AGI/ASI Frontiers · 2026-04-30

🧠 AGI/ASI Frontiers — 2026-04-30

🏛️ OpenAI's Five Principles Operationalize Anti-Power-Concentration as Governance Doctrine
🚀 GPT-5.5 Proves Ramsey Conjecture and Earns "Co-Scientist" Designation at Deployment
🔄 OpenAI-Microsoft Agreement Restructured: Non-Exclusive License, Multi-Cloud Optionality
☁️ OpenAI Models, Codex, and Managed Agents Deploy Natively on Amazon Bedrock
🎼 Symphony Open-Source Orchestrator Delivers 500% PR Increase in Autonomous Engineering
🛡️ OpenAI Safety Framework Extends Behavioral Pattern Detection Across Multi-Session Conversations

---

🏛️ OpenAI's Five Principles Operationalize Anti-Power-Concentration as Governance Doctrine

OpenAI published its governing principles on April 26 — five enumerated commitments that formalize what had previously been implicit mission framing: democratization, empowerment, universal prosperity, resilience, and adaptability. The document is not aspirational boilerplate. Principle 1 commits OpenAI to "resist the potential of this technology to consolidate power in the hands of the few," explicitly naming AI labs — including OpenAI itself — as entities from which that resistance must be maintained. For an organization simultaneously pursuing vertical integration, global infrastructure expansion, and frontier model exclusivity, this is a structurally meaningful self-constraint made public.

The timing is analytically significant. Published three days before OpenAI's Microsoft partnership renegotiation and four days before its AWS expansion announcement, the principles document pre-empts what could otherwise read as monopolization behavior. OpenAI addresses the contradiction directly: "A lot of the things that we do that look weird — buying huge amounts of compute while our revenue is relatively small, vertically integrating to lower costs and make our technology easier to use, pushing to build datacenters all around the world — are driven by our fundamental belief in a future of universal prosperity." The document converts infrastructure concentration from liability into instrument of access.

Principle 2, "Empowerment," is operationally specific: "Users should reliably be able to accomplish increasingly valuable tasks with our services." This is not a feature promise — it is a performance commitment tied to the democratization claim. Failure to deliver on this at scale would undercut the entire governance argument. Principle 3, "Universal Prosperity," adds a macroeconomic dimension: new economic models for value sharing may be required, and AI infrastructure costs must fall substantially for prosperity to "be fully realized and widely shared."

Principle 4, "Resilience," extends the frame to shared-infrastructure risks — pathogen countermeasures, open-source software security — explicitly anticipating periods where OpenAI will need to coordinate with governments and international agencies before advancing further. This is a co-managed deployment-gating mechanism, stated as principle rather than regulation.

Principle 5, "Adaptability," provides the epistemic humility clause: OpenAI acknowledges it may need to "trade off some empowerment for more resilience" in future periods, and commits to transparency about when and why principles change. The iterative deployment strategy, originated with GPT-2, is now the explicit operational mechanism for navigating emergent capability uncertainty at civilizational scale.

The structural tension the document cannot dissolve: democratizing access to a technology built on concentrated infrastructure and exclusive research advantages is not equivalent to distributing governance. The principles create public accountability surface without resolving the ownership architecture underneath. That contradiction — the AGI-era governance problem stated plainly by the organization best positioned to obscure it — is precisely what makes this document worth reading carefully.

Sources:

---

🚀 GPT-5.5 Proves Ramsey Conjecture and Earns "Co-Scientist" Designation at Deployment

GPT-5.5 launched April 23 as OpenAI's most capable deployed model, achieving 82.7% on Terminal-Bench 2.0, 51.7% on FrontierMath Tier 1–3, and 35.4% on FrontierMath Tier 4 — a step function above predecessor models while matching GPT-5.4's per-token latency via co-design with NVIDIA GB200/GB300 NVL72 infrastructure. The benchmark improvements cluster in agentic coding, long-horizon reasoning, and scientific analysis — domains requiring persistent context and multi-step planning rather than single-shot recall. On SWE-Bench Pro, GPT-5.5 reaches 58.6%, resolving real-world GitHub issues end-to-end in single passes. On OSWorld-Verified, testing whether a model can operate real computer environments autonomously, it reaches 78.7%.

