Ntụgharị nkọwa usoro
Nkọwapụta usoro bụ mbọ iji weghachi-njinịnụ ngụkọ ime nke netwọkụ akwara n'ime algọridim mmadụ nwere nghọta.
Nchịkọta
Rather than asking 'which input mattered,' it asks 'what is this network actually computing, circuit by circuit?'
Ime miri emi
Ebe usoro dị ka SHAP na-akọwa ntinye na ntinye, nkọwa nhazi na-emepe igbe ahụ ma mụọ ihe ọ̀tụ̀tụ̀ na ihe n'onwe ha. Ndị na-eme nchọpụta (ọkachasị na Anthropic, OpenAI, na agụmakwụkwọ) na-emeso onye ngbanwe dị ka mmemme a ga-agbakọta, na-achọpụta 'circuits': mpempe akwụkwọ nke neurons na ndị isi nlebara anya na-arụ otu ọrụ. Nchọpụta akara ala gụnyere 'isi induction,' ndị isi nlebara anya na-edegharị ụkpụrụ iji mee ka mmụta ihe dị n'ime, na nchọpụta na otu neurons na-abụkarị 'polysemantic,' na-agbapụ maka ọtụtụ echiche na-enweghị njikọ n'ihi na ihe nlereanya ahụ nwere atụmatụ karịa akụkụ (n'elu). A na-eji ndị na-akpachapụ anya eme ihe ugbu a iji wepụ ihe ndị a ka ọ bụrụ ihe dị ọcha, 'njirimara' monosemantic, dị ka ntụzịaka na-arụ ọrụ na àkwà mmiri Golden Gate.
Nghọta nka nka
Ihe mgbochi bụ isi bụ n'elu: netwọk nwere d akụkụ nwere ike ịnọchite anya karịa atụmatụ d site na ịchekwa ha ka ọ fọrọ nke nta ka ọ bụrụ ntụzịaka orthogonal, ya mere neurons onye ọ bụla na-ere ọkụ maka echiche enweghị njikọ. Ndị na-edochaghị anya autoencoders na-elebara nke a anya site n'ịmụ akwụkwọ ọkọwa okwu zuru oke nke na-eji naanị nkeji ole na ole na-arụ ọrụ arụkwa ọrụ n'otu oge, na-egosipụta njirimara nwere ike ịkọwa. Ndị ọrụ nyocha wee kwado sekit site na iji ntinye aka na-akpata, ịgbachapụ ma ọ bụ 'patching' rụọ ọrụ iji gosi na otu akụrụngwa na-arụ ọrụ ngụkọ echepụtara n'ezie.
Mmetụta atụmatụ
Ọnụ ego na mmefu ego
Mkpebi ihe owuwu ụlọ na-akwalite arụmọrụ yana ọnụ ahịa ọrụ ruo ọtụtụ afọ.
Mkpebi doro anya
Nkà mmụta nka na-enyere ndị otu egwuregwu aka ịhọrọ nchịkọta ziri ezi, ọ bụghị naanị nke kachasị ọhụrụ.
Quality akara
Nhọrọ injinia ka mma na-ebelata ihe omume ntụkwasị obi na mmepụta.
Ọdịnihu nke Ntụgharị Asụsụ Mechanistic
Nkọwapụta usoro ihe dị n'etiti nchekwa AI: ịghọta ndị na-arụ ọrụ nwere ike ime ka anyị nyochaa ụdị aghụghọ maka aghụghọ, chọpụta ike dị ize ndụ, na iduzi omume site na idezi atụmatụ ozugbo. Ọrụ na-adịru nso na-elekwasị anya n'ịkọba ndị autoencoders na-adịghị ahụkebe n'ụdị oke, ime nchọpụta sekit, na iwulite akwụkwọ ọkọwa okwu njiri mara nke a pụrụ ịdabere na ya. Ebumnuche ebumnuche bụ 'MRI maka netwọkụ akwara ozi', ụzọ isi gụọ echiche nke ihe nlereanya tupu ebuga ya, n'agbanyeghị na iji ikwesị ntụkwasị obi na-akọwa sistemụ ijeri-parameter n'ogo ka bụ nnukwu ihe ịma aka mepere emepe.
