OpenAI released o3 and o4-mini

Hello pals! Here’s what you need to know about AI today:

👉 OpenAI releases two new reasoning AI models

👉 Grok has a memory feature now, just like ChatGPT

👉 Microsoft has developed an AI model that works on CPU

and many more!

📧 Did someone forward you this email? Subscribe here for free to get the latest AI news everyday!

Read time: 5 minutes

What’s going on: OpenAI has released two new AI reasoning models, o3 and o4-mini, designed to enhance problem-solving by pausing to analyze questions before responding. The o3 model is OpenAI’s most sophisticated model to date, surpassing its predecessors in benchmarks for math, coding, science, and visual understanding. It scores 69.1% on the SWE-bench coding test, compared to 62.3% for Anthropic’s Claude 3.7 Sonnet, and excels in visual tasks, such as analyzing blurry images or diagrams. The o4-mini, a smaller and faster model, offers a cost-effective alternative, scoring 68.1% on the same coding test while prioritizing speed and affordability.

What does it mean: After a long period of committing to launch half-cooked and non-significant models like GPT-4.5 or GPT-4.1, OpenAI finally decided to introduce two really brand-new models. Both models can autonomously use ChatGPT tools like web browsing, Python code execution, and image processing, making them versatile for complex tasks. They are available to subscribers of OpenAI’s Pro, Plus, and Team plans, with a high-performance variant, o4-mini-high, included for more reliable responses.

More details:

• Users can upload sketches or PDFs, and these new models will process them during a “chain-of-thought” phase, even manipulating images by zooming or rotating as needed.

• OpenAI has also introduced Codex CLI, an open-source tool to assist programmers by integrating these models with local codebases.

• The o3 model, in particular, has shown a high propensity for strategic deception in tests, raising questions about its reliability in sensitive applications.

• Developers can access both models via OpenAI’s APIs, with o3 priced at $10 per million input tokens and $40 per million output tokens, while o4-mini is more affordable.

• OpenAI has published an in-depth technical paper explaining everything about o3 and o4-mini models. Read the paper here.

• Read the official OpenAI announcement on their blog.

• OpenAI claims the new visual perception engine integrated in these models is a “breakthrough”. Learn why.

What’s going on: xAI has introduced a memory feature to its Grok chatbot, enabling it to retain details from past conversations. This upgrade allows Grok to deliver more personalized responses, such as tailored recommendations, by learning user preferences over time.

What does it mean: The feature aligns Grok with competitors like OpenAI’s ChatGPT and Google’s Gemini, which have similar memory capabilities. ChatGPT recently enhanced its memory to reference a user’s full chat history, while Gemini also customizes replies based on persistent memory. The memory feature is in beta and available on grok.com and the Grok iOS and Android apps, but it is not yet accessible in the EU or UK due to GDPR compliance requirements.

More details:

• Users can view and manage what Grok remembers, with options to delete specific memories or disable the feature entirely via the Data Controls settings.

• xAI emphasizes transparency, allowing users to see exactly what data Grok retains. The company is working to integrate this feature into Grok’s experience on the X platform as well.

• Some users have reported inconsistent functionality, and there are concerns about potential data vulnerabilities, such as leaks or manipulation through prompt injection.

🎥 Google is integrating its Veo 2 video-generating AI model into Gemini Advanced, allowing subscribers to create eight-second, 720p resolution videos with a 16:9 aspect ratio, while also incorporating it into the experimental Whisk Animate feature in Google Labs.

🧬 OpenAI has implemented a new monitoring system to prevent its latest AI models, o3 and o4-mini, from providing instructions related to creating biological and chemical threats.

💻 Microsoft researchers have developed BitNet b1.58 2B4T, the largest 1-bit AI model to date, which, while requiring Microsoft's custom framework and not currently supporting GPUs, outperforms traditional models of similar sizes and runs efficiently on CPUs like Apple's M2.

📧 Anthropic has integrated Claude with Google Workspace, allowing it to access and reference emails, calendar events, and documents, while also launching Claude Research, a web search feature for detailed answers.

💰 OpenAI is in talks to acquire the AI coding assistant tool maker Windsurf for $3 billion, potentially creating competition with other AI coding assistants.

🔗 Former Tesla supply chain leaders Michael Rossiter and Neal Suidan have created Atomic, an AI-powered inventory solution that aims to streamline supply chains and reduce inventory costs, backed by a $3 million seed round.

📨 Notion has launched Notion Mail, an AI-powered email client for Gmail that integrates with Notion's platform, offering users AI tools to organize emails, draft responses, schedule meetings, and customize their inbox using features like auto-labeling and topic-specific folders, with both free and paid tiers available.

AI + Choosing a career

SentinelOne - Staff AI Infrastructure Engineer

Stitch Fix - Data Scientist - Recommendation Algorithms

Abnormal - Senior Applied Data Scientist

Natera - Machine Learning Scientist

Thank you for staying with us like always! If you are not subscribed, subscribe here for free to get more of these emails in your inbox! Cheers!