05 — The Staying-Current System¶
In 2025, approximately 100–200 new ML papers are posted to arXiv every single day. That is one paper every 8 minutes, globally, continuously. The naive response — “I need to read more” — is the wrong response. Reading faster does not solve a selection problem. What you need is a system: a curated, time-boxed, sustainable information diet that surfaces signal without generating anxiety.
This file is a precise specification of that system. It is not a list of nice-to-have sources. It is a weekly operating rhythm with time allocations. If you follow it at full-time-job pace (~10-15 hours/week total), you will stay in the top 10% of applied ML engineers by awareness without burning out.
The Core Principle: Asymmetric Information Sampling¶
Do not try to be comprehensive. Try to be well-positioned. The goal is not to have read everything — it is to have a system that reliably surfaces the 5% of new work that materially affects your domain, filters it from the 95% that does not, and gives you enough depth on the filtered set to form and update opinions.
This means accepting that you will miss things. The question is: which things can you afford to miss?
Affordable to miss:
Papers outside your 2-3 active sub-domains
Incremental follow-up work on established methods
Workshop papers without corresponding experiments
Preprints with no downstream adoption in 6 months
Not affordable to miss:
New architectures with >500 citations in 3 months
Papers from major labs (Google DeepMind, Meta AI, OpenAI, Anthropic, DeepSeek) with immediate practical application
Papers that change the default approach to something you use daily (attention, fine-tuning, quantization, evaluation)
Daily System (20–30 minutes/day)¶
Twitter/X: Primary Discovery Channel¶
Twitter/X is, empirically, the fastest way to surface important new papers. The community has pre-filtered by engagement. The accounts below are the highest signal-to-noise ratio in applied ML:
Core follow list (non-negotiable):
Account |
Why |
|---|---|
@karpathy (Andrej Karpathy) |
Technical depth, honest assessments, implementation intuition |
@srush_nlp (Sasha Rush) |
NLP/systems intersection, critical paper reading, teaching clarity |
@ylecun (Yann LeCun) |
Architecture debates, long-horizon ML perspectives |
@GuggerSylvain (Sylvain Gugger) |
HuggingFace internals, practical transformers |
@rasbt (Sebastian Raschka) |
Weekly ML paper summaries, reproducible benchmarks, applied focus |
@giffmana (Andreas Giffhorn) |
Flash Attention, efficient training, CUDA-level work |
@timdettmers |
Quantization, memory efficiency, practical 4-bit methods |
@marktenenholtz |
ML system design, interview prep, practitioner perspective |
@swyx |
AI engineering meta-layer, newsletter synthesis |
@npew (Nathan Lambert) |
RL alignment, RLHF, policy optimization |
@cwolferesearch |
Interesting paper threads with critical commentary |
Accounts to follow with skepticism (high hype ratio):
Any account whose primary output is “X new paper just dropped — here’s what you need to know” threads without original analysis
Any account with >500K followers whose posts are consistently non-technical announcements
Daily Twitter/X routine (10 minutes):
Open the list — not the main feed
Scan for paper links with substantive commentary from the accounts above
If a paper appears in 3+ posts from independent accounts in one day: add to weekly deep-read queue
Do not click clickbait thread summaries as a substitute for papers
arXiv Daily Alerts (10 minutes)¶
Set up email or RSS alerts for:
Optional secondary alerts based on your sub-domain:
Daily arXiv routine:
Scan titles only — takes 5 minutes
Read abstracts for ~10 that seem relevant — takes 5 minutes
Add 1-2 to weekly queue if the abstract suggests a genuine mechanism (not just a new benchmark with marginal improvement)
What most people get wrong: Adding too many papers to a queue that never gets processed. If your queue grows faster than you drain it, the queue is not a system — it is a backlog that generates guilt. Cap the weekly queue at 3 papers. Three well-read papers beat twenty skimmed ones by an order of magnitude.
Weekly System (2 hours/week)¶
Papers With Code: Trending (20 minutes, Monday)¶
Visit paperswithcode.com/trending.
This is the most reliable signal for papers gaining traction in the practitioner community. Filter by: last 7 days, your primary domain. The papers appearing here have code — which means they can be reproduced, not just cited.
Routine:
Scan the top 10 trending papers
For any with >50 GitHub stars in the first 7 days: add to queue
Star repos you want to return to
Deep-Read: 1 Paper Per Week (60 minutes, Saturday)¶
The one-paper habit is the highest-leverage practice in this entire system.
From your weekly queue (capped at 3), select 1 paper for a full 3-pass read. Execute the reading protocol from 01_how_to_read_papers.md. Take notes in your paper management system. Answer the 5 questions. Note the failure modes. If the paper has code: run the minimal reproduction.
Track this in a simple log:
## Weekly Paper Log
| Week | Paper | arXiv ID | 3-Pass Complete? | Can Reproduce? | Key Insight | Updated Opinion? |
|------|-------|----------|------------------|----------------|-------------|-----------------|
| 2025-W01 | DeepSeek-R1 | 2501.12948 | ✅ | Partial | GRPO + verifiable reward = emergent reasoning | Updated: GRPO > PPO for reasoning tasks |
After 13 months of this system: 52 deeply-read papers. That is more rigorous paper coverage than most PhD students in years 1-2.
Monthly System (2-3 hours/month)¶
Monthly Review: What Changed? (1 hour)¶
On the last Sunday of each month, review your weekly paper log. Ask:
What are the 3 most important things that changed in ML this month?
Did any of my predictions get confirmed or falsified?
Does my mental model of any sub-domain need updating?
What is the single most important thing to read next month?
This 1-hour investment compounds over 13 months into an extremely high-resolution picture of how the field evolves in real time. Most engineers cannot answer “what changed in ML last month?” with specifics. You should be able to answer that question in 90 seconds with three concrete claims.
Monthly State of the Field Summary (1-2 hours)¶
Write a 500-word private summary of the month’s developments. This does not need to be published. The act of writing forces synthesis. After 6 months, these summaries become the raw material for blog posts, talks, and interview preparation.
Format:
## ML Field Summary: [Month Year]
**Architecture**: [1-2 sentences on notable architecture papers]
**Training/Alignment**: [1-2 sentences]
**Inference/Efficiency**: [1-2 sentences]
**Surprising result**: [1 finding that changed your prior]
**Prediction update**: [Any prior prediction updated this month?]
**Next month focus**: [What to watch]
Quarterly: Calibration Audit (1 hour/quarter)¶
Review your prediction log from 04_developing_technical_opinion.md. Grade each resolved prediction. Calculate calibration: for predictions you gave 70% confidence, how many were correct? It should be roughly 70%. If your 70% predictions are correct 95% of the time, you are under-confident and should state stronger claims. If they are correct 40% of the time, you need to ground your opinions more carefully in evidence.
The Information Diet Rules¶
Accept these constraints:
Daily reading ≤ 30 minutes. Anything longer is procrastination with intellectual branding.
Queue ≤ 3 papers at any time. Add one only when you complete one.
Follow only accounts that make you think, not accounts that make you feel informed.
One deep-read per week. Non-negotiable. Even if it’s a short paper.
The signal sources, ranked by reliability:
Primary source: arXiv paper itself
Peer with known technical taste who posts analysis (not just a link)
Newsletter from a practitioner with a track record of correct priors
Twitter/X thread from an account on your curated list
General tech news site writing about AI
Read in that order. Never substitute tier 5 for tier 1.
Return to README.md · Previous: 04_developing_technical_opinion.md · Next: 06_interview_and_assessment_mastery.md