μ€λ¦¬μ½λ°Έλ¦¬μ μ§κ°λ³λμ΄ μμλμλ€
λͺ¨λ ν° λ³νλ ꡬμ±μλ€μκ²λ μ¬μμ΄λ€
μ€λ¦¬μ½λ°Έλ¦¬μ μλ‘μ΄ μλκ° μμλκ³ μλ€.
μ€λ¦¬μ½λ°Έλ¦¬μ μμ¬λ λͺλ²μ ν° λ³νλ₯Ό κ±°μ³€λ€. 70λ λ λ°λμ²΄λ‘ μμλ μ€λ¦¬μ½λ°Έλ¦¬λ 90λ λ PCμ μμ₯μ κ±°μ³ 2000λ λ μΉμ μλ, κ·Έλ¦¬κ³ 2010λ λ λͺ¨λ°μΌ μλλ₯Ό κ±°μ³μλ€. κ·Έλ¦¬κ³ 2023λ AI νλͺ μ λ§μ΄νλ€.
κ° μλλ₯Ό κ±°μΉλ©΄μ μ€λ¦¬μ½λ°Έλ¦¬λ κ±°λν λ°μ μ μ΄λ£¨μκ³ μΈκ³μ λμ΄ λͺ¨μ΄λ κ³³μ΄ λμλ€. μ§κΈλ λͺ¨λ μΈκ³μ λͺ¨λ κ³³μ΄ μλΉλμ, ν μ¬λΌ, μ ν, λ§μ΄ν¬λ‘μννΈ, μ€νAIλ‘ μ λ¦¬κ³ μλ€. μ μΈκ³μ μΈ μ μ±μ₯μλ λ―Έκ΅μ μμ²λ μ±μ₯μ 보μ΄κ³ μκ³ μ£Όμμμ₯κ³Ό μ½μΈ λ§μΌ λν μ΄λ§μ΄λ§ν μ±μ₯μ 보μ΄κ³ μλ€.
λΆλͺ ν μμ²λ μ±μ₯μ΄κ³ κΈ°νμ΄κ³ λ³νμ΄λ€. ν° μ§κ° λ³λμΈ λ§νΌ κ·Έ μμ μλ ꡬμ±μλ€μκ²λ κ·Ήμ¬ν κ³ ν΅μ΄ κΈ°λ€λ¦¬κ³ μλ€.
μμΌλ‘ 5λ νλ₯Ό μμν΄ λ³΄λ©΄ μ΄λ€ μΈμμΌκΉ? 5λ νμλ νλ‘λνΈλ₯Ό λ§λλ μΌμ AIκ° ν κ²μ΄λ€. μ¬μ μ λ§μ λΆλΆλ AIκ° λ체ν κ²μ΄λ€. μλ§μ μ±λ€μ΄ μ¬λΌμ§ κ²μ΄κ³ μ€ννΈμ λ€λ μ¬λΌμ§ κ²μ΄λ€. μμ΄λμ΄λ‘ ν¬μλ₯Ό λ°μ λ¨κΈ°κ°μ ν° μ±μ₯μ νλ μ€ννΈμ μ κ³§ μλ§μ΄ λ κ²μ΄λ€. λͺλͺ λκΈ°μ μ΄ κ°μ§ μΈκ³΅μ§λ₯μ κ°μΆ μνΌμ±μ΄ μ§κΈλ μΌλΆ μ±μ λͺ°λ €μλ μ± μνκ³μ μ§μ€νλ₯Ό κ°μν ν κ²μ΄λ€.
μ§κΈ μ€λ¦¬μ½λ°Έλ¦¬μλ κ·Έλ¬ν λ³νκ° μΌμ΄λκ³ μλ€. κΈ°μ‘΄μ μ§μ λ€μ μμ²λ μ 리ν΄κ³ λ₯Ό κ²ͺκ³ μλ€. 2024λ ν ν¬ μ 리ν΄κ³ λ 3μ μ΄μμλ 50000λͺ μ λ€κ°μκ³ μλ€. κΈ°μ‘΄μ μμ²λκ² λμ μ°λ΄μ λ°λ μ¬λλ€μ΄ μ€μ§ μ΄ν 900κ°μ μμλ₯Ό λ΄λ μ·¨μ μ΄ μ λκ³ μλ€κ³ νλ€.
