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 291
Subscribers
-324 hours
-167 days
-4530 days
Posts Archive
βMachine learning costs, talent and chip shortagesβ¦ any AI and machine learning company faces at least one of these challenges, and most face a few at a time,β Pekhimenko told TechCrunch in an email interview. βThe highest-end chips are commonly unavailable due to the large demand from enterprises and startups alike. This leads to companies sacrificing on the size of the model they can deploy or results in higher inference latencies for their deployed models.β
With spending on AI-focused chips expected to hit $53 billion this year and more than double in the next four years, according to Gartner, Pekhimenko felt the time was right to launch software that could make models run more efficiently on existing hardware.
βTraining AI and machine learning models is increasingly expensive,β Pekhimenko said. βWith CentMLβs optimization technology, weβre able to reduce expenses up to 80% without compromising speed or accuracy.β
βFor one of our customers, we optimized their Llama 2 model to work 3x faster by using Nvidia A10 GPU cards,β
CentML isnβt the first to take a software-based approach to model optimization. It has competitors in MosaicML, which Databricks acquired in June for $1.3 billion, and OctoML, which landed an $85 million cash infusion in November 2021 for its machine learning acceleration platform.
βThe CentML platform can run any model,β Pekhimenko said. βCentML produces optimized code for a variety of GPUs and reduces the memory needed to deploy models, and, as such, allows teams to deploy on smaller and cheaper GPUs.β
μμ΄ν° μΆμ ν 첫 4~5λ
μ λλ μ±μ μΆμνλ κ²λ§μΌλ‘λ μ μ μ wowλ₯Ό μ΄λμ΄ λΌ μ μμλ μμ μ΄ μμλ€. μ΄λνλ©° μΈν°λ·λ ν μ μκ³ ν°μΉμ€ν¬λ¦°μ νμ©ν μ μλ μ€λ§νΈν°μμ ꡬλμ΄ λλ€λ κ²λ§μΌλ‘λ wow μλ κ²μ΄λ€.
κ·Έ μ΄ν κΈ°μ‘΄ λ κ°μ μ°μ
μ ν¨λ¬λ€μμ λ°κΎΌ μλΉμ€λ€μ΄ λ±μ₯νμ¬ wow λ₯Ό μ μ¬νλ€. μ ν¬λΈ/λ·νλ¦μ€λ TV/λ―Έλμ΄ μ°μ
μ λ°κΏ¨κ³ , νλΆ/ꡬκΈμ λ―Έλμ΄/ν¬νΈμ°μ
μ λ°κΏλμμΌλ©°, μλ§μ‘΄μ μΌν/μλ²μμ€ν
μ, ν
μ¬λΌλ μλμ°¨ μ°μ
μ, μ°λ²λ νμ μ°μ
μ, μμ΄λΉμλΉλ₯Ό νΈν
μ°μ
μ ν° μν₯μ 쀬λ€.
μμ κ°μ 2000λ
μ΄λ° μ΄ν 격λμ 20λ
μ λλ₯Ό κ²ͺμΌλ©°, νλμ¬νμ λ§μ μ¬λλ€μ λͺ¨λ°μΌ λ° μΉ μ€μ¬μ μλΉμ€μ κ½€ λ§μ΄ μ΅μν΄μ§ λ― νλ€. κ·Έλμ, μλ‘ μΆμλλ μ±μ λ΄λ κ·Έ κ°ν₯μ΄ κ³Όκ±° λλΉ λ¬΄λμ§κ³ μμμ λλλ€.
μ€νλ €, μμ¦μ κ³Όκ±°μλ λ§μ‘±νλ©° μ¬μ©νλ Big Tech μ¬μ μλΉμ€ μ‘°μ°¨λ μ¬μ¬ 'μ§κ²¨μμ§λ€' λλΌλ λΆλ€μ΄ μ¦κ°νλ λ―νλ€. λμμ 'κ³Όκ±°λΆν° λ§μ΄ μ΄μ©νλ μλΉμ€λ€μ΄ μ΅κ·Όμλ μμ λ§ λͺ»ν΄μ μ€λ§κ°μ΄ ν¬κ³ , μ΄μ© μ μμ΄ μ΄μ©νκ³ μλ€'κ³ λ§νλ λΆλ€λ λ§μμ§λ λ―νλ€. OpenAI κ° chatGPT λ₯Ό μΆμνλ©° μλ‘μ΄ νλ ₯μ λΆμ΄λ£κ³ μμ§λ§, κ·Έ νκΈλ ₯μ΄ κ³Όκ±°μ μμ΄ν° λͺ¨λ©νΈμ²λΌ μμ²λμ§ μμ κ²λ μ¬μ€μ΄λ€. (chatGPTλ μλΉμ 보λ€λ μ€ννΈμ
μ νλ ₯μ λΆμ΄λ£κ³ μλ λ―νλ€. κ·Έλμ κ±±μ μ΄κΈ°λ νλ€. μλΉμκ° λ°μνμ§ μμΌλ©΄ κ²°κ΅ imapct κ° μ νμ μΌ μλ°μ μκΈ° λλ¬Έμ΄λ€) μμ¦μ μ μ λΆλ€μ΄ λλ€μ μλΉμ€μ κ³Όκ±° λλΉ λ¨λ¨λ―Έμ§κ·Όν λ°μμ 보μ΄λ μκΈ°μΈ κ² κ°κΈ°λ νλ€.
