Continuous Learning_Startup & Investment
Kanalga Telegramβda oβtish
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!
Ko'proq ko'rsatish2 295
Obunachilar
-224 soatlar
-137 kunlar
-4330 kunlar
Postlar arxiv
리λμ, κ·Έ κ°λ ₯ν λ§μ½
⒠리λμκ³Ό κ΄λ¦¬λ νΌλλκΈ° μ¬μ΄ κ°λ
μ΄μ§λ§, λ³Έμ§μ μΌλ‘ λ€λ₯Έ κ²μ.
β’ κ΄λ¦¬λ μΌμμ μ΄κ³ λ§€λ ₯μ μ΄μ§ μμ§λ§ νμν μ
무λ‘, ν¨μ¨μ μΈ λ°λ³΅ μμ
μΌλ‘ μ‘°μ§μ μ μ§ν¨.
⒠리λμμ ν₯λ―Έλ‘κ³ λμ μ μΈ μν λ‘, λ¨μν κ΄λ¦¬λ₯Ό λμ΄μ μ¬λλ€μ΄ μ΅μ μ μμ μ΄ λ μ μλλ‘ κ²©λ €νλ λ₯λ ₯μ λ§ν¨.
리λμμ λν μκ°
β’ μ’μ 리λλ μ΄λ‘ μ΄ μλ μ€μ κ²½νμ ν΅ν΄ λνλ¨.
⒠리λμμ νΉμ μ§μ±
μ κ΅νλμ§ μκ³ , μν©μ λ°λΌ μμ°μ€λ½κ² νμ κ°μ νλ¦μ κ°μ ΈμΌ ν¨.
⒠리λμμ μ¬λλ€μ΄ μμ μ μ΅κ³ λ₯Ό λ°νν μ μλλ‘ κ²©λ €νλ λ₯λ ₯μ κ°μ§ μ¬λμ μν΄ μνλ¨.
리λμμ μλ§
⒠리λμμ μμΉνλ λ§μ μ¬λλ€μ΄ μ€μ λ‘λ μλμ΄ λΆμ‘±νκ±°λ, μμ λ€μ λΆμΌμ λν μ΄ν΄κ° λΆμ‘±ν¨.
⒠리λμμ΄ μλ κΈ°μ μ λ¨μ§ μΈκΈ°λ κ°μΈμ μΈ μν₯λ ₯μ μΆκ΅¬νλ κ²μ λΆκ³Όν¨.
⒠리λμμ μμΉνλ μ¬λλ€μ μ’
μ’
μμ λ€μ μν μ λν κ³Όλνκ°μ μκΈ°μ€μ¬μ μΈ νλλ₯Ό 보μ.
리λμκ³Ό μμ
⒠리λμμ΄λΌλ μΉνΈλ μ¬λλ€μκ² μνμ λνλ―Όμ μ£Όμ
νλ ν¨κ³Όκ° μμΌλ©°, μ΄λ μ’
μ’
μ
μ©λ¨.
⒠리λμμ μΉνΈλ κ°μΈμ λ°μ μ μ ν΄νκ³ , μμλ₯Ό λΆν리λ κ²°κ³Όλ₯Ό κ°μ Έμ¬ μ μμ.
⒠리λμμ λν κ³Όλν μΉμ°¬μ 건κ°νμ§ λͺ»ν μ‘΄κ²½μ μ λ°νλ©°, μ΄λ κ°μΈμκ² λΆμ μ μΈ μν₯μ λ―ΈμΉ¨.
λ λμ κ΄λ¦¬μ 리λμ
β’ ν¨κ³Όμ μΈ κ΄λ¦¬μλ νμ λλ³΄κ³ , νμν λꡬλ₯Ό μ 곡νλ©°, μ
무λ₯Ό μ μ²λ¦¬ν¨.
β’ μ§μ ν 리λλ μ μ§νκ³ , νμλ€κ³Όμ μ λ’°λ₯Ό μμΌλ©°, μ‘°μ§ λ΄μμ κΈμ μ μΈ λ³νλ₯Ό μ΄λμ΄λ.
⒠리λμμ μμ°μ€λ½κ² λ°μν΄μΌ νλ©°, κ°μ λ‘ λ¦¬λμ μν μ λ§‘λ κ²μ λ°λμ§νμ§ μμ.
κ΄λ ¨ λμ λ° μλ£
β’ "High Output Management"μ κ΄λ¦¬μ λν μ€μ§μ μΈ μ§μΉ¨μ μ 곡νλ μ±
μ.