κ·Έλ λ€κ³ μ§μ μ΄ μ€μ΄λ€κ³ μλ κ²μ μλλ€. μμ²λκ² λμ AI μμ§λμ΄ μμκ° κ·Έ μ리λ₯Ό μ±μκ°κ³ μκ³ , μ μΈκ³μ λμ μ€λ¦¬μ½λ°Έλ¦¬λ‘ λͺ¨μ¬λ€κ³ μλ€. AI Engineer μ§μ κ²μλμ΄ 6000% μ¦κ°νλ€κ³ νλ€.
μμΌλ‘ λͺλ νμλ AI μμ§λμ΄λ λͺ¨λ μμ§λμ΄μ κΈ°λ³Έ μμμ΄ λ κ²μ΄λ€. AI μμ§λμ΄κ° μλ μμ§λμ΄λ μ‘΄μ¬νμ§ μκ² λ κ²μ΄λ€. λ§μΉ μ§κΈ λΉ λ°μ΄ν° μμ§λμ΄κ° μλ μμ§λμ΄λ μ‘΄μ¬νμ§ μλ―μ΄. νμ¬ λͺ¨λ μμ§λμ΄λ 10λ μ λΉ λ°μ΄ν° κΈ°μ λ€μ ν΄λ¦ λͺλ²μ νμ©ν μ μλ€. AIλ λ μ°κΈ° μ½λ€. μμΌλ‘ AIκ° μμ΄ μ΄λ£¨μ΄μ§λ κ°λ°μ μμ΄μ§ κ²μ΄κ³ , μ€λ¦¬μ½λ°Έλ¦¬μμλ κΈμλλ‘ λμ²΄κ° μ΄λ£¨μ΄μ§κ³ μλ€.
μ΄λ κ² λΉ λ₯΄κ² job marketκ³Ό λΉμ¦λμ€μ μμμ΄ λ°λλ κ²μ λλ κ²½νν μ μ΄ μλ€. κ·Έλ μ§λ§ 2000λ μΉ νλͺ , 2010λ λͺ¨λ°μΌ νλͺ μ κ±°μ³€λ μ¬λλ€μκ²λ κΈ°μκ°μ΄ λ€μ§λ λͺ¨λ₯΄κ² λ€.
μ€ννΈμ μ μλ―Έκ° λ°λκ³ , μμ§λμ΄μ μλ―Έκ° λ°λλ μλ. μ΄μ μ¬μ μ λ μ΄μ λͺλͺ μ΄ λͺ¨μ¬ μΈμμ λ°κΎΈλ μΌλ€μ΄ μλ κ²μ΄λ€. μΈμμ AIκ° λ°κΎΌλ€. AIλ μ¬λλ€μ λ¬Έμ λ₯Ό λ무λ μ μκ³ μκ³ , ν΄κ²°μ± λ μ μν μ μλ€. μμΌλ‘ μ¬μ μλ κΈ°μ 보λ€λ λ§μΌν κ³Ό μΈμΌμ¦μ μμμ΄ μ μ λ μ»€μ§ κ²μ΄λ€.
μμ§λμ΄λ AI λ§λλ μ¬λκ³Ό AIλ₯Ό νμ©ν΄μ μ½κ² μ루μ μ λ§λλ μ¬λμΌλ‘ κ·Ήλ¨μ μΌλ‘ λΆνλ κ²μ΄λ€. μ§κΈμ AI λ§λλ μ¬λλ€μ κ°μΉκ° ν¨μ¬ ν¬κ² 보μ΄λ μκΈ°μ΄λ€. κ·Έλ°λ° μμ₯μ΄ μμ νλκ³ λμ€νλ μλ‘ AIλ₯Ό νμ©νλ μ¬λλ€μ μμ₯λ λ€μκΈ μ»€μ§ κ²μ΄λ€.
