AI Rarities, M&A Drama, & More

TLDR;

Welcome back! Today, we take a look at a shocking research result on AI productivity, a roundup of finance news (including the weirdest M&A drama), new cool products, and the realization that China has very much caught up.

Enjoy!

CODING

AI Coding Makes You Slower

An interesting piece of research by METR has shown that AI coding models make developers slower, at least for top coders like those who participated in the study.

Interestingly, not only does this go against every single intuition you’ve heard about the use of AI in coding, but even the developers themselves thought AI had made them work faster when, in reality, it took them longer to solve the problems.

TheWhiteBox’s takeaway:

If you’re a regular of this newsletter, maybe you aren’t that surprised. We’ve repeated time and time again that AI models are trained by minimizing average loss. That is, they are trained to reduce the loss expectation across all problems in their training dataset.

Therefore, unless your dataset is of extreme, high-level coding quality, which never is, it’s logical to expect the outcome to be an average coder, one that surely elevates the productivity of vibe coders or average programmers like myself (I’ve never been a software developer in any paying job) but will of course slow down top coders.

That said, it’s also important to note several caveats:

Most of the developers in the list were rare users, if anything, of AI coding tools, meaning they were using the tools with little experience with them.

These coders were familiar with the repos they were working on. As many developers have noted, one of the best things about AI coders is that they get you up to speed with new repos much faster, a part of the job that was not evaluated here.

They used Claude 3.5 and 3.7 Sonnet, models that were SOTA at the time of the study but by no means SOTA today. Thus, there’s a chance that, as Labs create higher-quality datasets during post-training, future models will become less average over time, approximating high-level coding behavior more accurately.

Thus, to me, the biggest takeaway here is the spread between perceived speed and actual speed. The study touches on that very nuanced reality I had not considered before: one thing is the perceived productivity, another is the actual productivity when measured against your non-AI self.

CODING

Mistral’s Small and Medium Coders are Great

Mistral has released two new coding models, Devstral Small 1.1 and Devstral Medium, which set state-of-the-art performance relative to size.

The most impressive of the two is the smaller one.

With just 24 billion parameters, it achieves an impressive 53% on SWE-Bench Verified (OpenAI’s version of a challenging coding benchmark), surpassing other popular open models, including the flagship Qwen 3 model and DeepSeek R1. Notably, these two models are general-purpose, whereas Devstral is specifically designed for coding.

TheWhiteBox’s takeaway:

If it weren’t for the release of Kimi K2, it would have been the open-weight best coding…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.