I think this is mostly right.
- LLMs created a whole new layer of abstraction and profession.
- I've so far called this role "Prompt Engineer" but agree it is misleading. It's not just prompting alone, there's a lot of glue code/infra around it. Maybe "AI Engineer" is ~usable, though it takes something a bit too specific and makes it a bit too broad.
- ML people train algorithms/networks, usually from scratch, usually at lower capability.
- LLM training is becoming sufficently different from ML because of its systems-heavy workloads, and is also splitting off into a new kind of role, focused on very large scale training of transformers on supercomputers.
- In numbers, there's probably going to be significantly more AI Engineers than there are ML engineers / LLM engineers.
- One can be quite successful in this role without ever training anything.
- I don't fully follow the Software 1.0/2.0 framing. Software 3.0 (imo ~prompting LLMs) is amusing because prompts are human-designed "code", but in English, and interpreted by an LLM (itself now a Software 2.0 artifact). AI Engineers simultaneously program in all 3 paradigms. It's a bit ๐ตโ๐ซ
https://twitter.com/karpathy/status/1674873002314563584
- LLMs created a whole new layer of abstraction and profession.
- I've so far called this role "Prompt Engineer" but agree it is misleading. It's not just prompting alone, there's a lot of glue code/infra around it. Maybe "AI Engineer" is ~usable, though it takes something a bit too specific and makes it a bit too broad.
- ML people train algorithms/networks, usually from scratch, usually at lower capability.
- LLM training is becoming sufficently different from ML because of its systems-heavy workloads, and is also splitting off into a new kind of role, focused on very large scale training of transformers on supercomputers.
- In numbers, there's probably going to be significantly more AI Engineers than there are ML engineers / LLM engineers.
- One can be quite successful in this role without ever training anything.
- I don't fully follow the Software 1.0/2.0 framing. Software 3.0 (imo ~prompting LLMs) is amusing because prompts are human-designed "code", but in English, and interpreted by an LLM (itself now a Software 2.0 artifact). AI Engineers simultaneously program in all 3 paradigms. It's a bit ๐ตโ๐ซ
https://twitter.com/karpathy/status/1674873002314563584
[์ด๋ป๊ฒ AI๋ฅผ ์ฌ์ฉํด์ 10๋ฐฐ ์ข์ ์ ํ์ ๋ง๋ค ์ ์์๊น?_Github Copilot]
Copilot์ ์ฌ์ฉํ์๋ ๋ถ๋ค ์์ผ์ ๊ฐ์? GitHub Copilot์ GitHub์ OpenAI์์ ๊ฐ๋ฐํ AI coding assistance์ธ๋ฐ์. ์ ์ฃผ๋ณ์ Copilot์ ํ๋ฒ ์ฌ์ฉํ์ ๋ถ๋ค์ ๋๋ถ๋ถ ๊พธ์คํ ์ฌ์ฉํ์๋ฉด์ ๋ง์กฑํ์๋๋ผ๊ณ ์.
๋ด๊ฐ ์์ฑํ๋ ์ฝ๋ ๋ฒ ์ด์ค์ ๋งฅ๋ฝ์ ์์ธํ ์ดํดํ๊ณ ์ค์๊ฐ์ผ๋ก ์ ํฉํ ์ ๋ณด๋ฅผ ์ถ์ฒํ๋ ๊ฒ์ด Copilot์ด ๊ฐ์ง ๊ฐ์ฅ ํฐ ์ฅ์ ์ด๋ผ๊ณ ์๊ฐํ๋๋ฐ์. ์ด๋ฐ ๊ธฐ๋ฅ๋ค์ ์ด๋ป๊ฒ ์ง์ํ ์ ์์๊น์? AI๊ฐ ์์์ ๋ค ์ถ์ฒํด์ฃผ๋ ๊ฑธ๊น์?
์ต๊ทผ์ Parth Thakkar์ด ์ด Copilot Internals์ ์ฝ๊ณ ์๋กญ๊ฒ ๋ฐฐ์ด ๋ด์ฉ๊ณผ ์ ์๊ฐ ํ ์คํผ์ ๊ณต์ ํฉ๋๋ค.
Copilot์ด ๋ด๊ฐ ์์ฑํ๋ ์ฝ๋์ ๋งฅ๋ฝ์ ์ดํดํ๊ณ ์ค์๊ฐ์ผ๋ก ๋ต์ ์ฃผ๋ ๋ฐ์๋ ํฌ๊ฒ 3๊ฐ์ง ๋น๋ฒ ์์ค๊ฐ ์กด์ฌํฉ๋๋ค.
๋น๋ฒ ์์ค 1: ํ๋กฌํํธ ์์ง๋์ด๋ง
- ํด๋ผ์ด์ธํธ๊ฐ ํ๋กฌํํธ๋ฅผ ๋ณด๋ด๋ฉด(์ฝ๋๋ฅผ ์์ฑํ๋ฉด), ์ฝ๋์ ๊ด๋ จ๋ ๋งฅ๋ฝ(์ ๋์ฌ[์ฝ๋ ์์น, ๊ด๋ จ ์ฝ๋/ํ์ผ์ ์ค๋ํซ], ์ ๋ฏธ์ฌ(์์ฑ๋ ์ฝ๋๊ฐ ๋ค์ด๊ฐ ์ฅ์์ ๋ํ ๋งฅ๋ฝ), PromptElementRanges(ํ๋กฌํํธ๊ฐ ์ ์๋ํ๊ธฐ ์ํ ๊ธฐ๋ณธ ์ ๋ณด๋ค)์ AI Model(Codex)์ ๋ณด๋ ๋๋ค.