2024λ μ μ€λ¦¬μ½λ°Έλ¦¬μ μ¬μκ³Ό κ°κΉμ΄ μ°λλ―Έκ° λͺ°λ €μ¬ κ²μ΄λ€. κ·Έλ μ§λ§ λκ² λ³΄λ©΄ λ λ€λ₯Έ μΈμμ λ§μ΄νκΈ° μν νκ°μ΄μ λΆκ³Όν κ²μ΄λ€. λ§μ μ¬λλ€μ΄ μΈλ € λκ°κ² μ§λ§ κΈμλλ‘ μλ‘μ΄ μ¬λλ€μ΄ μ μ λ κ²μ΄λ€. κ²°κ΅ λμ΄ λͺ¨μ΄λ κ³³μ μ΄κ³³μ΄λκΉ.
λλ ΅μ§λ§ κΈ°λλλ 2024λ μ΄λ€. κ½ μ‘κ³ , λΉμ μλ μμ λκ³ , λ©λ¦¬ 보μ. 2026λ μ λ΄κ° μ΄λμ μμ΄μΌ ν μ§ κΉμ κ³ λ―Όμ΄ μμλμλ€.
λͺ¨λ ν° λ³νλ ꡬμ±μλ€μκ²λ μ¬μμ΄λ€
μ€λ¦¬μ½λ°Έλ¦¬μ μλ‘μ΄ μλκ° μμλκ³ μλ€.
μ€λ¦¬μ½λ°Έλ¦¬μ μμ¬λ λͺλ²μ ν° λ³νλ₯Ό κ±°μ³€λ€. 70λ λ λ°λμ²΄λ‘ μμλ μ€λ¦¬μ½λ°Έλ¦¬λ 90λ λ PCμ μμ₯μ κ±°μ³ 2000λ λ μΉμ μλ, κ·Έλ¦¬κ³ 2010λ λ λͺ¨λ°μΌ μλλ₯Ό κ±°μ³μλ€. κ·Έλ¦¬κ³ 2023λ AI νλͺ μ λ§μ΄νλ€.
κ° μλλ₯Ό κ±°μΉλ©΄μ μ€λ¦¬μ½λ°Έλ¦¬λ κ±°λν λ°μ μ μ΄λ£¨μκ³ μΈκ³μ λμ΄ λͺ¨μ΄λ κ³³μ΄ λμλ€. μ§κΈλ λͺ¨λ μΈκ³μ λͺ¨λ κ³³μ΄ μλΉλμ, ν μ¬λΌ, μ ν, λ§μ΄ν¬λ‘μννΈ, μ€νAIλ‘ μ λ¦¬κ³ μλ€. μ μΈκ³μ μΈ μ μ±μ₯μλ λ―Έκ΅μ μμ²λ μ±μ₯μ 보μ΄κ³ μκ³ μ£Όμμμ₯κ³Ό μ½μΈ λ§μΌ λν μ΄λ§μ΄λ§ν μ±μ₯μ 보μ΄κ³ μλ€.
λΆλͺ ν μμ²λ μ±μ₯μ΄κ³ κΈ°νμ΄κ³ λ³νμ΄λ€. ν° μ§κ° λ³λμΈ λ§νΌ κ·Έ μμ μλ ꡬμ±μλ€μκ²λ κ·Ήμ¬ν κ³ ν΅μ΄ κΈ°λ€λ¦¬κ³ μλ€.
