Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

Andrej Karpathy says when AI agents fail, it's usually a skill issue, not a capability issue You didn't write good enough instructions, didn't set up the right memory tool, or didn't parallelize correctly "the real shift is working in macro actions" One does research, one writes code, one plans,... show more

Haider.

69,544 subscribers

369,573 просмотров • 6 месяцев назад •via X (Twitter)

Наука и технологии

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

karpathy on Dwarkesh in Oct. '25: > AI agents don't work > They are not intelligent enough > They are not multi-modal enough > They can't do computer use karpathy on No Priors in March '26: > AI agents fail but it's a skill issue for the humans > You didn't give them the right instructions AI agents absolutely hit an inflection point in December of 2025 and there's no looking back.

karpathy on Dwarkesh in Oct. '25: > AI agents don't work > They are not intelligent enough > They are not multi-modal enough > They can't do computer use karpathy on No Priors in March '26: > AI agents fail but it's a skill issue for the humans > You didn't give them the right instructions AI agents absolutely hit an inflection point in December of 2025 and there's no looking back.

Dan McAteer

136,855 просмотров • 6 месяцев назад

New Andrej Karpathy interview Says AI agent failures stem from user skill, not model capability. Poor instructions cause errors. He suggests delegating 20-minute macro actions like coding and research to parallel agents and reviewing their work. --- "I think everything, like so many things, even if they don't work, I think to a large extent you feel like it's a skill issue. It's not that the capability is not there; it's that you just haven't found a way to string together what's available. Like, I didn't give good enough instructions to the agents in the file, or whatever it may be. I don't have a nice enough memory tool that I put in there, or something like that. So, it all kind of feels like a skill issue when it doesn't work to some extent. You want to see how you can parallelize them, and you want to be a 'Pierce tender,' basically. Pierce famously has a funny photo where he's in front of lots of these Codex agents behind the monitor. They all take about 20 minutes if you run them correctly and use high effort. You have multiple—you know, 10 or 20—pull requests checked out. It's just like you can do much larger macro actions. It's not just, 'Here's a line of code, here's a new function.' It's like, 'Here's a new functionality, delegate it to agent one. Here's a new functionality that's not going to interfere with the other one, give it to agent two.' Then, you try to review their work as best as you can, depending on how much you care about that code. You look for these macro actions that you can manipulate your software repository by. Another agent is doing some research, another agent is writing code, another one is coming up with a plan for some new implementation. Everything just happens in these macro actions over your repository. You're just trying to become really good at it and develop a muscle memory for it. It's very rewarding when it actually works, but it's also a new thing to learn. Hence, the psychosis." --- From No Priors YT channel (link in comment)

New Andrej Karpathy interview Says AI agent failures stem from user skill, not model capability. Poor instructions cause errors. He suggests delegating 20-minute macro actions like coding and research to parallel agents and reviewing their work. --- "I think everything, like so many things, even if they don't work, I think to a large extent you feel like it's a skill issue. It's not that the capability is not there; it's that you just haven't found a way to string together what's available. Like, I didn't give good enough instructions to the agents in the file, or whatever it may be. I don't have a nice enough memory tool that I put in there, or something like that. So, it all kind of feels like a skill issue when it doesn't work to some extent. You want to see how you can parallelize them, and you want to be a 'Pierce tender,' basically. Pierce famously has a funny photo where he's in front of lots of these Codex agents behind the monitor. They all take about 20 minutes if you run them correctly and use high effort. You have multiple—you know, 10 or 20—pull requests checked out. It's just like you can do much larger macro actions. It's not just, 'Here's a line of code, here's a new function.' It's like, 'Here's a new functionality, delegate it to agent one. Here's a new functionality that's not going to interfere with the other one, give it to agent two.' Then, you try to review their work as best as you can, depending on how much you care about that code. You look for these macro actions that you can manipulate your software repository by. Another agent is doing some research, another agent is writing code, another one is coming up with a plan for some new implementation. Everything just happens in these macro actions over your repository. You're just trying to become really good at it and develop a muscle memory for it. It's very rewarding when it actually works, but it's also a new thing to learn. Hence, the psychosis." --- From No Priors YT channel (link in comment)

Rohan Paul

23,207 просмотров • 6 месяцев назад

Claude Code is no longer one AI helping you write code. It now creates full agent teams. Multiple AIs. Different roles. Shared tasks. Constant communication. All working at the same time. Like a real dev team… but automated.

