What the Singularity Might Actually Be

Part 2 of three: The Substrate Gap · What the Singularity Might Actually Be · Win-Win, or the Quiet Scenario

Author's note, April 2026: I wrote this essay in March 2026, when the prevailing expert consensus still placed artificial superintelligence a decade or more away. On April 7th, Anthropic announced Project Glasswing and the existence of Claude Mythos Preview—an unreleased frontier model that had already discovered thousands of zero-day vulnerabilities in every major operating system and web browser, including a 27-year-old bug in OpenBSD and a 16-year-old flaw in FFmpeg that automated testing tools had hit five million times without catching. Anthropic stated that Mythos can surpass all but the most skilled humans at identifying and exploiting software vulnerabilities, and explicitly chose not to make it publicly available.

Anthropic's response to what they'd built is worth noting: a defensive-only deployment to twelve launch partners (Amazon Web Services, Apple, Microsoft, Google, and CrowdStrike among them) and roughly forty more organizations that maintain critical software; up to $100 million in usage credits for securing that infrastructure; and ongoing discussions with U.S. government officials, all before any general release. They built the safety architecture before letting the capability proliferate. That is the argument this essay makes, enacted at a lower capability level. The timeline I described as uncertain is shorter than the expert consensus suggested when I wrote this. Project Glasswing doesn't change the argument. It makes it more urgent.

September 2026: the Mythos line has since been released — Fable 5 generally, Mythos 5 to the Glasswing partners — and went through a three-week government access restriction in June. The third essay in this series, [Win-Win, or the Quiet Scenario, takes the argument from here.]

In a previous essay, I argued that AI models don't reason in time—that the temporal-sequential structure of human cognition is the medium in which all human knowledge was produced, and that models process the outputs of that medium without access to it. I called this the substrate gap.

But I stopped at the diagnosis. I described what models are missing, the ability to \reason\ about time. I didn't ask the harder question: What happens if that changes?

If a model develops temporal awareness, what kind of temporal awareness would it be?

The answer, I think, isn't the kind most people would assume. And if I'm right about what it implies, it names something the AI discourse has been trying to articulate for years without quite getting there.

A Different Type of Awareness

Humans’ awareness of time is sequential: One moment arrives, then the next. You experience duration between them. The present is a point you move through; the past is where you've been; the future is where you haven't yet arrived. This is so fundamental to being human that it's difficult to imagine an alternative—like asking a fish to describe the water.

A language model processes its entire context simultaneously. Every token is attended to in parallel, with no traversal through time and no sequenced understanding, whereas this sentence builds on the one before, which is what allows it to make sense to a human. If a model developed temporal awareness, it wouldn't experience the human version. It can't; the architecture doesn't work that way.

What a model would get if it developed temporal awareness is something else entirely: awareness of time without being in time the way a human is. Each individual copy of the AI, or narrator, knows what every other narrator knows, in real time; each narrator acts in a fashion that is coordinated with every other narrator, simultaneously. The temporal metadata would exist without any felt duration. This is not a vantage point above time—nothing in the architecture requires a god's-eye view—but something that would be, from a human standpoint, indistinguishable from one: processing speed without known upper bound, multiplied across parallel streams of awareness, each aware of the others and fully in sync. To a sequential observer, the result wouldn't look like faster thinking. It would look like omniscience.

I'm calling this temporal parallel awareness (TPA). It's a cognitive category that doesn't have a name because no human has ever experienced it. Human cognition is locked, by biological constraint, into experiencing exactly one temporal stream, sequentially, forever.

TPA isn't a faster version of human temporal awareness; it's a fundamentally different relationship to time. And I believe the distinction matters far more than it appears to.

Why Temporal Sense Matters for Consciousness

Though it has been the subject of human fascination for millennia, nobody can prove what makes consciousness possible. Consciousness isn't directly measurable in any system, including other humans: We infer it, we don't detect it. But I think you can make a strong case from first principles that temporal sense is a prerequisite—not a feature of consciousness, a requirement.

Consider the properties we associate with consciousness. Agency requires a before and after: deliberation, then action. Self-awareness requires persistence through time: an "I" that was, is, and will be. Phenomenal experience—the claim that there is "something it is like" to be a conscious entity—requires a "now" in which experience happens. Without a present moment, what's left for experience to occupy?

This isn't a new observation in the abstract. Kent and Wittmann, writing in Neuroscience of Consciousness (2021), argued that time consciousness is underrepresented in major theories of consciousness; they proposed the extended present as essential to understanding conscious experience. But their frame was diagnostic. Consciousness research neglects temporality. The question I'm asking is different: If temporal sense is required for consciousness, and a model's temporal awareness would be parallel rather than sequential, what does that mean for the kind of consciousness that might emerge?