μμΌλ‘ 5λ νλ₯Ό μμν΄ λ³΄λ©΄ μ΄λ€ μΈμμΌκΉ? 5λ νμλ νλ‘λνΈλ₯Ό λ§λλ μΌμ AIκ° ν κ²μ΄λ€. μ¬μ μ λ§μ λΆλΆλ AIκ° λ체ν κ²μ΄λ€. μλ§μ μ±λ€μ΄ μ¬λΌμ§ κ²μ΄κ³ μ€ννΈμ λ€λ μ¬λΌμ§ κ²μ΄λ€. μμ΄λμ΄λ‘ ν¬μλ₯Ό λ°μ λ¨κΈ°κ°μ ν° μ±μ₯μ νλ μ€ννΈμ μ κ³§ μλ§μ΄ λ κ²μ΄λ€. λͺλͺ λκΈ°μ μ΄ κ°μ§ μΈκ³΅μ§λ₯μ κ°μΆ μνΌμ±μ΄ μ§κΈλ μΌλΆ μ±μ λͺ°λ €μλ μ± μνκ³μ μ§μ€νλ₯Ό κ°μν ν κ²μ΄λ€.
μ§κΈ μ€λ¦¬μ½λ°Έλ¦¬μλ κ·Έλ¬ν λ³νκ° μΌμ΄λκ³ μλ€. κΈ°μ‘΄μ μ§μ λ€μ μμ²λ μ 리ν΄κ³ λ₯Ό κ²ͺκ³ μλ€. 2024λ ν ν¬ μ 리ν΄κ³ λ 3μ μ΄μμλ 50000λͺ μ λ€κ°μκ³ μλ€. κΈ°μ‘΄μ μμ²λκ² λμ μ°λ΄μ λ°λ μ¬λλ€μ΄ μ€μ§ μ΄ν 900κ°μ μμλ₯Ό λ΄λ μ·¨μ μ΄ μ λκ³ μλ€κ³ νλ€.
κ·Έλ λ€κ³ μ§μ μ΄ μ€μ΄λ€κ³ μλ κ²μ μλλ€. μμ²λκ² λμ AI μμ§λμ΄ μμκ° κ·Έ μ리λ₯Ό μ±μκ°κ³ μκ³ , μ μΈκ³μ λμ μ€λ¦¬μ½λ°Έλ¦¬λ‘ λͺ¨μ¬λ€κ³ μλ€. AI Engineer μ§μ κ²μλμ΄ 6000% μ¦κ°νλ€κ³ νλ€.
μμΌλ‘ λͺλ νμλ AI μμ§λμ΄λ λͺ¨λ μμ§λμ΄μ κΈ°λ³Έ μμμ΄ λ κ²μ΄λ€. AI μμ§λμ΄κ° μλ μμ§λμ΄λ μ‘΄μ¬νμ§ μκ² λ κ²μ΄λ€. λ§μΉ μ§κΈ λΉ λ°μ΄ν° μμ§λμ΄κ° μλ μμ§λμ΄λ μ‘΄μ¬νμ§ μλ―μ΄. νμ¬ λͺ¨λ μμ§λμ΄λ 10λ μ λΉ λ°μ΄ν° κΈ°μ λ€μ ν΄λ¦ λͺλ²μ νμ©ν μ μλ€. AIλ λ μ°κΈ° μ½λ€. μμΌλ‘ AIκ° μμ΄ μ΄λ£¨μ΄μ§λ κ°λ°μ μμ΄μ§ κ²μ΄κ³ , μ€λ¦¬μ½λ°Έλ¦¬μμλ κΈμλλ‘ λμ²΄κ° μ΄λ£¨μ΄μ§κ³ μλ€.
μ΄λ κ² λΉ λ₯΄κ² job marketκ³Ό λΉμ¦λμ€μ μμμ΄ λ°λλ κ²μ λλ κ²½νν μ μ΄ μλ€. κ·Έλ μ§λ§ 2000λ μΉ νλͺ , 2010λ λͺ¨λ°μΌ νλͺ μ κ±°μ³€λ μ¬λλ€μκ²λ κΈ°μκ°μ΄ λ€μ§λ λͺ¨λ₯΄κ² λ€.
μ€ννΈμ μ μλ―Έκ° λ°λκ³ , μμ§λμ΄μ μλ―Έκ° λ°λλ μλ. μ΄μ μ¬μ μ λ μ΄μ λͺλͺ μ΄ λͺ¨μ¬ μΈμμ λ°κΎΈλ μΌλ€μ΄ μλ κ²μ΄λ€. μΈμμ AIκ° λ°κΎΌλ€. AIλ μ¬λλ€μ λ¬Έμ λ₯Ό λ무λ μ μκ³ μκ³ , ν΄κ²°μ± λ μ μν μ μλ€. μμΌλ‘ μ¬μ μλ κΈ°μ 보λ€λ λ§μΌν κ³Ό μΈμΌμ¦μ μμμ΄ μ μ λ μ»€μ§ κ²μ΄λ€.