λ€λ§, μ μ λ€μ μ¬μ ν λ λμ μΆμ μ΄κ³ μΆμ΄νλ μκ΅¬κ° κ°νλ€. λ μ±μ₯νκ³ μΆμ΄νκ³ , λ ν볡ν μΆμ μ΄κ³ μΆμ΄νκ³ , λ μ¦κ±°μ΄ μκ°μ 보λ΄κ³ μΆμ΄ νλ€. λ³ΈμΈ μΈμμ κ°λ ¬νκ² λ€κ°μ¬ μ μλ μλΉμ€λ₯Ό λ§λ, ν΄λΉ μλΉμ€μ ν¨κ» λ λ§μ‘±μ€λ¬μ΄ μΆμ μ΄κΈ°λ₯Ό ν¬λ§νλ λΆλ€μ΄ κ½€ μλ€.
κ·Έλμ, λ―Έλμλ μ΄λ€ μλΉμ€κ° λ§μ μ μ λ‘λΆν° wowλ₯Ό μ΄λμ΄ λΌ μ μμμ§... κΆκΈνλ€. μ μ μ μ±μ₯κ³Ό λ§μ‘±μ λ λ§μ΄ μ§μ°©νλ μλΉμ€κ° κ·Έ wow λ₯Ό μ΄λμ΄ λΌ μ μμΌλ¦¬λΌ λ―Ώμ§λ§, κ³Όκ±° λλΉ μ μ μ wow λ₯Ό μ΄λμ΄ λ΄κΈ° μν λμ΄λκ° λμμ Έμ κ³Όκ±° λ§νΌ 'κΈμ±μ₯νλ μ€ννΈμ
'μ΄ λ§μ΄ λμ¬ κ² κ°μ§λ μλ€. κ·Έλλ, μ΄λ° λμΈλ₯Ό μ΄κ²¨λ΄λ κ°μ₯ μ’μ λ°©λ²μ 'λ νλ€μ΄μ§λ νκ²½'μ λ°μνκΈ° 보λ€λ 'μ μ 'μ μ§μ€νμ¬ ν루 ν루 λ μ’μ μλΉμ€λ₯Ό λ§λ€μ΄ λ΄λ κ²μ΄λΌ μκ°νλ€. λ¬Όλ‘ κ³Όκ±°μλ 1λ
λ
Έλ ₯νλ©΄ λ§μ‘±μ μ΄λμ΄ λΌ μ μμλ€λ©΄, μ΅κ·Όμλ 3~5λ°°λ λ ν΄μΌνλ μ΄λ €μμ΄ μ‘΄μ¬νμ§λ§ λ§μ΄λ€.
κ·Έλμ μμ¦ κ°μ μλλ Why? κ° λ μ€μν΄μ§λ κ² κ°λ€. μ μμνλμ§? 무μμ μν΄ μλΉμ€λ₯Ό λ§λ€μ΄ λ΄κ³ μλμ§?μ λν λͺ
ννκ³ μμ§ν λ΅μ κ°μ§κ³ μλμ§κ° λ μ€μν΄μ§λ μλκ° μλ€. κ·Έ μ΄μ μ λν λ΅μ μ 체 νμ΄ κ³΅μ νκ³ μμ΄μΌ, wow λ₯Ό λ§λ€μ΄ λ΄κΈ° μν λ
Έλ ₯/μκ°μ΄ μ¦κ°ν μλλ₯Ό 묡묡ν μ΄κ²¨λκ° μ μκΈ° λλ¬Έμ΄λ€.
