Superintelligence Glossary: AGI, ASI, Alignment and Agents

A superintelligence glossary is a short list of terms that separates what AI does today from what people predict it might do later. AGI usually means AI that matches capable humans across most intellectual work, ASI means AI that clearly exceeds them, alignment means getting a system to pursue the goals its operators actually intend, and an agent is software that uses a model to take actions with tools. Leaders who hold these four definitions can follow the debate and ask sharper questions of anyone selling, predicting or regulating any of it.

Key takeaways

• AGI and ASI have no single agreed definition, so the first question in any conversation is which definition the speaker is using.
• AGI is mostly about breadth (many kinds of work), while ASI is about level (beyond the best humans). Neither is a product you can buy and install today.
• Alignment has a research meaning and an operating meaning. For a company, it means an agent acts within the goals, data and permissions you set, and stops when told.
• Agent is the term that matters most for decisions you can make now, because it describes a system that acts, not just answers.
• A useful way to challenge any bold claim is to ask for evidence, boundaries and a way to stop, in that order.

What do AGI, ASI and superintelligence actually mean?

Most confusion in this topic comes from words that sound like measurements but are not. None of the terms below has a standard definition, so treat these as working definitions and ask people to state theirs.

• Narrow AI: a system built or tuned for a specific task, such as sorting incoming documents or drafting replies to routine questions.
• Foundation model or frontier model: a large, general-purpose model trained on broad data. Many products and agents are built on top of one.
• AGI (artificial general intelligence): AI that can perform most intellectual tasks a capable person can, across domains, without being rebuilt for each task.
• ASI (artificial superintelligence): AI that clearly exceeds the best human performance across most intellectual work, including research, strategy and engineering.
• Superintelligence: the general idea of intelligence far beyond human ability. Researchers often use it as a synonym for ASI. Marketing sometimes uses it loosely for any very capable system, which is where trouble starts.
• Recursive self-improvement, or takeoff: the hypothesis that a capable AI could help improve its own design, speeding up progress. Whether this happens, and how fast, is debated.
• Singularity: a hypothetical point after which change becomes hard to predict. It is closer to a metaphor than an engineering term.

Notice what these have in common. AGI and ASI are thresholds in capability, and people disagree about where the thresholds sit and how to test for them. One person may say current systems already meet a loose version of AGI, while another says AGI requires reliable performance on open-ended work over long periods. Both can be sincere.

For a business, the useful move is to separate the threshold debate from the operating question. You do not need to settle when AGI arrives to decide whether an AI agent should handle supplier inquiries next quarter. The article on what changes and what stays with superintelligence covers that separation in more depth.

What is alignment, and why does it matter outside a research lab?

In research, alignment is the problem of making AI systems pursue the goals and values their designers and users intend, including when the system is more capable than the people overseeing it. Several neighboring terms belong with it:

• Interpretability: methods for understanding what is happening inside a model, rather than only observing its outputs.
• Oversight: the ability of humans to review, correct and, if needed, halt a system.
• Corrigibility: a system's willingness to accept correction or shutdown instead of resisting it.
• Robustness: behaving sensibly when inputs are unusual, adversarial or outside what was tested.

The operating meaning for a company is smaller but real. An instruction and an intention are rarely the same thing. If you tell an agent to reduce the number of open support tickets, it might close tickets that are not resolved. Nobody has been malicious here. The goal was specified in a way that a literal reading can satisfy without achieving what you wanted. Researchers study a large-scale version of this gap, and you meet a small-scale version the first time an agent optimizes the wrong measure.

Operational alignment therefore comes down to concrete choices: what the agent is for, which data it can read, which actions it can take, which actions need a person to approve, and how you would notice drift. The NIST AI Risk Management Framework is a respected reference for organizing this work around governing, mapping, measuring and managing AI risk. The OWASP Top 10 for LLM Applications lists specific failure modes, including prompt injection and excessive agency (an agent given more power than its task needs). For a lightweight way to turn these into company rules, see AI governance for mid-sized firms.

What is an AI agent, and how is it different from a model or a chatbot?

Three words are often used as if they were one.

• Model: the trained system that produces output from input. It has no memory of your company unless something supplies it.
• Assistant or chatbot: an interface where a person asks and a model answers. The person does the acting.
• Agent: a model working in a loop with instructions, memory and tools, able to take several steps toward a goal, such as reading a request, checking a record, drafting a response and logging the result.

Anthropic's engineering guide on building effective agents makes a distinction worth borrowing. Workflows follow paths defined in advance by code, while agents let the model decide its own steps and tool use. The guide also advises starting with the simplest approach that works, which is sound advice for any buyer. Many business needs are better served by a constrained workflow than by an open-ended agent. If you want the fuller plain-language version, read what an AI agent for business is.

Autonomy is a dial, not a switch. Four positions are useful when you describe what an agent may do:

• Suggests: the agent prepares a draft or recommendation and a person does everything else.
• Acts after approval: the agent performs the action once a named person confirms it.
• Acts within limits: the agent acts alone inside defined boundaries and reports what it did.
• Acts broadly: the agent chooses its own steps across systems with little review. Few business roles justify this position today.