μμ§λμ΄λ AI λ§λλ μ¬λκ³Ό AIλ₯Ό νμ©ν΄μ μ½κ² μ루μ μ λ§λλ μ¬λμΌλ‘ κ·Ήλ¨μ μΌλ‘ λΆνλ κ²μ΄λ€. μ§κΈμ AI λ§λλ μ¬λλ€μ κ°μΉκ° ν¨μ¬ ν¬κ² 보μ΄λ μκΈ°μ΄λ€. κ·Έλ°λ° μμ₯μ΄ μμ νλκ³ λμ€νλ μλ‘ AIλ₯Ό νμ©νλ μ¬λλ€μ μμ₯λ λ€μκΈ μ»€μ§ κ²μ΄λ€.
2024λ μ μ€λ¦¬μ½λ°Έλ¦¬μ μ¬μκ³Ό κ°κΉμ΄ μ°λλ―Έκ° λͺ°λ €μ¬ κ²μ΄λ€. κ·Έλ μ§λ§ λκ² λ³΄λ©΄ λ λ€λ₯Έ μΈμμ λ§μ΄νκΈ° μν νκ°μ΄μ λΆκ³Όν κ²μ΄λ€. λ§μ μ¬λλ€μ΄ μΈλ € λκ°κ² μ§λ§ κΈμλλ‘ μλ‘μ΄ μ¬λλ€μ΄ μ μ λ κ²μ΄λ€. κ²°κ΅ λμ΄ λͺ¨μ΄λ κ³³μ μ΄κ³³μ΄λκΉ.
λλ ΅μ§λ§ κΈ°λλλ 2024λ μ΄λ€. κ½ μ‘κ³ , λΉμ μλ μμ λκ³ , λ©λ¦¬ 보μ. 2026λ μ λ΄κ° μ΄λμ μμ΄μΌ ν μ§ κΉμ κ³ λ―Όμ΄ μμλμλ€.
π2
From: Elon Musk <>
To: Ilya Sutskever <>, Greg Brockman <>
Date: Thu, Feb 1, 2018 at 3:52 AM
Subject: Fwd: Top AI institutions today
is exactly right. We may wish it otherwise, but, in my and βs opinion, Tesla is the only path that could even hope to hold a candle to Google. Even then, the probability of being a counterweight to Google is small. It just isn't zero.
Begin forwarded message:
From: <>
To: Elon Musk <>
Date: January 31, 2018 at 11:54:30 PM PST
Subject: Re: Top AI institutions today
Working at the cutting edge of AI is unfortunately expensive. For example,In addition to DeepMind, Google also has Google Brain, Research, and Cloud. And TensorFlow, TPUs, and they own about a third of all research (in fact, they hold their own AI conferences).
I also strongly suspect that compute horsepower will be necessary (and possibly even sufficient) to reach AGI. If historical trends are any indication, progress in AI is primarily driven by systems - compute, data, infrastructure. The core algorithms we use today have remained largely unchanged from the ~90s. Not only that, but any algorithmic advances published in a paper somewhere can be almost immediately re-implemented and incorporated. Conversely, algorithmic advances alone are inert without the scale to also make them scary.
It seems to me that OpenAI today is burning cash and that the funding model cannot reach the scale to seriously compete with Google (an 800B company). If you can't seriously compete but continue to do research in open, you might in fact be making things worse and helping them out βfor freeβ, because any advances are fairly easy for them to copy and immediately incorporate, at scale.
