Watch AI run a chess-commentary YouTube channel entirely on its own: the engine can play, the LLM can talk, and combining them is the holy grail.
Running a Chess YouTube Channel entirely by AI — Stephan Steinfurt, TNG · Stephan Steinfurt
16 min total·Actually worth watching closely: ~4 min·2 must-watch clips
- 0:00 – 1:04Listen
Opening: the holy grail of chess programming
The core premise: engines have been extremely strong for decades but can't explain chess, LLMs have the words but can't play — combine them and you get the holy grail, AI that explains chess as well as a human trainer.
Combining complementary systems is itself the product innovation; there's no need to wait for one model that does everything.
Purely spoken argument with nothing critical on screen — fine to listen to like a podcast.▶ Jump to 0:00Speaker · Stephan Steinfurt - 1:04 – 3:10Watch
The output: a fully auto-generated commentary video
A two-minute chess commentary created entirely by the AI engine is played live, running from tactical narration through variation analysis to a closing call to subscribe.
The output already reaches publishable, watchable quality — which is what makes the architecture discussion that follows worth having.
The first visual moment flagged by the LLM: the video demo carries its point visually in a way text can't replace, so watch it.▶ Jump to 1:04Speaker · Stephan Steinfurt - 3:10 – 7:03Listen
Tool design: productizing the human coach
A walk through the agent's toolbox: legal moves to keep it from thinking about anything illegal, a full board for playing moves, taking them back and exploring variations, a chess engine always on call, checks, captures and threats taken from the beginner's checklist, plus web search for historic context.
A good tool set isn't a pile of APIs — it translates a domain expert's way of thinking into constraints the model works within.
Mostly spoken design rationale with supporting lists on screen; following the logic is enough.▶ Jump to 3:10Speaker · Stephan Steinfurt - 7:03 – 9:20Watch
The architectural turn: who does the thinking, and predicting human moves
The project moved from Python scripts that analyzed positions and assembled information to reasoning models that think about the position themselves and call tools at the right moment; alongside that, the Maya engine predicts by rating what a human would really play, bad moves included, so commentary no longer revolves solely around the engine's best move.
Once reasoning models arrived, the orchestration logic belongs to the model rather than the code — "who should do the thinking" is the dividing line for agent architecture.
A visual moment is flagged at the 429-second mark: the tool shows the moves a human might consider, such as the obviously bad bishop capture on a4 and the non-obvious rook sacrifice on e3. The board example is what makes the design click.▶ Jump to 7:03Speaker · Stephan Steinfurt - 9:20 – 11:40Skim
Model choice and the production pipeline
How much chess the base model knows decides whether the approach works: Grok 4 was the strongest last autumn, while Gemini 3.1 Pro now understands chess markedly better. On the pipeline side, the analysis is turned into a special format that renders into video, ElevenLabs v3 does the speech with emotion tags, and the agent decides highlights and arrows by itself.
Whether an agent approach holds up depends first on the base model's domain ability, so model choice needs re-testing on a regular cadence.
Mostly slides listing model comparisons and pipeline components; skim the visuals for the conclusions.▶ Jump to 9:20Speaker · Stephan Steinfurt - 11:40 – 14:04Listen
Operating reality: numbers, error rate and trade-offs
Around 500K views and more than 4,000 subscribers with growth accelerating, but monetization hasn't been reached so it's net minus; with automatic uploading on, roughly every 20th video has an obvious problem, and the policy is to publish first and take the bad ones down afterwards.
The workable answer for fully automated content production is accepting a bounded error rate and trading publishing speed for iteration, rather than chasing zero errors.
The numbers and trade-offs are all spoken and there aren't many of them — no need to watch the screen.▶ Jump to 11:40Speaker · Stephan Steinfurt - 14:04 – 16:13Listen
Q&A: cost, human-like moves and audience tiers
Answers to questions about how much to prioritize cost optimization, making engine moves look more human, and generating different videos for different rating levels.
The same game can become different commentary versions for different audience strengths — personalization at scale is where this kind of product ends up.
Pure Q&A with nothing visual — fine to listen to on the go.▶ Jump to 14:04Speaker · Stephan Steinfurt