The firehose problem
arXiv publishes hundreds of ML papers per day. Twitter / X serves a constant stream of breathless threads. Hugging Face uploads dozens of new models every week. If you try to read everything, you'll read nothing carefully and ship nothing at all.
A sustainable filter
- Two or three trusted curators — Sebastian Raschka, Andrej Karpathy, ML Collective, Yannic Kilcher, the Hugging Face blog. Their picks save you 90% of the firehose.
- One or two domain newsletters — Ahead of AI, Import AI, the Gradient. Weekly summaries you can read in 15 minutes.
- One paper per week, read deeply — pick something relevant to a problem you're actively solving. Reproduce one experiment. The depth beats the breadth.
- Build, don't just read — every concept you learn from a paper, implement on a toy problem. The hands-on time is what makes the knowledge stick.
Tip: The compounding skill in ML isn't reading more papers — it's the habit of going from a paper to working code on a toy problem in a single afternoon. Practice that loop and you outpace people who only read.
What to ignore
Architecture papers that don't release code or weights. Benchmark wins of 0.1% over the previous SOTA. Anything claiming AGI in the next 18 months. Twitter threads with 'this changes everything'. The signal-to-noise on these is brutal.
The longer view
Deep learning's fundamentals (gradient descent, backprop, attention, normalization) haven't changed in years. The headlines change weekly. Invest your reading budget in fundamentals; skim headlines for the rare nuggets that affect your work.
Principle: Curiosity is fuel; FOMO is poison. Pick the curators, build the projects, and ignore the rest. The field rewards depth and shipping, not feed-scrolling.