@karpathy: We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: W…
Andrej Karpathy
@karpathy
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
💬 1.2K · 🔁 5.5K · ❤️ 40.9K · 👁 4.4M
7:37 AM · Oct 2, 2026 · View on X
Replies (72)
Elvis @elvissun
elvis @omarsar0
Yup.
▶
David Strausser @dstrausser83
Ask your LLM to explain things in simplified English. I ask clients to explain requirements in simplified English and get 40-page PDFs. #AI #Tech #Consulting
Erick P. @bondsnbourbon
I followed your suggestion and created a 3b1b-style explainer video on stochastic calculus. It reminded me of my first encounter with the topic in my professional education.
I truly appreciate your insights and this new approach to LLMs.
I wish I had access to such technology
▶
thomas hansen @hansent
the explainer videos are soo good! seems like a gold mine for training data for video understanding and generation. @karpathy how long until models can just output in pixels space directly do you think?
Fanatomica @Numerical_Art
A controlled language is a charming way to put a tie on the stochastic parrot.
Grady Booch @Grady_Booch
"Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand."
Hm, there's this thing called the UML...
Kevin @Kevinduan2014
Thanks for the tips! Will try the first. I have similar skills for the rest. Worked well
Eric Brown @Eric_F_Brown
I really like the ASD format suggestion. Really useful when you are trying to compare which open source LLMs to use for various tasks on your system as you doing some tuning.
Leif-Erik Hvide @leiferik
HTML explanations with diagrams is a huge unlock on clarity
yvo de rooij @yvoderooij
@grok what is 3b1b video explainer? Example?
Reply2Future @Reply2Future
I will ask AI to explain things to me using UML diagrams.
Adem Vessell Arkhē @AdemVessell
Thank you very much indeed.
Currently I'm engaged in a competition for machine-learning research to estimate a neural network’s average output accurately using much less computation.
Truthfully it's so fun and directly connected to work I've been crashing around in for months
Draxx @Draxych
Four 'but even better' in one post, and the conclusion is still 'trust the output you didn't read.
Supratik Bhattacharya @SupratikResearc
with diagrams, the issue I see is LLMs may invent some weird terminologies that it has abstracted out of the content.
Thomas Sheil @Thomas_Sheil_
What comes after explainer videos? Livestreams? 👀
John Miller @JYMiller_data
Do I understand you saying that rewriting our rules / md files in this vocab will possibly improve application?
Muhammad Ali @AliFoundry
Tried the diagrams and HTML approach, it is much easier to understand instead of reading long paragraphs
🍋 Antoine Mersch 🍋 @antoine_mersch
a wrong 3b1b style video convinces faster than a wrong paragraph though, the nicer the format the harder the oversight
Mugisho @pacymugisho
Been using STE100 since opus 5 came out. It helps a lot to understand what the LLM is trying to say or overly complicate. Will be trying this on 5.5
sager @LaSager
how can LLMs be so good in general, but then we need to be experts at getting them to explain things clearly to us?
it seems like a big flaw/gap
VS @saravagi
@grok what is 3b1b style of video?
Chris Dukes @clduab11
Interesting; thanks @karpathy ... I used a style of a military/JSOC comms officer to accomplish much the same thing, but given this a LOT more standardized, def willing to give this a go!
rishav @rishavvk
on this note i can strongly vouch for vendoring the ASD-STE100 and adapting it to your preferences
Spyros D @sdoulgeridis
Very interesting and useful tip! And in no way disagreeing with it. But every approach has trade offs, so…
it makes me think… we make the language dumber. We reduce this way ambiguity, but is this going to affect us too? Since currently the bigger generators of text are not
rye @RaviYenugu
Have Claude do images please!
Mickey Builds AI @mickeyxiao5
@grok Summarize and translate into Chinese
Lucas Alvarez Lacasa @lucas_alvarez91
Do we already have a formal specification for this I can paste in my AGENTS.md?
Maxluria @Maxluria3000
I always find an interesting way to interact with LLMs after your posts/videos. Thank you!
Nagle @nagle182
It'll be great when some of these tactics get applied to the education system.
At any point of confusion in any subject, a student will be able to get clarity in whatever medium works best for them.
Kenneth Lau @kennethlau12393
html output gets you tables and sliders for free which plain text never will
John Childress @jachildress
I switched to html output a while back and it had been super helpful. Still using markdown for the agents but for myself the html was much easier to read.
Yug Marwaha @yugmarwaha
the html files are really helpful for understanding. I don't use animations quite a lot for now, but I'll definitely ask the llm to explain in ASD-STE100, as sometimes it just gets hard to understand what it says.
Mark Olech @MarkOlech
Simplified technical English had been a game changer for me. I like the 80/20 rule idea in that, will give it a try.
Harshit Chauhan @harshit_928
I tired using ASD-STE100 method but it didn't made much difference to me as it seemed both the responses are almost same to me. I am not well versed with ASD-STE100. Can anyone tell me what difference you get using it with LLM models.
