Welcome to 2026
Estimated Reading Time: 3 minutes
TL;DR
We discuss some prospects for AI and more in 2026 after a several month hiatus.
Some Thoughts on the New Year
Welcome back to Deep in the Forest! We have been quiet over the last year as we have been focusing on internal projects (we can’t talk about everything we are working on, but for a public sampler, check out our recent blog post on LLM-guided retrosynthesis ).
The AI industry has been humming along, driven by the extraordinary buildout in datacenter capacity. I have publicly been skeptical in the past. See for example (https://deepforest.substack.com/p/how-fast-will-ai-advance, and https://deepforest.substack.com/p/towards-sentience-probably-not). The crux of my past arguments were that scaling had to end and there was only so much money people would be willing to put into AI.
I will acknowledge a failure of imagination on the scale of money people were willing to put into AI! I would have said at the time there’s no way anyone is putting in more than a billion dollars max into training a model. Given current trends, I would not entirely rule out something like a trillion dollar training run sometime in the next decade. This increased scale yields several differences. The first and most obvious is that three orders of magnitude additional investment will yield brute force scaling improvements. Second, this new scale will likely unlock new modalities, such as video, that are difficult to work with even today. Third, and perhaps most importantly, the ecosystem of people thinking about AI scaling and improvements has enormously ballooned. There are probably more people working on AI scaling than there were working on chip scaling in the peak of Moore’s law. This enormous investment of people and resources will yield a pathway to continuing improvements.
This new year, I partially renounce my position as a skeptic for these reasons. That said, I continue to think that AI rollouts will have a longer and more painful path to impact than commonly acknowledged. Rodney Brooks put out (https://rodneybrooks.com/predictions-scorecard-2026-january-01/) his recent, and excellent, analysis of his past technology predictions. To take one discussion in particular, he makes the point that despite Waymo’s (and more recently Tesla’s) advances in self-driving, these systems rely critically on teleoperated human intervention to handle difficult failure modes. Waymo’s recent systems failure during the SF power outage (https://waymo.com/blog/2025/12/autonomously-navigating-the-real-world) showed a real world example of a surge in human guidance requirements taking down their entire network. I think it would be fair to say that the average 16-year-old driver could probably still navigate tricky situations much better than the Waymo driver (without remote human intervention). That said, the 16-year-old is probably a less safe driver on the whole, so Waymo is clearly still a powerful and useful technology! The ability to take car rides with increased privacy and control has hit a note with consumers. We will likely see Waymo expand to most major metropolitan areas in the US over the next few years. More broadly, AI advances will likely follow the Waymo story. Success will depend on navigating a tricky corner between AI scaling and tactical human intervention for the long tail of difficult behavior.
On another note, for this blog for at least the next year, I am resolved to follow a “no-AI” rule. All of the writing here is purely artisanal, fueled only by caffeine and not transformers. I don’t personally see a moral failing in using AI; rather using AI feels like dumbing my thoughts down. I spend more time carefully having to vet AI generated sentences than creatively crafting ideas. I tried over the last year to use AI to assist in rewriting some intro/discussion sections for papers I was working on. The pace was painful since I could not trust references or comments to be correct. In the end, I gave up and just did it by hand. For me personally, for creative endeavors, I think this will likely not be the year of AI.
More broadly though, I am concerned about the effect of AI on intellectual ecosystems. To use a case study, DeepChem is an open source library I have maintained since my PhD. In the past, if a contributor put up a clean pull request (PR) with well documented code and tests, I could more or less believe in the correctness of the work. The only people who cared enough to make such contributions were usually serious scientists or students. Now, we are flooded with PRs that on the surface look correct. Most of these are from eager students leveraging the power of Claude Code or related tools. This is a double edged sword. On the one hand, I am genuinely happy students feel like they have a way to get involved more easily. I believe in democratizing science, and this is a powerful opening of the doors. At the same time, the flood has broken our review process. I can no longer trust in the superficial social markers of correctness. Instead, I, or another DeepChem maintainer, have to painstakingly check equations and benchmarks. Due to my limited time, in practice, we only merge PRs now from known contributors. This is sadly the opposite of democratization! This year, I will continue to experiment with ways to better navigate this new boundary.
Versions of this problem abound. Last year, for Google Summer of Code, DeepChem was flooded with nearly 250 applications. In previous years we would see something like 20-40. Most of these applications were AI-generated and it was painfully tricky to separate genuine candidates from fly-by-night AI submissions. The only thing that worked was talking to candidates on calls and vetting their ability to meaningfully answer questions about their applications live. This will likely hold for 2026, but I am dreading the day when people can use ChatGPT live to speak on their behalf. At that point, we may need to fall back to only considering candidates we can meet in the real world!
After writing about technology and geopolitics in public for the last several years, I feel that I have increasingly lost confidence in my ability to make long term predictions. Too much is in flux. This is both interesting and terrifying since it feels like we live in a world where the old orders are breaking down piece by painful piece. Where this leads, I don’t know, but I share my deepest wishes that this year brings better things to everyone reading!
Interesting Links from Around the Web
https://github.com/karpathy/nanochat/discussions/420: A very nice mini-blog post by Karpathy about nanochat scaling.
About
Deep Into the Forest is a newsletter by Deep Forest Sciences, Inc. Get in touch with us at partnerships@deepforestsci.com.
Credits
Author: Bharath Ramsundar, Ph.D.
Editor: Sandya Subramanian, Ph.D.
I'm even more broadly concerned about the future of social signaling. Subtle signals (e.g. embedded within text) disappearing as human promts are transformed into LLM output. Even more so when going back from LLM output to LLM summaries on the receiver end. I believe this will either reduce information density in general, resulting in more inefficient communication, or result in a new way of authentic signaling during promting - however that looks like.
Having studied chemical engineering, I'm blown away by DeepRetro... crazy!
by the time AI is good enough to fake a deeply technical zoom interview, I doubt we'll be hiring for those roles. We'll be focused on evals for which AIs to buy, not which humans. I still struggle with what humans will be doing in this future.