μλΉμ€λ₯Ό μ΄μ΄λκ°κ³ μλ λ³Έμ§μ μ΄μ λ 무μμΈκ°? μ°λ¦¬ μλΉμ€λ λꡬλ₯Ό μν΄ μ‘΄μ¬νλ©° μ μ‘΄μ¬ν΄μΌ νλκ°? κΉμ λ°€, λ§κΈμ λ μ±μ₯μμΌ λκ°κΈ° μν λ°©λ²μ κ³ λ―Όνλ€κ°, μλ‘ μ μΈ μ§λ¬Έμ λν΄ λ€μ ν λ² μκ°ν΄λ³Έλ€.
----
Navigating "corporate speak" isn't easy.
Here's a helpful guide I put together:
"Let me check with my team" = No
"Possibly" = No
"On my roadmap" = Not happening
"This will be done in Q4" = This will be done in Q2 next year
"Disagree and commit" = I hate you
"Per my last email" = Try reading, for once in your life
"Challenging landscape" = We're going out of business, quickly
"Digital transformation" = We're going out of business, slowly
"Let's circle back" = We'll never speak of this again
"Take it offline" = We'll never speak of this again
"30,000 foot view" = I don't know what I'm saying
"Low hanging fruit" = Easy promotion
"Open up the kimono" = HR violation
"We use AI" = We don't use AI
"We use machine learning" = We don't use machine learning
"All hands on deck" = Let's actually try for once, please
π
βμ°½μ
κ°μ κΈΈμ λ€μ΄μλ©΄ νμ¬κ° μ¬λ¬λΆμ κ°μ‘±λ³΄λ€ μ€μν΄μ Έμ. κ°μ‘±λ€μ λμμ¬λ₯Ό μ±κΈΈ μ μκ³ , μλ
λ€μκ² κ΅Ώλμ ν€μ€λ₯Ό ν μ μκ² λ κ±°μμ. λ νκΈ μμ°μ΄ νμν΄ μ’μ μ§κ³Ό μ’μ μ°¨λ ν¬κΈ°ν΄μΌν κ±°κ³ μ. μΉκ΅¬λ μ§μΈλ€μ μ¬λ¬λΆμ μ΄ν΄νμ§ λͺ»νκ³ , μΈμμμ κΈΈμ μκ³ λ°©ν©νλ€κ³ μκ°ν κ±°μμ. λΆλ λ°©ν©μ΄ μ§§μμΌν ν
λ° νκ² μ£ .β
βνμλ€μ μκΈμ μ€ μ μλ μκΈ°κ° μ¬ κ±°μμ. μ΄ μ¬μ μμ λκ΅°κ°λ λΉμ (μ°½μ
κ°)μ κ³ μν κ±°μμ. μ¬λ¬λΆμ κ³ μνλ μ¬λμ΄ κ³΅λ μ°½μ
μμΌ μλ μκ³ ν¬μμμΌ μλ μμ΄μ. λ μμ£Ό ν릴 κ²μ΄κ³ , μ΄ λ¬Έμ λ€μ νμλ€μ΄ λ€ λ³΄κ² λ κ±°μμ. μͺ½νλ¦° λ¨κ³λ₯Ό λμ΄μλ©΄ λ΄κ° νλ κ² λ§λ, (λνμ§μμ) λ΄λ €μμΌκ² ꡬλλΌλ μκ°λ λ€κ² λμ£ . νμλ€λ κ³μ μ¬λ¬λΆμ μ€λ§μν¬ ν
μ§λ§, κ·ΈλΌμλ κ³μ μμμ²λΌ μ¬λμ νΌμ€μΌ ν΄μ. νμλ€μ μν΄ κ²©λ¦¬ μμΌμΌ λ§λ
ν μ§μμκ²λ μμ΄μΌ νλ μκΈ°κ° μμ΅λλ€.β
(μ΄μΉκ±΄ λΉλ°λ¦¬νΌλΈλ¦¬μΉ΄ λν, μ μ£Όμ μ°½μ
κ²½μ§λν λ°λͺ¨λ°μ΄ 2023)
μ λ§ λ¦¬μΌν μ°½μ
μμ μΆ. μ§κΈ μν©μ΄ λ무 νλ λ° λ¬΄μ¨ λΆκ·μνλ₯Ό λλ¦¬κ² λ€κ³ μ΄κ±Έ νλ μΆμ§λ§ λ λλ¦μ μ΄μ κ° μ겨μ μ§μνλ κ²μ΄μ£ . κ°μΈμ μΌλ‘λ μ΄ μΆμ΄ λΉμ°ν기보λ€λ μ μ°¨ κ°μ λμΌλ©΄ μ’κ² μ΅λλ€. μ°½μ
μμ λν μ‘΄μ€κ³Ό 보νΈλ μ»€μ ΈμΌ νκ³ μ.