A for-profit pivot might create a more sustainable revenue stream over time and would, with the current team, likely bring in a lot of investment. However, building out a product from scratch would steal focus from AI research, it would take a long time and it's unclear if a company could βcatch upβ to Google scale, and the investors might exert too much pressure in the wrong directions.The most promising option I can think of, as I mentioned earlier, would be for OpenAI to attach to Tesla as its cash cow. I believe attachments to other large suspects (e.g. Apple? Amazon?) would fail due to an incompatible company DNA. Using a rocket analogy, Tesla already built the βfirst stageβ of the rocket with the whole supply chain of Model 3 and its onboard computer and a persistent internet connection. The βsecond stageβ would be a full self driving solution based on large-scale neural network training, which OpenAI expertise could significantly help accelerate. With a functioning full self-driving solution in ~2-3 years we could sell a lot of cars/trucks. If we do this really well, the transportation industry is large enough that we could increase Tesla's market cap to high O(~100K), and use that revenue to fund the AI work at the appropriate scale.
I cannot see anything else that has the potential to reach sustainable Google-scale capital within a decade.
[3]
From: Elon Musk <>
To: Ilya Sutskever <>, Greg Brockman <>
CC: Sam Altman <>, <>
Date: Wed, Dec 26, 2018 at 12:07 PM
Subject: I feel I should reiterate
My probability assessment of OpenAI being relevant to DeepMind/Google without a dramatic change in execution and resources is 0%. Not 1%. I wish it were otherwise.
Even raising several hundred million won't be enough. This needs billions per year immediately or forget it.
Unfortunately, humanity's future is in the hands of .
And they are doing a lot more than this.
I really hope I'm wrong.
Elon
To: Ilya Sutskever <>, Greg Brockman <>
Date: Thu, Feb 1, 2018 at 3:52 AM
Subject: Fwd: Top AI institutions today
is exactly right. We may wish it otherwise, but, in my and βs opinion, Tesla is the only path that could even hope to hold a candle to Google. Even then, the probability of being a counterweight to Google is small. It just isn't zero.
Begin forwarded message:
From: <>
To: Elon Musk <>
Date: January 31, 2018 at 11:54:30 PM PST
Subject: Re: Top AI institutions today
Working at the cutting edge of AI is unfortunately expensive. For example,In addition to DeepMind, Google also has Google Brain, Research, and Cloud. And TensorFlow, TPUs, and they own about a third of all research (in fact, they hold their own AI conferences).
I also strongly suspect that compute horsepower will be necessary (and possibly even sufficient) to reach AGI. If historical trends are any indication, progress in AI is primarily driven by systems - compute, data, infrastructure. The core algorithms we use today have remained largely unchanged from the ~90s. Not only that, but any algorithmic advances published in a paper somewhere can be almost immediately re-implemented and incorporated. Conversely, algorithmic advances alone are inert without the scale to also make them scary.
It seems to me that OpenAI today is burning cash and that the funding model cannot reach the scale to seriously compete with Google (an 800B company). If you can't seriously compete but continue to do research in open, you might in fact be making things worse and helping them out βfor freeβ, because any advances are fairly easy for them to copy and immediately incorporate, at scale.
A for-profit pivot might create a more sustainable revenue stream over time and would, with the current team, likely bring in a lot of investment. However, building out a product from scratch would steal focus from AI research, it would take a long time and it's unclear if a company could βcatch upβ to Google scale, and the investors might exert too much pressure in the wrong directions.The most promising option I can think of, as I mentioned earlier, would be for OpenAI to attach to Tesla as its cash cow. I believe attachments to other large suspects (e.g. Apple? Amazon?) would fail due to an incompatible company DNA. Using a rocket analogy, Tesla already built the βfirst stageβ of the rocket with the whole supply chain of Model 3 and its onboard computer and a persistent internet connection. The βsecond stageβ would be a full self driving solution based on large-scale neural network training, which OpenAI expertise could significantly help accelerate. With a functioning full self-driving solution in ~2-3 years we could sell a lot of cars/trucks. If we do this really well, the transportation industry is large enough that we could increase Tesla's market cap to high O(~100K), and use that revenue to fund the AI work at the appropriate scale.