๋น๋ฒ ์์ค 2: ๋ชจ๋ธ ํธ์ถ(Model Invocation)
- Copilot์ ์ธ๋ผ์ธ/๊ณ ์คํธํ ์คํธ ๊ทธ๋ฆฌ๊ณ Copilotํจ๋ ๋๊ฐ์ง ์ฑ๋์ ํตํด์ AI๋ชจ๋ธ์ ํธ์ถํฉ๋๋ค.
- GitHub Copilot์ ์ธ๋ผ์ธ/๊ณ ์คํธํ ์คํธ ์ธํฐํ์ด์ค๋ ์ ์ ์๋๋ฅผ ๋์ด๊ณ , ๋ฐ๋ณต์ ์ธ ๋ชจ๋ธ ํธ์ถ์ ์ค์ด๊ณ , ์ฌ์ฉ์์ ์ ๋ ฅ์ ๋ฐ๋ผ ์ ์์ ์กฐ์ ํ๋ฉฐ, ๋น ๋ฅธ ์ ๋ ฅ์ ์ฒ๋ฆฌํ๊ธฐ ์ํด ๋๋ฐ์ด์ฑ ๋ฉ์ปค๋์ฆ์ ์ฌ์ฉํฉ๋๋ค. ๋ฐ๋ฉด, ์ฝํ์ผ๋ฟ ํจ๋์ ๋ ๋ง์ ์ํ์ ์์ฒญํ๊ณ , ๋ก๊ทธ ํ๋ก๋ธ๋ฅผ ์ฌ์ฉํ์ฌ ์๋ฃจ์ ์ ์ ๋ ฌํฉ๋๋ค. ๋ ์ธํฐํ์ด์ค ๋ชจ๋๋ ๋์์ด ๋์ง ์๋ ์๋ฃ๋ฅผ ๋ฐฉ์งํ๊ธฐ ์ํด ๊ฒ์ฌ๋ฅผ ์ํํฉ๋๋ค.
๋น๋ฒ ์์ค 3: ์๊ฒฉ ์ธก์ (Telemetry)
- GitHub Copilot๋ ์๊ฒฉ ์ธก์ ์ ํตํด ์ฌ์ฉ์ ์ํธ์์ฉ์ ํ์ตํ๊ณ ์ ํ์ ๊ฐ์ ํฉ๋๋ค. ์ด๋ ์ ์์ ์๋ฝ์ด๋ ๊ฑฐ๋ถ, ์ฝ๋์ ๋จ์ ์๋ ์๋ฝ๋ ์ ์์ ์ง์์ฑ, ์ ์ ์๋ฝ ํ 30์ด ์ด๋ด์ ์บก์ฒ๋ ์ฝ๋ ์ค๋ํซ ๋ฑ์ ํฌํจํ๋ฉฐ, ์ฌ์ฉ์๋ ๊ฐ์ธ ์ ๋ณด ๋ณดํธ๋ฅผ ์ํด ์ด๋ฌํ ๋ฐ์ดํฐ ์์ง์ ๊ฑฐ๋ถํ ์ ์์ต๋๋ค.
AI๋ฅผ ์ฌ์ฉํด์ ๊ณ ๊ฐ์๊ฒ 10๋ฐฐ ์ข์ ์ ํ์ ๋ง๋๋ ์ฐฝ์ ์๋ก์ ๋ฌด์์ ๋ฐฐ์ธ ์ ์์๊น์?
- ๊ณ ๊ฐ์๊ฒ ๊ฐ์น๋ฅผ ์ ๋ฌํ๊ธฐ ์ํด์๋ ๋ชจ๋ธ ๊ทธ์์ฒด๋ก ์ถฉ๋ถํ์ง ์์ต๋๋ค. ๊ณ ๊ฐ์ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ธฐ ์ํด ๋ชจ๋ธ์ ์ ์ดํดํ๊ณ , ์ด๋ฅผ ์ ์ฌ์ฉํ ์ ์๋ Engineering ์ญ๋๊ณผ ๋น ๋ฅธ Iteration์ด ์ค์ํฉ๋๋ค.
- ์์ง, LLM ํน์ AI๋ฅผ ์ฌ์ฉํด์ ์ข์ ์ ํ์ ๋ง๋๋ ๋ฐฉ๋ฒ์ด ์ ์๋ ค์ง์ง ์์๊ณ ๋๊ตฌ๋ ๋ต์ ์๊ณ ์๋ค๊ณ ๋งํ๊ธฐ ์ด๋ ต์ต๋๋ค. ๋ฐ๋ผ์, AI๋ฅผ ์ ์ดํดํ๋ฉด์๋ ๊ณ ๊ฐ์ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ธฐ ์ํด์ AI ๋ชจ๋ธ๊ณผ ๋ค์ํ ์์ง๋์ด๋ง์ ๊ฒฐํฉํ๋ ค๋ ์คํํธ์ ์๊ฒ ๊ธฐํ๊ฐ ์์ต๋๋ค.
