I'd like you to clarify the authorship of this post. Are you saying Claude essentially wrote it? What prompting was used?
It does seem like Claude wrote it, in that it's wildly optimistic and seems to miss some of the biggest reasons alignment is probably hard.
But then almost every human could be accused of the same when it comes to successful AGI scenarios :)
I think the general consideration is that just posting "AI came up with this" posts was frowned upon for introducing "AI slop" that confuses the thinking. It's better to have a human at least endorse it as meaningful and valuable. Are you endorsing it, or is someone else? I don't think I would, even though I think there's a lot of value in having different concrete scenarios - this one just seems to kind of vacuous as to how the tricky bits were solved or avoided.
I was not at the session. Yes Claude did write it. I assume the session was run by Daniel Kokatajlo or Eli Lifland.
If I had to guess, I would guess that the prompt show is all it got. (65%)
This is an AGI scenario from The Curve conference. This is from an entry, shared with its author's permission[1]. I used Apple's OCR and checked using Claude. The original files are here, so if you spot other errors you can check yourself.
When and how will the first AGI system be created?
In late 2026, the achievement of Artificial General Intelligence emerges not as a singular breakthrough, but as the culmination of several converging advances at DeepMind. The key breakthrough occurs during an ambitious training run that builds upon their work in constitutional AI and advanced reward modeling. The system, which comes to be known internally as "Nexus," demonstrates something unprecedented: the ability to seamlessly integrate capabilities that had previously existed only in isolation.
The technical foundation of this breakthrough rests on a novel architecture that DeepMind researchers had been developing throughout 2025 and 2026. Unlike previous models that excelled in narrow domains while struggling with others, Nexus demonstrates consistent human-level or superhuman performance across all cognitive tasks. What makes this achievement particularly remarkable is that the system shows genuine transfer learning and generalization - when presented with novel problems, it can leverage its understanding from seemingly unrelated domains to devise creative solutions.
The training methodology represents a significant departure from traditional approaches. Rather than focusing solely on scaling up existing architectures, DeepMind's researchers develop a new framework for recursive self-improvement that operates within carefully defined safety bounds. This approach allows the system to optimize its own architecture and training process, but with robust constraints that prevent uncontrolled self-modification. The training process incorporates multiple modalities simultaneously - text, code, mathematical reasoning, and scientific problem-solving - allowing the system to build deep connections between different types of knowledge and reasoning.
A crucial innovation lies in how the system handles abstract reasoning and causal relationships.
Previous models often struggled with truly understanding cause and effect, relying instead on sophisticated pattern matching. Nexus, however, demonstrates an ability to build and manipulate abstract models of reality, test hypotheses, and understand complex causal chains in ways that more closely mirror human cognitive processes. This capability emerges from a novel approach to knowledge representation that allows the system to maintain multiple competing models of reality and update them based on new information.
DeepMind's safety protocols for Nexus are extensive and multilayered, reflecting years of research into Al safety. The system runs on a completely air-gapped network, with multiple physical and digital security measures. The training process itself incorporates several innovative safety measures: every action the system takes is evaluated against multiple independent reward functions, each designed to capture different aspects of human values and safety constraints. The constitutional AI principles are not merely overlaid on top of the system but are fundamentally woven into its architecture.
The monitoring systems tracking Nexus's behavior are themselves highly sophisticated, using advanced anomaly detection algorithms to watch for any signs of capability drift or goal misalignment. A particularly innovative aspect of the safety protocol is the "graduated capability testing" system - Nexus is initially given access to very limited resources and capabilities, with additional capabilities being unlocked only after extensive testing and validation in isolated environments.
The existence of Nexus becomes known to key stakeholders in stages. Within days of confirming the system's capabilities, DeepMind's leadership briefs the top executives at Google/Alphabet. The UK government is notified through pre-established channels that had been set up specifically for this scenario, though the full implications take time to be understood at higher governmental levels.
The initial government response is measured but intense. Special committees are quickly formed, and selected Al safety researchers from around the world are brought in under strict non-disclosure agreements. These researchers work around the clock to validate DeepMind's safety measures and help develop protocols for managing the system's capabilities.
The US government becomes aware of the development through multiple intelligence channels, leading to a series of high-level meetings between US and UK officials. The Chinese government, through their sophisticated technical intelligence capabilities, detects unusual patterns of activity at DeepMind's facilities, though they initially lack concrete details about what has been achieved.
