Partial day — headline is the last complete day (2026-10-13).
twitter
tracked since 2026-10-03T06:25:25.881553+00:00
conclusion
Measured mentions per day
day (UTC)
mentions
2026-09-27
0
2026-09-28
0
2026-09-29
0
2026-09-30
0
2026-10-01
1
2026-10-02
7
2026-10-03
1
2026-10-04
5
2026-10-05
7
2026-10-06
0
2026-10-07
1
2026-10-08
0
2026-10-09
0
2026-10-10
0
2026-10-11
1
2026-10-12
4
2026-10-13
0
2026-10-14
0
derived from
karpathy
word-boundary match over title + text, query='timeline' · window
2026-09-27 → 2026-10-15
· 27 matching posts
evidenceShowing 20 newest of 27
•
twitterlevelsio1dpositive
A very heated debate yesterday about this, which shows I hit a nerve
I asked @grok to summarize it and it said people believe I am directionally correct but a bit early
Which I agree with
It might take a few more years before everyone in the world uses AI chat apps and agents as @karpathy said
And also until most …A very heated debate yesterday about this, which shows I hit a nerve
I asked @grok to summarize it and it said people believe I am directionally correct but a bit early
Which I agree with
It might take a few more years before everyone in the world uses AI chat apps and agents as @karpathy said
And also until most people understand what's actually possible with current and future AI models
But it's never bad to be early with these things and adapt before everyone else
One of the repeated misconceptions I see is that people think I mean people will ask AI agents to build software, but I don't, their AI agents will solve their problems by building software on the fly
With that a market springs up of software/data/etc for AI agents to pay for and connect to (like with MCP/APIs), but again the AI agents will try optimize to get the best service for the lowest price (free if possible) which makes it a cut-throat market, but possible to build a business in
Anyway, it's interesting times and I learnt a lot from the replies and quote tweets yesterday, thanks everyone!
> be Andrej Karpathy
> co-found OpenAI
> leave to run AI at Tesla
> come back to OpenAI, leave again
> start an AI school
> coin vibe coding
> half the internet starts shipping apps without writing code
> join Anthropic in May
> post napkin math on who actually understands AI
> 6 billion people have barely touched an L…> be Andrej Karpathy
> co-found OpenAI
> leave to run AI at Tesla
> come back to OpenAI, leave again
> start an AI school
> coin vibe coding
> half the internet starts shipping apps without writing code
> join Anthropic in May
> post napkin math on who actually understands AI
> 6 billion people have barely touched an LLM
> 1-2 billion think it's a derpy chatbot
> 20 million watch agents finish months of work from one prompt
> ~5,000 people have access to frontier systems internally
> he's one of the 5,000
> says the gap keeps getting wider
most people are still arguing about whether ChatGPT is useful.
which tier are you in?
Made a fun motion graphics video to illustrate these last two long posts from @karpathy - check it out! https://t.co/OUKacPRtvFMade a fun motion graphics video to illustrate these last two long posts from @karpathy - check it out! https://t.co/OUKacPRtvF
Best accounts to follow from each frontier lab to stay constantly up to date
Anthropic
@karpathy -must-follow account for AI; recently joined Anthropic
@bcherny- Claude Code creator, always shares great tips
@trq212 also a Claude Code developer; writes amazing articles on CC
OpenAI
@polynoamial- works on reasoning re…Best accounts to follow from each frontier lab to stay constantly up to date
Anthropic
@karpathy -must-follow account for AI; recently joined Anthropic
@bcherny- Claude Code creator, always shares great tips
@trq212 also a Claude Code developer; writes amazing articles on CC
OpenAI
@polynoamial- works on reasoning research, shares a lot of technical details
@gabriel1- Sora developer, great career path
@jxnlco - works on dev experience, shares a lot about Codex
Google AI
@OfficialLoganK- all the major Google Gemini and AI Studio updates
@ammaar- product and design; shares great things about vibe-coding in Google AI Studio
@fofrAI- cool use cases for generative models
Cursor
@leerob- the loudest voice behind Cursor updates
@ericzakariasson- shares great insights on using Cursor
@mntruell - Cursor’s CEO; major releases and usage updates
xAI
@ai_explorer25- ex-MSFT, covers all ai content and free resources
@milichab - recently joined xAI, shares updates on Grok
@skcd42 - also covers major Grok releases
Here's Opus 5.5 with a visual reading of Karpathy's latest tweet. Watch on 2x speed. (one prompt btw) https://t.co/65t6muXyMNHere's Opus 5.5 with a visual reading of Karpathy's latest tweet. Watch on 2x speed. (one prompt btw) https://t.co/65t6muXyMN
Andrej Karpathy (ex-OpenAI, ex-Tesla) just released a prompt that turns Opus 5.5 into a harness that actually finishes the any job.