๋ ์์ธํ ๋ด์ฉ์ ๋ณด๊ณ ์ถ๋ค๋ฉด ์๋ ๋งํฌ๋ฅผ ์ฐธ๊ณ ํด์ฃผ์ธ์. https://bit.ly/copilotinternal
ํน์, Engineering ๊ฒฝํ๊ณผ ์ง์์ด ๋ฐ์ด๋ ๋ถ์ด๋ AI ๋ชจ๋ธ์ ๋ํ ์ดํด๊ฐ ๋์ผ์ ๋ถ๋ค ์ค์์ ๊ณ ๊ฐ์๊ฒ 10๋ฐฐ ์ข์ ๊ฐ์น๋ฅผ ๋ง๋ค์ด๋ด๋ ๋ฐ์ ๊ด์ฌ์๋ ๋ถ์ด ์์ผ์๋ค๋ฉด DM ํน์ minseok.kim0129@gmail.com์ผ๋ก ์ด๋ค ๋ฌธ์ ๋ฅผ ํด๊ฒฐํด์ค์ จ๊ณ ์์ผ๋ก๋ ์ด๋ค ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ณ ์ถ์ผ์ ์ง ํธํ๊ฒ ์๋ ค์ฃผ์ ์ ๐
์์ฆ AI ์๋น์ค๋ค์ ์ฌ์ฉํ๋ฉด์ PC, ์ธํฐ๋ท, ๋ชจ๋ฐ์ผ ์ด์ฐฝ๊ธฐ์ ๋น์ทํ๊ฒ ์ด๋ค ์๋น์ค๊ฐ ์ ์ ์๊ฒ ๊ฐ์น์์์ง ๋ชฐ๋ผ์ ๋ญ๋ ํด๋ณผ ์ ์๋์๊ธฐ๋ผ๊ณ ์๊ฐํฉ๋๋ค. ๋ง์น ๋ชจ๋ฐ์ผ ์ด์ฐฝ๊ธฐ์๋ LBS(Location Based Service)๋ผ๋ ๊ฐ๋ ์ด ์์์ง๋ง ์ง๊ธ์ ๋๋ถ๋ถ ๋ชจ๋ฐ์ผ ์๋น์ค์์ GPS๋ฅผ ์์ฃผ ๊ธฐ๋ณธ์ ์ผ๋ก ์ ๊ณตํ๋ ๊ฒ์ฒ๋ผ์.
๊ฒฐ๊ตญ ์ ์ ์ ๋ณํ์ง ์๋ ๋์ฆ๋ฅผ ๋ฐ๊ฒฌํ๊ณ ๋น ๋ฅด๊ฒ ๋ณํํ๋ ๊ธฐ์ ์ ์ ํ์ฉํด์ 10๋ฐฐ ์ข์ ์๋น์ค๋ฅผ ์ง์์ ์ผ๋ก ๋ง๋ค ์ ์๋ ํ์ด ์ข์ ์ ํ ๊ทธ๋ฆฌ๊ณ ์ข์ ํ์ฌ๋ฅผ ๋ง๋ค ์ ์๋ค๊ณ ๋ฏฟ์ต๋๋ค.
Copilot์ ์ฌ์ฉํ์๋ ๋ถ๋ค ์์ผ์ ๊ฐ์? GitHub Copilot์ GitHub์ OpenAI์์ ๊ฐ๋ฐํ AI coding assistance์ธ๋ฐ์. ์ ์ฃผ๋ณ์ Copilot์ ํ๋ฒ ์ฌ์ฉํ์ ๋ถ๋ค์ ๋๋ถ๋ถ ๊พธ์คํ ์ฌ์ฉํ์๋ฉด์ ๋ง์กฑํ์๋๋ผ๊ณ ์.
๋ด๊ฐ ์์ฑํ๋ ์ฝ๋ ๋ฒ ์ด์ค์ ๋งฅ๋ฝ์ ์์ธํ ์ดํดํ๊ณ ์ค์๊ฐ์ผ๋ก ์ ํฉํ ์ ๋ณด๋ฅผ ์ถ์ฒํ๋ ๊ฒ์ด Copilot์ด ๊ฐ์ง ๊ฐ์ฅ ํฐ ์ฅ์ ์ด๋ผ๊ณ ์๊ฐํ๋๋ฐ์. ์ด๋ฐ ๊ธฐ๋ฅ๋ค์ ์ด๋ป๊ฒ ์ง์ํ ์ ์์๊น์? AI๊ฐ ์์์ ๋ค ์ถ์ฒํด์ฃผ๋ ๊ฑธ๊น์?
์ต๊ทผ์ Parth Thakkar์ด ์ด Copilot Internals์ ์ฝ๊ณ ์๋กญ๊ฒ ๋ฐฐ์ด ๋ด์ฉ๊ณผ ์ ์๊ฐ ํ ์คํผ์ ๊ณต์ ํฉ๋๋ค.
Copilot์ด ๋ด๊ฐ ์์ฑํ๋ ์ฝ๋์ ๋งฅ๋ฝ์ ์ดํดํ๊ณ ์ค์๊ฐ์ผ๋ก ๋ต์ ์ฃผ๋ ๋ฐ์๋ ํฌ๊ฒ 3๊ฐ์ง ๋น๋ฒ ์์ค๊ฐ ์กด์ฌํฉ๋๋ค.
๋น๋ฒ ์์ค 1: ํ๋กฌํํธ ์์ง๋์ด๋ง
- ํด๋ผ์ด์ธํธ๊ฐ ํ๋กฌํํธ๋ฅผ ๋ณด๋ด๋ฉด(์ฝ๋๋ฅผ ์์ฑํ๋ฉด), ์ฝ๋์ ๊ด๋ จ๋ ๋งฅ๋ฝ(์ ๋์ฌ[์ฝ๋ ์์น, ๊ด๋ จ ์ฝ๋/ํ์ผ์ ์ค๋ํซ], ์ ๋ฏธ์ฌ(์์ฑ๋ ์ฝ๋๊ฐ ๋ค์ด๊ฐ ์ฅ์์ ๋ํ ๋งฅ๋ฝ), PromptElementRanges(ํ๋กฌํํธ๊ฐ ์ ์๋ํ๊ธฐ ์ํ ๊ธฐ๋ณธ ์ ๋ณด๋ค)์ AI Model(Codex)์ ๋ณด๋ ๋๋ค.
