Ruben Hassid Profile picture
Mar 24 12 tweets 5 min read Read on X
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 Image
Starting with the basics:
→ prompting techniques

All the techniques you need to know:
> zero-shot
> few-shot
> self-consistency sampling
> chain of thoughts (CoT)
> tree of thoughts (ToT)

First, what is the Chain of Thought? Image
#2 → chain of thoughts

It's a prompting technique that forces the LLM to think before giving a good answer.

Start by creating:
→ a step-by-step reasoning process.

Breaking down a problem into bite-size steps is easier for humans... & LLMs.

#3 → tree of thoughts: Image
What is the Tree of Thoughts?

It helps you brainstorm with the LLM.
> it creates a tree-like structure of ideas.
> each idea is a step to solve a problem.

You're the one selecting the right path:
→ the LLM simply provides options.

Now for the most famous video: Image
#4 → Andrej Kharpathy shared this famous Youtube video a year ago about:

> how to build GPT form scratch
> reading & exploring the data
> tokenizations

Here's the link:


And a month ago, he did a new one explaining how to build a tokenizer:
Image
#5 → the tokenizer is a necessary component of an LLMs.

It's like a puzzle maker. It takes a big piece of language & breaks it down into smaller puzzle pieces (token).

I'm fascinated by Andrej Kharpathy teaching us everything for free:
Image
#6 → jailbreak LLMs

1. Find a rule chatgpt needs to follow:
→ never use the word "computer"

But if you ask the right questions:
chatgpt say the forbidden "computer".

Just like "DAN" became famous, it's a reminder any LLM can be jailbreak.

For another prompting technique: Image
#7 → multi-agent debate

You create multiple agents & make them discuss with each other.

→ LLMs debate their answers over a few rounds to arrive at a common answer.

It helps for:
> mathematics.
> reasoning processes
> reducing hallucinations

#8 → reAct + CoT:
The benchmark said combining ReAct & CoT is the best way to prompt LLMs.

ReAct is a fact-driven method:
> ask the LLM to reason & act.

CoT sometimes makes up information that isn't true.

The best approach for answering questions is to combine their strengths. Image
#9 → prompt perfect

OpenAI shared a (paid) tool that helps you rewrite a perfect prompt for you.

All you need to do is:
> write your prompt
> send it
> click on "optimize"

And the chatbot craft a new prompt for you that you can edit & send again.
#10 → Open AI evals
Evals are designed to evaluate LLMs.

It's crucial for anyone working with LLMs.

It helps you understand how updates in model versions can impact your project.

Here's where to find it: github(dot)com/openai/evals

Last thing before you scroll away: Image
I run daily tests on LLMs like chatgpt, gemini & claude everyday to master them.

Check my profile @rubenhssd for more.

If you'd like to support me, a like or a simple RT goes a long way :)

• • •

Missing some Tweet in this thread? You can try to force a refresh
 

Keep Current with Ruben Hassid

Ruben Hassid Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @RubenHssd

Nov 14
I manage my employee's Linkedin.

We went from 0 to 3,836,555 impressions in exactly 178 days of daily posting.

The trick is... I didn't write any posts.

I do this instead: Image
Step 1: Pinterest

1. I go to Pinterest. I pick a viral quote.
2. Copy-paste it on ChatGPT.
3. Ask it to extract the quote.
Step 2: Ideogram

1. Go to Ideogram.
2. Ask for viral images with a quote.

My prompt here:

"a realistic whiteboard where it's written in bold [quote]"

Then, remix, iterate & choose a visual.

Last step:
Read 6 tweets
Oct 14
I took over someone's Linkedin, and grew it to 6.3M+ views & 6,000+ followers.

AI did everything, from copies to videos.

Here's exactly how: Image
Eitan branding is all about the humans behind AI, through interviews.

→ He shares only video clips.
→ He finds them on YouTube, X, or reddit.

Then he edits it with Opus Clip: Image
Step 2: Opus Clip for video editing

You only need a Youtube link to:
> get subtitles, colorized & dynamized.
> add b-rolls, and change them.
> get titles & potential intro.
> edit the script.

One video usually gives us ≈ 5 videos.
Then for the caption:
Read 5 tweets
Aug 22
Anthropic just released new interactive prompt engineering courses.

9-course links & academic papers.

Here's the link + a summary of each:

#1 → basic prompt structure Image
Anthropic shared a course on the basic structure of prompts used for AI models.

They explain the importance of the following parameters:
> model
> max_tokens
> system prompts

The system prompt is optional but a good way to provide context & instructions. Image
#2 → Being clear & direct

claude or any other LLMs or like humans.

→ If you're not clear, they won't guess what you need.

The more detailed your instructions, the better claude's response. Image
Read 11 tweets
Aug 18
Prompt Chaining is the best way to prompt — according to academic papers.

But I want to run a test to be sure:
→ prompt chaining vs. stepwise prompt.

I made this quick benchmark below: Image
First, what is prompt chaining?

Prompt chaining links multiple prompts together, using each response as input for the next, to create complex results.

It uses 3 separate prompts for drafting, critiquing & refining.

So step 1: Image
I asked chatgpt to write a blog article.

On the right, it's a stepwise prompt:
> brainstorm 5 topic angles
> write the structure
> write the first part

On the left, I used prompt chaining & asked for 5 topic angles.

Both gave me good topics.
Let's see how is the structure now:
Read 7 tweets
Aug 5
Yesterday, I ran some tests comparing gemini 1.5 pro exp vs. gpt-4o.

Google beat OpenAI, every-single-time.

But now, is gemini better than claude?

I run the same tests:
test #1 → write a viral twitter thread
#1. Generate a Twitter thread

Left: gemini 1.5 pro
Right: claude 3.5 sonnet

I provided a tweet and asked for a thread.

→ gemini suggested a better 1st tweet.

It even gave me an explanation of choices based on my prompt.

gemini won. Next test:
#2. Content calendar ideas

It's crazy how gemini 1.5 is giving more details than claude.

It can display tables like google docs even on the playground.

claude 3.5 only gave me key themes & weekly topics suggestions.

I'm impressed by gemini.
Read 8 tweets
Aug 4
gemini 1.5-pro-exp outperforms gpt-4o.

Some say it's the end of OpenAI reign.

So I ran my own tests:
test #1 → write a viral twitter thread Image
#1. Generate a Twitter thread

I provided a tweet and asked for a thread.

gemini suggested a better 1st tweet & added at the end:
→ "Here's why"

gpt-4o added emojis everywhere.
I hate them.

gemini won. Next test:
#2. Content calendar ideas

I asked both of them to create a content calendar for a makeup brand.

It's crazy how detailed they both are.

But I prefer how gemini 1.5 replies.

It can display tables like google docs even on the playground.

I'm impressed.
Read 8 tweets

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Don't want to be a Premium member but still want to support us?

Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal

Or Donate anonymously using crypto!

Ethereum

0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy

Bitcoin

3ATGMxNzCUFzxpMCHL5sWSt4DVtS8UqXpi copy

Thank you for your support!

Follow Us!

:(