Claude Code is no longer one AI helping you write code. It now creates full agent teams. Multiple AIs. Different roles. Shared tasks. Constant communication. All working at the same time. Like a real dev team… but automated.

0xMarioNawfal

62,140 просмотров • 7 месяцев назад

LSP? What is it? Why is it? How is it? No one ever asks how it is, and I didn't in the video either but it's a pretty fun one. Also, we set it up in like 3 lines of code, you'll like that :)

LSP? What is it? Why is it? How is it? No one ever asks how it is, and I didn't in the video either but it's a pretty fun one. Also, we set it up in like 3 lines of code, you'll like that :)

teej dv 🔭

23,553 просмотров • 1 год назад

jeongin said they should ask kickflip kyehoon whether jeongin really gave jisung a kiss or not 😭 🐿: no, he kissed in the hallway and you kissed me in the room 🦊: in my memory, we didn't 🐿: your memory is wrong. the one who does it doesn't know and the one who gets kissed knows

Sensitive content

jeongin said they should ask kickflip kyehoon whether jeongin really gave jisung a kiss or not 😭 🐿: no, he kissed in the hallway and you kissed me in the room 🦊: in my memory, we didn't 🐿: your memory is wrong. the one who does it doesn't know and the one who gets kissed knows

비니

335,737 просмотров • 10 месяцев назад

Today we're launching Macro 1.0 — a new office suite for the intelligence age. Macro is a unified workspace for your entire company, your team and your agents. 1. Fully open source. Built in Rust + SolidJS. Self-host if you want. 2. Designed as a single system: email, messaging, docs, tasks, agents, calls, and CRM all @-linked together in one tab. 3. Unified team-level memory for agents built over emails, team chat, tasks, PR's and CRM. Access the memory from Macro or AI or external agents via MCP. We built Macro for ourselves because we wanted a single system for our whole company. We didn't want to get slowed down by rate limits and 2FA'ing into 17 different tabs. Let's take a look at each block 🧵👇

Today we're launching Macro 1.0 — a new office suite for the intelligence age. Macro is a unified workspace for your entire company, your team and your agents. 1. Fully open source. Built in Rust + SolidJS. Self-host if you want. 2. Designed as a single system: email, messaging, docs, tasks, agents, calls, and CRM all @-linked together in one tab. 3. Unified team-level memory for agents built over emails, team chat, tasks, PR's and CRM. Access the memory from Macro or AI or external agents via MCP. We built Macro for ourselves because we wanted a single system for our whole company. We didn't want to get slowed down by rate limits and 2FA'ing into 17 different tabs. Let's take a look at each block 🧵👇

Macro

165,654 просмотров • 1 месяц назад

Introducing Memory 101 🧠 Every ai tool you use is a goldfish. brilliant for one chat, close the tab and it forgets you completely. memory is the most underbuilt layer in ai right now. so we're explaining how it actually works, one episode at a time. coming up: → rag vs memory, why most "ai memory" is just search → the 5 things real memory needs: facts, profiles, graphs, time, forgetting → why good memory has to forget things follow along!

Introducing Memory 101 🧠 Every ai tool you use is a goldfish. brilliant for one chat, close the tab and it forgets you completely. memory is the most underbuilt layer in ai right now. so we're explaining how it actually works, one episode at a time. coming up: → rag vs memory, why most "ai memory" is just search → the 5 things real memory needs: facts, profiles, graphs, time, forgetting → why good memory has to forget things follow along!

supermemory

17,235 просмотров • 3 месяцев назад

Excited to present the LLM-Council skill. Initial idea by Karpathy. I just packaged it as a skill. You can easily spin up a council of LLMs or agents via Fireworks. Watch how the new GLM-5 model "deliberates" on other LLMs' thoughts on the big question, "Can LLMs reason?" Things worth paying attention to: New open models like GLM-5 have surprisingly improved on complex reasoning and long-running agentic tasks. The AskUserQuestion tool in Claude Code came in handy to select the council and chairperson. As Andrej Karpathy puts it, it's a really interesting way to get different perspectives from LLMs, which can lead to better decision-making on whatever task you are working on. You can use it for other agentic coding use cases, like evaluation, tool building, designing, and research.