๋น๋ฒ ์์ค 2: ๋ชจ๋ธ ํธ์ถ(Model Invocation)
- Copilot์ ์ธ๋ผ์ธ/๊ณ ์คํธํ ์คํธ ๊ทธ๋ฆฌ๊ณ Copilotํจ๋ ๋๊ฐ์ง ์ฑ๋์ ํตํด์ AI๋ชจ๋ธ์ ํธ์ถํฉ๋๋ค.
- GitHub Copilot์ ์ธ๋ผ์ธ/๊ณ ์คํธํ ์คํธ ์ธํฐํ์ด์ค๋ ์ ์ ์๋๋ฅผ ๋์ด๊ณ , ๋ฐ๋ณต์ ์ธ ๋ชจ๋ธ ํธ์ถ์ ์ค์ด๊ณ , ์ฌ์ฉ์์ ์ ๋ ฅ์ ๋ฐ๋ผ ์ ์์ ์กฐ์ ํ๋ฉฐ, ๋น ๋ฅธ ์ ๋ ฅ์ ์ฒ๋ฆฌํ๊ธฐ ์ํด ๋๋ฐ์ด์ฑ ๋ฉ์ปค๋์ฆ์ ์ฌ์ฉํฉ๋๋ค. ๋ฐ๋ฉด, ์ฝํ์ผ๋ฟ ํจ๋์ ๋ ๋ง์ ์ํ์ ์์ฒญํ๊ณ , ๋ก๊ทธ ํ๋ก๋ธ๋ฅผ ์ฌ์ฉํ์ฌ ์๋ฃจ์ ์ ์ ๋ ฌํฉ๋๋ค. ๋ ์ธํฐํ์ด์ค ๋ชจ๋๋ ๋์์ด ๋์ง ์๋ ์๋ฃ๋ฅผ ๋ฐฉ์งํ๊ธฐ ์ํด ๊ฒ์ฌ๋ฅผ ์ํํฉ๋๋ค.
๋น๋ฒ ์์ค 3: ์๊ฒฉ ์ธก์ (Telemetry)
- GitHub Copilot๋ ์๊ฒฉ ์ธก์ ์ ํตํด ์ฌ์ฉ์ ์ํธ์์ฉ์ ํ์ตํ๊ณ ์ ํ์ ๊ฐ์ ํฉ๋๋ค. ์ด๋ ์ ์์ ์๋ฝ์ด๋ ๊ฑฐ๋ถ, ์ฝ๋์ ๋จ์ ์๋ ์๋ฝ๋ ์ ์์ ์ง์์ฑ, ์ ์ ์๋ฝ ํ 30์ด ์ด๋ด์ ์บก์ฒ๋ ์ฝ๋ ์ค๋ํซ ๋ฑ์ ํฌํจํ๋ฉฐ, ์ฌ์ฉ์๋ ๊ฐ์ธ ์ ๋ณด ๋ณดํธ๋ฅผ ์ํด ์ด๋ฌํ ๋ฐ์ดํฐ ์์ง์ ๊ฑฐ๋ถํ ์ ์์ต๋๋ค.
AI๋ฅผ ์ฌ์ฉํด์ ๊ณ ๊ฐ์๊ฒ 10๋ฐฐ ์ข์ ์ ํ์ ๋ง๋๋ ์ฐฝ์ ์๋ก์ ๋ฌด์์ ๋ฐฐ์ธ ์ ์์๊น์?
- ๊ณ ๊ฐ์๊ฒ ๊ฐ์น๋ฅผ ์ ๋ฌํ๊ธฐ ์ํด์๋ ๋ชจ๋ธ ๊ทธ์์ฒด๋ก ์ถฉ๋ถํ์ง ์์ต๋๋ค. ๊ณ ๊ฐ์ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ธฐ ์ํด ๋ชจ๋ธ์ ์ ์ดํดํ๊ณ , ์ด๋ฅผ ์ ์ฌ์ฉํ ์ ์๋ Engineering ์ญ๋๊ณผ ๋น ๋ฅธ Iteration์ด ์ค์ํฉ๋๋ค.
- ์์ง, LLM ํน์ AI๋ฅผ ์ฌ์ฉํด์ ์ข์ ์ ํ์ ๋ง๋๋ ๋ฐฉ๋ฒ์ด ์ ์๋ ค์ง์ง ์์๊ณ ๋๊ตฌ๋ ๋ต์ ์๊ณ ์๋ค๊ณ ๋งํ๊ธฐ ์ด๋ ต์ต๋๋ค. ๋ฐ๋ผ์, AI๋ฅผ ์ ์ดํดํ๋ฉด์๋ ๊ณ ๊ฐ์ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ธฐ ์ํด์ AI ๋ชจ๋ธ๊ณผ ๋ค์ํ ์์ง๋์ด๋ง์ ๊ฒฐํฉํ๋ ค๋ ์คํํธ์ ์๊ฒ ๊ธฐํ๊ฐ ์์ต๋๋ค.
