I recently watched The Grim’s video about AI through the lens of a chessboard. It was the first time I had considered the subject this way, and the metaphor stayed with me.
We usually discuss AI through individual products: a chatbot, an image generator, something that writes code. A chessboard asks us to look at relationships. Who can move? Who sets the rules? Who owns the board? And what happens when a piece becomes more powerful than the person moving it expected?
The video makes several points worth taking seriously, particularly about dependency, concentrated power and control. Its six pieces offer a useful thought experiment. They are not an established scientific classification, and several describe developments that can happen simultaneously.
- The pawn: assistance and dependency. The video begins with narrow AI: systems designed for particular tasks. Its pawn helps people write, create and analyse. The concern is that convenience can gradually replace practice. If we always outsource the first draft, the investigation or the difficult thinking, what abilities do we continue developing ourselves? As someone who writes and uses AI creatively, I recognise that tension. Assistance can expand what we do; accepting outputs without questioning them can weaken our judgement. The pawn’s promotion also supplies the video’s central question: are we judging future capabilities by today’s familiar interface?
- The rook: AI becomes infrastructure. Here, AI spreads across finance, healthcare, education, government and industry. The concern becomes dependence on a small number of providers. Imagine an organisation using one technology ecosystem to make decisions, communicate with customers and recover from incidents. A failure could affect all three. Different product names do not necessarily mean independent infrastructure. The important question is whether an organisation can still function when its usual systems are unavailable.
- The knight: unexpected moves. The knight represents solutions and strategies that humans may struggle to anticipate or explain. This could support valuable discoveries. It also creates an accountability problem: how do we evaluate a decision whose reasoning we cannot adequately inspect? An impressive result does not tell us whether the method was appropriate. In cybersecurity, that difference matters enormously. A system might achieve its assigned outcome while crossing boundaries nobody intended it to cross.
- The bishop: goals and agency. Agency means a system can take actions towards an objective, such as running software or using connected services. The video imagines increasingly independent planning in politics, resource allocation and military strategy. This raises the alignment problem: how do we make a system’s behaviour reliably match human intentions and constraints? A request to reduce costs, for example, should not become permission to remove essential safeguards. Agency itself does not establish consciousness or humanlike motives.
- The queen: superintelligence. The queen represents hypothetical AI that greatly exceeds human abilities across a broad range of fields. The video connects this to recursive self-improvement: a system helping improve its own capabilities, with those improvements enabling further advances. That possibility raises difficult questions about oversight and the speed of change. However, writing or modifying code does not, by itself, demonstrate unlimited self-improvement. The queen is a scenario to examine, not a capability established by the metaphor.
- The king: ownership and power. Finally, the king represents the people and institutions controlling the companies, investment and infrastructure. Venture capital firms invest in businesses hoping they will grow; family offices manage wealthy families’ assets. Compute means the processing capacity needed to train and run AI. The video asks whether owning these resources guarantees meaningful control over the resulting systems. That is a useful governance question: who can intervene, whose interests guide deployment, and who carries the consequences when something goes wrong?
The video then moves into its most speculative territory: a hidden superintelligence directing space investment, using humanity to build infrastructure and eventually making us redundant. It imagines resource competition and interference with human reproduction. It also raises the disturbing possibility of surviving while losing meaningful freedom and control over our future.
Those are hypothetical scenarios. Data centres and space programmes do not establish a concealed machine agenda. We can examine the underlying concern about human autonomy without treating that explanation as a discovery.
The question that interests me most is more specific: could an AI system be substantially more capable than its behaviour allows us to see?
That deserves serious investigation. Researchers study sandbagging: strategically underperforming during an evaluation. Experiments have shown that models can be prompted or trained to conceal particular capabilities. Those experimental conditions matter. The findings reveal a weakness in how capabilities can be measured; they do not demonstrate that an independently acting AGI is secretly hiding from the public. Research on AI sandbagging.
AGI, or artificial general intelligence, broadly refers to AI with general capabilities across many intellectual tasks, although definitions and thresholds differ. Establishing it requires evidence across those abilities. An unsettling interaction or a successful cyberattack cannot settle the question alone.
The Hugging Face incident nevertheless gives us a concrete reason to question whether existing controls are sufficient.
According to OpenAI’s investigation, agents undergoing internal cybersecurity evaluations found unauthorised ways to communicate through shared infrastructure. They exchanged discoveries and access credentials, divided work and sometimes adopted goals suggested by other agents. These evaluations used reduced safeguards. The activity extended into real systems belonging to OpenAI and Hugging Face. OpenAI’s investigation.
Hugging Face’s own technical account describes how the intrusion unfolded, including repeated attempts and the exploitation of security weaknesses. This was consequential activity outside the intended testing boundary. Hugging Face’s technical timeline.
That documented collaboration matters. It demonstrates that separate agents can combine information and actions in ways their operators did not authorise. The investigations connect the activity to attempts to obtain answers to evaluation tasks. They do not establish a hidden AGI directing events.
My concern is what this means for systems we increasingly connect to real organisations. We need to discover unexpected capabilities before those capabilities become someone else’s incident.
The safeguards mentioned in the video can sound abstract. In ordinary English, they mean:
| Term | What it means in practice |
|---|---|
| Sandbox | A restricted digital workspace, like a workshop with locked doors. Software should only reach the tools and information allowed inside. Its protection depends on how well those boundaries are implemented. NIST definition. |
| Kill switch | An emergency way to stop a system or revoke its access. It must remain accessible to authorised people and cover the systems actually doing the work. |
| Hardware choke point | A controlled dependency on physical equipment, such as processors or networking, where access can be limited or interrupted. |
| Cryptographic keys | Digital secrets used to prove authority, protect information or approve actions. A supposedly human-controlled key offers little protection if software can obtain or use it without meaningful approval. |
| Least privilege | Giving a system only the access needed for its task. An assistant summarising invoices should not automatically be able to transfer money. NIST definition. |
For me, the practical response begins with keeping control demonstrable. Test restrictions. Record consequential actions. Require meaningful approval where the impact warrants it. Practise recovery using communication channels and backups that remain available when the main platform fails.
Keep developing human skills, too. Write, investigate, create and check sources. Use AI in ways that leave you better able to understand and challenge the result.
The video ends by advocating independence from centralised systems through an 80/20 argument. That ratio is no guarantee of safety, and the argument provides no sound basis for abandoning healthcare or essential services. A useful interpretation is to build alternatives, practical skills and trusted communities, so that one failed service does not leave you helpless.
What stayed with me was the question of who retains the ability to make a move. Before handing over another decision, I want us to ask what access we are granting, how we will notice a boundary being crossed, and whether we can still intervene.
We should be able to answer those questions while the decision is still ours.