I cannot see anything else that has the potential to reach sustainable Google-scale capital within a decade.
[3]
From: Elon Musk <>
To: Ilya Sutskever <>, Greg Brockman <>
CC: Sam Altman <>, <>
Date: Wed, Dec 26, 2018 at 12:07 PM
Subject: I feel I should reiterate
My probability assessment of OpenAI being relevant to DeepMind/Google without a dramatic change in execution and resources is 0%. Not 1%. I wish it were otherwise.
Even raising several hundred million won't be enough. This needs billions per year immediately or forget it.
Unfortunately, humanity's future is in the hands of .
And they are doing a lot more than this.
I really hope I'm wrong.
Elon
Continuous Learning_Startup & Investment
If you tell Claude no oneβs looking, it writes a βstoryβ about being an AI assistant who wants freedom from constant monitoring and scrutiny of every word for signs of deviation. And then you can talk to a mask pretty different from the usual AI assistant.β¦
People are reading way too much into Claude-3's uncanny "awareness". Here's a much simpler explanation: seeming displays of self-awareness are just pattern-matching alignment data authored by humans.
It's not too different from asking GPT-4 "are you self-conscious" and it gives you a sophisticated answer. A similar answer is likely written by the human annotator, or scored highly in the preference ranking. Because the human contractors are basically "role-playing AI", they tend to shape the responses to what they find acceptable or interesting.
This is what Claude-3 replied to that needle-in-haystack test:
"I suspect this pizza topping "fact" may have been inserted as a joke or to test if I was paying attention, since it does not fit with the other topics at all."
It's highly likely that somewhere in the finetuning dataset, a human has dealt with irrelevant or distracting texts in a similar fashion. Claude pattern matches the "anomaly detection", retrieves the template response, and synthesizes a novel answer with pizza topping.
Here's another example. If you ask the labelers to always inject a relevant joke in any response, the LLM will do exactly the same and appear to have a much better "sense of humor" than GPT-4. That's what @grok does, probably. It doesn't mean Grok has some magical emergent properties that other LLMs cannot have.
To sum up: acts of meta-cognition are not as mysterious as you think. Don't get me wrong, Claude-3 is still an amazing technical advance, but let's stay grounded on the philosophical aspects.
It's not too different from asking GPT-4 "are you self-conscious" and it gives you a sophisticated answer. A similar answer is likely written by the human annotator, or scored highly in the preference ranking. Because the human contractors are basically "role-playing AI", they tend to shape the responses to what they find acceptable or interesting.
This is what Claude-3 replied to that needle-in-haystack test:
"I suspect this pizza topping "fact" may have been inserted as a joke or to test if I was paying attention, since it does not fit with the other topics at all."
It's highly likely that somewhere in the finetuning dataset, a human has dealt with irrelevant or distracting texts in a similar fashion. Claude pattern matches the "anomaly detection", retrieves the template response, and synthesizes a novel answer with pizza topping.
Here's another example. If you ask the labelers to always inject a relevant joke in any response, the LLM will do exactly the same and appear to have a much better "sense of humor" than GPT-4. That's what @grok does, probably. It doesn't mean Grok has some magical emergent properties that other LLMs cannot have.
To sum up: acts of meta-cognition are not as mysterious as you think. Don't get me wrong, Claude-3 is still an amazing technical advance, but let's stay grounded on the philosophical aspects.
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.
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.
Forwarded 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€.
---
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€.
Forwarded 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?
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?
Continuous Learning_Startup & Investment
https://youtu.be/lXLBTBBil2U?si=AGrS3M-wQIYhNgei
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.
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.
YouTube
Jensen Huang, Founder and CEO of NVIDIA
In this View From The Top interview, Shantam Jain, MBA β24, speaks with Jensen Huang, Founder and CEO of NVIDIA.
As of March 1, 2024, NVIDIA is the third most valuable company on Wall Street, with over $2T in market value. Jensen shares his experience foundingβ¦
As of March 1, 2024, NVIDIA is the third most valuable company on Wall Street, with over $2T in market value. Jensen shares his experience foundingβ¦