# Is It AGI, or Is It Memorex? > Source: > Published: 2026-09-08 10:22:46+00:00 TL;DR — Key Takeaways - **OpenAI’s Astra has intensified the debate over whether AGI has arrived** , but impressive benchmark performance alone does not settle the question. - **Astra’s ARC-AGI-3 results vary dramatically depending on configuration** , highlighting why benchmark numbers need context rather than headline treatment. - **For enterprises, practical reliability, permissions, transparency and task-specific performance matter more than the AGI label itself.** - **Greater intelligence does not automatically guarantee reliability or alignment** , making governance and clearly defined authority essential for autonomous agents. - **Organizations could become dependent on AI long before the AGI debate is settled** , potentially losing skills, capacity and operational alternatives in the process. Those of us who remember buying blank cassette tapes also remember the question: [“Is it live, or is it Memorex?”](https://www.chiefmarketer.com/is-it-live-or-is-it-memorex/) The advertising invited us to wonder whether we could distinguish the original performance from the recording. It sold a technical achievement through a question that stuck in your head. Watching the discussion around OpenAI’s Astra, I found myself borrowing it. According to [Axios](https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman), Greg Brockman personally believes AGI has arrived and that Astra may mark the milestone. Is it AGI, or is it Memorex? Before the researchers start throwing things, the analogy has limits. I am not suggesting these systems simply replay their training material. A machine that can learn an unfamiliar task, reason through it and produce a useful result deserves to be assessed on those abilities. Calling everything it accomplishes an imitation becomes less persuasive as the accomplishments grow. But the advertising question gets at something worth examining. What does an impressive performance establish about the system behind it? And what does calling that system artificial general intelligence actually change for the customer? ### Give the Technology its Due [OpenAI describes Astra](https://openai.com/index/gpt-6-astra/) as a substantial advance in computer use, science, software engineering and professional work. There is enough supporting evidence to take the announcement seriously. [ARC Prize’s evaluation](https://arcprize.org/blog/astra) reports a 62.7% score on ARC-AGI-3 under its standard setup and 99.9% using a provider adapter that preserves reasoning between requests and manages long conversations. Those are different configurations, with different reasoning settings behind the best scores. Both set records under their respective conditions. Reporting the near-perfect number without explaining the setup leaves out something material. The benchmark tests how agents learn to navigate unfamiliar environments. Yet the ARC team explicitly says these results do not establish AGI. Its environments have bounded rules and objectives; they do not encompass the open-ended demands of the world outside the test. That caution coexists with enthusiasm. ARC creator François Chollet has [suggested AGI could arrive sooner](https://x.com/fchollet/status/2095607046129463577) than his previous 2030 expectation because progress is exceeding his expectations. Even [Gary Marcus’s skeptical assessment](https://garymarcus.substack.com/p/hot-take-on-gpt-6-astra) acknowledges impressive capabilities, while questioning how well they will extend to open-ended work. Other evaluations complicate the picture. [Artificial Analysis](https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra) finds improvements in coding efficiency and knowledge work, but gives Astra the same overall Intelligence Index score as GPT-5.6 Sol, below Claude Fable 5.1. That index cannot settle the AGI debate either. It does tell us that progress can be substantial without being uniform. For a more tangible example, [Ethan Mollick reports](https://www.linkedin.com/posts/emollick_an-example-of-useful-knowledge-work-i-assigned-activity-7501372639584751616-lhrx) that Astra spent five days assembling a personal knowledge base from his emails, writings, calendar and other information without further intervention. That is his early-access experience, rather than an independently replicated result. Still, sustained useful work is precisely the kind of capability customers should investigate. ### What Does AGI Promise the Customer? OpenAI’s [charter](https://openai.com/charter/) defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” That is an ambitious standard. It also leaves a business leader with plenty of questions about the particular work sitting on the desk. Imagine asking an agent to reconcile customer records spread across contracts, spreadsheets and a CRM. Some records conflict. One contract has been amended. Two companies share similar names. An old spreadsheet contains information that looks authoritative but is no longer current. The value lies in resolving those inconsistencies, identifying what remains uncertain and producing something the business can use. The customer needs to know how the agent reaches its conclusions and what happens when it encounters an exception. Access to the records also raises a separate question: Is the agent allowed to change them? An AGI designation answers none of that. It does not specify an error rate, establish access permissions or tell you whether checking the output will consume more time than doing the assignment yourself. Suppose