๋ ์์ธํ ๋ด์ฉ์ ๋ณด๊ณ ์ถ๋ค๋ฉด ์๋ ๋งํฌ๋ฅผ ์ฐธ๊ณ ํด์ฃผ์ธ์. https://bit.ly/copilotinternal
ํน์, Engineering ๊ฒฝํ๊ณผ ์ง์์ด ๋ฐ์ด๋ ๋ถ์ด๋ AI ๋ชจ๋ธ์ ๋ํ ์ดํด๊ฐ ๋์ผ์ ๋ถ๋ค ์ค์์ ๊ณ ๊ฐ์๊ฒ 10๋ฐฐ ์ข์ ๊ฐ์น๋ฅผ ๋ง๋ค์ด๋ด๋ ๋ฐ์ ๊ด์ฌ์๋ ๋ถ์ด ์์ผ์๋ค๋ฉด DM ํน์ minseok.kim0129@gmail.com์ผ๋ก ์ด๋ค ๋ฌธ์ ๋ฅผ ํด๊ฒฐํด์ค์ จ๊ณ ์์ผ๋ก๋ ์ด๋ค ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ณ ์ถ์ผ์ ์ง ํธํ๊ฒ ์๋ ค์ฃผ์ ์ ๐
์์ฆ AI ์๋น์ค๋ค์ ์ฌ์ฉํ๋ฉด์ PC, ์ธํฐ๋ท, ๋ชจ๋ฐ์ผ ์ด์ฐฝ๊ธฐ์ ๋น์ทํ๊ฒ ์ด๋ค ์๋น์ค๊ฐ ์ ์ ์๊ฒ ๊ฐ์น์์์ง ๋ชฐ๋ผ์ ๋ญ๋ ํด๋ณผ ์ ์๋์๊ธฐ๋ผ๊ณ ์๊ฐํฉ๋๋ค. ๋ง์น ๋ชจ๋ฐ์ผ ์ด์ฐฝ๊ธฐ์๋ LBS(Location Based Service)๋ผ๋ ๊ฐ๋ ์ด ์์์ง๋ง ์ง๊ธ์ ๋๋ถ๋ถ ๋ชจ๋ฐ์ผ ์๋น์ค์์ GPS๋ฅผ ์์ฃผ ๊ธฐ๋ณธ์ ์ผ๋ก ์ ๊ณตํ๋ ๊ฒ์ฒ๋ผ์.
๊ฒฐ๊ตญ ์ ์ ์ ๋ณํ์ง ์๋ ๋์ฆ๋ฅผ ๋ฐ๊ฒฌํ๊ณ ๋น ๋ฅด๊ฒ ๋ณํํ๋ ๊ธฐ์ ์ ์ ํ์ฉํด์ 10๋ฐฐ ์ข์ ์๋น์ค๋ฅผ ์ง์์ ์ผ๋ก ๋ง๋ค ์ ์๋ ํ์ด ์ข์ ์ ํ ๊ทธ๋ฆฌ๊ณ ์ข์ ํ์ฌ๋ฅผ ๋ง๋ค ์ ์๋ค๊ณ ๋ฏฟ์ต๋๋ค.
KIM MINSEOK's Notion on Notion
How does copilot work?
Github Copilot์ Reverse Engineeringํ๋ ๊ธ์ ๊ณต๋ถํ๋ฉด์ ๋ฒ์ญ/์์ญ, ์ถ๊ฐ์กฐ์ฌํ ๋ด์ฉ์ ์ ๋ฆฌํ ๊ธ์
๋๋ค.
Forwarded from ์ ์ข
ํ์ ์ธ์ฌ์ดํธ
"Snowflake announced a new container service and a partnership with Nvidia to make it easier to build generative AI applications making use of all that data and running them on Nvidia GPUs."
https://techcrunch.com/2023/06/27/snowflake-nvidia-partnership-could-make-it-easier-to-build-generative-ai-applications/?utm_source=bensbites&utm_medium=newsletter&utm_campaign=ai-partnerships-acquisitions-and-funding&guccounter=1
https://techcrunch.com/2023/06/27/snowflake-nvidia-partnership-could-make-it-easier-to-build-generative-ai-applications/?utm_source=bensbites&utm_medium=newsletter&utm_campaign=ai-partnerships-acquisitions-and-funding&guccounter=1
TechCrunch
Snowflake-Nvidia partnership could make it easier to build generative AI applications
Snowflake has always been about storing large amounts of unstructured data in the cloud. With two recent acquisitions, Neeva and Streamlit, it will make it easier to search and build applications on top of the data. Today, the company announced a new containerโฆ
https://youtu.be/ajkAbLe-0Uk
Major Takeaways:
Product Differentiation: Perplexity AI focuses on providing accurate and trustworthy search results with citations, thereby positioning itself as a superior alternative to AI models like ChatGPT and Bart in terms of search accuracy. They differentiate themselves further by leveraging reasoning engines in combination with a well-ranked index of relevant content to generate quick and accurate answers.
Technology Utilization and Development: Perplexity AI's strategy relies on utilizing well-established AI models such as ChatGPT and Bart, but also developing their own models to address specific aspects of their product. This allows them to create a competitive and unique search experience. Moreover, the company orchestrates various components in their backend to ensure they work together efficiently and reliably.
Business Model and Advertising: The company considers advertising within a chat interface, which could provide relevant and targeted ads based on user profiles and queries, as a promising potential business model. The need for transparency and ethical advertising practices is emphasized.
AI Integration: The future vision for Perplexity AI involves the seamless integration of language models into everyday devices, which will enable natural conversations and immediate responses. The speaker acknowledges the existing limitations but expresses confidence in the continual advancements of the technology.
Data Quality and Training: The quality of training data is highlighted as a key factor in achieving higher levels of reasoning and intelligence in AI models. This is seen as a factor contributing to the lead of OpenAI in the AI market.
