AI Engineer Europe 2026
Running a Chess YouTube Channel Entirely by AI — Stephan Steinfurt, TNG Technology Consulting
Loading slides…
Read the talk
Running a Chess YouTube Channel Entirely by AI
Stephan Steinfurt’s chess-video pipeline combines engine analysis, model-directed exploration, synthesized narration, and board annotations to explain real human games.
From a talk by Stephan Steinfurt
Before you start: Basic chess notation helps: R means rook, N knight, Q queen, x a capture, + check, and an ellipsis marks Black’s move.
Can AI explain chess like a human trainer?
Can an AI explain why a chess move works well enough to teach a human player? That is the ambition behind Stephan Steinfurt’s automated chess videos. After taking Swyx’s advice to avoid boring titles, he borrowed a more ambitious description from a German newspaper: the Holy Grail of chess programming. The newspaper attributed that phrase to Wilhelm Weber and suggested that explanations matching a human trainer could still be five years away. Steinfurt’s Munich team had already shown it entirely AI-created videos; he opens with one of those generated lessons.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
One sacrifice, two ways to lose the queen
The generated lesson begins with Black’s queen under attack from a white rook on the h-file. Black retreats with ...Qg4, apparently reaching safety. The narration calls this a crushing blunder: leaving the h-file allows Rxh5, offering the rook in a trap built around the knight’s next move.
First, the video explores the pawn capture. After ...gxh5, White plays Nf6+. The knight on f6 attacks both the king on g8 and the queen on g4. In the demonstrated continuation, Black’s king moves to f8 and White takes the queen with Nxg4. The video then rewinds to consider ...Qxh5: capturing with the queen draws it onto another square attacked by the same knight jump.
Black’s capture after Rxh5 | White’s reply | Demonstrated continuation |
|---|---|---|
...gxh5 | Nf6+, attacking king and queen | ...Kf8 Nxg4 |
...Qxh5 | Nf6+, attacking king and queen | ...Kf8 Nxh5 |
The shared mechanism is a fork with check: f6 attacks g8, g4, and h5. The check makes the king the immediate problem while the queen remains attacked. The second branch adds attraction—the sacrificed rook draws the queen onto h5. This is the explanatory work the video must do beyond announcing that Rxh5 is strong.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From nightly games to a finished lesson
This particular video was an overnight result. Steinfurt had prepared a different example, but his morning check for embarrassing output turned up a lesson he liked better. The system was already creating videos and uploading them to YouTube automatically each night.
The production sequence has four stages:
- Download human games from Lichess each night.
- Analyze them in the background, then let an agent investigate more deeply.
- Convert the analysis into a special intermediate format for video creation.
- Produce a video that explains variations and identifies brilliant moves and blunders.
The intermediate format separates the analysis from the final audiovisual presentation. The agent’s investigation must yield enough structure to show a variation while explaining why its moves matter.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Give the language model a board it can investigate
The underlying mismatch is straightforward: chess engines have played extremely well for decades, but do not explain their conclusions well. Language models can produce explanations, but cannot reliably play chess. The agent joins verbal reasoning to tools that supply chess facts.
At the time of the talk, Steinfurt considers Gemini 3.1 Pro the best model he has used for chess. He sees stronger chess understanding in its reasoning traces than in earlier models. His suggestion that this reflects chess-specific post-training is speculation, rather than a disclosed training detail.
The tools provide several distinct capabilities:
- Legal moves: Constrain exploration to moves allowed in the current position.
- Board state: Play a move, take it back, and investigate another variation.
- Engine evaluation: Ask a chess engine to analyze the position reached during exploration.
- Checks, captures, and threats: Expose tactical candidates the model might otherwise overlook.
- Historical context: Use web search for videos where the background of a game helps the explanation.
The board is therefore a working environment, not merely a position pasted into a prompt. The model can test an idea, inspect its consequences, and return to an earlier branch. The accompanying architecture slide places this model–tools loop alongside LangGraph / Deepagents.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Explain the moves a person might consider
The checks, captures, and threats tool turns a familiar beginner’s scanning method into information the agent can request. In the complicated position shown on the slide, the model receives more than the engine’s best move: it can inspect forcing candidates and explore why they succeed or fail. Among the checks, Steinfurt points to the queen taking a bishop on a4 as an obviously bad choice. Other queen checks are more reasonable.
The best move, he explains, is instead a rook sacrifice on e3. In another position, captures might be the more productive candidates to investigate. Returning these different possibilities through one tool broadens the agent’s exploration. A useful lesson must sometimes explain a bad move because a human would be tempted to play it. Supplying only the strongest continuation would omit that teaching opportunity.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Move the investigation into the agent
The first implementation put most of the analytical decisions in Python scripts. Those scripts inspected positions, assembled checks and engine evaluations, and passed the collected information to a language model for description. Reasoning models changed where that work could happen: the agent could decide which positions to investigate and which tools to call.
