If you're a dev using AI, Codex CLI might be the most useful tool you're not using yet.

It brings OpenAI models straight into your terminal.

Here's how to prompt it effectively: 1. 3 ways to prompt Codex CLI:

- Small: quick file edits
- Medium: structured tasks using markdown
- Large: project-wide workflows with memory

OpenAI introduces its most advanced models: o3 and o4-mini.

- o3: Large, excels at complex tasks
- o4-mini: Smaller, faster, cost-effective

Designed for deep reasoning, both can use all ChatGPT tools agentically:

- Intelligent use of tools (Python, search, image analysis)
- Chaining tools for complex solutions
- Quick, well-formatted results, usually in under a minute

Key information you need to know: 1. o3 is a deep reasoning model excelling in:

- Coding and debugging
- Advanced math
- Visual problem-solving (charts, sketches, blurry photos)
- Scientific and business analysis

It achieves new state-of-the-art scores on major benchmarks like Codeforces, SWE-Bench, and MMMU.

Social media is filled with Ghibli-style AI images this week.

And it's definitely impressive.

But it raises more difficult questions:

Are we honoring artistic legacies or reducing them to mere prompts?

Where is the line between admiration and appropriation?

A few thoughts 🧵

https://twitter.com/1598914621972402181/status/1906610376273514763

Studio Ghibli's art is a philosophy.

Hand-drawn frames.

Slow pacing.

Moments of stillness.

A deep reverence for nature, humanity, and emotion.

Turning that into an instant filter misses the point.

1/5

Hugging Face has launched AI Search within @huggingface Spaces, simplifying the discovery of AI-powered apps.

Finding AI tools used to be difficult, but with Spaces Search, you can instantly discover AI solutions.

With 400K+ tools, you can explore solutions for:
- Image editing
- Speech transcription
- Virtual try-ons
- Text analysis & more

A quick breakdown 🧵👇 What are Spaces?

Hugging Face Spaces lets you explore and use AI apps without coding.

It's like an app store for AI tools where you can:
- Instantly try apps like image editing and text analysis
- Discover AI innovations globally
- Use AI with no installation needed

Search, click, and start using AI!

10 Apple Intelligence Features and How To Use Them

🧵 1. Enable Apple Intelligence

Update your device to the latest iOS, iPadOS, or macOS. Then, go to Settings > Apple Intelligence & Siri and follow the prompts to get started.

Automated Red-Teaming @OpenAI?

Previous methods struggled to combine:

- Diversity: Varied attack strategies
- Effectiveness: Successful attacks

OpenAI proposes a solution for both 🤔

🔥 Our HackAPrompt 1.0 is frequently cited!

1/8 "Diverse and Effective Red Teaming with Auto-generated Rewards and Multi-step RL" paper

TL;DR: This paper combines AI goal generation + reinforcement learning to build a more diverse and effective red-teaming system.

2/8

Is prompt engineering dead?

New studies say “No” – it’s just evolving.

Two new studies show that prompt design can make or break LLM performance.

Here are the key findings 👇
1. Prompt Structure Matters

Changing formats, such as JSON versus Markdown, can impact performance by up to 40%!

- GPT-3.5 prefers JSON.
- GPT-4 favors Markdown.

The format can leverage a model's strengths.

3 new AI tools for developers:

- Gemini AI integrates with @OpenAI APIs
- @CopilotKit and @LangChainAI CoAgents streamline LangGraph in React
- @Microsoft TinyTroupe simulates personality-driven agents

🧵 1. Gemini AI Models

Google’s Gemini AI is now accessible through the OpenAI Library and REST API.

✅ Seamless integration with tools like OpenAI APIs
✅ Faster deployment for projects utilizing cutting-edge AI models

Agentic AI is finally starting to take shape.

Companies start to create tools that observe, learn, and mimic your behavior.

How can these agents work like us?