one system completes this work dependably and another performs brilliantly on abstract reasoning tests but struggles with the records. Your choice should follow the assignment. You do not owe a product your business because its maker believes it represents a new chapter in human history. We should apply that standard fairly. Humans make mistakes, too. Demanding that AI be flawless before recognizing general capability would be an excellent way to ensure the argument never ends. The useful comparison involves relevant performance, comparable conditions and an honest accounting of the help each side receives. ### There is Considerable Panache in Those Three Letters For the company making the announcement, AGI carries value beyond the performance of any particular task. It confers the prestige of pursuing, and potentially achieving, a major scientific milestone. It can attract talent, command attention and position a vendor as the company everyone else is chasing. There is considerable panache in being able to say you built it. A product launch ordinarily invites questions about features, pricing and competitive advantages. A declaration of AGI invites people to debate your place in history. As a publisher, I can appreciate the difference in the conversation those announcements generate. None of that establishes that the claim is cynical or false. Researchers can sincerely believe they have reached a milestone while their company benefits from the attention. Genuine progress and ambitious marketing have always been comfortable traveling companions in technology. But customers are entitled to ask what additional promise is being made. If AGI means broader adaptability, demonstrate that breadth. If it means useful autonomy over longer assignments, show the completed work, the interventions and the failures. If the term represents a scientific judgment whose implications remain uncertain, explain the uncertainty. The designation should add information. Otherwise, the panache is doing more work than the definition. ### When the Label Changes Trust The AGI debate becomes relevant to users when the label changes their behavior. Someone impressed by an agent’s success on a difficult research assignment may assume it can also manage purchasing, adjust production systems or communicate with customers. Each responsibility introduces different consequences and requires its own evidence. The phrase “general intelligence” can encourage people to skip that examination. Consider the difference between investigating an outage and making changes to restore service. An agent might correctly identify a problem, then choose a remedy that violates an organizational constraint. Its technical competence does not establish that it understood the limits of its authority. This is why I favor moving from human in the loop to human at the helm. Having someone click approval on every step can become a ritual that adds delay without adding much judgment. Leadership means establishing the destination, deciding which actions can be delegated and knowing when intervention is required. For an agent reconciling records, that might mean broad access to inspect information and propose corrections, with narrower authority to change the system of record. The appropriate boundaries should evolve as the organization gathers evidence about performance. Intelligence, reliability and alignment are related questions, but they are separate questions. A more capable system may understand an assignment better. That does not automatically establish that it will execute dependably or respect every boundary. The responsibilities of the person deploying it remain. ### Indispensable Before Anyone Agrees There is another threshold businesses may cross while the researchers are still arguing. An organization can become dependent on AI without ever deciding whether it qualifies as AGI. People begin using it because it saves time. Workflows form around it. Staffing decisions assume its availability. Eventually, the organization may discover it has lost the skills, capacity or alternatives needed to operate without it. That is the concern I explore in my forthcoming book, *The Indispensability Trap*. Useful technology becomes essential infrastructure, and dependence can narrow the choices of the people who rely on it. There is no need for a ceremonial declaration of AGI along the way. The transition happens through ordinary decisions about budgets, staffing, software and who is responsible for getting the work done. A business that becomes unable to explain its own processes, challenge its AI’s conclusions or continue through a service interruption has a concrete management problem. Winning an argument about the definition of intelligence will not solve it. AGI may prove to be a consequential scientific milestone. Astra may eventually be remembered as part of that achievement. We should remain willing to revise our judgments as the evidence develops, including when the evidence challenges our skepticism. But users have decisions to make now. They need to understand what they are buying and retain the ability to govern what they deploy. A claim of historic significance does not relieve the vendor of explaining the product, or the customer of evaluating it. Memorex gave us a memorable question. OpenAI has given us reason to ask a version of it again. Before we put the three letters on every presentation and hand over more responsibility, I would like a straightforward answer: Tell us what it can do, show us where it fails, and explain what calling it AGI changes about either.