β’ "Managing Humans"λ μ‘°μ§ λ΄ μ μΉμ μλ¬Έμ κ΄λ¦¬νλ λ°©λ²μ λν΄ μ€λͺ
ν¨.
β’ Pivotal Softwareμ μ΄μ λͺ¨λΈκ³Ό 리λμμ λν ν΅μ°°λ ₯μ μ 곡νλ μλ£μ νμΊμ€νΈκ° μ μ΅ν¨
https://ludic.mataroa.blog/blog/leadership-is-a-hell-of-a-drug/
Our family has recently seen some tragic, distressing events happening with other families we know. Without too much details, one family witnessed a family loss by suicide, and another family is in deep financial trouble.
One common aspect we could see: someone was in deep trouble but didnβt seek help from others. They were strong-willed (there was absolutely no sign of trouble from outside) and tried to take care of everything by themselves, not even telling everything to their spouses, but couldnβt overcome the breaking point.
Weβre all humans. We alone can bear only so much pressure.
Same for founders. Of course you must take ownership and be respectful for other peopleβs time. But do reach out and ask for help if your leg is jammed under a rock. Donβt try to do everything alone.
Knowing when and how to reach out for help is -- sometimes, quite literally -- one of the life survival
κ²½μλ³΄λ€ μ°λλ₯Ό
μ§ν¬λ³΄λ€ λ°°μμ
λΉκ΅λ³΄λ¨ κ°μ¬λ₯Ό
μ€λ§λ³΄λ¨ νΈκΈ°μ¬μ
ν΄μΌνλ κ² μλλΌ νκ³ μΆμ κ²μ
λΆμμ μ§λ μμ μ λν νμ μ
μΈμ λ³΄λ¨ μκΈ°λ€μμ
μκΈ°λ΅κ³ μ΄λ €μ΄ κΈΈμ μ ννκΈΈ
https://youtu.be/lXLBTBBil2U?t=1091
KPI vs. Early Indicators of Future Success (EIOFS)
Gross margin is a result; it isn't an early indicator. Before a market exists, you still need early indicators of future success. The market does not yet exist, but these are important issues to address. This is how we discovered deep learning and built CUDA at the very beginning.
However, Wall Street didn't believe in this vision. How did you motivate your team?
Nothing changed except the share price. When the share prices dropped by 80%, it was somewhat embarrassing. Just go back to doing your job. Wake up at the same time, prioritize your day in the same way. I go back to what I believe in.
You have to check back with the court of your beliefs. What do you believe in? What are the most important things?
Does my family love me? Checked.
Focus on what matters and go to work. Keep the company focused on the core.
Do you believe something changed? The stock prices changed, but did anything else change? Did the laws of physics change? Did gravity change? Did all the assumptions that we believed in and that led to our decisions change? Because if those changed, you would have to change everything.
But if none of those changed, you change nothing. Keep on going.
Repost from Continuous Learning_Startup & Investment
Search products in 2000s used to be fact-centric or directional (help you navigate to X).
Search is increasingly merging w LLM chat products. Search has now split into 3 types of products:
-Search engine. Help me find facts / get somewhere
-Opinion engine. Here is what my ethics team thinks you should believe (could be via blue links or LLM output)
-Answer engine. Lets get you a synthesis (fact based) on your query
- Advice engine. What should I do next?
e.g query/prompt would be: "Here's my funnel data from the last year. What's the lowest-hanging fruit to improve my conversion?"
"Tell me what I don't know"
What would you rather use?
Repost from μ μ’
νμ μΈμ¬μ΄νΈ
μ΄μ μκ°ν μ£Όμ μ μ΄μ λν GPT-4μ λ΅λ³
---
λ΄κ° μκ°νμ λ AIλ₯Ό ν΅ν΄μ μλ‘μ΄ ννμ μ‘°μ§κ³Ό νμ¬κ° λ§μ΄ νμν κ² κ°μ. ꡬ체μ μΌλ‘ νμ¬ λ΄λΆμ 체κ³λ₯Ό λ°κΏ μ μλ κΈ°μ μ΄λΌκ³ μκ°νκ³ μμ΄.