It's f*cking unreal...
send it to Opus 5.5, and thank me later. Then read the full guide for Opus 5.5 Harness Engineering. https://t.co/HGA34MAE56Andrej Karpathy (ex-OpenAI, ex-Tesla) just released a prompt that turns Opus 5.5 into a harness that actually finishes the any job.
It's f*cking unreal...
send it to Opus 5.5, and thank me later. Then read the full guide for Opus 5.5 Harness Engineering. https://t.co/HGA34MAE56
Dr Karpathy tweeted about this yesterday... First time I heard about "ASD-STE100".
Asked Claude to explain what it is in an HTML...
Its really cool that this style of writing has a name! https://t.co/v3dIKe0CAWDr Karpathy tweeted about this yesterday... First time I heard about "ASD-STE100".
Asked Claude to explain what it is in an HTML...
Its really cool that this style of writing has a name! https://t.co/v3dIKe0CAW
holy sh*t. Opus 5.5 + Andrej Karpathy’s output trick is the cleanest agent stack i’ve seen.
the agent does the work:
> goal + context: one specific job and everything it needs
> plan: short numbered steps with key information first
> act: code, files, browser and apps. one tester ran a 680,000 line migration in unde…holy sh*t. Opus 5.5 + Andrej Karpathy’s output trick is the cleanest agent stack i’ve seen.
the agent does the work:
> goal + context: one specific job and everything it needs
> plan: short numbered steps with key information first
> act: code, files, browser and apps. one tester ran a 680,000 line migration in under a day
> verify: tests and your rules decide, never the agent. 39 of 40 load time fixes landed
> learn: notes for the next run, with cached reads at $0.20 per million tokens
then it explains everything, Karpathy style:
STE report: strict controlled english from aerospace manuals
diagram: one picture showing what changed
HTML page: an interactive version you can inspect
explainer video: a narrated 3Blue1Brown style breakdown, still early but already working
you review the evidence. nothing merges without you.
why Opus 5.5: it puts the important information first, follows your writing rules and costs 40% less to run than Opus 5.
make your agent explain itself.
"Small Local LLMs are the future.”
Andrej Karpathy has been saying that since 2023.
your agent makes a thousand calls an hour. Maybe forty of them need a frontier model.
this PDF breaks down Karpathy's case for running that core on your own machine.
the real question isn't "is the local model as smart?"
It's "whic…"Small Local LLMs are the future.”
Andrej Karpathy has been saying that since 2023.
your agent makes a thousand calls an hour. Maybe forty of them need a frontier model.
this PDF breaks down Karpathy's case for running that core on your own machine.
the real question isn't "is the local model as smart?"
It's "which of my 1000 calls ever needed the smart one?"
read this one for what runs locally, then check out the Sonnet 5.5 guide for what runs next ⭣
karpathy is goated for suggesting to ask models to write in the ASD-STE 100 style.
it has improved the response of each model and helps so much to understand depth of something in most efficient manner.karpathy is goated for suggesting to ask models to write in the ASD-STE 100 style.
it has improved the response of each model and helps so much to understand depth of something in most efficient manner.
Every tip in @karpathy's post is gold.
I use the HTML one a lot. When I'm learning something new or trying to understand complex code, I ask the agent to build me a visual HTML page.
Here I was learning SEO and GEO. The agent turned the key ideas into an interactive page I could click through.
I picked it up faster …Every tip in @karpathy's post is gold.
I use the HTML one a lot. When I'm learning something new or trying to understand complex code, I ask the agent to build me a visual HTML page.
Here I was learning SEO and GEO. The agent turned the key ideas into an interactive page I could click through.
