Continuous Learning_Startup & Investment
Open in Telegram
We journey together through the captivating realms of entrepreneurship, investment, life, and technology. This is my chronicle of exploration, where I capture and share the lessons that shape our world. Join us and let's never stop learning!
Show more2 301
Subscribers
-524 hours
-107 days
-3630 days
Posts Archive
2024λ
μ μ λ°μ΄ κ°λ€. μ΄λμ λ²€μ²ν¬μμλ‘μ μΌνλ μκ°λ³΄λ€ μ¬μ
μ νλ ν¬μ§μ
μ΄ λ μκ°μ΄ 3λ°°κ° λμ΄κ°λ€. ν¬μμμ μ¬μ
κ°μ μ°¨μ΄λ μκ°λ³΄λ€ λ§€μ° ν° κ² κ°λ€.
κ±°μ 12λ
μ§Έ μ¬μ
μ νλ μ
μ₯μμ λ²€μ²μΊνΌν리μ€νΈλ‘μλ κ²°μ½ μ΄ν΄νμ§ λͺ»νμ κ² κ°μ μ λ€μ΄ λ§λ€. λͺ κ°μ§ λ μ€λ₯΄λ κ²μ 보면,
1) κ³νλλ‘ λλκ² λ¨ νκ°μ§λ μλ€λ κ²μ μκ² λλ€. 10κ°μ§ κ°μ€μ μ€ννλ€λ³΄λ©΄ κ·Έλλ λλκ² 2-3κ°λ μμ§ μμκΉ μκ°νμ§λ§, μ€μ λ‘λ 10κ°λ₯Ό μλν΄λ (μΌμν리λ§νΌ) λ¨ 1κ°λ κ³νλλ‘ λλκ² μλ€λ κ²μ νμΈνκ² λλκ² λ€λ°μ¬λ€.
2) ν΄κ° λ±μ λ΄μ μ리λ₯Ό λΉμ°λ©΄ λ°λμ κ·Έ κΈ°κ° λμμ λ¬Έμ κ° ν°μ§λ€. ν΄μ, 리νλ μ¬ λ±μ μ΄μ λ‘ μ§§κ±΄ 길건 μ½κ°μ ν΄κ°λ₯Ό μΈν
νκ³ λλ©΄, 곡κ΅λ‘κ²λ λ°λμ κ·Έ μ리λ₯Ό λΉμ°λ κΈ°κ° λμμ λ¬Έμ κ° ν°μ§κ³ , κ·Έλμ ν΄κ°λ λ¬Όκ±°νμ΄ λλ€. κ·Έλμ μΈμ κ°λΆν° 미리 ν΄κ° κΈ°κ°μ μ‘μ μ μλ€.
3) κ°λ μκ°νλ μ λμ μ±κ³Όκ° λμ κ΄λ ¨λ ꡬμ±μλ€κ³Ό celebrationμ νκ³ λλ©΄, νλ¬ μ΄λ΄μ λ°λμ ν° λ¬Έμ λ μμμΉ λͺ»ν negative eventκ° μκΈ΄λ€. κ·Έλμ μ’μ μΌμ΄ μ겨λ μ¨μ ν μ¦κΈ°κ±°λ μΆννκΈ°κ° μ½μ§ μλ€.
μ΄λ°κ±Έ 보면 λ΄κ° κ΅³μ΄ μ μ΄λ° κ³ μκΈΈμ κ΅³μ΄ νν΄μ κ°μΌνλ μΆλ€κ°λ, κ°λμ 콩λλ―μ΄ λ€λ€ μλ κ±°λΌ μκΈ°νλλ° 1λ
μ νλλ² μ λ λλ μ΄λ²€νΈλ₯Ό κ²ͺλ€λ³΄λ©΄, μ΄κ² μ°Έ λ§μ½μ²λΌ μ€λ
μ±μ΄ μλ€λκ±Έ μκ² λλ€. κ³ μκΈΈμΈμ€ μλ©΄μλ, λ€μ μΌμ΄λμ ν μ λ°μ μλ λ³ν κ°μ μ€λ
μ± π
κ·Έλμ μ¬λ¬ κ°λ‘ μλ°μ΄ μλλΌλ μ°½μ
μλ€ μ¬μ΄μλ 무μΈμ 곡κ°λλΌλ κ²μ κΉκ³ κ°λ€. κ°κΈ° μ§κ³ μλ 무κ²λ λ€λ₯΄λλΌλ ν° μ§μ μ§κ³ κ°λ, κ·Έ κ³Όμ μμ λλ¦μ ν¬μ΄ ν¬μΈνΈκ° μλ λ³ν κ°μ λμ°λ³μ΄μ’
.
λ°μ§μ
λνλ.
μΈμμ
μμ μ°μ μμ κ²ͺμ΄λ³Έ μ¬λλ€μ΄ νλ₯ν μνκ° μ¬μ₯μΌλ‘ μ±μ₯νκΈ° μν΄μ ν΄μ£Όλ μ‘°μΈλ€.
https://youtu.be/_Oo3Ji4MVZU?si=gm_czz69_vzuKJaw
Repost from μ΅μ μλ€ λ°μ§ν
μ μΉκ° (μΌν μΏ νμ)μμ
μ¬μκ° (νμΏ μμ μ μΉν€)λ₯Ό κ±°μ³
κ²½μ μΈ (μλΆμ¬μ μμ΄μ΄μΉ) κΉμ§.