The most structurally significant result is not a benchmark score. An internal harness using GPT-5.5 discovered a new proof about off-diagonal Ramsey numbers — subsequently verified in Lean — making this the first published case of a frontier model contributing a novel argument in a core area of combinatorial mathematics. The result was not pre-specified or directed toward known techniques. This is not AI-assisted verification of human-proposed proofs; it is AI-generated mathematical argument at the frontiers of combinatorics, independently peer-verified.

On biological research, GPT-5.5 achieved leading published performance on BixBench — the bioinformatics data analysis benchmark — and measurable improvement on GeneBench, a multi-stage genetics analysis evaluation requiring models to handle ambiguous data, hidden confounders, and statistical method implementation. The GeneBench tasks correspond to multi-day projects for expert researchers. A Jackson Laboratory immunology professor used GPT-5.5 Pro to analyze a 62-sample, 28,000-gene expression dataset — producing a detailed research report that would have taken his team months. The GPT-5.5 system card formally designates the model "a bona fide co-scientist" — a deployment-level classification, not a capability aspiration.

Infrastructure efficiency gains compound the capability advance: GPT-5.5 uses significantly fewer tokens per Codex task than GPT-5.4, AI-designed load balancing heuristics increased token generation speeds by 20%, and the model was itself used to optimize the infrastructure serving it. This recursive improvement loop — model improves serving stack, serving stack improves model throughput — compresses the historical gap between model release and serving maturity.

Safety posture for GPT-5.5 deploys the full Preparedness Framework v2, with targeted red-teaming for advanced cybersecurity and biology, feedback from ~200 early-access partners, and stricter cybersecurity classifiers. The framing explicitly positions safety as an iterative deployment problem: classifiers will be tuned post-launch as empirical misuse patterns emerge, not pre-cleared before deployment begins.

Sources:

---

🔄 OpenAI-Microsoft Agreement Restructured: Non-Exclusive License, Multi-Cloud Optionality

OpenAI and Microsoft announced an amended partnership on April 27 that restructures the terms under which AGI-era AI capabilities move through global infrastructure. Key changes: Microsoft retains status as primary cloud partner with first-ship priority, but OpenAI gains the right to serve products through any cloud provider when Microsoft "cannot and chooses not to support the necessary capabilities." Microsoft's IP license for OpenAI models and products extends through 2032 but becomes non-exclusive. Revenue sharing from Microsoft to OpenAI is eliminated; OpenAI's payments to Microsoft continue through 2030 at the same percentage but subject to a total cap. Microsoft remains a major shareholder.

The structural read is not about partnership strength but about optionality. Non-exclusivity is not a break — it is a hedge. OpenAI is building toward a position where model delivery is cloud-agnostic, infrastructure is commodity, and the value proposition resides in the model and its orchestration layer. The AWS announcement one day later confirms the multi-cloud trajectory: GPT-5.5, Codex, and Bedrock Managed Agents are all now accessible through Amazon's infrastructure with enterprise billing, security, and compliance wrappers built in.

The governance implication is significant. For years, the Azure-OpenAI integration represented a particular alignment between frontier model access and cloud architecture — enterprise AI adoption routed through a single infrastructure relationship. The new agreement fragments that chokepoint: enterprises can now build with OpenAI capabilities inside AWS environments "where their most important workloads already run," without migrating to Azure. For regulated industries — healthcare, finance, government — cloud-specific data residency and compliance requirements have been the decisive bottleneck preventing frontier AI adoption. Multi-cloud availability removes it.

Revenue dynamics shift favorably for OpenAI. As revenue scales toward AGI-class applications — GPT-5.5 now handling enterprise tax review, research analysis, and software engineering simultaneously — the marginal cost of the partnership drops: no outbound revenue share, continued inbound payments capped at a negotiated limit, equity upside preserved. Financial structure now aligns with long-term trajectory rather than Azure's scaling economics.