Open-source vs. Closed Models: The speaker discusses the implications of open-source models and closed models like Google and OpenAI, noting that the progress in the field depends on algorithmic efficiencies and talented researchers. The dynamics of this will be influenced by whether organizations continue to publish their techniques or opt to stay closed.
Lessons for AI Startup Founders:
Differentiation is Key: In a competitive field, providing a unique value proposition is crucial. This might involve creating more accurate or trustworthy results, or delivering them in a more efficient manner.
Leverage and Develop Technology: While it's beneficial to leverage established AI models, developing your own models to address specific aspects of your product can create a competitive edge.
Backend Efficiency: The success of your startup doesn't only rely on the end product but also how well the backend processes and components are orchestrated.
Ethical Business Practices: In implementing advertising or other monetization methods, maintaining transparency and ethical practices is essential to avoid the risk of alienating users.
Quality of Training Data: As an AI startup, the quality of your training data is paramount. Efforts should be made to curate high-quality data to achieve superior models.
Open Source vs. Closed Debate: The choice between operating with open-source models or closed ones can have implications on your company's future. Founders should consider the pros and cons of each, taking into account factors such as collaboration, progress speed, and knowledge sharing.
Major Takeaways:
Product Differentiation: Perplexity AI focuses on providing accurate and trustworthy search results with citations, thereby positioning itself as a superior alternative to AI models like ChatGPT and Bart in terms of search accuracy. They differentiate themselves further by leveraging reasoning engines in combination with a well-ranked index of relevant content to generate quick and accurate answers.
Technology Utilization and Development: Perplexity AI's strategy relies on utilizing well-established AI models such as ChatGPT and Bart, but also developing their own models to address specific aspects of their product. This allows them to create a competitive and unique search experience. Moreover, the company orchestrates various components in their backend to ensure they work together efficiently and reliably.
Business Model and Advertising: The company considers advertising within a chat interface, which could provide relevant and targeted ads based on user profiles and queries, as a promising potential business model. The need for transparency and ethical advertising practices is emphasized.
AI Integration: The future vision for Perplexity AI involves the seamless integration of language models into everyday devices, which will enable natural conversations and immediate responses. The speaker acknowledges the existing limitations but expresses confidence in the continual advancements of the technology.
Data Quality and Training: The quality of training data is highlighted as a key factor in achieving higher levels of reasoning and intelligence in AI models. This is seen as a factor contributing to the lead of OpenAI in the AI market.
Open-source vs. Closed Models: The speaker discusses the implications of open-source models and closed models like Google and OpenAI, noting that the progress in the field depends on algorithmic efficiencies and talented researchers. The dynamics of this will be influenced by whether organizations continue to publish their techniques or opt to stay closed.
Lessons for AI Startup Founders:
Differentiation is Key: In a competitive field, providing a unique value proposition is crucial. This might involve creating more accurate or trustworthy results, or delivering them in a more efficient manner.
Leverage and Develop Technology: While it's beneficial to leverage established AI models, developing your own models to address specific aspects of your product can create a competitive edge.
Backend Efficiency: The success of your startup doesn't only rely on the end product but also how well the backend processes and components are orchestrated.
Ethical Business Practices: In implementing advertising or other monetization methods, maintaining transparency and ethical practices is essential to avoid the risk of alienating users.
Quality of Training Data: As an AI startup, the quality of your training data is paramount. Efforts should be made to curate high-quality data to achieve superior models.
Open Source vs. Closed Debate: The choice between operating with open-source models or closed ones can have implications on your company's future. Founders should consider the pros and cons of each, taking into account factors such as collaboration, progress speed, and knowledge sharing.
YouTube
No Priors Ep. 9 | With Perplexity AIโs Aravind Srinivas and Denis Yarats
With advances in machine learning, the way we search for information online will never be the same.
This week on the No Priors podcast, we dive into a startup that aims to be the most trustworthy place to search for information online. Perplexity.ai is aโฆ
This week on the No Priors podcast, we dive into a startup that aims to be the most trustworthy place to search for information online. Perplexity.ai is aโฆ
Based on the available data, the usage of ChatGPT in the selected countries is as follows:
1. United States: The United States accounts for 15.32% of the total audience using ChatGPT
2. India: India accounts for 6.32% of the total audience using ChatGPT.
3. Japan: Japan accounts for 3.97% of the total audience using ChatGPT.
4. Canada: Canada accounts for 2.74% of the total audience using ChatGPT.
5. Other countries: The rest of the world accounts for 68.36% of visits to ChatGPT's website.
1. United States: The United States accounts for 15.32% of the total audience using ChatGPT
2. India: India accounts for 6.32% of the total audience using ChatGPT.
3. Japan: Japan accounts for 3.97% of the total audience using ChatGPT.
4. Canada: Canada accounts for 2.74% of the total audience using ChatGPT.
5. Other countries: The rest of the world accounts for 68.36% of visits to ChatGPT's website.
๐โโ๏ธ How to Play Long Term Games:
Systems > Goals
Discipline > Motivation
Trust > Distrust
Principles > Tactics
Writing > Reading
Vulnerability > Confidence
North Stars > Low Hanging Fruit
Trends > News
Habits > Sprints
Questions > Answers
Problems > Solutions
People > Projects
Systems > Goals
Discipline > Motivation
Trust > Distrust
Principles > Tactics
Writing > Reading
Vulnerability > Confidence
North Stars > Low Hanging Fruit
Trends > News
Habits > Sprints
Questions > Answers
Problems > Solutions
People > Projects
AI๊ฐ ๊ฒ์์ ์ ์๋ถํฐ ๊ฒ์์ UI/UX๊น์ง ๋ง์ ๋ถ๋ถ์ ๋ณํ์์ผ๋์ ๊ฑฐ๋ผ๊ณ ์๊ฐํฉ๋๋ค.
