Most people who want to understand AI are stuck in a bad spot. The academic papers are impenetrable. The Twitter takes are either hype or doom. The mainstream tech journalism is three weeks late and usually misses the actual interesting part. That’s the gap AI Explained — the YouTube channel run by an anonymous creator who goes only by the initials AE — has quietly filled since around 2022. As of early 2026, it sits at roughly 700,000+ subscribers and has become one of the most-cited “how I actually understood that paper” sources among developers, researchers, and curious generalists alike. If you’ve ever wanted to understand what a model architecture actually does, why a benchmark matters or doesn’t, or what the real-world implications of a new release are — this is the channel that does it without either dumbing it down or burying you in notation.
What AI Explained Actually Is (And What It Isn’t)
Let’s be precise. AI Explained is not a news channel. It doesn’t do weekly roundups, product reviews, or interview-style content. It publishes infrequently — sometimes only a handful of videos a month — and each one is a deep, structured breakdown of either a specific paper, a model release, or a conceptual question that the broader AI conversation is getting wrong. The editing is clean but minimal. There’s no studio setup, no dramatic music drops, no clickbait thumbnails with shocked faces. The creator seems to understand that the content itself is interesting enough if you explain it correctly.
What makes it useful is specificity. When GPT-4o was released, AI Explained didn’t just say “it’s multimodal and fast.” It walked through what real-time audio processing actually requires architecturally, what the difference is between cascaded systems (separate speech-to-text, reasoning, text-to-speech) and a true end-to-end model, and why that distinction matters for latency and emotional tone. When Anthropic published the Claude 3 model card with its interpretability findings, AI Explained covered what the sparse autoencoder results actually showed — not just that Claude was “thinking” but what features were being activated and why that’s interesting or limited depending on your priors.
It’s not a podcast. It’s not a newsletter. It’s closer to having a grad student friend who reads every major paper and then explains it to you on a whiteboard — except the whiteboard is a screen recording and the friend is extremely online.
The Videos That Define the Channel’s Approach
If you want to understand what AI Explained does well, start with a few specific examples rather than the vague “high quality content” claim everyone makes.
The Gemini 1.5 Pro breakdown is one of the best explanations of long-context models available anywhere. The video doesn’t just report the 1 million token context window as a stat — it walks through what that actually enables, where the “lost in the middle” problem still exists even with long context, and why retrieval-augmented generation isn’t necessarily obsolete just because context windows got bigger. These are the kinds of nuanced calls that most coverage gets wrong because most coverage is working from the press release.
The o1 and reasoning models series is where AI Explained earned significant credibility. When OpenAI released o1 in late 2024 and the discourse immediately split between “AGI is here” and “it’s just chain-of-thought prompting,” AI Explained took the harder middle path: explaining what “test-time compute” actually means, why the distinction between training-time and inference-time scaling matters, and what the real benchmark results showed versus what people were projecting onto them. This was the same conceptual territory Andrej Karpathy was covering on his own channel and posts around that time, and the two creators occupy similar intellectual space — technically rigorous, willing to say “we don’t fully know yet,” and uninterested in scoring points for either the bulls or the bears.
The interpretability content — particularly around Anthropic’s mechanistic interpretability work and what it does and doesn’t tell us about model internals — is genuinely hard to find explained well anywhere else. This is research that matters for alignment, for safety, and for anyone trying to understand whether current approaches to understanding LLMs are working. AI Explained treats it seriously without overstating what’s been proven.
How It Compares to Other AI Education Channels
There’s a real ecosystem of AI education content now, and understanding where AI Explained sits requires being honest about what the alternatives offer.
| Channel / Creator | Depth Level | Update Frequency | Best For | Limitation |
|---|---|---|---|---|
| AI Explained (AE) | High — paper-level | Low — 2–6/month | Understanding specific releases and research deeply | Doesn’t cover everything; slow cadence |
| Andrej Karpathy | Very High — implementation-level | Very Low — sporadic | Building intuition from first principles | Long-form; not suited for quick updates |
| Yannic Kilcher | High — paper-walkthroughs | Medium | Developers and researchers wanting full paper reads | Can be dense; less narrative structure |
| Two Minute Papers | Low — summaries only | High | Awareness of what’s being published | Surface-level; lots of enthusiasm, little depth |
| Lex Fridman | Medium — interview-dependent | Low | Long-form conversation with major figures | Varies wildly; often more philosophical than technical |
| Matthew Berman | Low-Medium — demo-focused | Very High | Seeing tools in action quickly | More product demo than conceptual understanding |
AI Explained is not trying to compete with any of these. It occupies a specific niche: the person who wants more than a summary but isn’t trying to implement the paper from scratch. That’s a large audience that was genuinely underserved before this channel existed.
What the Channel Gets Right That Most AI Coverage Gets Wrong
There are a few recurring patterns in how AI Explained handles topics that are worth naming explicitly, because they explain why the channel is useful rather than just popular.
It takes benchmarks seriously but doesn’t treat them as gospel. When a new model drops and its benchmark scores look impressive, AI Explained typically asks: what is this benchmark actually measuring? Is it saturated? Has this model likely been trained on the test set? This is the kind of critical framing that Francois Chollet has been pushing for years with his