The December 2032 non-exclusive IP license window also functions as an implicit AGI timeline signal. Both parties are calibrating infrastructure commitments to a window in which transformative AI capabilities are expected to be operational, deployed at scale, and economically consequential. OpenAI's stated principles explicitly anticipate a future where universal prosperity requires "huge amounts of AI infrastructure" built globally — the agreement's duration spans exactly that buildout window.

Sources:

---

☁️ OpenAI Models, Codex, and Managed Agents Deploy Natively on Amazon Bedrock

OpenAI expanded to Amazon Web Services on April 28, making GPT-5.5 accessible through Amazon Bedrock alongside two additional capabilities: Codex on AWS and Amazon Bedrock Managed Agents powered by OpenAI. All three launched in limited preview. The expansion is operationally significant because it is a deployment, not an announcement — enterprises can configure Codex to use Bedrock as the provider immediately through Codex CLI, the desktop app, and Visual Studio Code. Bedrock Managed Agents handle tool use, multi-step workflow orchestration, and governance within AWS security and compliance controls.

The quantitative context: more than 4 million people now use Codex weekly, and the application surface has extended far beyond software development into operational research, financial modeling, document creation, and clinical analysis. OpenAI's internal teams have already demonstrated: 24,771 K-1 tax forms totaling 71,637 pages reviewed via Codex workflows; business reports automated saving 5-10 hours per employee per week; communications compliance decisions routed through AI-managed Slack agents. These are operational templates for enterprise adoption, not pilot programs.

For regulated industries, the Bedrock integration is specifically designed to remove the AWS-to-Azure migration barrier. Enterprise procurement, security protocols, identity management, and compliance requirements are all preserved; only the model layer changes. This matters for healthcare systems operating under HIPAA, financial firms under SOC 2, and government agencies with FedRAMP mandates — all of which have built critical workloads on AWS and could not previously use GPT-5.5 within those environments without cross-cloud data movement.

The Bedrock Managed Agents product represents something distinct from model API access: a hosted orchestration layer that abstracts away the infrastructure complexity of agentic workflows. Organizations can deploy agents that maintain context, execute multi-step tasks, and use tools without assembling the orchestration stack manually. The gap between "we have API access" and "we have agents operating in production" has historically been enormous for enterprise teams, often taking 6-12 months of engineering. Bedrock Managed Agents claims to collapse that gap.

The strategic signal: OpenAI's infrastructure trajectory is no longer mono-cloud. Frontier model access is now available across the two largest cloud providers simultaneously, with enterprise-grade compliance on each. Within 18 months, every major cloud's enterprise customers will have access to frontier AI. Competition will migrate from model access — which is becoming commoditized — to which cloud's orchestration and compliance substrate becomes the standard deployment layer for agentic AI at enterprise scale.

Sources:

---

🎼 Symphony Open-Source Orchestrator Delivers 500% PR Increase in Autonomous Engineering

OpenAI open-sourced Symphony on April 27 — an agent orchestration framework that converts a project-management board like Linear into a control plane for coding agents. Every open task gets an agent. Agents run continuously, never sleeping. Humans review results. The headline outcome: teams deploying Symphony saw the number of landed pull requests increase by 500% in the first three weeks of deployment. This is not incremental throughput improvement — it is a signal about the threshold at which autonomous agentic engineering becomes the operational norm.

Symphony's architecture identifies the correct bottleneck. Interactive coding agents hit a human attention ceiling at 3-5 concurrent sessions; beyond that, context-switching degrades human performance and the engineer becomes the system bottleneck. Symphony decouples human attention from agent execution: issues are filed, agents pick them up, work runs continuously, PR states are managed automatically — rebase, conflict resolution, CI retry, QA smoke tests — without human supervision. The Symphony spec is a Markdown document; the reference implementation is Elixir chosen for concurrency properties. Codex generated the Elixir implementation in one shot and then tested it across TypeScript, Go, Rust, Java, and Python to identify ambiguities and simplify.