์ง๋ ๋ช๋ ๊ฐ AI ๋ชจ๋ธ์ ์์ฒญ๋ ์๋๋ก ๋ณํํด์๋๋ฐ์. ๊ฐ์ฅ ์ต์ ์ AI ๋ชจ๋ธ์ ๋ฐ์ ์ญ์ฌ์ ์์ผ๋ก ์์๋๋ AI ์ฐ๊ตฌ์ฃผ์ ๋ฅผ ๋ฐํ์ผ๋ก ๋ฏธ๋์ ๊ฒ์์ ์์ํด๋ด ๋๋ค.
Stable Diffusion ๋ชจ๋ธ์ด ๋น ๋ฅด๊ฒ ํ์ ํ๋ฉด์, ๊ฒ์ ์ํธ์ ๊ด๋ จํด์ ๋ค์ํ ์คํ์ด ์ด๋ฃจ์ด์ง๊ณ ์์ต๋๋ค. ๊ฒ์ ์ํธ๋ฅผ ๊ธฐํํ๊ณ ๊ฐ๋ฐํ๋ ๊ณผ์ ์์ AI๋ฅผ ์ ์ฌ์ฉํ ํ๋ก์ธ์ค๋ ๋ญ๊น์?
์ด ๋๊ฐ์ง ์ง๋ฌธ์ ๋ํด์ ๊ถ๊ธ์ฆ์ด ์๊ธฐ์ จ๋ค๋ฉด ์๋ ๊ตฌ๊ธํผ์ ์์ฑํด์ฃผ์ธ์ ๐
https://forms.gle/RFJjwqELL9juekP66
์ง๋ ๋ช๋ ๊ฐ AI ๋ชจ๋ธ์ ์์ฒญ๋ ์๋๋ก ๋ณํํด์๋๋ฐ์. ๊ฐ์ฅ ์ต์ ์ AI ๋ชจ๋ธ์ ๋ฐ์ ์ญ์ฌ์ ์์ผ๋ก ์์๋๋ AI ์ฐ๊ตฌ์ฃผ์ ๋ฅผ ๋ฐํ์ผ๋ก ๋ฏธ๋์ ๊ฒ์์ ์์ํด๋ด ๋๋ค.
Stable Diffusion ๋ชจ๋ธ์ด ๋น ๋ฅด๊ฒ ํ์ ํ๋ฉด์, ๊ฒ์ ์ํธ์ ๊ด๋ จํด์ ๋ค์ํ ์คํ์ด ์ด๋ฃจ์ด์ง๊ณ ์์ต๋๋ค. ๊ฒ์ ์ํธ๋ฅผ ๊ธฐํํ๊ณ ๊ฐ๋ฐํ๋ ๊ณผ์ ์์ AI๋ฅผ ์ ์ฌ์ฉํ ํ๋ก์ธ์ค๋ ๋ญ๊น์?
์ด ๋๊ฐ์ง ์ง๋ฌธ์ ๋ํด์ ๊ถ๊ธ์ฆ์ด ์๊ธฐ์ จ๋ค๋ฉด ์๋ ๊ตฌ๊ธํผ์ ์์ฑํด์ฃผ์ธ์ ๐
https://forms.gle/RFJjwqELL9juekP66
Google Docs
AGI Town in Seoul 4ํ์ฐจ(6/23 ๊ธ) ๋ฐํ์๋ฃ ์ ์ฒญ
I don't have to check hacker news on a daily basis anymore! Thanks for the service!
https://share.snipd.com/show/a7f48397-d9ed-458a-9bda-51b504acddee
https://share.snipd.com/show/a7f48397-d9ed-458a-9bda-51b504acddee
Snipd
Hacker News Recap
A podcast that recaps some of the top posts on Hacker News every day. This is a third-party project, independent from HN and YC. Text and audio generated usingโฆ
What era do we live in?
A wide range of AI tasks that used to take 5 years and a research team to accomplish in 2013, now just require API docs and a spare afternoon in 2023.
Not a single PhD in sight. When it comes to shipping AI products, you want engineers, not researchers.
Microsoft, Google, Meta, and the large Foundation Model labs have cornered scarce research talent to essentially deliver โAI Research as a Serviceโ APIs. You canโt hire them, but you can rent them โ if you have software engineers on the other end who know how to work with them. There are ~5000 LLM researchers in the world, but ~50m software engineers. Supply constraints dictate that an โin-betweenโ class of AI Engineers will rise to meet demand.
Fire, ready, aim. Instead of requiring data scientists/ML engineers do a laborious data collection exercise before training a single domain specific model that is then put into production, a product manager/software engineer can prompt an LLM, and build/validate a product idea, before getting specific data to finetune.
Letโs say there are 100-1000x more of the latter than the former, and the โfire, ready, aimโ workflow of prompted LLM prototypes lets you move 10-100x faster than traditional ML. So AI Engineers will be able to validate AI products say 1,000-10,000x cheaper. Itโs Waterfall vs Agile, all over again. AI is Agile.
A wide range of AI tasks that used to take 5 years and a research team to accomplish in 2013, now just require API docs and a spare afternoon in 2023.
Not a single PhD in sight. When it comes to shipping AI products, you want engineers, not researchers.
