jetnew.io

What is Jet New?

Jet New is a personal notes site: a place where its author records what they are thinking about, learning and working on. The writing is presented as continuous works-in-progress rather than finished articles, so the value is in following ideas as they develop.

What you'll find there

• Short, dated notes and essays. The archive includes pieces such as "100 Lessons from 1 Year of AI Research" (December 2021), "Getting Started with Reinforcement Learning" (March 2021) and "Why You Should Conduct a Technical Workshop" (March 2021).
• Work in public. The site describes an in-progress project, Atlas, as "a tool for understanding research," with a planned piece on why it is being built.
• Two clear sections. An "Upcoming" area for announced writing and an "Archive" for published notes, which makes it easy to skim by date.

Who it suits

It is most useful for readers curious about AI research practice, reinforcement learning basics and building tools for thought, who prefer informal field notes over polished tutorials. Someone starting out in RL, for example, could use the "Getting Started" note as a pointer to topics worth pursuing, then check the archive for later reflections.

A practical next step

Open the archive and read the most recent note first, since the author frames everything as evolving; then read the earlier AI-research lessons piece to see how their views have shifted. If you want a broader map of the field, compare with established educational resources such as DeepLearning.AI or OpenAI.

What is Atlas, the AI tool for thought Jet New is building?

Jet New describes Atlas as "a tool for understanding research" and, in a planned October 2026 essay, "an AI tool for thought." On the site itself, that is nearly all the detail given: the project is presented as something he is building, with the writing around it framed as continuous works-in-progress rather than finished product documentation. The site does not describe Atlas's interface, how it processes papers, or what it outputs, so any claim beyond "an AI workspace for research" would be speculation.

What the page supports

• Purpose: helping a reader understand research material, per Jet New's own description.
• Category: an AI research workspace, listed alongside his AI research and consultancy work.
• Status: in development, with the explanatory essay still upcoming.

A practical reading

The phrase "tool for thought" suggests the emphasis is on thinking with sources rather than just summarizing them — for example, a graduate student working through a stack of papers who wants help connecting arguments, or a practitioner trying to extract a method from dense literature. That is a reasonable inference from the wording, not a documented feature set. If you are evaluating Atlas for real use, treat the upcoming essay as the first place to look for concrete workflow examples, and compare it against established reference managers and literature-mapping tools once those details appear.

For context on the builder's background, his archive includes notes on AI research lessons and getting started with reinforcement learning, which you can read at Jet New.

What topics does Jet New write about in his notes and archive?

Jet New writes a personal notes archive about AI research, learning in public, and building tools for thinking. The site describes the writing as continuous works-in-progress, so the notes read more like an evolving working notebook than finished essays.

Recurring topics in the archive

• AI research lessons: "100 Lessons from 1 Year of AI Research" (Dec 2021) collects practical takeaways from a year of research work, aimed at people starting or reflecting on their own research practice.
• Reinforcement learning: "Getting Started with Reinforcement Learning" (Jul 2021) is an entry-point piece for readers new to the field.
• Technical workshops: "Why You Should Conduct a Technical Workshop" (Mar 2021) covers teaching and knowledge-sharing as a professional habit.
• Building AI tools: the upcoming piece "Paper, Wood, Leather – Why I'm building Atlas, an AI tool for thought" (Oct 2026) connects his work on Atlas, described as a tool for understanding research, to his wider interest in AI-assisted thinking.

Who it suits

The archive is most useful to readers who want short, experience-based notes on doing AI research and learning rather than polished tutorials. If you are weighing whether to run a technical workshop or are new to reinforcement learning, the dated posts give you a concrete starting point; if you follow tools for research and thought, the Atlas piece is the one to watch for.

A practical next step: open the archive in date order and read the two 2021 AI posts together, since the research-lessons piece builds on habits the reinforcement learning introduction assumes. For broader context on AI research and tooling, see OpenAI and Google DeepMind.

What can I learn from Jet New's 100 Lessons from 1 Year of AI Research?

Jet New's "100 Lessons from 1 Year of AI Research" (dated December 2021 in his archive) is best read as a practitioner's retrospective: a numbered list of observations from someone who spent a year doing AI research and later went on to build Atlas, an AI research tool. The page itself only lists the title, so treat the value as coming from the format and the author's vantage point rather than from any summary on the site.