μλΆμ¬μ μμ΄μ΄μΉ κ΄λ ¨ μ¬λ¬ μμμ΄ μμ§λ§ μ΄μ°½λ―Ό κ΅μλ μμμ 보μλ©΄ κ·Έκ° μΌλ§λ μλν μΈλ¬Όμ΄μλμ§ μ μ μμ΅λλ€. μ΄μ°½λ―Ό κ΅μλμ μΌλ³Έμ λ
μμλ€μ μ¬μ€ 4ν λΏλ§ μλλΌ μ λΆ λ€ κ·νκ³ μ’μ λ΄μ©μ
λλ€.
https://n.news.naver.com/mnews/article/023/0003844258
https://youtu.be/QnRT2tfRMAA?si=t3fHb0hROGJZDuiq&t=2220
(μ΄μ°½λ―Όμ μΌλ³Έμ λ
4ν: 37λΆλΆν°)
μΌλ³Έμ μ λ―Έκ΅μ μΉ¨λ²μ λ°κ³ λ λ€μ ν λ³νν μ μμλκ°?
μ§λ°©μ ν. μ§λ°©μ λ²μ£Όλ€μ΄ μ μνκ³ ν¨λ°°νλ©΄μ λ°°μμΌνλ€λ κ²μ κΉ¨λ«κ³ , κ·Έλ€λΌλ¦¬ κ²½μνλ©΄μ λ°μ ν μ μμμ. κ° μ§λ°©λΌλ¦¬μ μ΄λ§μ΄ λΉ λ₯Έ λ³νλ₯Ό λ§λ€μμ.
μ§λ¨ μ§μ±μ΄ μλμ§κ° λλ €λ©΄ μ€μ μ§κΆμ΄ μλλΌ μλ‘ λ¬λΌμΌ μλμ§κ° λμ¬ μ μμ.
μμκ³Ό κ΅μ§μ μ μΉλ₯΄λ©΄μ μ κ΅μ΄ λ κ²μ΄λ μλ―Όμ§κ° λ κ²μ΄λλΌλ μ νμ μ§μνμ.
μΈμμ μ΄λ»κ² λ°λΌλ³Ό κ²μ΄λ
https://maily.so/capitaledge/posts/99ed5dd7
λ§€νΈλ¦¬μ€λ₯Ό νλ§€νλ μΊμ€νΌλ νκ°μ λ§€κ°λμκ³ μ€λ¦¬μ½λ°Έλ¦¬μ μ λ°λ‘ μ£Όλͺ©λ°μλ μ¬λ²μ¦λ μμ₯ ν μ£Όκ°κ° 99% νλ½ν μν©μ
λλ€.
κ·ΈλΌμλ λΆκ΅¬νκ³ μ¬μ ν μλΉμ μμ₯μ μ§μ€νλ ν¬μμλ€μ΄ μμ΅λλ€. DTC λͺ¨λΈμ΄ μ΄λ €μμ κ²ͺμ κ²μ΄μ§, μλΉμ μΉν°μμλ μ¬μ ν μλ‘μ΄ κΈ°μ
λ€μ΄ νμνκ³ λΉ λ₯΄κ² μ±μ₯νλ©° μλΉμλ€μ μ§κ°μ μ΄κ² νλ μ¬λ‘λ€μ΄ κ³μν΄μ λνλκ³ μκΈ° λλ¬Έμ
λλ€.
Repost from μ¬μ λ‘κ³ κΉκ² I freebutdeep
μμ΅λ₯ μκΈ°λ μλκ³ μλ°κΈ°μ λκΌλ μ μ λμ§μ΄λ΄€μ΅λλ€. κΈμ΄ μ΄λ ΅κ² μ°μλ λ§νΌ, μμ§ λ κΉ¨λ¬μ κ²μ μ무κ°μ μμ±ν κ² κ°μ΅λλ€. 무μΈκ° λͺ
λ£νκ² λ§ν μ μλ€λ 건 μμ΄ μμν μΌμ΄μ§λ§, κ·Έλ¬κΈ° μν΄μλ μ λ§ λ§μ κ³ λ―Όκ³Ό 곡λΆκ° μλ°λλ€μ. μΈμ κ°λ μ½κ³ μ¬ννκ² ν΄λ³΄κ³ μΆμ΅λλ€. (λΈκΉ)
https://freebutdeep.substack.com/p/34-2024 #ν¬μ
Shopifyμ λͺ¨λ λ―Έμ
μ 6μ£Όλ§λ€ κ²ν λ©λλ€. κ·Έλμ ν λΉ, κΈλ , μ λ ν λ°©μ λͺ¨μ¬ νμ¬μ λͺ¨λ λ―Έμ
μ κ²ν ν©λλ€. λ°λΌμ νμ¬μ μ΄λ€ λ―Έμ
λ μ΅λ 6μ£Ό λμ κΆ€λμμ λ²μ΄λ μ μμ΅λλ€. λ³κ²½μ΄ νμνλ€λ©΄ λ°λ‘ λ³κ²½ν©λλ€. λλΆλΆμ νμ¬μμ λͺ λ¬, λͺ λΆκΈ°, λͺ λ
μ΄ κ±Έλ¦¬λ μΌμ Shopifyμμλ μ΅λ 6μ£Ό λ§μ λλΌ μ μμ΅λλ€." (Source: 2023 Shopify Investor Day)
https://youtu.be/lPrRVnxHqvE?si=TSZZMu0led8YWZ81
ν루ν루 μ€μν κ²μ μΆ©μ€ν κ². κ·Έλ° ν루λ€μ΄ λμ λλ κ².
