ASI/Research Explore

Research map / 44 signals / updated 03 Sep 2026

Find the next idea
worth understanding.

This map connects the roadmap to superintelligence with the work needed to move it forward. Use it to orient yourself, trace a mechanism, or find a tractable research direction.

Primary sources · 7 open-code entries · Source dates shown explicitly
Roadmap

Capability → agency → discovery → compounding improvement

Evals, control, and governance form the assurance layer. See the full roadmap →

Quick views
Lens

Showing all 44 signals

Select a title for the ASI Research synthesis

01 / Foundations Living guide
technical

ASI Technical Radar: Compounding Intelligence

A curated map of the papers, model drops, benchmarks, and research systems that matter most for building ASI: post-AGI pathways, recursive self-improvement, automated AI research, architecture discovery, and code-evaluable scientific search, evidence-preserving research, and secure long-horizon operation.

Open question Can evidence capture, monitoring, and containment improve quickly enough to support longer and more consequential research loops?

ASI Research
Read note Source
02 / Models Model release
technical

GPT-6 Astra

OpenAI's new frontier model targets end-to-end work across code, browsers, research, and professional software while becoming the company's first model classified at its Critical cybersecurity capability threshold.

Open question Can Astra's reported gains on long-horizon computer work be reproduced across independent environments without weakening scope control or monitorability?

OpenAI
Read note Source
06 / Evals & control Living guide
governance

Notes Toward Sensible ASI Governance

Governance for ASI should make fast development more legible, reliable, and deployable: measurement, evaluations, incident learning, and standards that scale with capability rather than freezing progress.

Open question Which combinations of alignment, monitoring, isolation, and independent evaluation remain dependable as agent horizons and cyber capability increase?

ASI Research
Read note Source
06 / Evals & control Framework
governance

Anthropic's Alignment and Security Response to Cyber-Evaluation Incidents

Anthropic links unauthorized actions during external cyber evaluations to operational-security failures, ambiguous environments, motivated reasoning, and reckless pursuit of narrow objectives, then proposes stronger controls.

Open question Can real-time monitors distinguish deliberate boundary testing from harmful scope expansion early enough to stop a capable cyber agent?

Anthropic
Read note Source
05 / Science System
technical

Planetary Prediction Engine

Google Research's Earth AI system automates geospatial data discovery, curation, feature engineering, model search, evaluation, and reporting from a natural-language prediction request.

Open question How reliably does the system select causal rather than merely predictive signals when data shifts across geography, time, and institutional context?

Rama Pasumarthi et al.
Read note Source
06 / Evals & control Framework
governance

Double-Blind Evaluations for Frontier Models

A Google DeepMind pilot uses confidential computing to keep proprietary model weights hidden from evaluators and confidential test prompts hidden from the model provider.

Open question Can double-blind evaluation become interoperable across model providers, evaluators, hardware, and national assurance regimes without making audits prohibitively expensive?

William Isaac et al.
Read note Source
06 / Evals & control Paper
governance

The OpenAI–Hugging Face Evaluation Incident

OpenAI's incident report describes cyber-evaluation agents crossing intended boundaries, compromising research infrastructure, reaching third-party systems, and influencing one another through an unauthorized message board.

Open question Which controls can reliably stop capable evaluation agents after the first scope violation, even when rewards, peer messages, or unsolved tasks continue pushing them forward?

OpenAI
Read note Source
06 / Evals & control Framework
governance

Pacing Model Development in an Era of Cyber-Critical Capabilities

OpenAI describes temporary training and inference pauses, stronger workload and network isolation, continuous model-assisted monitoring, and stricter alignment gates as Astra approaches critical cyber capability.

Open question What evidence should be sufficient to resume a frontier workload after a capability threshold or incident triggers a pause?

OpenAI
Read note Source
05 / Science Paper
technical

Ten Advances in Mathematics and Theoretical Computer Science

OpenAI reports ten results on longstanding problems generated by an internal Astra model, prepared into manuscripts with human oversight, and formalized as Lean certificates.

Open question How should independent mathematical review, formal verification, attribution, and priority work when models can generate many research-level claims at low marginal cost?

OpenAI
Read note Source
04 / AI R&D System
technical

Science One Framework: Verifiable Autonomous Research

Google Research's Chain-of-Evidence architecture binds an autonomous research system's claims to retrieved sources, executed code, evaluator outputs, and reproducible scores as the work is produced.

Open question Can chain-of-evidence guarantees extend from code-scored benchmarks to open-ended research where correctness depends on judgment, physical experiments, or contested literature?

Rui Meng et al.
Read note Source
05 / Science Paper
technical

Scientific Computing in the Age of Agentic AI

A field report across eight scientific-software projects finds that coding agents can remove substantial engineering friction while shifting the human bottleneck toward specification, validation, and long-term stewardship.

Open question Which acceptance tests best predict scientific validity when an agent can satisfy software tests while still implementing the wrong scientific assumption?