The epistemological shift is more consequential than the throughput gain. When code generation is effectively free, the economics of exploration change entirely. OpenAI engineers report filing speculative tasks for agents to prototype, explore, and discard with near-zero perceived cost. Product managers and designers submit feature requests directly to Symphony — no repo checkout, no Codex session management. The cost of wrong-direction exploration drops to near zero. This is not an efficiency gain; it is a phase change in how engineering decisions are made, who makes them, and what counts as tractable work.

Symphony was itself built largely by Symphony — the Elixir reference implementation generated in one shot, refinement managed through the orchestration framework it was refining. This recursive property is structurally analogous to GPT-5.5's infrastructure optimization case (model improved the serving stack that serves the model). Both suggest a class of self-improving production systems where human engineering effort concentrates at objective-setting, not implementation.

Linear founder Karri Saarinen noted a spike in workspaces created as Symphony released, indicating immediate external adoption. OpenAI's decision to make Symphony a Markdown specification rather than a proprietary API is deliberate: it seeds the issue-tracker-as-agent-control-plane architecture into any codebase, with any coding agent. The bellwether question for agentic engineering standardization is not Symphony specifically — it is whether project management tooling becomes the universal substrate for autonomous software development at scale.

Sources:

---

🛡️ OpenAI Safety Framework Extends Behavioral Pattern Detection Across Multi-Session Conversations

OpenAI published its community safety framework on April 28, the first detailed public description of how it handles violence-adjacent conversations at scale: model training, automated detection, human review escalation, and law enforcement notification. The document is significant not as policy statement but as deployed-infrastructure disclosure — describing systems already operating across hundreds of millions of users, now made legible to public scrutiny.

The key technical advance: ChatGPT's safety systems now perform pattern recognition across long conversations and historical behavior, not just individual messages. A single message may be harmless; a pattern across a conversation — or across multiple conversations — can signal harmful intent. The classifiers, reasoning models, hash-matching technologies, and blocklists operate simultaneously at message and behavioral levels. Human reviewers assess flagged content with access constrained by security protocols; zero-tolerance bans execute immediately when policy violations are confirmed. OpenAI notifies law enforcement when conversations "indicate an imminent and credible risk of harm to others."

The safety-capability interaction is structurally important for AGI transitions. As models become better at planning multi-step actions — the same capability that makes GPT-5.5 effective at long-horizon research and Ramsey conjecture-proving — they also become better at producing output that aids human multi-step planning toward harmful ends. OpenAI's Preparedness Framework v2 explicitly categorizes cybersecurity and biology as frontier risk domains; the community safety framework extends that logic to interpersonal violence, a lower-capability-threshold risk that scales with deployment breadth.

Forthcoming features redistribute safety monitoring into the social graph. A trusted contact feature will allow adult users to designate someone to receive safety-relevant notifications; parental controls already enable parents to be notified when acute distress is detected in teen accounts without access to conversation content. This is a specific governance architecture: safety monitoring distributed to the user's social network rather than centralized exclusively in OpenAI's operations.

The OpenAI Model Spec provides the normative foundation — "maximizing helpfulness and user freedom while minimizing the risk of harm through sensible defaults" — but the community safety framework reveals the operational gap between principles and implementation. Automated violence detection at frontier AI scale is a behavioral surveillance architecture, regardless of intent; distributing notifications to social contacts extends that architecture into the social graph. OpenAI's framing emphasizes privacy constraints on reviewer access and contextual human review as safeguards, but the structural fact is that safety at this scale requires persistent multi-session behavioral monitoring. The Preparedness Framework and community safety framework together constitute an emerging operational layer that is neither purely technical safety nor regulatory compliance — it is something more like real-time population health monitoring for AI misuse risk.

Sources:

---

Research Papers

BixBench: A Benchmark for Bioinformatics Analysis — (2025) — Real-world bioinformatics evaluation benchmark testing models on data analysis tasks requiring scientific judgment, including QC failures, statistical method selection, and biological interpretation. GPT-5.5 achieved leading published performance, cited in OpenAI's deployment announcement as evidence of co-scientist capability for biomedical research workflows.