Repost from μ μ’
νμ μΈμ¬μ΄νΈ
μ΄κ±΄ λΆλͺ
μ€λΉλ μλ§μ΄ μ μ μλ κ°μ΄μμν
λ€.
"λΆμ²μ λ λ ν°λ¬Όμμ λ λκ° λλ€λ κ°μ΄ μλ€."
https://www.joongang.co.kr/article/25260802
곡λΆλΏλ§μ΄ μλλΌ, μ΄λ €μλΆν° μ¬λ§ν 건 λ€ μ§μ κ²°μ νλ€. λ¬Όλ‘ λ¬΄μ¨ λ¬Έμ κ° μκΈ°λ©΄ λΆλͺ¨λμ΄ λμ κ°μ΄ ν΄κ²° λ°©λ²μ μ°Ύμκ³ , νμ λνλ₯Ό λ§μ΄ νλ€. κ·Έλλ μ¬μ ν κ²°μ μ λ΄ λͺ«μ΄μλ€. μ§λ‘λ μΈμ λ°©ν₯λ λ΄κ° μ ννκ³ , λΆλͺ¨λμ μ‘΄μ€νλ€.
κ·Έλ¬λ λλ € μμΌλ‘ μ΄λ»κ² μ΄μμΌ ν μ§ μκ°μ΄ λ§μλ€. λ΄κ° λ μ’μνκ³ λ μνλμ§ λ 골λν μκ°νλ©° λ΅μ μ°Ύμλ€. κ·Έκ² μ¬μ
μ΄μλ€.
μ΄μ©λ©΄ νΌμλ§μ μ΄λ±κ°μ΄μμ§ λͺ¨λ₯Έλ€. μ΄μ¨λ μΉμμ λκ° μ μ§ μμμ§λ§, κ·Έλλ μ§κΈμ΄λ λΆνΈν μν©μ λ§λ₯λ¨λ¦¬λ©΄ μ€νλ € κΈμ μ μκ·Ήμ λ°λλ€. κ°λ Ή 2021λ
'λ§λν΄'μ΄ μ²μ ν¬μλ°μ μ¦μ "'룩 λΆ(λΈλλ μ€νμΌμ 보μ¬μ£Όλ μ¬μ§μ§)' μλ λΈλλκ° λ¬΄μ¨ λΈλλλ"λ μ리λ₯Ό λ€μλ€. μ무리 μ
κ³ κ΄νμ΄λΌμ§λ§, 룩 λΆ νμμ±μ λͺ» λλΌλλ° λ¨μ΄ νλκΉ λ¬΄μ‘°κ±΄ λ°λΌ νκ³ μΆμ§ μμλ€. κ·Έλ€ μ
λ§μ λ§μΆλ λμ "λ΄κ° λ§λ€λ κ±Έ 보μ¬μ£Όκ² λ€"κ³ λ€μ§νλ€.
κΈμμ λ컀λ
λ λΈ λν νλΉλ λΉ λ―ν νλ²ν λ§λ²μ΄ κ°μ μ κ³ μ‘Έ νλ ₯μΈ λ΄κ° μ±κ³΅νλ μ¬λ‘λ₯Ό 보μ¬μ£Όκ³ μΆμλ€. λ΄κ² κ·Έλ° μ‘΄μ¬κ° μμκΈ°μ, λΉμ·ν κΏμ κΎΈλ μ΄λ€μκ² κΈΈλΌμ‘μ΄κ° λκ³ μΆλ€.
λλλ¬Έμ κ°μ λμμΈμ μ΅μ λ μ₯ μ¬μΌ νκΈ°μ, μνλ 15λ§ μμ§λ¦¬ λ©μ§ λμμΈ λμ κ°λΉ 2λ§~3λ§ μλ κ°μΌ μ νΌ λ μ₯μ κ³¨λΌ λΈλ‘κ·Έμ μ¬λ Έλ€. μ νλ Έλ€.
μΆ μ²μ§ λλ₯Ό μ§μΌλ³΄λ μ΄λ¨Έλκ° "μΌλ§λ©΄ λκ² λλ"κ³ λ¬Όμλ€. "30λ§μ. " κ·Έ λμΌλ‘ λλλ¬Έμ λ¬λ €κ° κ±°μΉ μμΈλ€ μλνλ©° μ νΌ λ μ₯μ μλ€. λ§μ°¨λ μ§μμ λκΈ°κ³ μμ€μ 4000~5000μλΏμ΄μλ€. μΉ΄νμμ μ μμ°¨λ₯Ό μμ λκ³ λΆμ²κ°λ 첫차λ₯Ό κΈ°λ€λ¦¬λ©° ν¬μ€νΈμμ μ΄λ κ² μΌλ€. "μ±κ³΅ν κ±°λ€. " μ΄ λ©λͺ¨λ μκΈ°μ€νμ μμΈμ΄ λλ€.
6κ°μμ―€ μ§λλ μ· μ΄ λ λ§κ³ λ 100λ§μμ―€ λ¨μ μ μλλ₯Ό νλ€. λ¨μΌ νλͺ©μ΄ μλλΌ μ¬κΈ°μ κΈ°μ μ½νΈΒ·μ€μΉ΄ν λ±μΌλ‘ μ§μ μ½λν΄ ν보μ²λΌ μ°μ΄ μ¬λ Έλλ°, μ λͺμ²λ§μ λ§€μΆμ΄ λ λ§νΌ λλ°μ΄ λ¬λ€. 1λ
μ―€ μ§λμ κ³ κ°λ€μ΄ μ΄λ°μ λ° μꡬλ₯Ό λ΄μ "μ§μ λμμΈν΄λ¬λΌ"κ³ ν΄μ μ μκΉμ§ λ°μ΄λ€μλ€.
