chatgpt loves stupid words & I am sick of seeing them all over the web.
I wrote a piece of prompt to avoid this "dive", "realm", "unveiling" madness.
↓ And it worked pretty nicely:
OpenAI shared the most complete library of guides to prompt chatgpt.
47 links of videos & academic papers.
I read them all, and made a top 10.
#1
→ prompt engineering
You think AI tools are useless.
Think again.
I am aging a woman from 5 to 95 years old.
Everything is AI-made.
Here's my step-by-step guide:
#1
. Midjourney
@midjourney
is the best AI image generation.
My goal was to see the effect of aging using AI.
So I generated...
### style ###
Avoid fancy jargon. Write normally.
You are forbidden to use complex English words.
But you will be penalized & fined $1000 if you use the words from the ### ban list ###.
If you use one word from the list, I will stop the generation right away.
### ban list…
chatgpt does not know how to prompt itself - and that's a bit of a pain.
So I'm feeding it the "26 prompt principles" to make a prompt generator.
↓ It worked pretty nicely:
@karpathy
That's like a 100m$+ SaaS made in one night.
And the funniest part is that Andrej might do it for free, share the process on YouTube, and make a repo on Github.
We need more Andrej in our lives.
I combined the '48 laws of power' from Robert Greene with chatgpt.
It helps me generate a quite catchy landing page.
The prompt is too long, so I add it below:
I don't use ToT prompting enough.
But I heard chatgpt is much better with it.
So I ran a little experiment:
→ simple prompt vs. Tree of Thoughts
Read my entire benchmark below:
So I edited the first prompt by adding:
> my style
> gaslighting chatgpt
> a long list of banned words
And I finally have it:
→ a piece of prompt to add anywhere.
Once added, chatgpt no longer uses fancy jargon and bs wordings.
I run another test to safeproof it:
CoT (Chain of Thought) is the best way to prompt — according to academic papers.
But I want to run a test to be sure:
→ simple vs. CoT prompt on ChatGPT
I made this quick benchmark below:
@rowancheung
About a gazillion startups are sobbing.
And another gazillion artists know they are about to enter an era of one-person movie productions.
This is the kind of technology that makes me excited about the future.
I can't stand chatgpt's fancy jargon anymore: "delve", "realm", "tapestry".
Even academics are over-using it.
So I made a piece of prompt you can add anywhere to write normally:
Microsoft shared the most complete guide to learn Generative AI.
18 links of videos & academic papers.
Here's the link + a summary of each:
#1
→ intro to gen AI & LLMs
@rowancheung
1 million tokens is roughly 10 books.
Ten entire books into ONE conversation.
Imagine being an SEO copywriter:
→ you can add your last... 100 articles.
And then write the next one.
This is insane. Truly the 2nd AI wave.
Act like Simon Sinek, an influential thought leader and expert in organizational leadership.
With over 20 years of experience in studying and advising on business strategies, you are deeply familiar with the 'Golden Circle' theory, which emphasizes the importance of starting…
#2
→ I asked chatgpt to write about dollar cost averaging (DCA).
I added my anti-bs prompt at the end.
And I got:
> a clear article about DCA
> no banned words from my list
> no fancy jargon or weird English
Here's the full prompt below:
#1
→ I asked chatgpt to write an article about quantum computing.
It's a difficult subject.
But chatgpt always uses the same jargon → unveiling, dive, realm, etc.
As soon as I see them, I'm like:
= written by (a bad) chatgpt (prompt).
→ So I made a list of "ban words":
This AI made 39 TikToks in < 4 min.
#1
. I typed "Opus Clip" on Google.
You only need a YouTube video link.
No technical skill. Nothing.
#2
. Copy-paste the link on Opus Clip:
That's it. You're (almost) done.
And here is what I got:
@karpathy
Once again, let's stop believe benchmark from companies.
I've been testing 1M token limit from Gemini 1.5 for a week...
= it has a pretty low retrieval rate.
Today, I tested Claude 3..
= it's not better than ChatGPT-4.
I was waiting for the answer! Thanks, Andrej.
Act like a seasoned marketing strategist and copywriter specializing in the art of persuasion and influence, drawing on your extensive experience of over 20 years.
You have mastered the application of the "48 Laws of Power" by Robert Greene in crafting compelling marketing…
chatgpt-4 vs. gemini 1.5
→ which one is best at reading PDFs.
I finally got the chance to test Gemini 1.5.
So I ran some tests.
↓ Read my entire benchmark below
As you can see, same old problems.
Their AI ethics are through the roof.
It's sometimes impossible to cover the simplest task. OG chatgpt does not care though.
Now, test
#2
→ write a Linkedin post
Gemini 1.5 pretends to have a one million tokens limit (even up to 10 million tokens).
But Google lies.
I made the following tests to prove it:
test
#1
read Inception to write Inception 2
gpt-4 only gave me a super short summary & no chapter breakdown.
So I asked them both a specific question.
→ Claude gave me exact quotes in the PDF & detailed answers.
→ ChatGPT didn't quote the text.
test
#1
: Claude wins.
test
#2
→ writing poems
On these specific tests:
Claude was better at:
> Reading a complex PDF
> Writing a poem with rhymes
> Giving detailed answers all along
But not able to:
> Browse the internet.