Microsoft, Google, Meta, and the large Foundation Model labs have cornered scarce research talent to essentially deliver โAI Research as a Serviceโ APIs. You canโt hire them, but you can rent them โ if you have software engineers on the other end who know how to work with them. There are ~5000 LLM researchers in the world, but ~50m software engineers. Supply constraints dictate that an โin-betweenโ class of AI Engineers will rise to meet demand.
Fire, ready, aim. Instead of requiring data scientists/ML engineers do a laborious data collection exercise before training a single domain specific model that is then put into production, a product manager/software engineer can prompt an LLM, and build/validate a product idea, before getting specific data to finetune.
Letโs say there are 100-1000x more of the latter than the former, and the โfire, ready, aimโ workflow of prompted LLM prototypes lets you move 10-100x faster than traditional ML. So AI Engineers will be able to validate AI products say 1,000-10,000x cheaper. Itโs Waterfall vs Agile, all over again. AI is Agile.
์๋ก์ด ๊ฒ์ด ๋ฑ์ฅํ๋ฉด ๊ทธ ๋๊ตฌ๋ ์ ๋ฌธ๊ฐ๊ฐ ๋ ์ ์๋ ์๊ธฐ๊ฐ ์์ต๋๋ค. ๊ทธ์ ๊ด์ฌ ์๋ ์ฌ๋๋ค๋ง ๊ด์ฌ์ ๊ฐ๊ณ ๊ฐ์ง๊ณ ๋๋ฉฐ ์๋ก ์ด์ผ๊ธฐํ ๋ฟ์
๋๋ค. ํ์ง๋ง ๊ฒฐ๊ตญ์๋ ๊ทธ ์ผ์ด ์ฑ์ํด์ง๊ณ ๊ทธ ์ฐฝ์ด ๋ซํ๋๋ค. ์ง์
์ฅ๋ฒฝ์ด ํจ์ฌ ๋์์ง ํ์๋์.
๋น์ ์ AI๋ก ์ ํํ๊ธฐ ์ํด ๋๋ฌด ๋์ง ์์์ต๋๋ค.
https://www.latent.space/p/not-old
๋น์ ์ AI๋ก ์ ํํ๊ธฐ ์ํด ๋๋ฌด ๋์ง ์์์ต๋๋ค.
https://www.latent.space/p/not-old
www.latent.space
You Are Not Too Old (To Pivot Into AI)
Everything important in AI happened in the last 5 years and you can catch up
AI x Design: https://www.figma.com/blog/ai-the-next-chapter-in-design/
ํน์ Design ์ชฝ ์ปค๋ฆฌ์ด๋ฅผ ๊ฐ์ ธ๊ฐ๊ณ ์๋ ๋ถ๋ค์ค ์ค๋ ฅ๊ณผ ๊ด์ฌ ๋๊ฐ์ง๊ฐ ๋ค ์๋ ์ง์ธ ๋ถ๋ค์ด ์์ผ์ค๊น์?~ ใ ใ
5๋ช ์ ๋๋ง ๋ชจ์ฌ๋ ์ฌ๋ฐ๋ ์ด์ผ๊ธฐ ๋ง์ด ํ ์ ์์ ๊ฒ ๊ฐ์๋ฐ์!
ํน์ Design ์ชฝ ์ปค๋ฆฌ์ด๋ฅผ ๊ฐ์ ธ๊ฐ๊ณ ์๋ ๋ถ๋ค์ค ์ค๋ ฅ๊ณผ ๊ด์ฌ ๋๊ฐ์ง๊ฐ ๋ค ์๋ ์ง์ธ ๋ถ๋ค์ด ์์ผ์ค๊น์?~ ใ ใ
5๋ช ์ ๋๋ง ๋ชจ์ฌ๋ ์ฌ๋ฐ๋ ์ด์ผ๊ธฐ ๋ง์ด ํ ์ ์์ ๊ฒ ๊ฐ์๋ฐ์!
Figma
AI: The Next Chapter in Design | Figma Blog
AI is more than a product, itโs a platform that will change how and what we designโand who gets involved.
We need to understand function call Open AI recently announced.
https://www.latent.space/p/function-agents#details
https://www.latent.space/p/function-agents#details
www.latent.space
Emergency Pod: OpenAI's new Functions API, up to 75% Price Drop, 4x Context Length (w/ Simon Willison, Riley Goodside, Roie Schwaberโฆ
Listen now | Leading AI Engineers from Scale, Microsoft, Pinecone, Huggingface and more convene to discuss the June 2023 OpenAI updates and the emerging Code x LLM paradigms. Plus: Recursive Function Agents!
Wow he is earning meony $1 mrr bwith two ai services.
๐ธ http://PhotoAI.com $62K MRR
๐ผ http://InteriorAI.com $52K MRR
๐ธ http://PhotoAI.com $62K MRR
๐ผ http://InteriorAI.com $52K MRR
Photo AI
AI Video Generator & Image Generator by Photo AI
Generate photorealistic images and videos of people with AI. Take stunning photos of people with the first AI Photographer! Generate photo and video content for your social media with AI. Save time and money and do an AI photo shoot from your laptop or phoneโฆ
I found GitHub to be the best organizer of AI-related newsletters and podcasts. Eureka!!!
https://github.com/swyxio/ai-notes/blob/main/Resources/Good%20AI%20Podcasts%20and%20Newsletters.md
https://github.com/swyxio/ai-notes/blob/main/Resources/Good%20AI%20Podcasts%20and%20Newsletters.md
GitHub
ai-notes/Resources/Good AI Podcasts and Newsletters.md at main ยท swyxio/ai-notes
notes for software engineers getting up to speed on new AI developments. Serves as datastore for https://latent.space writing, and product brainstorming, but has cleaned up canonical references und...