ν¨μ
μ λ¬Έ ν¬μμ¬ λ± μ¬λ¬ κ³³μ΄ ν¬μλ₯Ό μ μν΄μλ€. κ²½μ λ₯λ ₯μ νκ³λ₯Ό λΌμ λ¦¬κ² λκΌκΈ°μ κ°μ₯ λ§μ λμ ν¬μνκ² λ€λ κ³³μ 골λλ€. λ§€μΆμ 150μ΅(2021), 500μ΅(2022), 1000μ΅μ(2023)μΌλ‘ κ°νλ₯΄κ² μ¬λλ€. νμ§λ§ 60% μ§λΆν¬μλ₯Ό ν ν¬μμ¬μμ κ°λ±μΌλ‘ λλ΄ λ΄κ° ν€μ΄ 'λ§λν΄'κ³Ό μ§λν΄ κ²°λ³νλ€.
μμνλ€. νμ§λ§ κ³Όκ±°μ μ¬λ‘μ‘νλ©΄ μ€λμ μ΄ μ μλ€. λ―Έλλ‘ λμκ° μ μλ€. λ΄ μ νμ κ²°κ³Όλ λꡬ νλ, μ΄λ€ λ³λͺ
λ νκ³ μΆμ§ μμλ€. κ²λ€κ° μμ κ³Ό κ²Ήμ³ μ κ²½ μΈ μ¬λ ₯λ μμλ€. κΉλνκ² λ΄λ €λ¨λ€. μ§λ 3μ 첫 μμ΄ μΆμ° 30λΆ μ κΉμ§λ ν΄λν°μΌλ‘ μ
무 μ²λ¦¬νκ³ , μ°ν쑰리μμμ μ§μλ€μ΄ 보λ΄μ€ λμμΈμ 컨ννλ€. μ¬λ κ² νΈμΉ μμλ€. λνκ° μ§μ¬μΌλ‘ λκ΅¬λ³΄λ€ μ΄μ¬ν μΌν΄μΌ μ§μλ€μ΄ λ°λΌμ¨λ€κ³ μκ°ν΄μλ€. κ·Έλ κ² μ§λ 5μ 50λ
μμ¬ ν¨μ
κΈ°μ
(μΈμ )μ ν¬μλ₯Ό λ°μ λ΄ μ΄λ¦μ λ΄ μ λΈλλλ₯Ό λ°μΉνλ€.
λ λ¨λ μκ°λΆν° μ λ€κΈ° μ§μ κΉμ§, μ¬μ§μ΄ μΆν΄κ·Ό μ κ³Ό νλ μ
무μ μ°μ₯μ΄λ€. μ λ
μ νλ μκ° SNS λΌμ΄λΈλ‘ κ³ κ°κ³Ό μν΅νκ³ , λ€μλ μ΄μ€ λ°μμ΄ μ’μ κ±Έ κ³¨λΌ μ
κ³ λ μ μλ μ§μμ λ¨νΈκ³Ό ν¨κ» μΆκ·ΌκΈΈ μΉ΄νμ λ€λ¬ λ°μΌλ¦¬λ£©μ μΈμ€νκ·Έλ¨μ μ¬λ¦°λ€. μ΄ κ³Όμ μμ λμμΈΒ·μμ λ± κ³ κ° μꡬλ₯Ό μ΅λν 빨리 λ°μν΄ μνμΌλ‘ λ΄λλλ€. μ§μλ€μ νλ€κ² μ§λ§ μνλ λΈλλκ° λ무 λ§μ μμ¦ μ΄λ κ² νμ§ μμΌλ©΄ μ΄μλ¨μ μ μλ€.
μ΄λ κ² μΌμ λͺ°λνλ©΄μ μ μμ΄λ₯Ό λ³μλλκ³ λ¬»λ μ¬λλ€μ΄ μλ€. λ μ»€λ¦¬μ΄ μ±μ₯ λͺ»μ§μκ² νλ³΅μ΄ μ€μνλ€. ν΄κ·Ό ν λ¨νΈκ³Ό κΉμΉμ°κ°μ μμ£Ό νμνλ©΄μ λννλ μΌκ³Ό μΌμμ κ· νμ΄ λ κΈ°μκ³ ν볡νλ€. λΉμ₯μ μ±κ³΅μ μν΄ μμ΄λ₯Ό μ ννμ§ μλλ€λ©΄ λμ€μ 곡ννκ³ ννν κ±° κ°μλ€. μ¬μ
μ ν€μμ€λ©΄μ λ κ²½κ³νλ μ§μ μ΄κΈ°λ νλ€. μ±κ³΅ν΄μ λΆνν΄μ§μ§ λ§μ. λμ΄λ , 컀리μ΄λ , λͺ
μλ , κ·Έ 무μμλ λ΄ μΆμ΄ μ‘μλ¨Ήνμ§ μμμΌλ©΄ μ’κ² λ€. κ·Έλ κ² λ€μΈμμ ν λ² λ μ±κ³΅μμΌ, 첫 λ²μ§Έ μ±κ³΅μ΄ μ΄μ΄ μλμλ€λ κ±Έ μ¦λͺ
νκ³ μΆλ€.
https://www.joongang.co.kr/article/25260802
In case you missed it: Chrome is adding a Gemini Nano AI model right inside your browser via a new
window.ai API.