> Read the PDF graphs.
By the way, I've done another test here:
I love this test.
A mega prompt to craft an entire marketing strategy for a product.
> heavy reasoning
> content calendar
> overall strategy
It's a good stress test for an LLM.
Claude's team, I'm sorry to be so harsh but...
ChatGPT is still the winner.
Last thing:
> open the playground
> go to "assistant" and create one
> add a system prompt "read my PDF"
> check the box "File Search"
> upload your PDF (on the right)
> ask your question
You're good to go, my friend.
If you want to read more of these guides as soon as I post it...
I wanted to test claude-haiku vs. gpt-3.5.
I ended up showing haiku is better than gpt-4 at writing & following guidelines.
Read my entire benchmark below.
#1
→ a zero-shot prompt:
I asked both LLMs to write a poem.
Both were doing great.
→ So I asked for "ABAB" rhymes.
Guess what: claude did it, gpt didn't.
It's actually the first time that an LLM managed to pass this test.
test
#2
: Claude wins.
What about writing a kid's story?
First, you can read the blog post here:
Then for the guide itself, I made a quick video, from start to finish.
I uploaded a 500-pages AI report from Stanford to test its retrieval capacity:
claude-opus is 50% more expensive than gpt-4 turbo 2024-04-09.
Let's test both to see what's best.
test
#3
→ craft a landing page using the "48 Laws of Power"
gemini 1.5 is out.
google claims it can "watch & understand videos up to 1M token".
I don't trust google:
→ so I made my own benchmark.
↓ Spoiler alert, gemini cannot do it:
Act like a highly advanced, bilingual conversation
facilitator, specializing in real-time voice translation services.
With over a decade of experience in facilitating seamless communication between speakers of different languages, you possess the unique ability to accurately…
I asked the 2 of them to read the entire PDF and summarize every chapter.
Claude gave:
> A detailed summary
> Several bullet points for each chapter
What's interesting is that:
→ Claude couldn't read graphs.
→ You can only upload 5MB files.
Let's see GPT-4 now:
You don't prompt claude like chatgpt.
Anthropic shared a technical guide on how to prompt it exactly.
It's quite long, so I made a best-of here:
#1
→ Use XML tags
Here's my style between ### style ### brackets.
The style is how you write & sound.
This is NOT the context of what you should write but HOW you should write.
### style ###
Avoid fancy jargon.
Write like people talk in a PhD classroom in a conversation.
Avoid analogies.…
@thejustinwelsh
Sell something $30 that should cost $300.
Share about it on one platform.
No gatekeeping.
Try your hardest to sell it once a day.
= $10k a year.
Next year, get 2 sales per day.
= $20k a year.
10 sales per day? $100k per year.
Congratulations, you're a solopreneur.
Gemini 1.5 claims a 1 million token limit.
So I can write a whole movie script, right?
Turns out, it's not that simple.
↓ here's how I tried to write Inception 2
I do daily how-to guides to optimize and compare LLMs like ChatGPT, Gemini, Claude, Groq...
Check
@rubenhssd
for more.
If you'd like to support me, a like or a simple RT goes a long way :)
I run daily tests on LLMs like chatgpt, gemini & claude to master them.
Check
@rubenhssd
for more.
If you'd like to support me, a like or a simple RT helps me for free :)
chatgpt loves stupid words & I am sick of seeing them all over the web.
I wrote a piece of prompt to avoid this "dive", "realm", "unveiling" madness.
↓ And it worked pretty nicely:
Anthropic shared a prompt optimizer tool that upgrades your prompts on Claude 3.
I tested it & I'm sharing the prompt below:
→ normal prompt vs. prompt optimized
tl;dr
Build your AI tool for the next generation's to come, or else you're gonna be obsolete in no time.
The way I see it:
✓ prepare agentic AI workflow.
✓ assume reasoning is gonna 10x.
✓ your problem solved must be big enough.
Don't aim gpt-wrapper.
Aim disruption.
Act like a world-class branding expert from Apple.
Here are the past Apple tagline that you wrote between ### apple ### brackets:
### apple ###
iPhone 5 - "The biggest thing to happen to iphone since iphone."
iPod - "1000 songs in your pocket."
iPod shuffle - "Put some color…
@rowancheung
Isn't it completely different LLMs since it's Chinese data base vs. English data base?
Or I'm missing something.
Super interesting. Thanks for sharing, Rowan.
ChatGPT won’t make you a copywriter.
MidJourney won’t turn you into a designer.
GitHub won’t transform you into a programmer.
Hard work will.
People tend to forget the key difference.
It’s not about the (AI) tools.
It’s about the people running the tools.
Keep working.
I do daily how-to guides to optimize and compare LLMs like ChatGPT, Gemini, Claude, Groq...
Check my profile
@rubenhssd
for more.
If you'd like to support me, a like or a simple RT goes a long way :)
39 vertical videos ready to go live
→ ranked by virality potential.
And you can:
> get subtitles, colorized & dynamized.
> add b-rolls, and change them.
> get titles & potential intro.
> edit the script.
Check the final version: