AI headlines have a way of turning every new model into a prophecy. One release becomes “the step toward AGI.” Another demo becomes “proof of superintelligence.” The language spreads quickly because it’s catchy, because it compresses uncertainty into a single, dramatic label—and because investors, researchers, and journalists all operate under time pressure.
But Alexandre LeBrun, CEO of AMI Labs—an AI startup associated with the world-model approach championed by Yann LeCun—doesn’t want his company’s work described in those terms. In a recent discussion highlighted by TechCrunch, LeBrun pushed back on the habit of calling today’s systems “AGI” or “superintelligence,” arguing that the words do more than describe capability. They shape expectations, influence how people interpret progress, and can quietly steer the conversation away from what matters: measurable performance, clear boundaries, and honest uncertainty.
That stance might sound like semantics. In practice, it’s a philosophy about how to talk about AI development when the field is moving fast and the public narrative is often moving faster.
A world-model company, resisting world-model hype
AMI Labs is part of a broader shift in AI research: the push toward systems that learn models of the world rather than only pattern-match within a narrow dataset. The “world model” framing is important because it implies something different from classic supervised learning. Instead of treating intelligence as a bag of skills stitched together, world-model approaches aim to build internal representations that can support prediction, planning, and interaction—at least in principle.
LeBrun’s reluctance to use AGI or superintelligence labels doesn’t contradict that ambition. If anything, it reflects a belief that the most meaningful milestones are not branding milestones. A world model is not automatically “general intelligence” just because it can predict or simulate. It’s a component of a larger system, and it still needs to be evaluated in context: What tasks does it handle? Under what constraints? How robust is it when the environment changes? How does it behave when it’s wrong?
In other words, LeBrun’s argument is that the field should treat capability claims like engineering claims, not like marketing slogans.
Why “superintelligence” is a loaded word
“Superintelligence” is one of those terms that sounds precise but isn’t. It suggests a qualitative leap: an intelligence that surpasses humans across domains, perhaps permanently, perhaps rapidly, perhaps inevitably. But the term doesn’t specify what “surpass” means, how it would be measured, or which domains count. It also tends to imply a trajectory—one that can make current progress feel like evidence of a future outcome.
LeBrun’s critique, as presented in the report, is less about denying that AI could become extremely capable and more about rejecting the framing that treats the future as already settled. When people say “superintelligence,” they’re often not talking about a specific system’s demonstrated abilities. They’re talking about a category of outcomes. That category framing can be useful for debate, but it becomes misleading when it’s used as a substitute for evaluation.
There’s also a psychological effect. Once a label like “superintelligence” enters the conversation, it becomes harder to walk it back. If a system is called superintelligent and then later fails at something basic—reasoning under distribution shift, long-horizon planning, reliable tool use, or safe behavior—the failure doesn’t just correct a technical claim. It undermines trust in the entire narrative. LeBrun’s preference for precision is partly a response to that dynamic: don’t inflate the story early, because the story will be judged later.
The AGI problem: a definition gap that keeps widening
“AGI” is even trickier. Artificial General Intelligence is supposed to mean generality: competence across a wide range of tasks, with transfer and adaptability that resembles human flexibility. But in practice, AGI has become a moving target. Different people define it differently. Some treat it as “can do most things a human can do.” Others treat it as “can learn new tasks quickly.” Still others treat it as “can reason reliably across domains.”
The result is that AGI becomes a rhetorical umbrella. When a model performs well on benchmarks, it’s sometimes treated as evidence of generality—even if the performance is narrow, benchmark-specific, or dependent on carefully curated evaluation settings. When a model fails, the narrative shifts again: maybe it’s not AGI yet, but it’s close; maybe the missing piece is just around the corner; maybe the next scaling run will unlock the rest.
LeBrun’s position is essentially: if we can’t agree on what the label means, we shouldn’t pretend the label is informative. Better to talk about what the system actually does today, and what it can do tomorrow under realistic conditions.
This is not a call for pessimism. It’s a call for clarity.
