AI (ML and Robotics) in 2026

I am following these curriculums:

• CMU's new PhD curriculum
• MIT's new research Agentic Investigator
• MIT's current Robotics curriculum

The plan is to learn the models and use free rented GPU access for the software because then we as plebs can modify them and do research because who has access to GPUs and memory anymore in 2026. Nvidia doesn't even sell consumer GPUs anymore it's all going to data centers on the moon or something.

We will learn:

• The mathematical model of Machine Learning which hasn't changed.
  • All lectures are taught by the author of the book and there's undergrad notes (MIT).
• 30 Lectures (March 2026) by MIT's leading Neuroscience researcher on the foundations of Deep Learning and his conjecture why it works (sparsity) and what isn't working.
  • Some optional short lectures from the perspective of theoretical neuroscience.
  • A book Understanding Deep Learning to help understand the lectures.

Then we will build systems:

• 10-714 DL Algorithms & Implementation (CMU)
  • Build from scratch an entire framework for deep learning
  • Generative AI audit some slides for instruction fine-tuning

Now we will fuse it with multimodal capabilities:

• MAS.S60 How2AI (MIT)
  • How to AI (Almost) Anything
  • YouTube lectures

However it will require throwing out almost all math learned in undergrad because as per Tomaso Poggio in his 30 lectures on Deep Learning the Real Numbers don't exist. Everything must be done by numerical analysis which means no limits, no Riemann integral, no pivots or echelon forms, no inverting matrices, limited use of eigenwhatever, and exclusively using matrix factorization for linear algebra. Congrats you no longer need to read Mathematics for Machine Learning we are dumping 4 years of undergrad in the trash.

• MIT Numerical Methods and large scale linear algebra.
  • Infinitesimal calculus book
    • We will only ever use dy/dx
  • SciML recorded lectures to learn discretized differential eq
  • Matrix calculus

New papers on robotics are being dumped just as rapidly as AI papers but they all assume you know this:

• The mathematical model of Robotic Manipulation
  • Robot hands/arms
• The mathematical model of Underactuated Robotics
  • Walking, Swimming, Flying

We can utilize Evan Chen's Napkin Project to figure out the math. Your job is then to sit with ChatGPT and the Agentic Investigator to come up with a better model.

Reverse engineering LLMs to make sure they are doing what they claim they are doing:

• Mechanistic Interpretability (Neel Nanda@Google DeepMind)
  • Complete guide to doing research yourself in mech interp
  • Example material is here and 90% programming
    • Chapter 0
    • Chapter 1
    • Chapter 2
    • Chapter 3

Many companies will be very interested if you can do this. This competition is still running as Apr 9 2026 but there will be more. I didn't exaggerate when I said there's a window where noone (available to be hired) knows these topics.

Any future medical AI, any drone defense AI, any space exploring AI, anything cool and futuristic is going to need Causal AI models. If do X what will happen to Y? To paraphrase Glenn Shafer the basic idea is to bring back the probability tree to represent a step-by-step evolution of an observer's knowledge. If that observer is nature, then the steps in the tree are causes, and the probabilities in the tree express nature's limited ability to predict the causes.

• Elements of Causal Inference Foundations and learning algorithms
  • Whatever else I can find when we get here because this is taking off right now.

Conformal Prediction or confidence intervals are also wide open to research for example you want to know how much money some junky AI that Anthropic peddles like Claude Code is going to charge you to generate some slop. You can learn this using conformal prediction and write a tool.

• Algorithmic Learning in a Random World using classic techniques like support vector machines

It's possible to completely throw out stochastic math and do statistics and probability purely in the field of game theory. Most human activities involve someone else and none of it is random so if you want to make a pokerbot this is how you do it. This including Causal AI is my primary interest because I've always thought (most) probability was bullshit but I'm a heretic so we'll still learn a very good Probability for Computing book.

• Game-Theoretic Foundations for Probability and Finance