2025 Interconnects year in review

Three years writing every week about AI.

Dec 18, 2025

A lot has happened in AI this year, which is the new normal. Since I last wrote a year in review, before going on Lex’s Podcast and writing timely coverage of the DeepSeek models, the Interconnects audience was below 20K readers. It’s more than doubled this year in viewership (and influence), so it’s a good time to reintroduce how I see the value in Interconnects to everyone.

My main day job is being deep in the weeds of the research needed to train better Olmo models. Interconnects serves as an outlet for these ideas to help me stay on top of having an elite vision and plan for said research.

The best way to view my content is a raw expression of my worldview as a leading AI researcher reacting to the world of AI. A lot of my writing is raw, sometimes off the cuff, and often technical. I lean into this as it gives readers a very cutting-edge view of my thinking, which often can translate into the unique feeling of “getting it” at the cutting-edge of research.1

One of my favorite things while wrapping up the year is I’m regularly using three frontier models (ChatGPT, Claude, & Gemini) for slightly different things. The landscape is as competitive as it ever has been. These products I love are also facing serious pressure from open models close on their heels. In 2026 I want to use more open models directly.

This year brought my long-term focuses of reinforcement learning and open models to the forefront of the entire AI ecosystem, so it makes sense that it was the busiest year on Interconnects yet. Through 2025, there were 80 total posts of which 19 were for prominent model releases, 11 were our roundups of open models (Artifacts Log), 4 interviews, 6 of my talks, and the remaining 40 spread across analysis of the world of AI. In total I earned about 3.5 million pageviews on Substack.

This is well over my goal of one per week and likely will be a local optimum over this few-year phase of AI where I’m working full-time on research, with a major side gig in the AI media and policy spheres. I don’t expect there to be another increase in year-over-year amount of content, but we should be able to make every piece of content even more valuable. We’re learning a lot.

It’s arguable that I’m burning out this year, but with how motivated and aligned I am to my work, it feels more like just hitting hard limits. I can’t do everything. I’m lucky to have a life where I can shape a lot of my time, energy, habits, and goals around shaping trends in AI. Working at the Allen Institute for AI has been great for this, but they should be because they obviously benefit a lot from my work too.

I’m tired but feeling very content with my contributions to the AI world. I still think this is the early years of my work and getting the right practice in is more important to me than perfection of every piece. If you can, upgrading a paid subscription helps me pay for AI tools or operational assistance and just helps me stay motivated in the project (plus, you get access to the excellent Discord Server too).

This year I launched two major projects that are closely intertwined with my goals for Interconnects: The ATOM Project, the case for immediate investment in centers building American open models, and the SAIL Bundle, a paid bundle of my favorite writers on Substack sold only to teams (we’ve started publishing a bunch of content on the SAIL YouTube channel). I intend to invest further in both of these in 2026, so you’ll keep hearing about these. It is a risk to invest so much time in these things, so it’s a relief when they land.

I’ll see you all in January. We have no planned content until then, but we’d chime in if DeepSeek V4 (or an equivalent) comes out and is causing international ripples. I’m off to finish up my RLHF book, which you should pre-order in print now!

Previous years in review: 2024 | 2023

You can see the weekly views this year below.

In this post I’ve organized all the posts of the year into sections. Throughout, I’ve marked my favorite posts with ★ if you’re looking to catch up on what I felt is the mostly lasting content from the year.

Comment below what your favorite posts were or any topic you wish I covered more. Contents:

Of all the sections I listed, the subsection of Analysis posts on Reasoning & RL is my highlight for the year. The average quality there is excellent, especially given how much uncertainty and confusion exists when new paradigms emerge.

Of all the posts, 26 of them directly cover or discuss the role of Chinese open models on the AI ecosystem, which will surely continue into 2026 (hint: We already have some great interviews lined up or recorded).