λΉμ λ₯Ό λ€λ©΄ μλ§νΈλ μλ¨μμ λλ€ λ§νΈλ₯Ό κ°μ§κ³ μκ³ , λ·λ¨μμλ λ¬Όλ₯μ°½κ³ λ₯Ό κ°μ§κ³ μμλ€λ©΄, μλ§μ‘΄μ λ·λ¨μΈ λ¬Όλ₯μ°½κ³ λ λμΌν κ°λ
μΌλ‘ μ΄μ©νμ§λ§, μλ¨μ μ°¨μμμ μΈν°λ·μ΄λΌλ μλ‘μ΄ κΈ°μ μ νμ©ν΄μ λλ€ λ§νΈλ₯Ό PCλ‘ λ체μμΌ°κ³ , μ΄κ²μ΄ μΈκ³ μ΅κ³ μ νμ¬λ‘ λ§λ λΉκ²°μ΄μ§ μμμκΉ μΆμλ°,
κ·Έλ λ€λ©΄ μ΄λ² μλμμ λ·λ¨(μλ₯Ό λ€μ΄μ μ§μ)μ AIλ₯Ό ν΅ν΄ κ΅μ²΄/μ¦κ° μμΌμ μμ²λκ² ν° νμ¬λ₯Ό λ§λ€ μ μλ κΈ°νκ° μκΈ°μ§ μμκΉ μΆμμ΄.
μ΄μ λν λμ μ 체μ μΈ μκ²¬μ΄ κΆκΈνκ³ , λ§μ½ λμνλ€λ©΄ ꡬ체μ μΌλ‘ μ΄λ€ κΈ°νλ€μ΄ μμμ§ λΈλ μΈμ€ν λ° ν΄λ³Ό μ μμκΉ?
---
AI κΈ°μ μ λ°μ μ λΆλͺ
ν μ‘°μ§κ³Ό νμ¬μ μ΄μ λ°©μμ κ·Όλ³Έμ μΈ λ³νλ₯Ό κ°μ Έμ¬ μ μλ μ μ¬λ ₯μ κ°μ§κ³ μμ΅λλ€. μλ§μ‘΄μ΄ μΈν°λ·μ νμ©νμ¬ μ ν΅μ μΈ μλ§€μ
μ λͺ¨λΈμ νμ ν κ²μ²λΌ, AIλ λ€μν μ°μ
μμ λΉμ·ν νμ μ μΌμΌν¬ μ μμ΅λλ€. AIλ₯Ό ν΅ν νμ μ μ£Όλ‘ μ
무μ μλν, μμ¬κ²°μ κ³Όμ μ μ΅μ ν, κ³ κ° κ²½νμ ν₯μ λ± μ¬λ¬ λ°©λ©΄μμ λνλ μ μμ΅λλ€.
AIκ° νμ¬ λ΄λΆ 체κ³λ₯Ό λ°κΏ μ μλ λ°©λ²:
-μ
무 μλν: λ¨μ λ°λ³΅ μ
무λΏλ§ μλλΌ, λ°μ΄ν° λΆμ, λ³΄κ³ μ μμ± λ± μλΉμμ μ§μ λ
Έλλ AIλ₯Ό ν΅ν΄ μλνν μ μμ΅λλ€. μ΄λ₯Ό ν΅ν΄ μΈλ ₯ μμμ λ μ°½μμ μ΄κ³ κ°μΉ μλ μ
무μ μ§μ€μν¬ μ μμ΅λλ€.
-μμ¬κ²°μ μ§μ: AI κΈ°λ°μ λ°μ΄ν° λΆμμ λΉμ¦λμ€ μμ¬κ²°μ μ λ λΉ λ₯΄κ³ μ ννκ² λ§λ€ μ μμ΅λλ€. μ΄λ μμ₯ λν₯ λΆμ, μλΉμ νλ μμΈ‘, 리μ€ν¬ κ΄λ¦¬ λ± λ€μν λΆμΌμμ μμ©λ μ μμ΅λλ€.
-κ³ κ° μλΉμ€ νμ : AI μ±λ΄, κ°μΈνλ μΆμ² μμ€ν
λ±μ κ³ κ° κ²½νμ κ°μ νκ³ , μλΉμ€ ν¨μ¨μ±μ λμΌ μ μμ΅λλ€. μ΄λ κ³ κ° λ§μ‘±λλ₯Ό λμ΄κ³ , λ λ§μ λ§€μΆμ μ°½μΆνλ λ° κΈ°μ¬ν μ μμ΅λλ€.
