AI Daily: Mistral Devstral 2 Arrives, OpenAI Launches Official Certification Courses

Major updates in AI this week, ranging from developer tools to educational certifications. Mistral AI launched the powerful Devstral 2 model and Vibe CLI, attempting to change the development experience for engineers; OpenAI partnered with Coursera to officially launch AI skill certifications, aiming to equip millions with practical AI capabilities in the coming years. Additionally, Google Cloud released AlphaEvolve, taking algorithm optimization to a new level, while designers welcomed Google Stitch’s heatmap prediction feature. Let’s look at how these updates affect various industries.

Mistral AI Strikes Again: Devstral 2 and Vibe CLI Define New Development Standards

The most heated discussion in the developer community over the past two days has undoubtedly been Mistral AI’s announcement of Devstral 2. This French AI unicorn seems to always drop bombshells in the open-source world. This time, they brought Devstral 2, a new model family designed specifically for code generation, including a main model with 123B parameters and a lighter Devstral Small 2 (24B).

The emergence of these two models is undoubtedly great news for enterprises that do not want to be constrained by closed-source models. Devstral 2 achieved an accuracy of 72.2% in the SWE-bench Verified test, which means its ability to solve actual software engineering problems is already quite amazing. It is also worth mentioning that its operating cost is much lower than competitors like Claude Sonnet, with efficiency up to 7 times higher.

But what’s more interesting is actually the Mistral Vibe CLI. For engineers accustomed to typing commands in a terminal window, this is simply a tool tailored for them. It is not just a chatbot, but more like a senior colleague living in the command line. Vibe CLI can perceive the context of the project, Git status, and even perform code refactoring across multiple files. Imagine not having to switch windows, directly using natural language in the terminal to let AI help you fix bugs or rewrite modules; this smoothness is indeed desirable.

A New Workplace Essential: OpenAI Launches First Batch of Official Certification Courses

If Mistral is providing weapons for engineers, then OpenAI is providing instruction manuals for the public on how to use weapons. As AI tools become more popular, many people are actually not sure if they really “know how to use” AI. To bridge this gap, OpenAI officially released the first batch of certification courses, marking the stage where AI skills officially enter standardized assessment.

This set of courses named “AI Foundations” was launched in cooperation with platforms like Coursera, and the content is very pragmatic. It does not talk about profound machine learning theories but focuses on how to use tools like ChatGPT to solve practical problems in work. Through these courses, students can practice in the ChatGPT environment and receive immediate feedback.

For job seekers, this may be an important turning point. OpenAI’s goal is to help 10 million Americans obtain certification by 2030. It is foreseeable that “OpenAI Official Certification” appearing on resumes in the future may be as common and important as English proficiency proof is now. In addition, they also specifically designed “ChatGPT Foundations for Teachers” to help educators integrate AI into the classroom, which is crucial for cultivating the AI literacy of the next generation.

Google AlphaEvolve: Evolving AI Algorithms with AI

In the fields of science and engineering, many difficult problems can actually be boiled down to “optimization problems,” such as how to make logistics paths the shortest or how to make chip heat dissipation the best. Google Cloud’s newly released AlphaEvolve was born to solve such challenges.

The concept of this technology sounds very sci-fi: it uses the Gemini model to “evolve” algorithms. Simply put, it doesn’t give the answer directly, but generates multiple code variants (mutations) like biological evolution, and then tests the performance of these variants (natural selection). Code that performs well will be kept and undergo the next round of optimization. This loop repeats continuously, eventually producing efficient algorithms that human engineers might never come up with.

Google has already used this technology internally to achieve amazing results, such as increasing the training speed of Gemini models by 1% or optimizing data center task scheduling. Now, this technology is open to enterprises through Google Cloud, meaning biotech companies can use it to accelerate drug development, and logistics companies can use it to reduce fuel costs. This demonstrates how AI is transforming from a mere auxiliary tool into a core engine driving scientific discovery.

Accelerator for Design and Creation: Stitch Heatmaps and Veo 3.1

For designers and creators, there are also exciting tool updates this week. First is the Predictive Heatmaps launched by Google Stitch.

In the past, if UI/UX designers wanted to understand user visual focus, they usually had to wait for lengthy user testing data. But now, through Stitch’s new feature, designers can conduct an “attention audit” on design drafts before writing a single line of code. This AI model (Nano Banana Pro) has learned to watch screens like real users and can instantly tell designers where users’ eyes will look. This not only saves time but also makes design decisions more data-supported.

On the other hand, the threshold for video creation is further lowered. Google is currently testing Veo 3.1’s template feature on a small scale. Users only need to choose a style template or upload their own pictures for customization to quickly generate high-quality videos. For marketers or creators who need a lot of short video content, this is undoubtedly a sharp tool that can significantly increase productivity.

Frequently Asked Questions (FAQ)

Q1: Is Mistral Devstral 2 completely free? Devstral 2 adopts a modified MIT license and is an open-weights model; developers can download and run it on their own hardware. At the same time, Mistral also provides this service through an API. Currently, the API is free to use during the promotion period, but in the future, it will be charged based on Token usage (Devstral 2 is $0.40/$2.00 per million Token input/output).

Q2: Is OpenAI’s certification course suitable for people with no programming background? Yes, very suitable. The original intention of the “AI Foundations” course design is to improve the AI literacy of general workers, focusing on how to “use” tools to solve problems and improve productivity, rather than exploring the underlying program development or mathematical principles.

Q3: How is Google AlphaEvolve different from general code generation tools? General code generation (such as GitHub Copilot) mainly assists engineers in completing functions. AlphaEvolve is a system focused on “optimization.” It actively explores and improves the efficiency of algorithms through the combination of evolutionary algorithms and Gemini models, usually used to solve extremely complex mathematical or logic optimization problems, not just writing code that works.

Q4: Can I use Google Stitch’s heatmap feature now? Yes, according to official news, the Predictive Heatmaps feature is now online in Stitch’s “Generate” menu, and designers can directly use this feature to detect the visual focus of design drafts.