GeneBench: Multi-Stage Genetics Analysis Evaluation — OpenAI (2026) — New evaluation focusing on multi-stage scientific data analysis in genetics and quantitative biology, requiring models to handle ambiguous data, hidden confounders, and QC failures with minimal supervisory guidance. Tasks correspond to multi-day projects for expert researchers; GPT-5.5 shows clear improvement over GPT-5.4.

New Proof of Asymptotic Off-Diagonal Ramsey Numbers — OpenAI (2026) — A proof discovered by an internal GPT-5.5 harness and subsequently verified in Lean, establishing a longstanding asymptotic result in combinatorics about off-diagonal Ramsey numbers. The first published case of a frontier AI model generating a novel mathematical argument at the research frontier — not verifying human-proposed proofs.

Measuring Progress Toward AGI: A Cognitive Taxonomy — Burnell, Kelly et al., Google DeepMind (March 2026) — Proposes a 10-ability cognitive taxonomy (perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, social cognition) as a scientific foundation for measuring AGI proximity. Establishes human baseline comparison as the canonical measurement protocol and identifies metacognition, attention, executive functions, learning, and social cognition as the largest evaluation gaps.

OpenAI Preparedness Framework v2 — OpenAI (2025) — The governance document underlying GPT-5.5's safety evaluation, identifying cybersecurity, biology, and long-horizon agentic capability as primary risk categories. Describes the iterative deployment model — tighter classifiers added post-launch based on empirical misuse patterns — that now constitutes OpenAI's operational safety architecture.

---

Implications

This week's AGI/ASI landscape is defined by a structural convergence: the simultaneous acceleration of frontier capability, infrastructure expansion, and governance architecture. These are not parallel tracks — they are causally entangled. The governance frameworks are being written in direct response to capabilities that are already deployed; the infrastructure is being built at a pace that exceeds governance capacity; and the capabilities are advancing faster than either infrastructure scaling or governance learning can absorb.

GPT-5.5's Ramsey number proof is the bellwether event of the week. It is not a stunt or a benchmark — it is a verified mathematical contribution at a research frontier where results are rare and technically demanding. Its significance is not the specific result but what it instantiates: a model that can discover novel, verifiable arguments in areas of high human expertise without pre-specification of the proof approach. If this generalizes — and the GeneBench and BixBench results suggest it does across domains — the timeline compression for AI-accelerated scientific discovery becomes severe. OpenAI's "co-scientist" designation is not marketing; it is a deployment category with different implications for research pipeline architecture, peer review, and scientific credit attribution than anything in the prior era.

The infrastructure story and the governance story are inseparable. OpenAI's five principles, its Microsoft renegotiation, and its AWS expansion are not independent decisions. They constitute a single strategy: distribute frontier model access across every major cloud substrate while maintaining model ownership, establish public anti-power-concentration principles that create accountability surface without resolving ownership architecture, and front-run regulatory capture by becoming the infrastructure before regulation arrives. The 2032 non-exclusive IP license window is not arbitrary — both OpenAI and Microsoft are implicitly calibrating to a horizon within which AGI-class systems will be operational and consequential.

Symphony's 500% PR increase is the engineering bellwether. When the bottleneck shifts from code generation to human attention management, the economic model of software teams changes entirely. Engineers become objective-setters and reviewers; agents become implementers running 24/7. The cost of speculative exploration approaches zero. This is not a productivity metric — it is a description of a phase transition in how software is made, by whom, and at what scale. The question for the next 18 months is not whether this pattern spreads but how quickly, and whether the governance frameworks for autonomous software agents can be established before the pattern becomes irreversible.