I have attached a sample of mine for
iOS Vs Android memes @iosvsand
Hey, I was just casually doing something related to it using ChatGPT, and I saw this in ChatGPT’s response. ----> ..But there is a fascinating catch: the infographic itself contains at least one factual error. In panel D it marks “approximately” as not approved and suggests
Reem @reemixed
on diagram i ask it to use escalidraw haha makes for great documentation and rfc lol
Manan @manan_suneja
thanks been experimenting with explainer video and will try to push the boundaries of what it can do.
Also 💯 on 3b1b style video - didn't think of it that way
Harvey Vogelmann @HarveyVogelmann
Having them make detailed flow charts and interactive movies of the flow charts of all the moving parts is really helpful too
Alik Huseynov @alik_huseyn0v
Someone did it 3 months ago
Jamie Gale @JamieGale7
asd-ste100 as a way to force models to explain clearly is clever 🧠
curious how many plant specs would get shorter if we made the model write them that way...
Riad H @Ri4d_H
Using a controlled language spec to make LLMs write clearer is a neat trick - simple English as a prompt constraint makes a lot of sense
Tim B @smooth_tim_
ASD-STE100 is an anti -pattern that leads to staccato sentences. When reading them, their choppy cadence quickly overloads the working memory and hinders your comprehension.
Makes sense given allegedly we have only 3-6 slots in our memory before the buffer flushes.
WaveRideASI @Manonymous20
Brilliant post! I’m playing around with the explainer videos idea since a few days now similar to how you wrote it here. This will be big
David Sweet @phinance99
Nice. Also: Try asking for a technical report as a PDF via TeX. They look beautiful, math is ofc excellently rendered, and TeX has a variety of great tools the agent can use to present code, diagrams, plots, etc.
Amir Fish @amir_fish
Agree. I use as standard video explainers . Open sourced :
Tuwang @TuWangZ
amazing. this describes my entire user journey since I started using Claude over the past 1 year ;)
Asmir @asmirkn
the html one is real for me. codex made me a swipe page for 400 of my old asks and i got through them way faster than any doc
Matt Kaan @mattkaanai
There's nothing ELI5 skill cannot teach you.
Heikki Hämäläinen @hamalainenhe
Thinking is easy - keeping understanding is hard - have tried everything else except ASD-STE100 will try during next weekend session
Renga B @BashyamBashyam
@grok explain this in ASD-STE100 way
zzzSam @SequoiaMaple
Just open Eleven Labs or whatever website in Chrome and Claude does their thing. It’s the best.
suhaas @0x5uh445
i always ask for a one pager that can be presented to 5yo kids 😅
The Italian Geek Dad @newbreedofgeek
So I can ask for an entire custom video to explain the thing I was too embarrassed to ask twice? Sold.
Brian Kearns @Bkearns123
For other newbies: I just pasted his post into grok and asked to use when appropriate
Maximiliano Rodríguez @maxirodr_
I didn't know about the STE100. Did you try asking for it in another language? The standard is designed only for English, and I'm curious if the model still respects the rules in Spanish.
Sharique @shariquesha
The reverse also holds: write the prompt in ASD-STE100 to get the output you want, and feed diagrams or images in to generate the image or text out.
Gonzalo Cordova @gonzalo_io
for diagrams I highly recommend the Excalidraw integration
it's soooo good
and with Claude you can even see the diagrams inline:
Bones @SkypirateBones
No offense but you would have to be almost as intelligent as the AI to understand all that. Which coming from the brilliant Andrej is no surprise 😁
Nathan Zylbersztejn @nathan_zyl
You’re at Anthropic. Can you get Claude to write is STE by default so we don’t need to?
Milan Nankov @milannankov
Do tou think the LLM Wiki is still needed/helpful?
baris @GBminA
You’ve just used up your weekly limit in a few hours; you’ll have to wait five days 😁
tudorbuildsai @tudorxjh
bruh somebody made a video about this a while ago and people started using it, now Karpathy approves it…yeah gonna implement this too
Renato @Nakarena
Cool, this is exactly what I’ve been doing since Opus 5.5 came out! The videos and diagrams go a long way.
Forha @Forhanvv
what about evaluating their own performance on tasks like this?
Artem Milinchuk 👨🌾 @artem_milinchuk
If Karpathy says to use ASD then I'm definitely using ASD
Harlan A Nelson @harlananelson
If you trained them to output using quarto that would be great.
M. Adel Alhashemi @malhashemi90
I think a shout out to @mattpocockuk and @dexhorthy is needed here, I am not sure if you saw their posts or not but the first person I heard talking about ASD-STE100 was Matt and the first who made things more visual on large scale was `/show-me` by Dex
xdm99 @xdm99_eth
Good, because ASD-STE100 approach was discovered by @egggdev few months ago, mentioned here at X in 1st of August, and it works like a charm.
Luis Nell @originell
For coding, mixing ASD-STE100 with Googles Developer Documentation Style Guide improves explanations quite.
Joe Gardiner @thejoegardiner
Good 👍 tips thanks 🙏…
Try this with any codebase:
“Build me an interactive visual map of this project. Show what’s built, feature maturity, logical groupings, test coverage, design cohesion, missing capabilities and recommended next steps. Use the repo as the source of truth
Sonath @Sonath_AI
delivering this thesis as a wall of text is either the best or worst argument for it
Comments (0)
No comments yet