranoid) d tin i d-yeqqaren belli yeṭṭef-ik akken ad yelmed d acu i txeddmeḍ.

Ḥulfaɣ-as ula d nekk. Ur zmireɣ ara ad t-id-sbeggneɣ, yerna ttbut n tidet yemmal ɣer umkan-nniḍen. Deg yulyu, medden ṭṭfen s taḍfirt (fingerprinted) tadrimt n uxemmem uffir (hidden reasoning budget) n umudam ifazen n OpenAI, walan-t yeɣli seg 960 ɣer 128 deg yiwen yiḍ, i yal yiwen. OpenAI teqbel abeddel-nni kan ass-nni. Semman-as "aseggem n yiɣewwaren (parameter tuning) n lawan". Amwehhi (router) i yettgezziren ma ad tawiḍ amudam azerbi neɣ amudam n uxemmem d ayla-nsen, mačči d ayla-k. Tilisa n yimalas llant. Yal yiwen seg yibuṭunen-agi yettezzi ɣef ssuma.

Reasoning budget · top model · JulyOBSERVED FROM OUTSIDE

960

was 960now 128

Overnight, for everyone. Only then did they admit a change.

Ihi amudam ahat ur yettaẓay ara axaṭer telhiḍ. Yettaẓay axaṭer tettesɛiḍ aṭas. Maca ur zmireɣ ara ad friqeɣ gar-asen seg yidis-iw n ugdil, ula d kečč ur tezmireḍ ara, yerna d wagi i d asentel n umagrad-agi meṛṛa. Asefled, almad d usmudem ur ttbanen ara seg beṛṛa. Tettafeḍ-ten kan mi ara yesteqsi yiwen umusnaw n tusnakt yesɛan aẓbayer deg Lean, yerna tiririt d tasusmi.

D acu ara nexdem

Kraḍ n tɣawsiwin, ula d yiwet ur tewɛir.

Sens abuṭun. Ass-a. Deg ChatGPT, Settings, "Data Controls", "Improve the model for everyone". Deg wallal n tira n tengalt, weḥd-s. Deg Claude, "Help improve Claude". Deg Gemini, armud (activity). Yettawi snat n tesdatin yerna yekkes abrid akk ishilen.

Ḥrez axeddim n tidet anda ur zmiren ara ad t-walin. Iḥricen imagnuyen n uxeddim zemren ad ddun ɣer yal API. Aḥric i k-yerran d kečč, tikti, isefka, abrid ideg qrib ula d yiwen ur yelli, ilaq ad yili ɣef umudam i tezmreḍ ad d-tsidreḍ (download) d tmacint i tḥekmeḍ. Wagi mačči d paranoya. D ayen ara k-yini Buckmaster tura.

Ssufeɣ-d zik. Ayen isellken sin-agi d aẓbayer deg Lean yesɛan azemz, d ulɣu ɣef uqeddac (server) n tesdawit. Izemzen (timestamps) rennun tirebbaɛ. Ma teqrebeḍ ɣer kra, err azemz deg usekles uqbel ad k-id-yaweḍ usiwel n tilifun.

LLM-ik tessefled. Tlemmed. Tesmudum-ik. Ula d yiwet seg tigi mačči d ayen yeffren. Atan deg umuɣ n Settings d yineqqisen n tafrawant (transparency reports). Asteqsi kan i yeqqimen d acu i tkemmleḍ tettakeḍ-as.

Yessufeɣ-d ɣas akken. Akken daɣen ilaq ad txedmeḍ kečč.