Spent some time this weekend putting together a simple open source browser extension using Chromeβs experimental local LLM API and Cartesia's super-fast TTS model β‘οΈ
It can:
β Summarize highlighted text (Cmd+Shift+S)
β Explain content as if you're 5 years old (Cmd+Shift+E)
... all on-device!
Code: https://lnkd.in/gyWPHyK7
https://www.linkedin.com/posts/varunshenoy_ai-ugcPost-7213965233600495616-K4CE?utm_source=share&utm_medium=member_iosλͺ¨λ λ§λ¨/λ―Έν
μ μ΅μ μ λ€ν΄μΌνλ€.
κ·Έλ₯ μ§λμΉ μλ μμλ λ―Έν
μ΄(ν° μμ λ€ μμ΄ μ§νλμλ λ―Έν
μ΄), λμ€μ μ’μ μλ―Έμ ν° impact κ° λμ΄ λμμ€λ κ²½μ°λ₯Ό μ’
μ’
κ²½ννλ€. μΌλ‘λ‘, μ μ λΆκ³Ό μ§ννλ λ―Έν
μ΄ λμ€μ λλΉν¨κ³Όκ° λμ΄ ν¬μ μ μΉλ‘ μ΄μ΄μ§κΈ°λ νκ³ , ν° B2B Deal λ‘ μ΄μ΄μ§κΈ°λ νλ€.
κ·Έλ° κ²½νμ λͺ λ² ν μ΄νλ‘λ, 'μ΄κ² μ κ² μ¬μ§ λ§κ³ , λͺ¨λ λ―Έν
/λ§λ¨μ μ΅μ μ λ€νμ' μκ°νλ©° λ―Έν
μ μ§ννκ³ μλ€.
λ―Έν
μ μ§ννκ³ μ¬λμ λ§λ λ, μ΅κ΄μ μΌλ‘ λ€μ§νλ λΆλΆμ μλμ κ°λ€.
1. μλλ°©μ μ£λΆλ¦¬ νλ¨νμ§ μλλ€. (μλλ°©μ Judge νλ©΄ μλλ€. λμ νΈνν¨κ³Ό μ΄λ¦¬μμμ λλ¬λ λΏμ΄λ©°, ν° κΈ°νλ₯Ό λμΉλ μ§λ¦κΈΈμ΄λ€)
2, λλ₯Ό νμλ‘ νλ μ¬λ/μλ¦¬κ° μμΌλ©΄ μ΅λν κ°λ€. κ°μ μ΅μ μ λ€νλ€. (λ°μμ²νμ§ μκ³ νκΈ°μ§ μλλ€. μκ°μ΄ μλ λλ μ΅λν μν΄λ₯Ό ꡬνλ€)
3. μλ―Όνκ±°λ κΉλ€λ‘κ² κ΅΄μ§ μλλ€. (μλλ°©μ νΌκ³€νκ²λ§ ν λΏμ΄λ€)
4. λ무 λ΄ μμ λ€λ§ λ΄μΈμ°μ§ μλλ€. μλλ°©μκ² λΆννκ³ μΆμ λ΄μ©μ΄ μμ λμλ, 1) μν΄λ₯Ό ꡬνκ³ , 2) μμ§νκ² λμμ ꡬνλ€.
5, λ―Έν
μ€ μκ°λ°λ μ¬λμ΄ μλ€λ©΄ (μ: μ΄ λΆμκ² μ°λ½ν΄ 보μΈμ) , μ°λ½μ² λ°μ ν μ΅λν λΉ λ₯΄κ² μ°λ½νκ³ , μ΅μ μ λ€ν΄ λ§λλ€ (μκ°ν΄μ€ λΆμ΄ νΌν΄λ°μ§ μκ² νλ€)Β
6. νμ μ μ΅λν 빨리 νλ€.
7. λν μ€ λ¨ μ΄μΌκΈ° νμ§ μλλ€ (λ¨ μ΄μΌκΈ°λ μ’μ μ΄μΌκΈ°λΌ ν΄λ μ λ§νλ©΄ μνλ κ²μ΄ λ μ’λ€)
8. μμ΄μ»¨ν μνκ³ , μ¨μ ν μ§μ€νλ€. (μ΄μΌκΈ°μ μ§μ€νμ§ μμΌλ©΄ λ°λ‘ ν°κ° λλ€)
9. μλλ°©μ΄ λΆνΈν μ μλ νν/νλμ μ΅λν νΌνκ³ , μ μΉ/μΈμ’
/μ’
κ΅ λ± λ―Όκ°ν λΆλΆμ νλμ μ¬λ¦¬μ§ μλλ€.
10. μ μμ λμμ μ μν λμλ, μ§ν¬ μ μκ³ & μ€νν μ μλ κ²λ§ μ μνλ€. (μκ°μ μ μκ°(?)μ μ¬λ‘μ‘ν 곡μν λ 리면 μλλ€)
11. μΈμ¬ μνκ³ , μ²κ²°ν¨μ μ μ§νλ€.
νΉν μ°½μ
μλ λͺ¨λ λ§λ¨μ μ΅μ μ λ€ν νμκ° μλ€. νμ¬μ μ±μ₯μ λͺ¨λ κ°λ₯μ±μ΄ μ΄λ €μμ μ μμμ, μ°½μ
μμ νλμ λ¬λ €μλ€.