Mmejuputa n'ezie n'ụwa
Anthropic weputara ọtụtụ nde atụmatụ a pụrụ ịkọwa na Claude wee gosi na ime ka otu akụkụ 'Golden Gate Bridge' pụta ìhè mere ka ihe nlereanya ahụ jiri nlezianya kpọọ akwa mmiri ahụ, na-egosipụta nduzi omume ozugbo.
Ndị nchọpụta chọpụtara 'isi induction' na transformers na-edegharị ma na-aga n'ihu na usoro token ugboro ugboro, na-akọwa isi usoro dị n'azụ mmụta n'ime ọnọdụ.
A na-eji patching rụọ ọrụ iji wepụta ebe ihe nlereanya na-echekwa eziokwu (dịka, isi obodo nke obodo), na-ekpughe ọkwa ndị akọwapụtara na ihe mejupụtara ya.
Ndị otu nchekwa na-enyocha njirimara dị n'ime iji chọpụta ma ihe nlereanya na-anọchite anya echiche dị ka aghụghọ ma ọ bụ ntuziaka adịghị mma, na-enyere nleba anya ma ọ bụ ntinye aka.
Ihe ize ndụ & okporo ụzọ nche
Ịkwalite otu akara ngosi nwere ike zoo adịghị ike sistemụ sara mbara.
A na-eledakarị ihe akụrụngwa na ụgwọ ọrụ anya.
Ọdịiche nchekwa na nleba anya nwere ike itolite ka sistemu na-adịwanye mgbagwoju anya.
Map mmejuputa
Kọwaa latency, ịdịmma na ebumnuche ọnụ ahịa tupu mmejuputa ya.
Benchmark n'okpuru ibu dị adị na ọnọdụ data.
Nleba anya akụrụngwa maka mperi, ịkpafu na mmetụta onye ọrụ.
Kwadebe ụzọ nzaghachi azụghachi azụ na ihe omume tupu ịchachaa.
Nọgide na-eme nchọpụta
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Mechanistic Interpretability quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ntuziaka na-esote
Mpempe akwụkwọ akpaaka maka nkọwa
Ajụjụ a na-ajụkarị
What is Mechanistic Interpretability?
Nkọwapụta usoro bụ mbọ iji weghachi-njinịnụ ngụkọ ime nke netwọkụ akwara n'ime algọridim mmadụ nwere nghọta. Kama ịjụ 'Olee ntinye dị mkpa,' ọ na-ajụ 'gịnị bụ ihe netwọk a na-agbakọ n'ezie, sekit site na sekit?'
Gịnị bụ isi ebumnobi nke nhazi nkọwa?
Ntugharị usoro ihe na-achọ ịghọta algọridim n'ezie nke netwọkụ na-eme n'ime, ọ bụghị naanị mmekọrịta ntinye-mmepụta.
Kedu ihe 'nkwalite elu' na-ezo aka na netwọkụ akwara?
Superposition na-eme ka netwọk tinye koodu echiche karịa ka ọ nwere neurons / akụkụ site na ịchekwa ha dị ka ntụzịaka na-ekpuchi ihe fọrọ nke nta ka ọ bụrụ orthogonal, na-eme ka neurons polysemantic.
Kedu ihe bụ 'isi induction'?
Ndị isi mmalite na-achọpụta ihe mere n'oge gara aga nke token dị ugbu a wee detuo ihe sochiri ya, usoro bụ isi na-enye ohere mmụta n'ime ọnọdụ.
Olee otú ndị nchọpụta si egosi na sekit a chọtara na-eme n'ezie mgbakọ e chepụtara echepụta?
Ụzọ mkpata dị ka patching activation ma ọ bụ ablation nwale ma ịgbanwe ihe mejupụtara na-agbanwe n'ezie omume dị mkpa, na-akwado ọrụ ya.
Kedu ihe kpatara eji ewere nkọwa ọrụ igwe dị mkpa maka nchekwa AI?
Ịghọta njirimara dị n'ime na sekit nwere ike ime ka nyocha maka aghụghọ ma ọ bụ ikike dị ize ndụ ma kwe ka ntinye aka ezubere iche, na-akwado ibunye nchekwa.