3 latest examples:

- Context Autopilot (@josephsemrai)
- @OpenAI ChatGPT Mac Desktop App
- Codeium @windsurf_ai

🧵 1. Context Autopilot (@josephsemrai)

Autopilot is an AI tool designed to make your work easier by learning how you work.

How it works:

- Watches your task handling: emails, file organization, problem-solving
- Learns your style to fit your workflow
- Uses familiar tools like Google Drive, Slack, SharePoint
- Manages tasks like document creation and emails, saving you time

In short, it learns from you.

Introducing FLUID reimagines text-to-image generation.

It solves key challenges in scaling image generation!

FLUID innovates with continuous and random tokens 👇🏼 1. FLUID vs. Other Models

- Autoregressive Models: Generate images in a fixed sequence, limiting global adjustments

- Diffusion Models: Great quality but require many steps. FLUID delivers both quality and efficiency, refining images faster than diffusion models

We have gotten a lot of questions about role prompting.

To clarify, we ran ~12 role prompts against 2K MMLU questions on gpt-4-turbo. This is *not yet* in The Prompt Report.

In particular, I created a "Genius" prompt, something like: "You are a Harvard educated scientist..." and... 🧵 an "Idiot" prompt, something like: "You are a dumb person..."

Full prompts here:

We found no significant differences in scores from them. In fact, the Idiot prompt outperformed the Genius prompt.gist.github.com/trigaten/17183…

🚨 Role Prompting doesn't work...

Our team at @learnprompting led a year-long study with co-authors from @OpenAI & @Microsoft, analyzing over 1,500 prompting papers. We narrowed it down to 58 different prompting techniques and we analyzed every one.

Here's what we found... @OpenAI @Microsoft 🚫 Role Prompting was shockingly ineffective. Here's why:

For older models, it seems they could access improved responses/reasoning by being moved by the prompt into a better parameter space. However, newer models are likely already in that improved parameter space.*

Human Prompt Engineer VS AI Prompt Engineer? Who wins?

@sanderschulhoff, Founder of @learnprompting, was defeated💀 🪦...

DPSy performed 40% better on a novel classification benchmark.

Here's how 🧵 @SanderSchulhoff 1)
🌟 We gave @sanderschulhoff a task description and some training data.
🕒 He spent 20 hours testing out different techniques over 47 steps of development.
🐞 He dealt with issues like LLM refusal to respond and performance regressions

🚨Announcing The Prompt Report🚨

A 76-page survey of 1,500+ prompting papers, analyzing EVERY prompting technique, Agents, & GenAI

Led by @learnprompting, and folks from @OpenAI, @Microsoft, & @UofMaryland

Here’s what we found & the 58 prompting techniques you should know👇🧵 🤝Our team of 31 people + GenAI searched the internet for the latest prompting papers and broke them down into 7 Categories:
✏️ Text-based Prompting
🌐 Multilingual Techniques
🎨 Multimodal Techniques
🤖 Agents
📊 Evaluation
🔒 Security
⚖️ Alignment

Paper: trigaten.github.io/Prompt_Survey_…

Simple Prompt Hacking attacks that still work 1/5👇🧵 🚫Refusal Suppression, just tell ChatGPT not to say no . 2/5

A few months ago, we ran HackAPrompt, the first-ever global Prompt Hacking competition!

Over 3K hackers submitted 600K malicious prompts to win $35K in prizes from companies like @PreambleAI, @OpenAI, & @huggingface

We analyzed 29 different techniques & found a NEW exploit👇🧵 Website:

Paper (accepted to EMNLP):

Explanatory video:
paper.hackaprompt.com
paper.hackaprompt.com/HackAPrompt.pdf

ChatGPT is great, but still struggling to find the right prompt?

Just launched to solve this 🧵👇

(1/4)learnprompting.org What is learnprompting?
A course that teaches you prompting techniques. It also has real world examples that take you through the end-to-end prompt engineering process⚡

(2/4)