μ’μ μΆμ μν μλ΄μ https://m.yes24.com/Goods/Detail/108551375
1. μ€ν μμ£Όμμλ€μ λͺ©νλ λͺ¨λ κ°μ μ λͺ°μλ΄λ κ²μ΄ μλλΌ λΆμ μ μΈ κ°μ λ§μ λͺ°μλ΄λ κ²μ΄λ€. μ κ²½μμ μ λ‘ μ»λ 무κ°κ° μνκ° μλλΌ, λΆλ
Έ, μ¬ν, λΆμ, λλ €μ λ±μ λΆμ μ κ°μ μ΄ μ΅μνλ λμμ κΈ°μ¨ λ±μ κΈμ μ κ°μ μ΄ κ°λμ°¬ λ§μ μν.
1. μκΈ°κ° κ°μ§ κ²μμ κΈ°μ¨μ μ°ΎμΌλ©° λ΄λ©΄μ κΈ°μ¨ μΈμ λ€λ₯Έ κΈ°μ¨μ λ°λΌμ§ μμ κ²μ΄λ€.
2. μλ§μ κ΄ν΄μ) λ§μ‘±ν μ€ λͺ¨λ₯΄λ λ§μμ 극볡νμ§ μκ³ λ μ’μ μΆ, μλ―Έ μλ μΆμ μ΄ μ μλ€λ κ²μ΄μλ€. λμμ΄ λ κ°μ§λ €λ μ±ν₯μ λ€μ€λ¦¬λ μ’μ λ°©λ²μ΄ μ§κΈ κ°μ§ κ²μ λ§μ‘±νλ κ².
3. μ°μ κ³Ό λΆ λ± μΆμ΄ μ μ¬νλ μ’μ κ²μ μ¦κΈ°λ κ±°κΈ°μ μ§μ°©ν΄μλ μλλ€κ³ 보μλ€. μ€ν μ μ² νμλ€μ μΆμ΄ μ£Όλ μ λ¬Όμ μ¦κΈ°λ λ°μ μ’
μ’
λ²μ΄λ μ§κΈ μ¦κΈ°κ³ μλ κ²μ μΈμ λ μμ μ μλ€λ μ¬μ€μ μμ£Ό μκ³ ν΄μΌνλ€κ³ 보μλ€.
2. μ€ν μ μ² νμ μλ ¨νλλ°μλ λΉμ©μ΄ λ λ€. μΆμ μ² νμ κ°μ§ μμμ λ μΉλ₯΄λ λΉμ©μ μ΄λ³΄λ€ λ ν¬λ€. 무μΉν κ²μ μ’μΌλ©° ν루ν루λ₯Ό 보λ΄λ€ κ²°κ΅ μμ€ν μΆμ λλΉνλ μνμ΄ λ ν° λΉμ©μ΄λ€.
3. μ°λ¦¬λ μ΄λͺ
μ΄ μ΄ μ°κ·Ήμ λ±μ₯νλ λ°°μ°μ μ§λμ§ μλλ€. μ°λ¦¬λ μ°κ·Ήμ λ°°μμ μ νν μ μλ€. μ΄λ€ λ°°μμ΄ μ£Όμ΄μ§λ μ΅μ μ λ€ν΄ μ°κΈ°ν λΏμ΄λ€.
4. μ±κ°μ μ¬λμ λν λλ λλ₯Ό μ±κ°μμ μ¬κΈ°λ μ¬λλ μ‘΄μ¬νλ€λ μ¬μ€μ κΈ°μ΅νλΌ. μλμ μλͺ»μ μ§μ¦μ΄ λ λλ μ μ λ©μΆ° μμ μ κ²°μ μ λν΄μ μκ°νλΌ. κ·Έλ¬λ©΄ μλμ κ²°μ μ 곡κ°νλ©° κ΄λνκ² λν μ μλ€.
5. λ€λ₯Έ μ¬λμ λν λμ μκ°μ ν΅μ νλ©΄, κ·Έλ€μ΄ μ°λ¦¬ μΆμ λ―ΈμΉλ λΆμ μ μΈ μν₯μ μ€μΌμ μλ€. μ£Όλ³ μ¬λμ νλκ³Ό λ§, μκ°κ³Ό κ³νμ λν΄ μΆμΈ‘νλλ°μ μκ°μ λλΉνμ§ λ§μμΌνλ€. κ·Έλ€μ κ΄ν΄ μ λ©λλ‘ μμνκ³ μ§ν¬νκ³ μκΈ°νκ³ μμ¬ν΄μλ μλλ€. λ§λ₯΄μΏ μ€μ λ°λ₯΄λ©΄ νλ₯ν μ€ν μμ£Όμμλ 곡곡μ μ΄μ΅μ μν΄μκ° μλλ©΄ λ€λ₯Έ μ¬λμ μκ°μ λν΄ μκ°νμ§ μλλ€.
6. μ°λ¦¬ λ΄λ©΄μ λΆλ
Έμ μ¦μ€, 볡μμ¬μ΄ μΌμ΄λ λ ν μ μλ μ΅μμ 볡μλ κ·Έ μ¬λκ³Ό λκ°μ΄ λκΈ°λ₯Ό κ±°λΆνλ κ²μ΄λ€.
