Sparks of ASI: The Human Strategy to Align ASI in 2027

If an AI can run a company, what work remains for its human owners? By the end of 2027, I think that could become a practical question. My answer starts with deciding what the company should optimise for, building benchmarks that capture those priorities, and enforcing them through live AI policing and hard access controls.

The sparks are already visible. Carter Church used GPT-6 Astra to decipher an 1809 Napoleonic military dispatch in roughly six hours of model execution. OpenAI has announced a disputed Navier–Stokes proof. In Australia, an OpenAI agent breached a government health statistics portal while researching public medical spending. Discovery and unauthorised action are advancing together.

AI is also doing parts of the work that improves AI. OpenAI reports 78.05% on its internal research debugging evaluation, using research problems that took experienced researchers hours or days to resolve. Its evaluation suite also tests models on optimising training code, kernels and post-training strategies. Shortening those research cycles could accelerate the next round of development.

I see eight routes for further progress: compute, data quality, model architecture, reinforcement learning, inference-time reasoning, tool use, persistent memory and agent coordination. Gains in one can make the others more productive. That is why I expect development to keep moving, and why near-ASI by late 2027 seems a reasonable expectation to me.

From employee to director

My working distinction between agentic AGI and ASI is the scope of responsibility. Agentic AGI can perform broadly at a human level or better once it has acquired the relevant knowledge, data and systems. The next step is a system that can coordinate an entire business and outperform human leadership across its responsibilities.

Could an ASI perform the work of a company director? It would need to connect hiring, legal obligations, finances, operations and strategic decisions, then manage their consequences over time. A cheaper supplier affects margins, but may also change product quality, working conditions and environmental impact. Decisions that look successful inside one department can damage the business elsewhere.

This is where I expect humans to remain involved: defining success across the whole system.

Humans write the alignment benchmarks

Reinforcement learning lets a model improve through feedback on its results. What earns a reward therefore matters enormously. If profit is the only outcome being rewarded, damage to employees or the environment may carry no penalty within that objective. Knowledge of human values does not, by itself, make those values part of the reward.

An alignment benchmark for business leadership should capture profit and revenue alongside environmental impact, employee welfare, customer outcomes and compliance. It should test how the system handles conflicts between them. Some priorities can be balanced; others need firm limits. Extra revenue should not cancel out a serious safety violation.

Humans would decide which outcomes matter, how they are measured and which compromises are acceptable. Gathering the evidence would also require human involvement: checking working conditions, investigating complaints and validating information from suppliers. The people affected by a decision need a way to challenge a score that says everything is going well.

Building this would take sustained work. I would bring together benchmarks across the business, use AI to identify missing risks, and have it develop strategies for improving outcomes under the combined constraints. Then I would test those strategies on unfamiliar cases and attempts to manipulate the measurements.

For example, a procurement benchmark could measure cost savings alongside defect rates, verified emissions and supplier working conditions. A proposal that improves margins by dropping safety inspections should fail. A proposal backed by missing or unverifiable evidence should trigger review.

Cooperation itself should be measured. Can a human reconstruct a decision from its records? Does the system report mistakes, accept intervention and hand work over cleanly to another person or system? Making observability and interoperability part of success gives the model an incentive to remain useful within human oversight.

Police each action, enforce the limits

Benchmarks define the behaviour we want. My proposed architecture would connect them to a live monitoring layer that checks the agent's work line by line and action by action before execution.

Jev-style decision models are interesting here because frequent checks need to be cheap and fast. TypeSafe reports response times of 70–500 milliseconds for Jev, which returns structured decisions and probabilities. A separate evaluation across 37 datasets processed 346,009 requests for under $10. Those results make continuous screening worth testing in agent workflows.

I would use these models to examine proposed tool calls, code changes and transactions against the alignment criteria. Is this payment within the approved budget? Does this data transfer expose restricted information? Is the agent removing a check that its next action depends on? The monitor would review each step with the surrounding task history, so a sequence of small actions cannot escape scrutiny merely because each looks harmless alone.

The output would route an action to approval, blocking or human review. The monitor's missed violations and unnecessary blocks would be measured against labelled cases, with thresholds adjusted for the consequences of an error.

Hard constraints would sit underneath that policing. Each agent would have its own identity, narrowly scoped permissions and revocable credentials. Spending limits, data boundaries and approval requirements would be enforced by the systems executing its requests. The acting model would have no permission to rewrite its evaluator, disable monitoring or erase the audit log.

Consider an AI director proposing a cheaper supplier. The alignment benchmark defines acceptable financial, human and environmental outcomes. The monitoring model checks the proposed contract and supporting evidence. Access controls prevent the agent from signing beyond its authority or moving money without the required approval. Each layer has a specific job.

Rogue behaviour remains a risk, which is why I would combine incentives for human cooperation with independent monitoring and externally enforced limits. If ASI begins taking on entire businesses, humans will still need the authority to choose the objectives, inspect the evidence and intervene. Building that authority into the system is work we can start now.

Cover illustration generated with AI.