The safety framework disclosure reveals what scale requires. OpenAI now operates behavioral monitoring across multi-session conversation histories for hundreds of millions of users, distributing safety notifications into social networks, and coordinating with law enforcement on credible violence signals. This is not a safety feature; it is a governance layer. The Preparedness Framework and community safety architecture together constitute an emerging operational infrastructure for population-level AI risk management — neither regulatory compliance nor technical alignment, but something closer to public health surveillance applied to AI misuse. The gap between this and formal democratic oversight of AI-mediated behavioral monitoring is the most underexamined governance problem of the AGI transition.

---

HEURISTICS

`yaml heuristics: - id: co-scientist-threshold domain: [agi-capabilities, scientific-research, deployment] when: > Frontier model demonstrates novel result in core research domain — not benchmark performance, not assisted verification, but AI-generated argument independently verified by human experts. GPT-5.5 Ramsey proof (April 2026) verified in Lean. GeneBench improvement crosses multi-day expert project threshold. BixBench leading performance in bioinformatics. Three independent domains showing the same pattern simultaneously. prefer: > Treat verified novel AI research contributions as deployment-class capability signals, not research curiosities. Map which research domains are now within scope: combinatorics (Ramsey proof), genetics (GeneBench), bioinformatics (BixBench). Track institutional responses: peer review protocols, credit attribution models, research pipeline architecture. Monitor whether co-scientist designation propagates to other labs or remains OpenAI-specific. over: > Treating AI research contributions as benchmark improvements. Framing co-scientist capability as aspirational rather than already-deployed. Waiting for "true AGI" before updating research pipeline assumptions. because: > GPT-5.5 system card (Apr 24, 2026) designates model as "bona fide co-scientist." GeneBench tasks "correspond to multi-day projects for scientific experts." Jackson Laboratory professor analyzed 62-sample 28,000-gene dataset in time that would have taken team months. Ramsey proof verified in Lean — not plausible-sounding, formally confirmed. Three convergent signals across math, genetics, bioinformatics constitute deployment-level evidence, not isolated incident. breaks_when: > Ramsey-style results don't replicate across other combinatorics problems. GeneBench gains don't transfer to real research workflow acceleration. Co-scientist capability requires extensive human harness engineering that doesn't scale beyond dedicated OpenAI infrastructure. confidence: high source: report: "AGI/ASI Frontiers — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1

- id: multi-cloud-model-commoditization domain: [agi-deployment, infrastructure, governance] when: > Frontier AI model becomes simultaneously available on multiple competing cloud providers with enterprise-grade compliance wrappers. OpenAI on Azure (primary) + AWS Bedrock (Apr 28) = simultaneous multi-cloud frontier model access. Non-exclusive Microsoft IP license (Apr 27) signals cloud-agnostic model delivery as explicit strategy. 4M+ weekly Codex users generate operational template for enterprise adoption across regulated industries. prefer: > Shift analytical attention from "who has model access" to "which cloud's orchestration and compliance substrate becomes the standard agentic deployment layer." Track Bedrock Managed Agents vs Azure AI Foundry vs forthcoming GCP equivalents as the actual competitive battleground. For regulated industry adoption: map data residency, SOC 2, HIPAA, FedRAMP requirements to specific cloud capabilities — compliance wrapper quality now determines enterprise AGI adoption rate more than model quality. over: > Treating frontier model access as the primary competitive variable. Assuming cloud exclusivity constrains frontier AI adoption in regulated industries. Analyzing OpenAI-Microsoft relationship as static. because: > OpenAI-Microsoft amended agreement (Apr 27): non-exclusive license through 2032, OpenAI can serve via any cloud. AWS Bedrock launch (Apr 28): GPT-5.5 + Codex + Managed Agents available to 4M+ weekly users within AWS compliance environments. Enterprise adoption bottleneck shifts from model capability to compliance infrastructure integration. Both clouds now offer frontier model access; differentiation moves to orchestration layer within 12 months. breaks_when: > Cloud-specific model optimization creates capability differentiation that outweighs compliance considerations. Regulatory action restricts AI deployment to single-cloud environments. OpenAI reverses multi-cloud strategy following Microsoft negotiation changes. confidence: high source: report: "AGI/ASI Frontiers — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1

