You commented "upskill". Here it is. All ten references from the Reel, each one explained in plain English, with a single link that actually teaches it rather than a search result. Nine are other people's work, linked to the original. The tenth is mine, and it is not ready yet. No course, no upsell.
You'll also get the weekly letter: five useful things I made with AI, with the prompts and workflows inside. Unsubscribe any time.
Already on the list? Enter the same email and it unlocks without subscribing you twice.
Recognising ten names does nothing for you. Understanding two of them properly is what changes the quality of the decisions you make next month.
They fall into three groups. One, three, four, six, seven and nine are things you build or tune, so read those if you are hands-on. Two, five and eight are people and publications, so follow those if your job is staying ahead rather than shipping. Ten is mine and it is not out yet.
The framework you reach for when your agent needs to be a system rather than a loop with a prompt in it.
LangChain describes it as a low-level orchestration framework and runtime for long-running, stateful agents. The important word is stateful. It lets you mix deterministic hand-written logic with model-driven decisions inside one graph, which means you can force the parts that must be reliable to be reliable, and leave judgement to the model only where judgement is actually needed. Persistence and human-in-the-loop are built in rather than bolted on.
Most agent projects that stall do so because the control flow lives in prompt text. This is the answer to that.
The person to read when you want to know whether the AI money is actually going anywhere, rather than whether the demo was impressive.
Managing Partner and CIO of Atreides Management. He started covering NVIDIA as an investor around the turn of the millennium, which gives him a longer view of the compute cycle than almost anyone talking about it now. He is worth following specifically because he argues about supply chains, capex and margins rather than capability demos, and that is the layer where most predictions about AI actually get settled.
His running commentary is on X as @GavinSBaker. Start with the long-form conversation below.
Prebuilt inference microservices, so running a model yourself stops being an infrastructure project.
NIM microservices wrap a model in a GPU-accelerated, industry-standard API and tune the serving stack for each model and GPU combination. You get a container and an endpoint instead of a week of CUDA versions, batching strategies and latency tuning.
This matters if your organisation cannot send data to a hosted API. NIMs are the shortest route from "we have to self-host" to "it is serving traffic".
Continuing training on a small, specific dataset so the model gets good at your task instead of everyone's.
Mechanically it is identical to pretraining except you start from trained weights rather than random ones, which is why it needs far less compute, data and time than people assume. The Hugging Face tutorial below walks the whole loop: tokenise, load the checkpoint, set the training arguments, train, push.
The judgement call is not how to do it, it is when. Most problems people reach for fine tuning to solve are retrieval problems or prompt problems. Read this one alongside 09.
The interviewer who does enough homework that the researchers he talks to have to think on the spot.
The Dwarkesh Podcast is billed as deeply researched interviews, and that is the whole differentiator. Most AI interviews let the guest run their press tour. His questions are specific enough that the conversation goes somewhere the guest has not been fifty times that month, which is where the genuinely new information sits.
If you only have one podcast slot a week for AI, this is a better use of it than the news roundups.
The claim that most of your agent's performance is decided by everything wrapped around the model, not the model itself.
The harness is the tools you expose, how you describe them, the retry and error handling, the context you assemble, the guardrails, the loop structure. Swap a frontier model into a bad harness and it stays bad. Improve the harness and a cheaper model often clears the bar.
This is the single highest-leverage idea on this list for anyone already shipping agents, because it redirects effort from model shopping to the parts you actually control.
Yann LeCun's bet that predicting abstract representations, not pixels or tokens, is the road to models that understand the world.
I-JEPA is the first model built on that architecture. Instead of reconstructing missing pixels, it predicts the representation of a missing region from the representation of the rest of the image. It learns semantic features efficiently and without the heavy fine-tuning that comparable approaches need.
Worth knowing because it is the most credible technical argument that scaling current language models is not the only path, and because the phrase gets used as shorthand in that debate.
The newsletter and podcast written for the people building the systems, not the people reporting on them.
It describes itself as the AI Engineer newsletter and covers how leading labs build agents, models, infrastructure and AI for science. The distinction that matters: it is written at the level of someone who has to make the thing work, so the detail survives contact with your actual codebase.
Run by swyx and Alessio, with guests including Andrej Karpathy, Simon Willison and Greg Brockman.
Knowing which of fine tuning, retrieval or distillation your problem actually calls for, before you spend a month on the wrong one.
They solve different problems. Retrieval gives the model facts it did not have. Fine tuning changes how it behaves and what format it produces. Distillation makes a smaller model imitate a larger one so you can afford to run it. Most teams reach for fine tuning when they needed retrieval, then conclude the technology does not work.
The piece below frames it as the gym, the library, or cloning yourself, which is a crude analogy that happens to make the decision obvious.
The tenth one is mine, and it is not ready to show you yet.
Nine of these are other people's work and I can point you straight at the source. This one I am building. Rather than write a paragraph that oversells it, I would rather show you the thing.
You are on the list now, so you will hear about it there first.
Reading is not the outcome. Recognising one of these inside something you have already built is.
You're on the list now. Every week you'll get five useful things I made with AI, with the prompt or workflow behind each one. This list was one of them.
Back to the site