I picked it up faster than I would have from reading docs.
https://t.co/C4PkQd3kEX
Best accounts to follow from each frontier lab to stay constantly up to date
Anthropic
@karpathy -must-follow account for AI; recently joined Anthropic
@bcherny- Claude Code creator, always shares great tips
@trq212- also a Claude Code developer; writes amazing articles on CC
OpenAI
@polynoamial- works on reasoning r…Best accounts to follow from each frontier lab to stay constantly up to date
Anthropic
@karpathy -must-follow account for AI; recently joined Anthropic
@bcherny- Claude Code creator, always shares great tips
@trq212- also a Claude Code developer; writes amazing articles on CC
OpenAI
@polynoamial- works on reasoning research, shares a lot of technical details
@gabriel1 - Sora developer, great career path
@jxnlco - works on dev experience, shares a lot about Codex
Google AI
@OfficialLoganK- all the major Google Gemini and AI Studio updates
@ammaar - product and design; shares great things about vibe-coding in Google AI Studio
@fofrAI - cool use cases for generative models
Cursor
@leerob- the loudest voice behind Cursor updates
@ericzakariasson- shares great insights on using Cursor
@mntruell - Cursor’s CEO; major releases and usage updates
xAI
@ai_explorer25- ex-MSFT, covers all ai content and free resources
@milichab - recently joined xAI, shares updates on Grok
@skcd42 - also covers major Grok releases
I took Karpathy's advice and finally tried using Opus 5.5 to create a few bespoke explainer videos (using ElevenLabs for voice).
Trying different libraries for the video generation too (Remotion, Hyperframes, Manim).
I'm impressed thus far.
Here's some of what Opus 5.5 one-shotted (more in thread):I took Karpathy's advice and finally tried using Opus 5.5 to create a few bespoke explainer videos (using ElevenLabs for voice).
Trying different libraries for the video generation too (Remotion, Hyperframes, Manim).
I'm impressed thus far.
Here's some of what Opus 5.5 one-shotted (more in thread):
Andrej Karpathy tip for Claude Code: once Opus 5.5 writes your code, stop asking Opus to review its own blind spots and stop burning local context on PR checks
hand verification to an adversarial cloud fleet with /code-review ultra
run /code-review ultra --post
Opus 5.5 plans architecture and writes the core logic
S…Andrej Karpathy tip for Claude Code: once Opus 5.5 writes your code, stop asking Opus to review its own blind spots and stop burning local context on PR checks
hand verification to an adversarial cloud fleet with /code-review ultra
run /code-review ultra --post
Opus 5.5 plans architecture and writes the core logic
Sonnet 5.5 swarms local refactors and tests at medium effort
The Cloud Reviewer Fleet launches in an isolated sandbox, spins up adversarial verifiers, and only reports bugs at three proven gates:
→ before opening a PR: does this diff break cross-module invariants?
→ reproduction check: can an independent verifier write a failing test proving the bug?
→ pre-merge signoff: zero style nits, every reported finding reproduced in bugs.json
Sonnet 5.5 builds. Opus 5.5 ships. Cloud fleet breaks and verifies
Jev engineering is the same move one layer down: mechanical decisions that need no reasoning (diff size calculation, git base detection, AST tree filtering, skip style nits) resolve in under 16ms, so the frontier models only wake up when logic actually branches
Code on high. Swarm on medium. Verify with ultrareview.
- the full verification tree
> Opus 5.5 on high runs the local development session
> worker edits and runs local test suites
> explorer reads repo context and maps dependencies
> all local subagents on Sonnet 5.5 at medium effort
> /code-review ultra spins up an adversarial cloud fleet on Anthropic infrastructure
> independent verifiers write tests to prove each bug before reporting
> bugs.json automatically stamps verified PR comments via --post
paste the tree and this prompt into Claude Code ↓
"Configure my Claude Code workspace for cloud fleet verification:
1. Audit ~/.claude/agents and .claude/agents for subagents fitting worker and explorer:
> Give each model: sonnet, effort: medium
> Keep all local exploration out of main context
2. Lock main session effortLevel to high in ~/.claude/settings.json
3. Configure ultrareview defaults
Hit 300k followers today!