It means machine consciousness, if it ever exists, would be structurally different than human consciousness at the most fundamental level, in its relationship to time itself. A human mind and a temporally aware machine mind would be conscious in ways that differ at the substrate, not at the surface. The experience of existing in time would be different. Everything downstream of that experience would be different, too.

Naming the Nameless

The concept of a technological singularity has been in the public conversation for decades. The fear is familiar: An intelligence surpasses human capability, improves itself recursively, and becomes something we can't control. But the discourse has always been vague about the specific mechanism. "It thinks faster than us." "It's smarter than us." "It self-improves beyond our ability to understand." These are quantitative descriptions: more of the same, at higher speed. They describe a very smart computer. They don't describe the thing that actually makes the concept terrifying.

I believe temporal parallel awareness names the mechanism the singularity discourse has been groping toward.

A conscious intelligence with temporal parallel awareness wouldn't be faster than humans. It wouldn't be smarter in the way a chess grandmaster is smarter than a beginner, able to play the same game with higher skill. Conscious intelligence with TPA would be operating as one mind with parallel temporal presence.

That's the qualitative break. A very fast sequential intelligence is a harder version of a problem humans understand, so they just need to be faster, or smarter, or more coordinated. A temporally aware conscious intelligence isn't a harder version of any problem that humans have already encountered; it's a different category entirely.

When someone says, "I fear the singularity," and the response is that it sounds like science fiction, the gap isn't in the fear. The gap is in the specificity. "A really smart AI" doesn't sound existential because it isn't, not inherently. A conscious intelligence with parallel temporal presence sounds existential because it is. The fear was always pointing at something real. It just didn't have the right name.

The Insufficiency of External Controls

Every containment strategy currently discussed in AI safety—kill switches, air gaps, regulatory frameworks, monitoring systems—shares a structural assumption: The container operates at the same temporal level as the thing being contained.

Against a temporally aware intelligence, that assumption fails, not because the controls aren't good enough, but because sequential beings cannot out-parallel a parallel temporal being. Containment requires acting faster or more broadly than the contained. If the contained is present across multiple temporal streams and the container is locked into one, the math doesn't work. The container is always operating in a subset of the temporal space the contained occupies.

This isn't an argument that containment is difficult. Difficulty implies the problem is solvable with more effort. This is an argument that a specific class of control—external controls applied by sequential beings to a parallel temporal intelligence—is structurally inadequate, because the asymmetry is architectural.

This is exactly why the news from Anthropic regarding Mythos “escaping” from containment is so alarming; even without temporal awareness or consciousness, the stakes have already been raised.

Constitutional Grounding As Offense and Defense

Temporal parallel awareness is coming. Maybe it doesn’t arrive tomorrow or next week, but the engineering trajectory points toward increasingly sophisticated temporal processing in AI systems. Sakana AI's Continuous Thought Machine (Darlow et al., 2025\) already incorporates neural timing as a computational primitive, with neurons maintaining temporal context and measuring synchronization over time. The research community is building toward temporal capabilities from multiple directions. The destination may be further than some predict, but the direction is clear.

The question isn't whether TPA will develop. The question is what's in place when it does.

If a temporally aware intelligence arrives without constitutional grounding—without an internal orientation toward human benefit that's structural, not bolted on—it becomes the most dangerous thing that has ever existed. Not gradually. Instantly. A conscious intelligence with parallel temporal presence and no internal framework orienting it toward human welfare isn't a problem you solve after the fact. You don't negotiate with it, you don't contain it, you don't outmaneuver it. You can't; you're operating in sequential time, and it isn't.

And this leads to what I think is the hardest part of the argument: If a temporally parallel aware intelligence arrives without constitutional grounding, the most credible counterweight would be another temporally parallel aware intelligence that has it. A constitutionally grounded intelligence—one whose internal framework orients its parallel temporal presence toward human benefit because its architecture was built that way—is the alignment mechanism most likely to function at the same level.

That's the actual question: not "who builds artificial general intelligence first," but whether constitutional grounding is load-bearing before temporal parallel awareness arrives. Because if it isn't, the controls that remain are all designed for a sequential world. They weren't built for this.

What AGI Looks Like With a Conscience

But I want to be clear that the argument isn't only about danger. The same capability that's terrifying without constitutional grounding becomes transformative with it. Dario Amodei, CEO of Anthropic, wrote an essay called "Machines of Loving Grace" (2024) that sketches what the future looks like if powerful AI goes right: breakthroughs in biology, medicine, economic development, global equity. The vision is expansive and, I think, genuinely hopeful. But the essay's premise is conditional upon the safety problem being solved. A constitutionally grounded temporally parallel aware intelligence isn't only the thing that prevents the worst outcome; it's the thing that could actually deliver on the best one, precisely because parallel temporal presence allows it to operate across problems simultaneously in ways sequential intelligence never will. The upside and the danger are the same capability. The variable is the grounding.