Excited to present the LLM-Council skill. Initial idea by Karpathy. I just packaged it as a skill. You can easily spin up a council of LLMs or agents via Fireworks. Watch how the new GLM-5 model "deliberates" on other LLMs' thoughts on the big question, "Can LLMs reason?" Things worth paying attention to: New open models like GLM-5 have surprisingly improved on complex reasoning and long-running agentic tasks. The AskUserQuestion tool in Claude Code came in handy to select the council and chairperson. As Andrej Karpathy puts it, it's a really interesting way to get different perspectives from LLMs, which can lead to better decision-making on whatever task you are working on. You can use it for other agentic coding use cases, like evaluation, tool building, designing, and research.

elvis

39,593 просмотров • 7 месяцев назад

Everyone wants in on memory stocks, but they're already a trillion-dollar trade. You didn't miss the trade. You just found the wrong one. The real play is the toll road, companies that collect no matter what memory does. "And the toll roads are sold out." Kris Bullock

Everyone wants in on memory stocks, but they're already a trillion-dollar trade. You didn't miss the trade. You just found the wrong one. The real play is the toll road, companies that collect no matter what memory does. "And the toll roads are sold out." Kris Bullock

Real Vision

12,690 просмотров • 2 месяцев назад

THIS IS HOW A SENIOR ENGINEER ACTUALLY SCALES THEMSELVES WITH CLAUDE CODE the biggest change with AI isn't coding faster. it's where you actually spend your time now. more detailed prompts, more code review, more planning, less typing, etc. here's the workflow: this guy has been shipping code since the days of cgi and perl. he uses a compound engineering plugin that runs 5 separate agents on every task. one brainstorms, one plans the technical implementation, one executes, one reviews, one checks different verticals. every step is documented in markdown files. it's slow and way more waiting. but the output quality is way higher because each agent is focused on one thing. then the REAL multiplier is in git worktrees if Claude Code made you 10x faster, worktrees multiplies that again depending on how many agents you can manage in parallel his team runs 4-8 Claude Code sessions at the same time across different worktrees with each one working on a separate task. the skill is managing multiple AI agents in parallel without losing track, that's the next evolution of engineering

THIS IS HOW A SENIOR ENGINEER ACTUALLY SCALES THEMSELVES WITH CLAUDE CODE the biggest change with AI isn't coding faster. it's where you actually spend your time now. more detailed prompts, more code review, more planning, less typing, etc. here's the workflow: this guy has been shipping code since the days of cgi and perl. he uses a compound engineering plugin that runs 5 separate agents on every task. one brainstorms, one plans the technical implementation, one executes, one reviews, one checks different verticals. every step is documented in markdown files. it's slow and way more waiting. but the output quality is way higher because each agent is focused on one thing. then the REAL multiplier is in git worktrees if Claude Code made you 10x faster, worktrees multiplies that again depending on how many agents you can manage in parallel his team runs 4-8 Claude Code sessions at the same time across different worktrees with each one working on a separate task. the skill is managing multiple AI agents in parallel without losing track, that's the next evolution of engineering

Om Patel

159,776 просмотров • 5 месяцев назад

AI agents are useful, but agents alone are not the endgame. One agent can answer a question, write a post, scan data, generate code or complete a task. But a business is not one task. A business has roles, memory, customers, payments, permissions, delivery, support, reputation and decisions that build on each other over time. That is why the next step is not just AI agents. It is AI Organisations. Foundry launches this August, with a 100M VANRY reward pool for early Orgs. Are you ready? Whitelist:

AI agents are useful, but agents alone are not the endgame. One agent can answer a question, write a post, scan data, generate code or complete a task. But a business is not one task. A business has roles, memory, customers, payments, permissions, delivery, support, reputation and decisions that build on each other over time. That is why the next step is not just AI agents. It is AI Organisations. Foundry launches this August, with a 100M VANRY reward pool for early Orgs. Are you ready? Whitelist:

Vanar

97,746 просмотров • 1 месяц назад

This guy literally automated his entire social media with one free AI agent. No team. No scheduling tools. No VA. Here's exactly how he did it 👇 The agent is called Hermes. It's free and open source. Here's what it's running on autopilot right now: → Scans Reddit every 24 hours for trending topics → Writes posts based on breaking news → Generates custom images with Nano Banana 2 → Posts directly to X, TikTok, and Instagram with captions and hashtags → Pulls daily analytics and flags what's working → Researches competitors and shows what's getting engagement One post already hit 1,000 views. He didn't write it. He didn't post it. He didn't even check it. Here's the part most people skip: Day one it won't be good. You give it feedback. It writes notes on what went wrong. It improves the skill file. By day 7 it's better. By day 30 it knows your voice. It's called a closed learning loop. Every task makes the next one sharper. The setup takes an afternoon: Install Hermes from GitHub. Connect a free Qwen 3.6 API key from OpenRouter. Add Firecrawl for web research. Create your first skill. Give it feedback. Let it run. That's the whole system. Free. Open source. Gets better every single day. Save this. Then go install it.

