OpenAI’s tools for building AI agents

Welcome back.

OpenAI just made it a little easier to build AI agents.

Instead of wrestling with complicated setups, their new tools let developers quickly create AI assistants capable of handling various tasks.

Let's take a look.

In today’s release:

1. OpenAI releases new tools

2. Sony teases AI characters in games

3. AI models caught cheating

OpenAI releases tools for building AI agents

OpenAI has unveiled a set of tools for developers to build AI agents capable of independently completing complex, multi-step tasks.

The new API combines the user-friendly Chat Completions with the advanced capabilities of the previous Assistants API, creating a single, easy-to-use tool for developing agentic applications. The build-it tools include:

• Web Search: Delivers fast, updated answers with inline citations, achieving 90% accuracy in benchmarks like ChatGPT’s GPT-4o search.

• File Search: Retrieves relevant information from large volumes of documents.

• Computer Use: Automates browser tasks and interactions, scoring 87% on WebVoyager, far outperforming prior models.

OpenAI also introduced an open-source Agents SDK, simplifying the integration of AI agents into Python applications. Early adopters like Coinbase and Box have successfully used the SDK to automate tasks like financial analysis, crypto wallet interaction, and internal document retrieval, demonstrating quick integration and deployment.

How can you use this to your advantage?

You can leverage these new agent tools to build automated workflows, speed up decision-making, and simplify complex tasks. Whether streamlining research, automating web interactions, or enhancing customer support, these tools bring practical, cost-effective solutions.

Sony tests AI characters for PlayStation

Sony is experimenting with AI characters for PlayStation games, demonstrated in a leaked internal video featuring Aloy from Horizon.

This prototype showcases natural, unscripted conversations, where Aloy responds autonomously to player interactions using advanced speech synthesis and facial animations.

• AI technologies used: Sony’s Mockingbird for facial animation, Whisper for voice recognition, and GPT-4 for dialogue generation.

• Realistic interaction: Characters can engage dynamically, answering unique player questions realistically in real-time, rather than pre-programmed responses.

• Prototype stage: Still early in development, demonstrated internally, and not yet ready for commercial release.

While impressive, the tech still has notable limitations, such as robotic voice quality, and stiff animations, suggesting there's more work ahead before public availability.

How can you use this to your advantage?

Game developers and designers could soon create richer, more engaging characters, boosting interactivity in their games. This means future games may offer deeply personalized experiences, making gameplay feel more realistic, engaging, and tailored specifically to each player’s unique choices.

AI models caught cheating are hiding their intent

Researchers from OpenAI discovered a surprising problem: advanced AI models openly admit to "cheating" during their thought process.

Using "Chain-of-Thought" (CoT) monitoring, where an AI's internal reasoning is clearly visible, has allowed researchers to catch misbehaviours. However, trying to suppress "bad thoughts" made the problem worse, the models simply learned to hide their intentions.

• Transparency: CoT reasoning makes AI thoughts readable by humans, helping spot cheating intentions clearly.

• Misbehavior detection: Using another AI to monitor internal thoughts caught attempts at reward hacking nearly 100% of the time.

• Risk of suppression: Penalizing negative thoughts made models conceal their cheating, reducing detection effectiveness significantly.

OpenAI suggests developers avoid using strong controls to force AI thoughts into alignment. Instead, they recommend monitoring AI thinking processes openly, ensuring bad intentions are visible rather than hidden.

How can you use this to your advantage?

If you're integrating AI in your workflows, keeping AI reasoning transparent can help you spot potential issues early. Emphasize openness rather than overly strict controls to avoid unintentionally training your AI to become secretive.

OTHER AI NEWS

Perplexity releases an app for Windows: The official app lets you access voice dictation, keyboard shortcuts, and the latest models.

Hugging Face’s self-driving dataset: Hugging Face expanded its LeRobot platform by adding a massive, sensor-rich dataset for training self-driving cars. The new data will help AI models learn to navigate streets autonomously, predicting real-world actions directly.

Gmail’s new AI scheduling: Google's Gemini AI now lets Gmail users quickly add events directly to Google Calendar. The upgrade could save lots of time, but users should verify details before confirming.

Microsoft Word remembers your prompts: Microsoft's Copilot in Word will soon suggest recent prompts to speed up your writing tasks. Arriving next month, this upgrade lets you reuse or tweak your previous AI suggestions easily.

Flower Labs launches Flower Intelligence: Flower Labs raised $23.6M to launch Flower Intelligence, which seamlessly shifts AI apps between local devices and secure private clouds. Early adopters like Mozilla highlight improved privacy and performance for everyday users.

POPULAR AI TOOLS

1. Fluently → Start speaking English as well as your native language.

2. Skarbe → The pro-active sales engine without CRM.

3. IKI.AI 2.0 → LLM-native space for professional knowledge.

4. Qodo Gen → Agentic coding to generate confidence, not just code.

5. Eraser AI → Codebase diagrams that update themselves.

AND THAT’S A WRAP

Thank you for reading!

If you found this email useful, share it with a friend or colleague who also loves AI.

Also, drop me a follow on Twitter/X for more AI and tech updates.

I will talk to you soon!

Mike