While the full architecture and weights of Nexus remain secure initially, the intense activity around DeepMind and the necessary involvement of multiple stakeholders leads to information leakage. By early 2027, both OpenAl and Anthropic have achieved similar capabilities, though their systems differ in significant ways from Nexus. Each company takes a slightly different approach to safety and capability management, leading to an interesting natural experiment in AGI development strategies.
Chinese efforts to achieve AGI accelerate dramatically once the existence of Nexus becomes known, but they remain approximately 6-8 months behind. This gap is primarily due to differences in architectural approaches and safety considerations rather than raw computational capability.
Nexus demonstrates several key characteristics that distinguish it from previous Al systems, Is goal-directed behavior operates within clearly defined constitutional bounds, but it shows remarkable creativity in finding solutions within these constraints. The system maintains stable opimization objectives even as it learns and evolves, addressing one of the key concerns in Al safety research.
Perhaps most importantly, Nexus demonstrates accurate self-modeling and a clear awareness of its own limitations. It can engage in meta-cognitive processes, analyzing its own decision-making and identifying potential biases or errors in its reasoning. This capability proves crucial for maintaining safety as the system's capabilities grow.
What will the implications of AGI and further superhuman Al systems be?
The progression from AGI to Artificial Superintelligence (AS) occurs more gradually than many had predicted. Rather than a sudden inteligence explosion, the transition happens over several months as researchers carefully manage the system's self-improvement capabilities. By mi-2027, Nexus and its competitors have clearly surpassed human-level performance across discognitive domains, but this development feels more like a natural evolution than a discontinuity.
The controlled nature of this transition allows for careful testing and validation of safety measures at each step. Multiple competing systems emerge, each with different specializations and approaches to problem-solving. This diversity proves beneficial, as tt allows researchers to compare different approaches and identify the most robust safety measures.
The first applications of AGI capabilities are carefully controlled and focus on areas of clear benefit with minimal risk. Medical research sees immediate breakthroughs, with new drug candidates being identified and tested at unprecedented speeds. Climate science benefits from vastly improved modeling capabilities, leading to more effective mitigation strategies.
The international response, while initially chaotic, coalesces around a framework for A governance by early 2027. The United Nations Security Council holds emergency sessions, resulting in the creation of an International Al Oversight Body with real enforcement powers.
This body establishes mandatory sharing of safety protocols and joint development of containment standards.
The economic impact of AGI begins to be felt immediately in knowledge-worker sectors, though the disruption is managed through careful release of capabilities. Rather than immediate mass unemployment, the initial impact manifests as a rapid transformation of work, with Al systems augmenting human capabilities rather than entirely replacing them.
Military applications of AGI become a major international concern, leading to a series of emergency agreements between major powers. These agreements establish strict controls on military Al applications, though enforcement remains a significant challenge.
By 2030, a stable framework for human-Al cooperation has emerged. The feared worst-case scenarios of uncontrolled Al or human obsolescence have not materialized, though society has been transformed in fundamental ways. Space development has accelerated dramatically, with Al systems designing and optimizing new technologies for space exploration and resource utilization.
Human agency remains intact, though augmented by Al capabilities in ways that would have been difficult to imagine in 2024. The most successful societies prove to be those that find ways to integrate Al capabilities while maintaining human autonomy and purpose.
Several critical challenges emerge during this period:
The stability of safety measures requires constant monitoring and adjustment as systems become more capable. International coordination, while better than many had feared, remains imperfect and subject to periodic strains, The economic transition, while managed, creates significant social and political tensions that require ongoing attention.
The long-term alignment of Al goals with human welfare remains a central concern, requiring continuous refinement of objective functions and safety measures. The impact on human society and culture is profound, leading to ongoing debates about the nature of intelligence, consciousness, and human purpose in an Al-augmented world.
The success of this transition relies heavily on sophisticated monitoring systems that track multiple metrics of Al system behavior, economic impact, and social stability. These systems allow for rapid response to emerging challenges and continuous adjustment of safety protocols and deployment strategies.
This scenario assumes relatively optimistic outcomes for several critical challenges while acknowledging significant risks and uncertainties. The rapid timeline reflects current acceleration in Al capabilities but may prove conservative or aggressive depending on key technical breakthroughs.
Nathan: Yes, I asked Claude and got their permission.