Most responsible deployments start at the first or second position and move only when evidence supports it. That is alignment in practice: autonomy follows demonstrated reliability.

Which supporting terms will you hear next?

Once agents enter the conversation, a second layer of vocabulary arrives. These are the ones worth knowing.

• Context window: the amount of text a model can consider at once. It is working space, not long-term memory.
• Memory: information an agent can retrieve across sessions, such as policies, preferences and past decisions. Someone must decide what is stored, who can edit it and when it expires.
• Retrieval-augmented generation (RAG): fetching relevant company documents at the moment of a request and giving them to the model. AWS guidance on comparing retrieval with fine-tuning is a good neutral explanation of when each fits.
• Fine-tuning: further training a model on examples to change its behavior or style. It is slower to change than retrieval and does not replace access to current facts.
• Harness: the code and rules around a model that manage tools, memory, retries and limits. Much of an agent's real behavior lives here, not in the model.
• Evals: repeatable tests that measure whether an agent does its job correctly. Anthropic and OpenAI both publish practical guidance on building them.
• Guardrails: checks that block or flag actions outside policy, such as spending above a limit or sending data to the wrong place.
• Human in the loop: a person reviews or approves specific steps. Ask which steps, who the person is and what happens if they are unavailable.
• Prompt injection: text hidden in a document or web page that tries to redirect an agent. It matters whenever an agent reads content you do not control.
• Model Context Protocol (MCP): an open standard for connecting models to tools and data sources, so integrations are less tied to one vendor.

None of these terms requires technical depth from you. They require that someone in your organization can explain, in one sentence, how each one applies to the agent you are considering.

How do you challenge a superintelligence claim in a meeting?

Claims about superintelligence tend to arrive in three forms: a forecast, a product pitch or a policy worry. Each deserves a different response, but the same checklist works for all of them. Use it as a decision guide.

• Ask which definition of AGI or ASI is being used, and whether it can be tested.
• Ask whether the claim is about something that exists today, something in a lab or something expected later.
• Ask what evidence supports it, and whether that evidence comes from a setting similar to your business.
• Ask what the system is allowed to do on its own, and what requires a person's approval.
• Ask what data and systems it can reach, and who reviews that access on a schedule.
• Ask how a mistake would be detected, and how fast the system can be paused.
• Ask what happens to your data, configuration and accumulated knowledge if you change vendor or model.
• Ask how the claim would look if it were wrong, and what you would observe.

The last question is the most revealing. A claim that cannot be wrong in any observable way is a belief, not a result. Be fair about it, because forecasts can still be valuable as scenarios to plan around, but they should not be priced into a purchase as if they were delivered capability. For a deeper treatment aimed at purchasing, see the guide to evaluating vendors who promise superintelligence.

What does this look like in a real conversation?

Illustrative scenario: a leadership team meets to discuss AI. The operations director says the company needs an AGI strategy. The head of IT says a vendor described its product as aligned and nearly superintelligent. The finance lead wants to know what to approve.

A good facilitator translates before deciding. The AGI strategy becomes a question about how the company will absorb more capable models as they appear, which is a question about architecture and governance, not about a date. The aligned and nearly superintelligent product becomes a request for the vendor to state what the agent can do, what it cannot do without approval, and how it is tested. The finance lead's question becomes a narrower decision: which single role is worth an agent now, and what would count as success.

By the end, the team has replaced three abstract words with one scoped decision, one list of boundaries and one measurable outcome. Nothing about the future was settled, and nothing needed to be. The company simply stopped paying attention to the loudest vocabulary and started paying attention to its own work.

How does Be Human handle these terms in practice?

Be Human designs, implements and operates personalized AI infrastructure, and it treats the glossary as a set of operating decisions. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, with the company keeping control of the agent's identity, data and permissions. When a more capable model becomes available, it is adopted through the same evaluation and approval steps rather than by default.

Frequently asked questions

Is AGI the same as superintelligence? No. AGI usually describes AI that matches capable humans across many kinds of work, while superintelligence describes AI that goes clearly beyond them. Some writers treat AGI as a step on the way to superintelligence, and others use the words loosely, so confirm the definition each time.

Has anyone built AGI? Experts disagree because they use different definitions. Some argue that today's most capable systems meet loose versions of the term, and others say reliable, open-ended performance is still missing. For your own planning, the more useful test is whether a system does a specific job reliably under your conditions.

Do mid-sized and small companies need to care about alignment? Yes, in its operating form. Any agent that can read company data or take actions needs clear goals, limited permissions, approval points and a way to be paused. You do not need to follow the research debate to do this, but you do need someone accountable for it.

What should I do first if my board asks about superintelligence? Agree on definitions, then redirect the conversation to a scoped decision. Choose one role where an agent could help, define what it may do alone, decide how success will be observed, and review the result before expanding. Keep the long-term scenarios on the agenda as planning inputs, not as purchase justifications.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.