Never imagined this when I started tweeting in 2024. (I thought you had to be Andrej Karpathy to get this many followers)
Looking back, it’s been super fun:
- shared side projects like nanoGPT & llm.c runs, and Muon!
- brought a lot of users to my previous startup
- met many incredible peopl…Hit 300k followers today!
Never imagined this when I started tweeting in 2024. (I thought you had to be Andrej Karpathy to get this many followers)
Looking back, it’s been super fun:
- shared side projects like nanoGPT & llm.c runs, and Muon!
- brought a lot of users to my previous startup
- met many incredible people here IRL
I mostly just try to consistently post 1 tweet a day, whenever something comes to mind.
X is the best place for AI and tech.
Will keep yapping!
Andrej Karpathy built an agent that runs AI research autonomously. The whole setup is just 3 files
It's called autoresearch. You give Claude or Codex a small LLM training setup and walk away.
The loop:
> try one idea in https://t.co/293zgMo6Bh
> commit
> train for exactly 5 minutes
> if score improved - keep it. Wor…Andrej Karpathy built an agent that runs AI research autonomously. The whole setup is just 3 files
It's called autoresearch. You give Claude or Codex a small LLM training setup and walk away.
The loop:
> try one idea in https://t.co/293zgMo6Bh
> commit
> train for exactly 5 minutes
> if score improved - keep it. Worse - git reset
> log the result, repeat forever
The instructions literally tell the agent to never stop and never ask "should I keep going?", because the human might be asleep.
You don't touch the Python at all. You only edit program.md, the agent's instructions. Karpathy calls it the code of your research org.
There's even a simplicity rule. A tiny gain that adds 20 lines of hacky code gets thrown out. Same score with less code gets kept.
I think this pattern works way beyond LLMs. Anything with one number to beat, one file to edit and a fixed time per run can grind overnight like this.
92k stars, MIT. Runs on one NVIDIA GPU, local or rented, with community forks for Mac, Windows and AMD.
repo: https://t.co/79UWtuLyk1
Very underrated tip from @karpathy.
Just tried ASD-STE100 rule for model responses, and I found it to be surprisingly good at making explanations clearer. It can be bit too strict, so I agree that "80% of the way to ASD-STE100" feels like a better default.
A prompt you can try with your agent:
"Use ASD-STE100 as a g…Very underrated tip from @karpathy.
Just tried ASD-STE100 rule for model responses, and I found it to be surprisingly good at making explanations clearer. It can be bit too strict, so I agree that "80% of the way to ASD-STE100" feels like a better default.
A prompt you can try with your agent:
"Use ASD-STE100 as a guide for your responses. Write short sentences, use active voice, and keep terminology consistent. Relax the vocabulary rules when they make explanations awkward. Preserve technical precision and uncertainty."
@0xpili_ also made a nice skill for it:
https://t.co/UODgEeQy6z
list of 15 best AI researchers on X:
1. @karpathy — Andrej Karpathy
2. @rasbt — Sebastian Raschka
3. @AndrewYNg — Andrew Ng
4. @drfeifei — Fei-Fei Li
5. @demishassabis — Demis Hassabis
6. @GaryMarcus — Gary Marcus
7. @ai_explorer25 — AI Explorer
8. @maximelabonne — Maxime Labonne
9. @chipro — Chip Huyen
10. @i…list of 15 best AI researchers on X:
1. @karpathy — Andrej Karpathy
2. @rasbt — Sebastian Raschka
3. @AndrewYNg — Andrew Ng
4. @drfeifei — Fei-Fei Li
5. @demishassabis — Demis Hassabis
6. @GaryMarcus — Gary Marcus
7. @ai_explorer25 — AI Explorer
8. @maximelabonne — Maxime Labonne
9. @chipro — Chip Huyen
10. @ilyasut — Ilya Sutskever
11. @thsottiaux— Thibault "Tibo" Sottiaux
12. @bcherny— Boris Cherny
13. @geoffreyhinton — Geoffrey Hinton
14. @JeffDean — Jeff Dean,
15. @fchollet — François Chollet
@karpathy tried this for a week and fried my brain. it lasts a week or two then you have to switch off imo@karpathy tried this for a week and fried my brain. it lasts a week or two then you have to switch off imo