51% of Zendesk's revenue is international
47% of Cloudflareβs revenue is international
46% of HubSpotβs revenue is international
45% of MongoDB's revenue is international
38% of Asanaβs revenue is international
34% of Twilioβs revenue is international
22% of Oktaβs revenue is international
Go Global as soon as you have customers there
Pull ahead
Repost from μ μ’
νμ μΈμ¬μ΄νΈ
μ΄ μ¬λμ μκΈ°κ° νλ μ΄νκ³ μλ κ²μμ νμ΄ μ΄λ»κ² μ€κ³λμ΄μλμ§ λΆλͺ
νκ² μκ³ μλ€
https://www.youtube.com/watch?v=Czmkes4yAOA
Long λ² μ΄λΉλΆλ¨Έ vs Short MZ
BofA λ μ΅κ·Ό λ³΄κ³ μμμ Long λ² μ΄λΉλΆλ¨Έ, Short λ°λ λμΌ (MZ) ν¬μ μμ΄λμ΄λ₯Ό μ μν©λλ€.
μ§κΈμ κ³ κΈλ¦¬, κ³ λ¬Όκ° νκ²½μμ λ―Έκ΅λ΄ μλΉ μ§μΆμ μ£ΌλκΆμ 65μΈ μ΄μμ κ³ λ ΉμΈ΅μ΄ μλ²½ν μ₯κ³ μμ΅λλ€.
νκ΅μ λ² μ΄λΉλΆλ¨Έ μΈλμΈ μ ν¬ λΆλͺ¨λλ€μ 보λλΌλ λλΆλΆ μκ°λ‘ 1 μ£Όν μ΄μμ 보μ μ€μ
λλ€. λμΆ κ±°μ μμ΅λλ€. μμ° κ°κ²©μ μΈνλ μ Wealth effect 무μ λͺ»ν©λλ€.
λ°λ©΄, MZ μΈλλ μ£Όλ‘ μ°κ³ μ£½μλ μΈκ³κ΄, μ΅κ·Ό λͺλ
κ° λΆλμ° μλμ‘±, μ½μΈ/μ£Όμ λΉν¬μ‘±μ ν©λ₯λ‘ λΆμ± λΆλ΄μ΄ μλΉν©λλ€. μ¨λ§ μ¬μ΄λ μ΄μμ Snowball effect λ₯Ό κ²½ννκ³ μμ£ .
λμ±μ΄ λ² μ΄λΉλΆλ¨Έλ λμ΄ λ§λ μ λ μ¬μ ν μ μΆμ λν λ―Ώμμ΄ μμ΅λλ€. λ―Έκ΅λ Cash return μ΄ 5% λ₯Ό μνν©λλ€. Cash is king μΈ μκΈμ μν©μ μΈλκ° κ²©μ°¨λ₯Ό λλ리기 νλ€ μ λλ‘ λ²λ¦¬κ³ μμ΅λλ€.
μΉμμ ν¨μκ° λͺ
νν κ°λ¦¬λ κ²μμ΄λΌλ©΄, μΉν°μ μ°μ
λ΄ ν¬μ μμ΄λμ΄λ‘ νμ©ν΄ λ³Όλ§ ν©λλ€.
π’ Introducing Voyage AI
Founded by a talented team of leading AI researchers and me ππ.
We build state-of-the-art embedding models (e.g., better than OpenAI π).
We also offer custom models that deliver π―+10-20% accuracy gain in your LLM products. π§΅
We think embeddings are an under-loved, but incredibly important, part of everyoneβs retrieval stack. So we set out to build just better embeddings.
Voyage embeddings are SOTA on MTEB and 9 held-out benchmarks covering industry domains. Learn more here: https://lnkd.in/gk59ZE-T
We partnered with @LangChainAI to help enhance their official chatbot, live at chat.langchain.com. Both our base and fine-tuned embeddings improve both the retrieval and response quality, and the latter is deployed!
Check out https://lnkd.in/gZ2aH4KU for more details.
Try Voyage embeddings for free β your first 5000 documents are on us!
Create an account at voyageai.com and check out our API docs at https://docs.voyageai.com .
To fine-tune your own custom model, email βοΈ contact@voyageai.com for early access!