Words shape expectations—and expectations shape investment
In fast-moving fields, language is not neutral. It affects what people fund, what people prioritize, and what people assume will happen next. When “AGI” becomes the default descriptor, teams can feel pressure to chase the appearance of generality rather than the substance of reliability. Benchmarks can start to matter more than real-world behavior. Demos can start to matter more than failure modes.
LeBrun’s emphasis on capability and progress—rather than megacategory branding—implicitly argues for a different kind of accountability. If you describe a system as “AGI,” you invite a broad standard of proof. If you describe it as a model that improves prediction, planning, or interaction in certain environments, you can evaluate it in those dimensions without pretending it has already solved everything.
That distinction matters for both technical strategy and public understanding. Technical teams need room to iterate. Public narratives need room to be accurate. When the same word is used to mean different things, accuracy suffers.
A unique take: treat intelligence as a spectrum of competencies
One reason the AGI/superintelligence labels persist is that they offer a simple story: either we have it or we don’t. But intelligence—biological or artificial—is not binary. It’s a spectrum of competencies: perception, memory, reasoning, planning, learning, adaptation, and control. Each competency has its own bottlenecks and failure patterns.
LeBrun’s refusal to use the big labels can be read as an attempt to keep the field focused on that spectrum. Instead of asking, “Is this AGI?” the better question becomes, “Which competencies are improving, and how do we know?”
For example, a world-model system might show strong predictive accuracy in certain settings. That’s valuable, but prediction alone doesn’t guarantee robust planning. Planning requires the model to support counterfactual thinking and long-horizon consistency. Even if the model predicts the next frame well, it may still struggle to choose actions that lead to goals over time. Similarly, a system might appear intelligent in interactive demos while failing under adversarial prompts or distribution shifts.
By avoiding AGI and superintelligence labels, LeBrun is effectively encouraging a more granular evaluation culture—one where progress is tracked by capability components rather than by a single headline claim.
What “progress” should mean in a world-model context
World-model approaches are often discussed as if they naturally lead to general intelligence. But the path from “learns a model of the world” to “acts intelligently across domains” is not automatic. It depends on how the model is trained, how it represents uncertainty, how it handles partial observability, and how it integrates with decision-making and memory.
So what does progress look like?
It looks like improvements in:
1) Representation quality
A world model must capture relevant structure, not just surface correlations. Progress here might show up as better generalization to new scenes, better handling of rare events, or improved ability to infer hidden variables.
2) Temporal coherence and long-horizon stability
Many models can produce plausible short sequences. The harder part is maintaining consistency over time—especially when the environment changes or when the system must remember what happened earlier.
3) Planning and action selection
Prediction is one thing; choosing actions is another. Progress might involve better goal-directed behavior, fewer “hallucinated” plans, and more reliable tool use.
4) Robustness under distribution shift
Real environments are messy. A system that works in a controlled setting but collapses when conditions change is not yet “general.” Progress means the system degrades gracefully—or better, resists degradation.
5) Calibration and uncertainty awareness
If a model cannot express uncertainty, it can’t reliably decide when to act confidently and when to ask for more information. This is crucial for safety and reliability.
LeBrun’s language preference aligns with these kinds of metrics. It’s easier to discuss them without the gravitational pull of AGI branding.
The media incentive problem: why the labels keep showing up
Even if researchers dislike the labels, the media ecosystem rewards them. “AGI” and “superintelligence” are search-friendly, shareable, and emotionally resonant. They also create a sense of urgency: readers feel they’re witnessing history.
TechCrunch’s framing of LeBrun’s comments highlights a tension: the industry wants to communicate excitement, but it also needs to communicate accurately. When a CEO says he won’t call his AI AGI or superintelligence, it’s not just a personal preference. It’s a signal to the market that AMI Labs is trying to avoid the trap of premature certainty.
That doesn’t mean the company lacks ambition. It means the company is trying to separate ambition from assertion.
A more honest conversation about timelines
Another subtle point in LeBrun’s stance is the timeline issue. Labels like AGI and superintelligence often imply a near-term endpoint. They suggest that the end state is not only possible but imminent. That can distort how people interpret incremental improvements.
In reality, AI progress is likely to be uneven. Some capabilities may improve quickly; others may lag. Some breakthroughs may be domain-specific