In the preceding autumn, Grok 4 had been the team’s strongest model for this task. Steinfurt also found enough baseline chess knowledge in OpenAI models to support useful tool selection and synthesis. That knowledge matters because the agent must recognize when it needs another fact, not simply turn a fixed bundle of facts into fluent prose.
| Analysis source | What it contributes |
|---|---|
| Engine analysis | Strong moves and their evaluations |
| Human-move tools | Plausible alternatives worth explaining |
| Historical material | Moves actually played and existing commentary |
These sources need not point toward the same move. Steinfurt finds that conflicting information helpful: it balances the best continuation against the moves a person might consider. Historical commentary can also supply a claim or variation for the agent to investigate.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turn the analysis into narration and board instructions
After analysis, the system converts its findings into the intermediate video format and uses ElevenLabs v3 for text-to-speech. Audio tags can steer delivery—for example, making a passage sound excited. The agent also chooses which squares to highlight, which arrows to draw, and whether a move merits a brilliant-move label.
A small Python representation illustrates how narration and annotations can stay together for the opening lesson’s Nf6+ moment:
python
from dataclasses import dataclass
@dataclass(frozen=True)
class VideoBeat:
move: str
narration: str
highlights: tuple[str, ...]
arrows: tuple[tuple[str, str], ...]
beat = VideoBeat(
move="Nf6+",
narration=(
"[excited] The knight checks the king on g8 "
"and attacks the queen on g4."
),
highlights=("f6", "g8", "g4"),
arrows=(("f6", "g8"), ("f6", "g4")),
)
This illustrative record expresses the separation of concerns: the analysis selects the move and its explanation; the presentation instructions identify what to emphasize while that explanation is spoken. It is not the project’s undisclosed format.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Scale commentary without losing the chess
Once production is automated, publishing volume becomes a choice. Steinfurt rejects the characterization of the output as slop, but acknowledges that the team could generate far more videos than it should necessarily upload. The input is human games, which opens a more personal use: a player could receive a narrated video of their own game and share it with friends or family.
The team prioritizes chess quality over effects such as exploding kings at checkmate. A popular human chess creator supplies the useful comparison: a popular human creator can explain games brilliantly, but is unlikely to cover any particular ordinary player’s game. Automation could make that kind of commentary available beyond elite players.
At the time of the talk, Steinfurt reports approximately 500,000 channel views and more than 4,000 subscribers, with most subscribers joining within the preceding month.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Revenue, errors, and the cost of extra investigation
The audience questions expose the operating conditions behind the automated channel. It has not yet reached YouTube’s monetization stage, so it earns no revenue and remains net negative. Whether that changes is still an open question.
Automation had been enabled only a couple of weeks before the talk. Steinfurt initially still wanted to watch videos before letting them publish. He informally estimates that roughly every twentieth video has a very strange description, such as missing a checkmate. That estimate has no stated sample size or formal evaluation procedure.
An error can reveal a missing step in the investigation: perhaps the agent failed to make a final tool call that would have exposed the checkmate. Such failures become debugging evidence. Steinfurt describes becoming more willing to let the system publish automatically and potentially remove a bad video afterward, rather than reviewing every result first.
Steinfurt estimates generation costs at roughly 20–30 cents per video, with much longer videos sometimes costing euros. He does not specify the currency of the cent estimate, a duration baseline, or a cost breakdown. The team is not yet optimizing costs aggressively because it prefers to spend more on a good explanation. Obvious opportunities remain: redundant tool calls, or an agent walking through the same game twice, add work without necessarily improving the lesson.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Which human is the explanation for?
A human-like candidate move could mean a grandmaster’s choice or a beginner’s mistake. Asked which kind the system considers, Steinfurt answers broadly: moves across skill levels. He points to Maia, associated with the University of Toronto, as a way to obtain plausible human moves for a player rating. The deployed version is unspecified; original Maia uses separate rating-band models, while Maia-2 accepts rating inputs. These predict behavior associated with ratings rather than guarantee a corresponding playing strength.
Exact imitation is not necessary for this application. A predicted move can be useful simply because it deserves an explanation. The goal is to enrich the candidate set with alternatives a learner might consider, not to prove that a particular player would choose them. The questioner connects this to teaching learners and people returning to chess.
Rating-aware information could also help target the videos themselves. A mate-in-one lesson may look pointless to a strong player, yet a beginner may need to see exactly why the king has no escape. The team has not settled which of these videos to publish or how to balance audiences. The remaining editorial problem is therefore concrete: deciding whose missing knowledge a lesson should address.
The final question asks whether the approach has been tried on other games. It has not. Steinfurt considers extension possible, but the demonstrated system remains a chess project.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Human-move prediction models trained for different rating bands, with weights and instructions for running them.
Google's documentation of the model's capabilities, limitations, and evaluation context.
How bracketed cues direct emotion, delivery, and vocal reactions in Eleven v3 narration.
Further reading
A unified chess model that conditions human-move predictions on player ratings, with inference examples.
The research paper accompanying Maia-2's approach to modeling human chess play across skill levels.
- Lichess open databaseDataset
Public chess data for building game-analysis and training workflows.
Updates since the talk
- TNG's chess-video projectArticle
TNG's July 2026 account of its chess channel, narrated puzzle videos, and use of real games.
Full transcript
Loading transcript…