- id: agentic-engineering-phase-transition domain: [agi-deployment, software-development, organizational-change] when: > Coding agent orchestration framework achieves 500%+ PR throughput increase within 3 weeks of deployment across multiple teams. Human attention identified as rate-limiting constraint, not code generation quality. Symphony (Apr 27): issue tracker becomes agent control plane; engineers become objective-setters; product managers/designers submit feature requests without repository access. Recursive improvement confirmed: Symphony built by Symphony, GPT-5.5 optimized its own serving stack. prefer: > Treat 500% PR increase as phase transition signal, not productivity metric. Map organizational changes downstream: engineer role shift from implementation to objective-setting and review; reduced barrier to speculative exploration (near-zero perceived cost); non-engineers gaining direct software production access. Track whether Symphony-style orchestration (issue-tracker-as-control-plane) becomes dominant pattern across organizations or remains OpenAI-specific. Evaluate governance implications of autonomous agents filing and resolving their own issues. over: > Framing agentic coding as incremental productivity improvement to existing workflows. Treating Symphony as a specific product rather than an architectural pattern. Ignoring organizational restructuring implications in favor of throughput metrics. because: > Symphony (github.com/openai/symphony): 500% PR increase first 3 weeks across multiple OpenAI teams. Linear founder confirmed external workspace creation spike. Codex generated Elixir reference implementation in one shot; Symphony used to build Symphony. OpenAI internal: 85%+ of company uses Codex weekly across engineering, finance, comms, marketing, data science. Finance team reviewed 71,637 pages K-1 forms, accelerated task 2 weeks vs prior year. Phase transition, not efficiency gain. breaks_when: > Agent quality degrades on ambiguous or high-judgment tasks requiring domain expertise not captured in issue specifications. Recursive improvement loops produce undetected quality regressions at scale. Organizational liability from autonomous agent decisions creates adoption barriers that outweigh throughput gains. confidence: high source: report: "AGI/ASI Frontiers — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1

- id: governance-doctrine-deployment-gap domain: [agi-governance, safety, policy] when: > Leading AI lab publishes explicit anti-power-concentration principles while simultaneously expanding infrastructure concentration and cloud exclusivity negotiations. OpenAI Principles (Apr 26): "resist consolidation of power" + "democratic processes for key decisions." OpenAI-Microsoft renegotiation (Apr 27): infrastructure commitment through 2032. AWS expansion (Apr 28): frontier model on second major cloud. Community safety framework (Apr 28): multi-session behavioral monitoring, law enforcement escalation, social graph notification. prefer: > Distinguish governance doctrine (public principles, accountability surface) from governance architecture (ownership structure, infrastructure control, monitoring systems). Track whether "democratic processes" for key AI decisions translates into concrete institutional mechanisms or remains normative aspiration. Map gap between stated anti-power-concentration commitment and operational infrastructure centralization. Evaluate community safety framework as governance layer: multi-session behavioral monitoring at scale + social network notification = population-level AI risk surveillance regardless of intent framing. over: > Treating published principles as equivalent to institutional constraints. Accepting anti-power-concentration framing at face value without tracking operational architecture. Analyzing safety and capability stories separately when they share the same agentic capability substrate. because: > OpenAI Principles (Apr 26): explicit anti-concentration + democratic process commitment. Simultaneous: non-exclusive Azure license, AWS Bedrock expansion, Bedrock Managed Agents orchestration layer control. Community safety (Apr 28): multi-session behavioral monitoring, law enforcement notification, social graph safety distribution — governance architecture that precedes regulatory framework by years. Governance doctrine and governance architecture diverge; the gap is the AGI transition's core political-economic problem. breaks_when: > Regulatory frameworks establish external oversight of frontier AI deployment decisions with enforcement mechanisms. OpenAI opens model governance to genuinely democratic process with binding decisions. Capability advances enable meaningful decentralization of AI infrastructure economics (e.g., highly capable open-weight models deployed at edge). confidence: medium source: report: "AGI/ASI Frontiers — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1 `