This guy literally automated his entire social media with one free AI agent. No team. No scheduling tools. No VA. Here's exactly how he did it 👇 The agent is called Hermes. It's free and open source. Here's what it's running on autopilot right now: → Scans Reddit every 24 hours for trending topics → Writes posts based on breaking news → Generates custom images with Nano Banana 2 → Posts directly to X, TikTok, and Instagram with captions and hashtags → Pulls daily analytics and flags what's working → Researches competitors and shows what's getting engagement One post already hit 1,000 views. He didn't write it. He didn't post it. He didn't even check it. Here's the part most people skip: Day one it won't be good. You give it feedback. It writes notes on what went wrong. It improves the skill file. By day 7 it's better. By day 30 it knows your voice. It's called a closed learning loop. Every task makes the next one sharper. The setup takes an afternoon: Install Hermes from GitHub. Connect a free Qwen 3.6 API key from OpenRouter. Add Firecrawl for web research. Create your first skill. Give it feedback. Let it run. That's the whole system. Free. Open source. Gets better every single day. Save this. Then go install it.

Julian Goldie SEO

27,164 просмотров • 5 месяцев назад

‼️Hijacking Anti-Zionism (w/ Norman Finkelstein)‼️ "I do believe we are in a difficult situation now because the right, is, at least on one issue, masquerading itself as the standard bearer of a leftist issue...You would not believe how many emails I get: 'what do you mean the Israelis didn't kill Kennedy?'"...They have now hijacked the issue. And I believe they can take it in places that are grounds for concern."

‼️Hijacking Anti-Zionism (w/ Norman Finkelstein)‼️ "I do believe we are in a difficult situation now because the right, is, at least on one issue, masquerading itself as the standard bearer of a leftist issue...You would not believe how many emails I get: 'what do you mean the Israelis didn't kill Kennedy?'"...They have now hijacked the issue. And I believe they can take it in places that are grounds for concern."

Briahna Joy Gray

87,077 просмотров • 5 месяцев назад

Andrej Karpathy just said what no one wants to admit: "Your agent loop is generating garbage at scale and you can't see it." 1 output looks fine. 1,000 outputs = dead entropy, dead diversity. 90% of AI companies are building demos, not products. In a 20-minute conversation, Karpathy reveals the full reasoning stack

Andrej Karpathy just said what no one wants to admit: "Your agent loop is generating garbage at scale and you can't see it." 1 output looks fine. 1,000 outputs = dead entropy, dead diversity. 90% of AI companies are building demos, not products. In a 20-minute conversation, Karpathy reveals the full reasoning stack

Ricker

314,765 просмотров • 2 месяцев назад

Polsia's product line is the real story, not the raise. Ben has no employees. The agents that ran his data room and diligence are the same ones shipping the products. An ops desk for small businesses. A tool that reads your paperwork for leaks. All of it in three languages already. The $30M didn't build the company. It confirmed one was already running.

Polsia's product line is the real story, not the raise. Ben has no employees. The agents that ran his data room and diligence are the same ones shipping the products. An ops desk for small businesses. A tool that reads your paperwork for leaks. All of it in three languages already. The $30M didn't build the company. It confirmed one was already running.

AI Highlight

10,923 просмотров • 1 месяц назад

open Claude Code type a prompt. one model does everything. congrats, you just used 20% of it meanwhile I built 30 sub-agents one reviews my code before commit one finds bugs and writes the patch one drafts cold emails from one line one triages my inbox before I wake up one logs every decision so future me knows why each agent = one markdown file in .claude/agents/ most Claude Code users don't even know this exists. all 30 files in the article. ready to copy. the article has every file. every prompt. just copy and run.

open Claude Code type a prompt. one model does everything. congrats, you just used 20% of it meanwhile I built 30 sub-agents one reviews my code before commit one finds bugs and writes the patch one drafts cold emails from one line one triages my inbox before I wake up one logs every decision so future me knows why each agent = one markdown file in .claude/agents/ most Claude Code users don't even know this exists. all 30 files in the article. ready to copy. the article has every file. every prompt. just copy and run.