What the format suggests you'll get

• Breadth over depth. A hundred short lessons implies many small, concrete takeaways — habits, tooling choices, reading strategies, collaboration norms — rather than one long argument. Read it with a notebook open and mark the five or six that apply to your situation.
• A snapshot of a specific period. December 2021 was before the current wave of large language models reshaped research workflows. Lessons about literature review, experiment tracking, or choosing problems may still hold; lessons about specific tools will feel dated.
• Evidence of how the author thinks. Because Jet New now describes Atlas as "a tool for understanding research," the lessons are a useful key to what he believes is hard about research — likely reading, organizing, and synthesizing prior work.

How to use it

If you're starting a first research role or a thesis, read the list once for orientation, then return to it after a month to see which lessons you've already internalized and which you keep violating. For a team, pick two or three lessons and turn them into a short discussion at a lab meeting — that converts a personal list into shared practice.

A realistic caveat

A hundred-item list is easy to skim and easy to forget. The lessons are one researcher's experience, not tested guidance; where a lesson conflicts with your advisor's or team's norms, treat it as a question to raise, not a rule to adopt. Pair it with a longer-form account of research practice, such as Matt Might's advice for researchers, for structure the list format can't provide.

How should a beginner get started with reinforcement learning, based on Jet New's guide?

Begin with a small, self-contained problem and build the full loop before tuning anything. Jet New's July 2021 note, "Getting Started with Reinforcement Learning," sits in an archive that also includes his later reflections on a year of AI research, so the beginner path it implies is: learn the vocabulary, implement one environment end to end, then read results critically.

A practical first sequence

1. Fix the mental model. Agent, environment, state, action, reward, episode. Be able to explain why a reward signal differs from a label in supervised learning.
2. Start with a tiny environment. A gridworld or a classic control task (cart-pole style) is enough; the point is a loop you can print and debug, not a benchmark score.
3. Implement one tabular method first. Q-learning or policy iteration on a small discrete problem makes the update rule visible. Only then move to function approximation.
4. Add one deep method. Pick a single algorithm and one environment, and reproduce a known result before changing hyperparameters.
5. Read results skeptically. Plot reward curves across several seeds; a single lucky run is not evidence.

Choosing where to start

For a concrete reader: if you have a weekend, spend Saturday on a tabular agent in a gridworld and Sunday writing down what each hyperparameter did. That written record is more useful than a second algorithm.

As a next step, keep a short log per experiment—environment, algorithm, seed, reward curve, and one sentence on what surprised you. Jet New's own archive shows this habit of continuous, revisable notes, and the same practice makes RL debugging far less mysterious. For broader background, the official documentation at Gymnasium and the text at Sutton and Barto's Reinforcement Learning are standard references to consult alongside any single guide.

How can I run a technical workshop effectively, according to Jet New's advice?

Jet New's site points to a piece titled "Why You Should Conduct a Technical Workshop" (March 2021), but the page evidence here gives only that title and a one-line description of the site as notes on what he is thinking about, learning, and working on. It does not include the article's actual recommendations, so I can't summarize his specific advice without inventing it. What follows is general workshop guidance, not a report of his views.

A practical structure for a technical workshop

1. Pick one concrete outcome. Decide what attendees should be able to do afterward—set up a working environment, train a small model, deploy a service—and cut everything that doesn't serve it.
2. Design for hands-on time. Aim for roughly half the session on participants' own machines. Demos are easy to watch and easy to forget.
3. Prepare a tested environment. Pin versions, provide a starter repo or container, and include a fallback for people whose setup fails. Environment problems are the most common way a technical workshop loses its schedule.
4. Use short explain–do cycles. Explain a concept for five to ten minutes, then have people apply it immediately. This surfaces misunderstandings while there is still time to fix them.
5. Build in checkpoints. Give expected outputs at each stage so attendees can tell whether they are on track without waiting for a helper.
6. Staff for questions. One facilitator per 10–15 participants is a reasonable ratio for coding-heavy sessions; circulate rather than lecturing from the front.
7. Close with an extension path. Point to next exercises and reference material so the learning continues after the room empties.

Choosing the format

Next step

If you want Jet New's own reasoning, read the March 2021 post directly on Jet New; his notes are described there as continuous works-in-progress, so treat the article as a starting point and adapt the structure to your audience.