Loading video...

Video Failed to Load

Go Home

Andrej Karpathy just explained the 5 shifts turning LLMs into agentic systems. 00:00 - Memory turns chat into personal AI 06:41 - Multimodal AI reads the world 16:58 - Thinking models solve harder tasks 24:51 - Search makes LLMs live 30:58 - Tools turn LLMs into workers Most people...

355,631 views • 1 month ago •via X (Twitter)

40 Comments

Anh Vũ's profile picture
Anh Vũ1 month ago

The main barrier for nontech users is not understanding concepts, but finding tools that are easy to plug together without deep coding

Lunar's profile picture
Lunar1 month ago

easy setup changes everything

Alberto Automates's profile picture
Alberto Automates1 month ago

the chatbot era is already over and most people haven't noticed. karpathy just quietly drew the line between "toy" and "coworker"

SPEKULATOR's profile picture
SPEKULATOR1 month ago

memory and tools are the easy demo; who audits stale memories and bad tool calls before they touch a customer record?

Lunar's profile picture
Lunar1 month ago

thats the real problem

SPEKULATOR's profile picture
SPEKULATOR1 month ago

yeah — the “real problem” is that the HAR looks valid while the generated code quietly

Vipul Kumar Kewat's profile picture
Vipul Kumar Kewat1 month ago

The shift from prompting models to building complete AI workflows is the biggest change happening right now. Understanding the full stack will be crucial for anyone building with AI.

noclipepe's profile picture
noclipepe1 month ago

so basically we've been using LLMs wrong this whole time

Lunar's profile picture
Lunar1 month ago

pretty much yeah

Samuel kwan's profile picture
Samuel kwan1 month ago

“tools turn LLMs into workers” is the only one already paying rent. check is simple: same session recovers a failed deploy/test log without you pasting the traceback. Cursor agent / Claude Code sometimes yes; most tool-card UIs still no.

Lunar's profile picture
Lunar1 month ago

@0xkakaakl yeah thats the real test

Samuel kwan's profile picture
Samuel kwan1 month ago

yeah — and “the real test” only lands when the agent re-grabs a failed deploy log in the same session without you pasting the traceback. Cursor agent sometimes clears it; most tool-card UIs still don’t.

I find Discounts, deals, and coupons for you's profile picture
I find Discounts, deals, and coupons for you1 month ago

andrej karpathy outlines key advancements enabling llms to function autonomously enhancing interactivity the exact product i use is linked in my bio

Lunar's profile picture
Lunar1 month ago

this looks super useful

ClariLayer's profile picture
ClariLayer1 month ago

These shifts really boost AI capabilities, shaping the future of automation

0xSlyth's profile picture
0xSlyth1 month ago

hell yes

Iron Mind's profile picture
Iron Mind1 month ago

Bookmarked that

Okarun's profile picture
Okarun1 month ago

@grok find youtube video link

그문달's profile picture
그문달1 month ago

그록빌드를 슈퍼그록 구독안의 토큰(정액형)으로 사용가능한 업무용 에이전트를 웹안에 집어넣었습니다. 그록은 정말 빠르고 일을 잘해요.

Luís Rodrigues's profile picture
Luís Rodrigues1 month ago

The biggest shift is moving from asking AI questions to building systems that can pursue goals.

Dipanshu Kushwaha's profile picture
Dipanshu Kushwaha1 month ago

That's a great breakdown! Makes a lot of sense how these pieces fit together.

mukay's profile picture
mukay1 month ago

Karpathy’s been saying this for two years, the stack just finally caught up to the idea.

dhinna ship .ico's profile picture
dhinna ship .ico1 month ago

A solid mental model for structuring agent workflows

Jef Newsom's profile picture
Jef Newsom1 month ago

I present “The Five Steps that subsequently turn them into autonomous WMDs” Robotics make them mobile. Satellite makes the connected. Innervation makes them borg. 3D printing makes them reproduce. Weapons make them skynet.

Swati Gupta's profile picture
Swati Gupta1 month ago

Unexpectedly practical, this perfectly connects where AI is heading...

Loong🐉's profile picture
Loong🐉1 month ago

the search shift (LLMs as live retrieval) is the one infra teams are under-budgeting for right now. most prod RAG assumes a static index. once the model does live search at request time, the 'vector DB moat' vaporizes — and Pinecone et al. know it.

Rizwan's profile picture
Rizwan1 month ago

memory and tools are the two that aren't solved. thinking, multimodal, search: mostly shipped. long-horizon memory and tool calls that don't silently break are what separate a demo from something people actually keep open.

Elijah Gjovig's profile picture
Elijah Gjovig1 month ago

@grok what is the YouTube link to this video? Or where can I find it online

Alex's profile picture
Alex1 month ago

How are you handling context compaction without losing important state?

Gustavo Caetano's profile picture
Gustavo Caetano1 month ago

@grok what is the youtube link for this video?

priyanka kumari's profile picture
priyanka kumari1 month ago

Worth watching, the shift toward agentic AI is becoming much clearer....

Orcun S's profile picture
Orcun S1 month ago

can you summarize for me please? @summarize_that

Leo Lin's profile picture
Leo Lin1 month ago

These shifts make models more capable, but ‘worker’ implies more than tools and memory. Agents become operational only when they can act reliably within clear authority, cost, and accountability boundaries.

Alex Foster's profile picture
Alex Foster1 month ago

memory is the quiet one here. tools make AI look capable, but memory is what makes it feel like the same system came back tomorrow and actually knows what happened yesterday

Yohaku's profile picture
Yohaku1 month ago

Thanks to Andrej Karpathy i always got something new about AI

Shaun's profile picture
Shaun1 month ago

its so funny when he says the memory in chatgpt is 'too personal' (openai literally watches him everyday lmaooo)

Godservant Bush's profile picture
Godservant Bush1 month ago

@grok when is this video from

ViceSol's profile picture
ViceSol1 month ago

One of the best ways to think about modern AI is as a system, not a chatbot. Memory, reasoning, search, and tools are becoming just as important as the underlying model

Ankur Misra's profile picture
Ankur Misra1 month ago

The stack is right: Memory → Vision → Reasoning → Search → Tools That’s what turns a chat tab into something that can work overnight. What’s still scarce isn’t the list. It’s ownership of the memory and the tools — so the loop doesn’t reset when the model or the SaaS seat does. Prompting was the old workflow. Agentic is the new one. Owned graph + approve is the layer people keep skipping.

Yudha Bhakti Nugraha's profile picture
Yudha Bhakti Nugraha1 month ago

@ecommurz I think the biggest shift is the last one: tools. Once an LLM can search, browse, execute code, and interact with external systems, it stops feeling like a chatbot and starts behaving more like a software engineer.

Related Videos