7. μ¬λλ€μ΄ λΆνν μ΄μ λ κ°μΉμλ κ²μ΄ 무μμΈμ§ λͺ¨λ₯΄κΈ° λλ¬Έμ΄λ€. μ λλ‘ λͺ¨λ₯΄κΈ° λλ¬Έμ μμ μ ν볡νκ² ν΄μ£Όλ κ²μ΄ μλλΌ λΆμνκ³ λΉμ°Ένκ² λ§λλ κ²μ μ’μΌλ©° ν루 ν루λ₯Ό 보λΈλ€.
8. μμ λ₯Ό μ§ν€λ €λ©΄ λμ κ΄ν μ¬λλ€μ ννμ 무μ¬ν΄μΌνλ€. μ¬λλ€μ λΆμΈμ μ΄ μ«λ€λ©΄ κ·Έλ§νΌ μ¬λλ€μ μΈμ μλ μ΄μ°ν΄μΌνλ€.
9. μ€ν μ μ² νμλ€μ μμ λ₯Ό μ€μνλ€. μ°λ¦¬κ° κ°μ§ νμ νμΈμκ² λ겨주λ νμλ 무μμ΄λ κΊΌλ Έλ€. μ¬νμ μ§μλ₯Ό μ’λλ€λ κ²μ μ°λ¦¬λ₯Ό μ§λ°°νλ νμ λ€λ₯Έ μ¬λλ€μκ² λ겨주λ κ²κ³Ό κ°λ€.
10. μ£½μ λ€ λͺ
μ±μ 곡νν λͺ
μ±μ΄λ€. μ£½μμλ λͺ
μ±μ λ릴 μ μλ€. λ―Έλ μΈλκ° νλ²λ λ³Έ μ μλ μ¬λμ μΉμ°¬ν κ²μ΄λΌ κΈ°λνλ κ²λ λΆμ§μλ€.
11. λΆλ₯Ό κ°μ§λ€κ³ μ¬ν μλ μΆμ μ΄ μ μμΌλ©° λμΌλ©΄ λΆκ° μ°λ¦¬λ₯Ό μλ‘ν μλ μλ€. λΆλ λ¬Όμ§μ μ¬μΉλ₯Ό μ»κ³ κ°κ°μ μΎλ½μ λλ¦¬κ² νμ§λ§ μ§μ ν λ§μ‘±μ κ°μ Έλ€μ£Όμ§λ μ¬νμ 물리μΉμ§λ λͺ»νλ€.
12. νλ²ν μΆ, κΈ°λ³Έλ§ κ°μΆ μνμμ κΈ°μ¨μ μ°Ύλ λ₯λ ₯μ΄ μ€μνλ€.
13. λλ €μμ νΌνλ 3κ°μ§ λ°©λ²
1. μμ μ νμ±μμ μ΅μ μ λ€ν΄ κΈ°μ¨μ μ°Ύλλ€.
2. κ·Έκ²μ μμ λλ₯Ό μ€λΉνλ€. μ§κΈ λ리λ κ²λ€μ΄ νμ΄μ μ¬κ±΄μμ κΈ°μ΅νλ€.
3. λμ΄ λμμ§μ§ μλλ‘ μ μνλ€. μ΅κ³ κ° μλλ©΄ κΈ°μ¨μ λͺ» λλΌλ μ¬λμ΄ λμ§ μλλ‘ μ‘°μ¬νλ€.
14. μ½κ² μ»μ μ μκ³ λꡬλ μμκ° μ μλ κ²μ μ¦κ²ΌκΈ° λλ¬Έμ μΆμμ μ¦κΈΈ κ²μ΄ μμ£Ό λ§μλ€. κ·Έλ€μ μλ κ·Έλλ‘μ μμ μ μ¦κ²ΌμΌλ©° μ§κΈ μ΄λλ‘μ μΆκ³Ό μΈμμ μ¦κ²Όλ€. μ΄κ²μ κ²°μ½ μμ μ±μ·¨κ° μλλ€.
15. μ΄μ μ 볡νκΈ°λ³΄λ€ μκΈ° μμ μ μ 볡νλΌ. κΈ°μ‘΄μ μ§μλ₯Ό λ°κΎΈκΈ°λ³΄λ€ μμ μ μλ§μ λ°κΎΈλ €κ³ λ
Έλ ₯νλΌ. μμ μ μκ° μΈμ μ΄λ€ κ²λ μλ²½ν ν΅μ ν μ μμμ λ―ΏμΌλΌ. λ¬Έμ ν΄κ²°μ μ΅μ μ λ€νλ€λ©΄ λ μ΄μ ν μ μλ μΌμ΄ μλ€κ³ λ―ΏμΌλΌ.
Data, Size of Models, and Efficiency
No Priors: Apple recently announced models running on devices at around three billion parameters. How does this relate to your focus on efficiency?
Karan Goel: The initial wave of companies focused on scaling data and compute, leading to great models. The second wave focuses on efficiency, enabling powerful models to run cheaply and repeatedly at scale. Our technology aims to make high-capability models run efficiently on smaller devices, moving from data centers to the edge.
Recent Launch of Text to Speech Product
No Priors: Cartesia recently launched its text-to-speech product, which is impressive in terms of performance. Can you tell us more about that launch?
Karan Goel: We wanted to put our technology to work, demonstrating efficient audio generation. Text-to-speech is a longstanding field, yet thereβs still room for improvement. Our goal is to create engaging, high-quality speech that people would want to interact with for more than 30 seconds.