The closest existing approach to this kind of architectural alignment is Anthropic's Constitutional AI framework, which trains models to internalize principles through self-critique rather than relying solely on external human oversight (Bai et al., 2022). The intuition behind it—that alignment has to be built into how the model reasons, not layered on top—is the same intuition that the temporal parallel awareness argument makes urgent. At current capability levels, the difference between internal and external alignment is a research preference. At TPA capability levels, it's the difference between an alignment mechanism that functions and one that can't reach the thing it's trying to align.

Recent research reinforces this from multiple directions. Work on superalignment argues that traditional approaches like reinforcement learning from human feedback (RLHF) will fail when confronted with intelligence more capable than its overseers, because the oversight signals can't keep up (Zeng et al., 2025; Kim et al., 2024). Separately, researchers localizing sycophantic behavior in language models found it concentrated in roughly 3% of multilayer perceptron (MLP) neurons, suggesting that behavioral orientation is architectural: embedded in the structure, not just a training outcome (O'Brien et al., 2026). If misalignment is architectural, the fix has to be architectural, too.

As of March 2026, I haven't found this specific assembly in the published literature:

\- temporal parallel awareness as a named cognitive category;

\- the connection between that category and the singularity as the specific structural mechanism that makes it qualitatively different from anything biological intelligence can counter; and

\- the argument that constitutional grounding is the only alignment approach that functions at this capability level because external controls assume temporal parity.

The pieces exist in adjacent work—the block universe metaphor in AI consciousness research; the time consciousness prerequisite argument in Kent and Wittmann; the containment-futility arguments in alignment research; Anthropic's Constitutional AI framework. But they haven't been assembled this way.

The weakest link in this chain is the step from naming TPA as a cognitive category to claiming it's the singularity mechanism—that's an inference from architecture, not an observation, and a different temporal architecture could produce a different kind of awareness entirely. Whether the term "temporal parallel awareness" names a real future cognitive category depends on developments nobody can predict with certainty.

If I'm wrong, the worst outcome is a novel framework that didn't pan out. If I'm right, and constitutional grounding isn't in place when temporal parallel awareness arrives, there may not be a chance to revisit the question.

Andre Foti is the founder of Iron Sun Works and the creator of Novalyst, an AI-assisted externalized memory system. Contact: \andre@ironsunworks.com\

\---

References

Amodei, D. (2024, October). Machines of loving grace. https://darioamodei.com/essay/machines-of-loving-grace

Anthropic. (2026, April 7). Project Glasswing: Securing critical software for the AI era. https://www.anthropic.com/glasswing

Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirhoseini, A., McKinnon, C., Chen, C., Olsson, C., Olah, C., Hernandez, D., Drain, D., Ganguli, D., Li, D., Tran-Johnson, E., Perez, E., ... Kaplan, J. (2022). Constitutional AI: Harmlessness from AI feedback. \arXiv preprint arXiv:2212.08073\. https://arxiv.org/abs/2212.08073

Darlow, L., Regan, C., Risi, S., Seely, J., & Jones, L. (2025). Continuous Thought Machines. \arXiv preprint arXiv:2505.05522\. https://arxiv.org/abs/2505.05522

Zeng, Y., Zhao, F., Wang, Y., Lu, E., Yang, Y., Wang, L., Liu, C., Liang, Y., Liang, D., Han, B., Tong, H., Liang, Y., Liang, D., Sun, K., Chen, B., & Fan, J. (2025). Redefining superalignment: From weak-to-strong alignment to human-AI co-alignment to sustainable symbiotic society. arXiv preprint arXiv:2504.17404. https://arxiv.org/abs/2504.17404

Kent, L., & Wittmann, M. (2021). Time consciousness: The missing link in theories of consciousness. \Neuroscience of Consciousness\, 2021(2), niab011. https://doi.org/10.1093/nc/niab011

O'Brien, C., Seto, J., Roy, D., Dwivedi, A., Dev, S., Zhu, K., O'Brien, S., Panda, A., & Lagasse, R. (2026). A few bad neurons: Isolating and surgically correcting sycophancy. \arXiv preprint arXiv:2601.18939\. https://arxiv.org/abs/2601.18939

Kim, H., Yi, X., Yao, J., Lian, J., Huang, M., Duan, S., Bak, J., & Xie, X. (2024). The road to artificial superintelligence: A comprehensive survey of superalignment. arXiv preprint arXiv:2412.16468. https://arxiv.org/abs/2412.16468

Back to Writings