λ΄κ° μΌλ³Έμ 근무ν λΉμ λ°λ체 ν¬μκ° μ΄λ€μ‘κΈ° λλ¬Έμ λμΏ μ§μ¬μλ μ΄λ³μ² νμ₯μ μλ¬Έμμ ν΄μ£Όλ λ°λ체 κ΄λ ¨ μ λ¬Έκ°λ€μ΄ μμ£Ό λ°©λ¬Ένλ€. μλ¬Έν μΌλ³ΈμΈ λνκ΅μμκ² βμ°λ¦¬ νμ₯λμ΄ λ°λ체λ₯Ό μΌλ§λ μκ³ κ³μλκΉ?β λΌκ³ λ¬Όμ΄λ΄€λ€. κ·ΈλΆμ μμΌλ©΄μ βμ΄λ±νμ μμ€μ
λλ€βλΌκ³ λ΅νλ€. βκ·Έλ°λ° μ μ£Όμ μ¬λλ€μ λ°λλ₯Ό 무λ¦
μ°κ³ λ°λ체 ν¬μλ₯Ό λ°μ΄λΆμ΄μ€κΉμ?βνκ³ λ€μ λ¬ΌμΌλ κ·ΈλΆμ βνμ₯λμ μ¬μ
μ λν κ°κ°μ λλ¬Όμ λ³Έλ₯μ
λλ€. μ무λ λ§λ¦΄ μ μμ΅λλ€βλΌκ³ λ΅ν΄ ν¨κ» μμλ€. κ²°κ³Όμ μΌλ‘ νκ΅μ μ μμ°μ
μ΄ μΌλ³Έμ μΆμν μ μλ 건 μ€λ μμ£Όμ μ§λ°°κ΅¬μ‘°μμ μ μνκ³ μ₯κΈ°μ μΈ μμ¬κ²°μ μ νκΈ° λλ¬Έμ΄λ€.
κ³Όκ±° μ΄κ±΄ν¬ νμ₯μ μ£Όμ¬μμ΄ κ·κ΅ν λ λ°λμ μΌλ³Έμ μ μμ νμ νλμ© μ¬μμ λΆμνλΌκ³ μ§μνλ€. λμ²λΌ κΈ°μ λ¬ΈμΈνμΈ μ¬λμκ²λ κ³€νΉμ€λ¬μ΄ μΌμ΄μλ€. μΌμ±μ μκ° VTR κ°λ°μ λμ μ΄κΈ°μλ λΆλλ₯ μ΄ λ무 λμ μ΄λμ₯μ μ νμ μμμ λΆμ μ§λ₯΄κΈ°λ νλ€. λμλ°μ νμ‘°λ₯Ό μ»μ΄ λΌμΈμ μ‘°μ₯λ€μ λμλ° κ³΅μ₯μ λ³΄λ΄ μ€μ΅μ μν€κΈ°λ‘ ν μ μ΄ μλ€. νλ° μ€μ΅μ΄ μ§νλλ μ€ λμλ°μμ λλ νκ΅μ μ°μμμ λ°μ§ μκ² λ€κ³ ν΅λ³΄νλ€.
λμλ° κ°λΆλ‘λΆν° μ ν΄ λ€μ μ¬μ λ μ΄μ΄κ° μμλ€. 3μ°¨ μ€μ΅μλ€μ΄λ 1Β·2μ°¨ μ€μ΅μλ€μ΄ νλ μ§λ¬Έμ΄ λμΌνλ€κ³ νλ€. κ°μ μ§λ¬Έμ΄ λμ€λλ° λΉμΌ κ²½λΉλ₯Ό λ€μ¬ ν΄μΈ μ°μλ₯Ό ν νμκ° μλλλ μκΈ°μλ€. μ‘°κΈ λͺ¨μμ μΌλ‘λ λ€λ Έμ§λ§ νλ¦° μ§μ μ΄ μλμλ€. κ°μΈμ μ΄μ΅λ³΄λ€ μ§λ¨μ μ΄μ΅μ μ°μ μνκ³ , μλ‘μ΄ μ§μμ 곡μ νλ μΌλ³ΈμΈμ μ¬κ³ λ°©μμΌλ‘λ μ°λ¦¬ μ°μμλ€μ κ°μΈμ£Όμμ μ¬κ³ λ°©μμ΄ μ΄ν΄νκΈ° μ΄λ €μ μ κ²μ΄λ€. μμ¦ μν€νλ°λΌ λ§€μ₯μ μμ리λ₯Ό μ°¨μ§νκ³ μλ μΌμ±, LG TVλ₯Ό 보면 40λ
μ μ μ°½νΌνλ κ·Έλμ μΌμ΄ λ μ€λ₯΄κ³€ νλ€.
Repost from μ΅μ μλ€ λ°μ§ν
μ λμμλ κ°λ μ’μ κΈλ€μ΄ μλ€.
https://shindonga.donga.com/economy/3/03/13/4503748/1
Available now! Telegram Research 2025 β the year's key insights 
