Back to index
OtherOngoing, videos typically 15 to 30 minutes·Free, with a paid Patreon tier

AI Explained (YouTube)

4.3

The best of the AI news channels by a distance, because it actually reads the sources and refuses to inflate them. Still news rather than education.

What We Liked

  • Reads the papers and system cards rather than reacting to press coverage
  • Consistently separates benchmark results from marketing claims
  • Runs independent tests, including his own benchmark, instead of repeating vendor numbers
  • Calm delivery with none of the thumbnail hysteria the genre is known for
  • Videos are dense and edited tightly, so fifteen minutes carries real content

What Could Be Better

  • News coverage, not a curriculum, so there is no cumulative learning
  • Some analysis and the private benchmark sit behind Patreon
  • Pace is quick and technical terms arrive without explanation
  • Release driven, so coverage follows whatever the labs ship rather than what matters
  • Even careful AI news encourages consumption over practice

Detailed review

AI news on YouTube is a genre optimised for the worst possible outcomes. Capitalised thumbnails, claims of imminent transformation, videos published within an hour of an announcement by people who have plainly read only the tweet. Against that background, AI Explained is a genuine outlier, and the difference is straightforward: the host reads the actual paper, the actual system card, and the actual evaluation appendix before saying anything. The habit of separating what was demonstrated from what was claimed is the channel's defining quality.

When a model is released, the video will distinguish between the headline benchmark number, the conditions under which it was obtained, whether the comparison baselines were run fairly, and what the model does when he tries it himself. He maintains his own benchmark specifically to have a measure that has not been optimised against, which is a serious piece of work and a sensible response to the state of public evaluations. The tone helps more than it sounds like it should. Delivery is measured, conclusions are hedged where hedging is warranted, and uncertainty is expressed as uncertainty.

Compared to a genre built on breathless certainty, this reads as competence. He also covers negative results and disappointing releases, which nobody optimising for engagement would do. Coverage extends past model launches to safety and evaluation research, interpretability work, and papers that deserve attention but are not attached to a product launch. That breadth is welcome and is where several of the more interesting videos have come from.

The editing is tight. Fifteen to thirty minutes with almost no filler, which given the length inflation everywhere else is a real courtesy. What it is not is a course. This is news, and news does not accumulate into understanding.

Watching every video for a year would leave you extremely well informed about what happened and no more able to build anything than when you started. The correct use is as a current awareness layer on top of actual study, and the failure mode, which I have watched happen to people, is treating AI news consumption as a substitute for learning AI. Some of the deeper analysis and the private benchmark details sit behind Patreon. That is a fair way for an independent creator to make a living and it does mean the free tier is not the complete picture.

The pace assumes background. Terms are used at speed without definition, and a genuine beginner will follow the vibe and miss the substance. Start with foundational material and come here once you can place what is being discussed. Being release driven is an inherent constraint.

The agenda is set by whatever the labs ship, which means coverage tracks commercial activity rather than importance. Quieter research with more long term significance gets less attention than a model launch, and no news channel escapes this. Where it fits: AI Explained for what happened this week, Simon Willison for what a practitioner can actually do with it, Machine Learning Street Talk for whether the underlying assumptions hold. My four point three is high for a news channel, earned by intellectual honesty that is genuinely rare in the format, marked down because news is not education and this remains news.

[ final ]

The verdict.

The one AI news channel worth subscribing to, and worth about twenty minutes a week. Do not let it substitute for building anything.