Nav Toor

71,467 просмотров • 4 месяцев назад

Elon Musk’s Grok Bot turned my laptop into a mini Citadel and it has already made me over $1,600 on a live trading account. The first session made $74. A few sessions later, the total crossed $1,600. Meanwhile the research floor has produced: > 1,703 research tasks > 356 verified findings > 73 filings processed > $1,513 in research value No analyst team. No overnight staff. No 12 people waiting for instructions. Just 12 AI agents with actual jobs: > one reads filings, > one verifies the numbers, > one runs screens, > one attacks the thesis, > one watches risk, > one works overnight, and one decides what is important enough to reach me. They hand work to each other and keep going when I close the laptop. The market is closed right now. GROKSTREET is still working. Wall Street has spent decades turning information into expensive headcount. Elon just gave the rest of us agents. Full build below

Elon Musk’s Grok Bot turned my laptop into a mini Citadel and it has already made me over $1,600 on a live trading account. The first session made $74. A few sessions later, the total crossed $1,600. Meanwhile the research floor has produced: > 1,703 research tasks > 356 verified findings > 73 filings processed > $1,513 in research value No analyst team. No overnight staff. No 12 people waiting for instructions. Just 12 AI agents with actual jobs: > one reads filings, > one verifies the numbers, > one runs screens, > one attacks the thesis, > one watches risk, > one works overnight, and one decides what is important enough to reach me. They hand work to each other and keep going when I close the laptop. The market is closed right now. GROKSTREET is still working. Wall Street has spent decades turning information into expensive headcount. Elon just gave the rest of us agents. Full build below

AdiiX

194,413 просмотров • 1 месяц назад

Growing up in a volatile household teaches you one of two things. You replicate the violence or you become a conflict avoider. I became a conflict avoider. And here's what nobody tells you, conflict avoidance is also bad behavior. Violence is one extreme. Silence and suppression is the other. The goal is somewhere in the middle, where you can actually express what you like and dislike about something without the whole thing blowing up. If I'm being brutally honest about my first marriage, I didn't have the self-awareness, the communication skills, or the understanding of appropriate boundaries. I was working all the time and when things weren't working, I didn't say so. That took me a long time to learn and it cost me.

Growing up in a volatile household teaches you one of two things. You replicate the violence or you become a conflict avoider. I became a conflict avoider. And here's what nobody tells you, conflict avoidance is also bad behavior. Violence is one extreme. Silence and suppression is the other. The goal is somewhere in the middle, where you can actually express what you like and dislike about something without the whole thing blowing up. If I'm being brutally honest about my first marriage, I didn't have the self-awareness, the communication skills, or the understanding of appropriate boundaries. I was working all the time and when things weren't working, I didn't say so. That took me a long time to learn and it cost me.

Anthony Scaramucci

35,715 просмотров • 2 месяцев назад

A 28-year-old in Berlin replaced an entire software agency with 7 AI agents running on a VPS. No office. No desk. No fixed computer. Just a phone and a terminal window connected to agents already working while she sleeps. A client got quoted $74,000 and 6 weeks by a traditional agency. Her agents shipped the validated PR in 19 hours. The agency was still revising the proposal. That’s the scary part about what AI is becoming. Not “ChatGPT writes code.” Entire companies are starting to collapse into one person + automated systems. One agent maps the repo. One writes specs. Two build in parallel. One tests. One audits security. Everything loops automatically until the PR is clean. Copy: She approved the final checkpoint from her phone at midnight. 30 seconds later the system kept running without her.

A 28-year-old in Berlin replaced an entire software agency with 7 AI agents running on a VPS. No office. No desk. No fixed computer. Just a phone and a terminal window connected to agents already working while she sleeps. A client got quoted $74,000 and 6 weeks by a traditional agency. Her agents shipped the validated PR in 19 hours. The agency was still revising the proposal. That’s the scary part about what AI is becoming. Not “ChatGPT writes code.” Entire companies are starting to collapse into one person + automated systems. One agent maps the repo. One writes specs. Two build in parallel. One tests. One audits security. Everything loops automatically until the PR is clean. Copy: She approved the final checkpoint from her phone at midnight. 30 seconds later the system kept running without her.

Phantom_Defi

17,496 просмотров • 4 месяцев назад