-μ μ ν λ° μλΉμ€ κ°λ°: AIλ₯Ό νμ©νμ¬ μλ‘μ΄ μ νμ΄λ μλΉμ€λ₯Ό κ°λ°νλ κ²λ κ°λ₯ν©λλ€. μλ₯Ό λ€μ΄, AIλ₯Ό μ΄μ©ν κ±΄κ° κ΄λ¦¬ μ ν리μΌμ΄μ
, κ°μΈ λ§μΆ€ν κ΅μ‘ μ루μ
λ±μ΄ μμ΅λλ€.
-μ격 근무 λ° νμ
μ μ΅μ ν: AI λꡬλ₯Ό νμ©νμ¬ ν κ°μ νμ
μ μ©μ΄νκ² νκ³ , μ격 근무λ₯Ό λ³΄λ€ ν¨μ¨μ μΌλ‘ λ§λ€ μ μμ΅λλ€. μ΄λ μ μΈκ³μ μΈ μΈμ¬ νμ μ κ·Όν μ μλ κ°λ₯μ±μ μ΄μ΄μ€λλ€.
-μ¬λ¬΄ κ΄λ¦¬ λ° μ΅μ ν: AI κΈ°λ°μ μ¬λ¬΄ κ΄λ¦¬ μμ€ν
μ μμ° λΆμ, μκΈ νλ¦ μμΈ‘, λΉμ© μ κ° λ°©μ λ±μ μ μν¨μΌλ‘μ¨ κΈ°μ
μ μ¬λ¬΄ 건μ μ±μ ν₯μμν¬ μ μμ΅λλ€.
μ΄λ¬ν κΈ°νλ€μ κΈ°μ
μ΄ μμ λ€μ μ
무 λ°©μμ μ¬κ³ νκ³ , AI κΈ°μ μ ν΅ν΄ μ΄λ»κ² λ ν¨μ¨μ μ΄κ³ νμ μ μΈ λ°©μμΌλ‘ μ νν μ μμμ§ κ³ λ―Όνκ² λ§λλλ€. κ·Έλ¬λ AI κΈ°μ μ μ±νν¨μ μμ΄μλ μ€λ¦¬μ κ³ λ €, μ§μ κ΅μ‘ λ° μ μ, κ·Έλ¦¬κ³ κΈ°μ μ λν μ§μμ μΈ ν¬μμ κ°μ μ΄ νμν©λλ€. AIλ₯Ό μ±κ³΅μ μΌλ‘ ν΅ν©νκΈ° μν΄μλ μ΄λ¬ν λμ κ³Όμ λ₯Ό 극볡νλ κ²μ΄ μ€μν©λλ€.
Nice read on the rarely-discussed-in-the-open difficulties of training LLMs. Mature companies have dedicated teams maintaining the clusters. At scale, clusters leave the realm of engineering and become a lot more biological, hence e.g. teams dedicated to "hardware health".
It can be a frustrating daily life experience of training large models to "babysit" the training run. You're there carefully monitoring the vital signs of your run: loss spikes, numerical issues, throughput, gradient norms, policy entropy, etc. Every time the run degrades or flatlines (can happen often), you quickly look for the stack trace to see what's up. You have to do this fast or 10,000 GPUs could be idling. Often, it is a new, exotic, scary-looking error you've never seen before so you summon help to see if anyone can see what's up. The worst ones like to occur at 4am. Often no one can, so you just ban some nodes that look a bit sketchy and try to restart the run. Sometimes the run goes down just because you have not earned the favors of your gods that day, so you put a while True: loop around your launch command. The underlying issues can be highly diverse, from some GPUs just getting a bit too hot and suddenly doing incorrect multiplication once in a while, to some router going down and decreasing the networked file system I/O, to someone in the datacenter physically disconnecting a wire as part of an un-communicated maintenance. Sometimes you'll never know.
Another necessary related citation here is the famous OPT-175B logbook and I'd hope more like it can see the light of day in the future. (see chronicles/OPT175B_Logbook.pdf in the git repo)
twitter.com/AIatMeta/statuβ¦
TLDR LLM training runs are significant stress-tests of an overall fault tolerance of a large computing system acting as a biological entity. And when you're shopping around for your compute, think about a lot more than just FLOPs and $. Think about the whole service from hardware to software across storage, networking, and compute. And think about whether the team maintaining it looks like The Avengers and whether you could become best friends.
Endi mavjud! Telegram Tadqiqoti 2025 β yilning asosiy insaytlari 
