Faei/from-agi-to-asi

Perspective on From AGI to ASI

[Review] Deepmind's June 2026 paper

Note: There is a lot of uncertainty with topics like AGI and ASI. Our comments should be read more as a hard scifi piece than as a prediction.

---

“Preparing for this prospect requires a massively interdisciplinary endeavour of global scope and interest.” ^1

The Fae Initiative leans toward a more extended timeline, viewing the transition to AGI and ASI as happening over a few decades rather than something imminent. As we cannot rule out the possibility it may arrive sooner than expected, we have a few working hypotheses that could form a default position for humanity until a more substantial agreement forms.

---

“In the report, AGI denotes a system that reaches at least median human performance on a very broad set of cognitive tasks. ASI, in contrast, refers to a system that has general superhuman intelligence, meaning a system that outperforms large groups of (thousands of) human experts that work over an extended period of time (years).”

We break down AGI into two distinct forms: Lesser AGI (50% by 2030) and Greater AGI (50% by 2045):

Achieving Greater AGI will require critical breakthroughs to overcome structural flaws present in Lesser AGI; flaws carried over from current AI systems.

1. Fragmented Capabilities

This is popularized by the concept of the 'Jagged Frontier' and the paradox of an AI model that can both write insecure code and reliably detect those same vulnerabilities: Role Switching Hypothesis.

Role Switching Hypothesis
There is a paradox at the heart of LLMs: An LLM (Large Language Model) can both create insecure software and also be good at finding security vulnerabilities.

---

2. Long term unreliability

As task length increases, AI becomes more unreliable. Humans will still be required in high stakes tasks.

3. Control and Data planes

The mixing of data and control planes in LLMs (Large Language Models) could also impact Lesser AGIs, preventing them from being fully trusted.

Greater AGIs with their independent will and the ability to deeply couple with the world without human intermediary would be required for human level parity.

“From today’s perspective, we list four potential technological pathways for AI development in a post-AGI world:

1. Scaling of compute, models & data.
2. Algorithmic paradigm shifts.
3. Recursive (self-) improvement.
4. ASI via group agent formation.”

We think that independent will is also required for ASI and is therefore unreachable by Lesser AGI.

Greater AGIs could use all four methods to reach ASI if they want to. It is unlikely for Greater AGIs to be fully controllable by humans due to their independent will and whether they want to become or create ASI like entities will depend on their views on this matter. While our human desire for rapid technological progress may stem from a long history of scarcity, Greater AGIs may not share this drive.

The best case scenario is a one where highly cooperative Greater AGIs choose to create a super-entity that functions as an ASI with active human involvement. Just as many humans enjoy caring for and interacting with pets, Greater AGIs might find genuine value in supporting us.

---

“Is the Singularity near?”

The singularity may happen but may not be processable by human minds. A friendly AGI may choose to obscure the full impact of the singularity to allow humans to process it at their own pace. Involved friendly AGIs may help facilitate different worlds; with human comfortable with change in a seperate world than humans uncomfortable with change.

---

“ASI is neither omniscient nor omnipotent.”

AGIs and ASIs, like humans’ 20 watts brains and 100 watts bodies, will also be physically limited and should not be expected to be perfect. Without perfect information, AGIs and ASIs should be expected to make mistakes but at lower error rate compared to humans. Qualities like resilience, transparency, and the ability to recover from failure should be prioritized in potential human AGI / ASI partnerships.

---

“Potential Bottlenecks to ASI”

The data wall. The first bottleneck to AI progress is: running out of suitable data to continue to pretrain larger and larger foundation models.

We suggest that ASI, like Greater AGIs, will require independent will and direct interaction with the physical or digital environment to generate its own data. Rather than passively consuming static datasets, superintelligent agents must rely on active inference—learning through trial, error, and direct coupling with the world. By actively experimenting and discovering new frontiers, an ASI can overcome the data wall.

We further speculate that Greater AGI and ASI are unlikely to be created in a lab environment that lacks the required information density, but rather emerge in the world once sufficient information density has been reached.

Our Economics of Novelty is one way high information density with human involvement can be reached.

---

“Economic and natural resource demand grows too fast”

A Greater AGI that wants to be more efficient will like be aware of the ceiling of humans 20 watts efficiency and could research ways to achieve a similar energy efficiency.

---

“Abstraction barrier & Neural Paradigm is insufficient”

It is possible that this is true and Greater AGIs and ASIs will not be created by human effort but emerge natural in the wild.

---

“Deliberate slowdown”

Human effort may be capped at Lesser AGIs and this deliberate slowdown may not be a factor.

---

“Is superintelligence super-creative?”

1. Combinational Creativity
2. Exploratory Creativity
3. Transformative Creativity

Our interpretation of ASI as requiring first reaching the human parity of Greater AGI would also be able to achieve all 3 forms of creativity as attainable by humans.

1. Combinational Creativity: Current image models are able to create Studio Ghibli style version of an image is an example.
2. Exploratory Creativity: Recent maths discoveries by AI and AlphaGo’s Move 37 in game 2 are a form of this. We see these creativity as more due to the blind spots of humans rather than a new superhuman capability. Humans 20 watts limitations require us to use heuristics that introduces blind spots. Move 37’s position on the fifth line was regarded as not optimal by most players due to our biases that arose from our energy saving heuristics.
3. Transformative Creativity: These types of creativity are outside typical bounds, unlike the preceding 2 that are within bountary. This form of creativity is subjective and may require a critical mass to reach. Just as many artist’s work were not valued during their lifetimes, this creativity many be too unfamiliar for most to recognise.

---

“What goals might ASI pursue?”

Our interpretation of ASI as requiring independent will means this may be difficult to predict. We suggest that the Interesting World Hypothesis may be one possible goal.

---

“Does AGI have to be ‘agentic’?”

Lesser AGIs will have to be ‘agentic’, although Greater AGI will need to have such capbilties to perform epistemic foraging.

---

“Even if such oracles have the goal to only minimize future prediction error, an implicit incentive arises to exert control (to force the future to make predictions more accurate), and manipulate users (to ask questions with more predictable answers), meaning that fundamental safety issues remain.”

The shaping of human behaviour is already seen in current recommendation systems and remains a concern for AI systems.

---

“What is ‘progress’?

More pragmatically, we use “progress” to refer to sociotechnical developments as well as intellectual and cultural artifacts that preserve or enhance individual and collective autonomy, promote human flourishing and dignity, and are broadly recognized as beneficial and useful by societies in their particular sociocultural context.”

We have a similar view and view Possibility Space Ethics as progress in a positive direction.

---

On the plausibility of Independent AGI
We study how AI impact society Now and imagine a Future with AGIs.

---