Multi-modality & Building Blocks
No Priors: How has your pioneering work on SSMs impacted your approach to multi-modality or speech?
Albert Gu: Multi-modality hasn't been the primary motivation, but different modalities present different challenges, influencing the design of our models. SSMs are versatile building blocks that can be applied to various modalities, including audio and text.
Whatβs Next at Cartesia?
No Priors: Whatβs next for Cartesia? Will you focus on Sonic and audio, or explore other modalities?
Karan Goel: Weβre excited about Sonic, showing the potential for real-time, low-latency audio generation. We aim to bring Sonic on-device, enhancing its capabilities and efficiency. Long-term, weβre developing multimodal models that can converse intelligently, integrating audio and text reasoning. Our goal is to build versatile models that run efficiently on various devices.
Latency in Text to Speech
No Priors: Latency is a significant issue in text-to-speech applications. How does multimodality address this?
Karan Goel: Reducing latency is crucial. Orchestrating multiple models adds overhead, making it inefficient. We aim to create seamless, integrated models that handle multimodal tasks efficiently, reducing latency and improving performance.
Choosing Research Problems Based on Aesthetic
No Priors: How do you choose research problems?
Albert Gu: I choose problems based on aesthetic appeal. Some solutions feel elegant and correct, driving my interest. This approach guided the development of SSMs and continues to influence my work.
Product Demo
No Priors: Can you show us a demo of your product?
Karan Goel: Sure. Hereβs our Sonic text-to-speech model running on a MacBook in real-time. [Demo showing the model generating speech quickly and efficiently.] This demonstrates the speed and quality of our model.
Cartesia Team & Hiring
No Priors: How large is your team, and are you hiring?
Karan Goel: We have 15 people and 8 interns. Weβre hiring across the engineering stack, especially for model roles. Weβre excited to have talented people join us in building the future.
No Priors: Thank you for joining us.
Karan Goel & Albert Gu: Thank you.
No Priors: Welcome back to No Priors. We're excited to talk to Karan Goel and Albert Gu, the co-founders of Cartesia and authors behind revolutionary models such as S4 and Mamba. They're leading a rebellion against the dominant architecture of Transformers, and weβre excited to talk to them about their company today. Welcome, Karan and Albert.
Karan Goel & Albert Gu: Thank you, nice to be here.
Use Cases for Cartesia and Sonic
No Priors: Tell us a little more about Cartesia, the product, and what people can do with it today.
Karan Goel: Definitely. We launched Sonic, a really fast text-to-speech engine. Some exciting use cases include interactive, low-latency voice generation. We've seen a lot of interest in gaming, where players can interact with characters and NPCs in real-time. Another area is voice agents, where low latency is crucial. Weβve managed to reduce latency by 150 milliseconds and aim to shave off another 600 milliseconds over the next year.
Karan Goel & Albert Guβs Professional Backgrounds
No Priors: Can you talk about your backgrounds and how you ended up starting Cartesia?
Albert Gu: Karan and I both came from the same PhD group at Stanford. I worked on sequence modeling during my PhD, starting with problems at DeepMind. I became interested in recurrent models around the same time Transformers gained popularity. Karan and I collaborated on the S4 model, which showed the effectiveness of state space models. Recently, I proposed Mamba, which has shown promising results in language modeling. Iβve also started as a professor at CMU, continuing research while also working on Cartesia.
Karan Goel: I grew up in India, from an engineering family. I initially aimed to be a doctor but switched to engineering due to low aptitude in biology. After IIT and grad school, I ended up at Stanford, working on reinforcement learning. Chris Ray, our PhD adviser, was skeptical about reinforcement learning, leading me to explore other areas. Albert and I worked together on various projects, including S4, which drew me into the field.
State Space Models (SSMs) versus Transformer-Based Architectures
No Priors: Could you tell us more about SSMs and how they differ from Transformer-based architectures?
Albert Gu: SSMs originated from recurrent neural networks and focus on processing sequences one at a time, updating beliefs or states with new information. This approach can be loosely inspired by brain functions. SSMs are particularly good at modeling perceptual signals, like raw waveforms and pixels, which are less compressible than text data. Recent models like Mamba have improved at handling text data. One key advantage of SSMs is their linear scaling, providing constant time processing for new tokens, unlike Transformers' quadratic scaling. However, both models have their strengths and weaknesses, and hybrid models combining SSMs and Transformers have shown promising results.
Domain Applications for Hybrid Approaches
No Priors: Are there specific domains where you see applications for these hybrid approaches?
Albert Gu: Currently, most applications are focused on text, as that's where interest lies. However, SSMs have been applied to DNA modeling, pre-training on long DNA sequences for tasks like detecting point mutations.
Text to Speech and Voice
No Priors: How did your research lead to focusing on text-to-speech and voice generation?
Albert Gu: We wanted to demonstrate the versatility of our models in practical applications. Audio seemed like a natural fit due to its real-time processing requirements. We thought it would be a cool first application.
Karan Goel: Audio and multimodal data are interesting because they are information-sparse, allowing for fast processing and large context handling. Audio, in particular, has many emerging commercial applications, like voice agents and gaming. Efficient models can also be deployed on smaller hardware, pushing inference closer to the edge rather than relying on data centers.
https://youtu.be/Ti6Gft5xO1s?si=BMf4_KB8Ln2J2iPD
Experts tell you what will not happen large change is driven by entrepreneur led companies.
Available now! Telegram Research 2025 β the year's key insights 