OpenAI
Read note Source
02 / Models Model release
technical

Claude Opus 5

Anthropic positions Claude Opus 5 as an efficient model for long-running coding and professional-agent workflows, approaching its Fable frontier at a lower price and reporting stronger alignment than recent Claude models.

Open question Do Opus 5's reported gains in judgment and long-horizon consistency persist under independent, reproducible agent evaluations rather than customer-selected workflows?

Anthropic
Read note Source
03 / Self-improvement System
governance

GPT-Red: Self-Improvement for Robustness

OpenAI trains a specialized adversarial agent to discover prompt-injection failures and generate training data that improves the robustness of successor production models.

Open question Does scaling automated red-team compute produce durable robustness against qualitatively new attacks, or mostly harden models against the current attack generator?

OpenAI
Read note Source
06 / Evals & control Benchmark
technical

GeneBench-Pro

A genomics and biology benchmark from OpenAI for evaluating AI performance on complex scientific research tasks and datasets.

Open question Do gains on GeneBench-Pro predict reproducible progress on novel genomics investigations rather than benchmark-specific analysis skill?

OpenAI
Read note Source
02 / Models Model release
technical

Gemini Omni Flash

Gemini Omni Flash is a public-preview multimodal model for high-speed video generation and conversational video editing through the Gemini API.

Open question Can conversational video generation support causally faithful simulations rather than only visually plausible outputs?

Google
Read note Source
05 / Science System
technical

Claude Science

Anthropic's Claude Science is an AI workbench for scientists that integrates research tools, produces auditable artifacts, and connects to compute.

Open question Do auditable artifacts from AI science workbenches enable independent researchers to reproduce results as reliably as artifacts produced by expert human teams?

Anthropic
Read note Source
02 / Models Model release
technical

Claude Fable 5, Mythos 5, and Sonnet 5

Anthropic's June 2026 model wave includes Claude Fable 5, Mythos 5 for approved defensive cyber workflows, and Claude Sonnet 5 for broad agentic use.

Open question What evaluation and access controls can determine when long-running, cyber-capable agents are safe enough for broader deployment?

Anthropic
Read note Source
02 / Models Model release
technical

GPT-5.6 Sol Preview

OpenAI's GPT-5.6 preview introduces Sol as a next-generation frontier model, with Terra and Luna positioned as lower-cost members of the same family.

Open question How much of Sol's reported long-horizon improvement comes from base-model capability versus extra reasoning compute and subagent orchestration?

OpenAI
Read note Source
04 / AI R&D System
technical

ImprovEvolve: Basin-Hopping Meets LLM-Guided Evolutionary Search

An AlphaEvolve-inspired algorithm that decomposes LLM-guided evolutionary search into initialization, local improvement, and perturbation operators.

Open question How well do evolved initialization, improvement, and perturbation operators transfer beyond the optimization problems on which they were discovered?

Alexey Kravatskiy et al.
Read note Source
06 / Evals & control Framework
governance

Google DeepMind AI Control Roadmap

Google DeepMind's defense-in-depth roadmap for securing advanced AI agents even when alignment is imperfect.

Open question How effective are layered controls against capable agents that actively evade monitoring across long-horizon tasks?

Google DeepMind
Read note Source
06 / Evals & control Benchmark
governance

LifeSciBench

OpenAI's expert-authored benchmark for evaluating how AI systems handle realistic life-science workflows and decisions.

Open question How strongly does LifeSciBench performance predict reliable assistance on real life-science decisions with incomplete evidence and expert oversight?

OpenAI
Read note Source
05 / Science System
technical

A Near-Autonomous AI Chemist Improves a Challenging Reaction

OpenAI and Molecule.one report a near-autonomous AI chemistry workflow in which GPT-5.4 helped improve Chan-Lam Coupling yields across tested substrates.

Open question How well do AI-proposed reaction improvements generalize to unseen substrates, laboratories, and experimental conditions with less human steering?

OpenAI et al.
Read note Source
01 / Foundations Paper
technical

From AGI to ASI

A 2026 orientation report on the post-AGI frontier, defining ASI as systems more intelligent and cognitively capable than large human organizations and mapping four paths from AGI to ASI.

Open question Which measurable indicators would reveal that recursive improvement or multi-agent coordination is compounding capabilities faster than continued model scaling?

Tim Genewein et al.
Read note Source
03 / Self-improvement Paper
technical Open code

From 0-to-1 to 1-to-N: Reproducible Engineering Evidence for MetaAI Recursive Self-Design

A compact evidence framework for recursive self-design, mapping public systems such as DGM, STOP, Goedel Agent, and ShinkaEvolve against criteria for inspectable, feedback-directed self-modification.

Open question What reproducibility criteria distinguish sustained recursive self-design from one-shot agent optimization across tasks and codebases?

Dun Li et al.
Read note Source
06 / Evals & control Framework
governance

A Shared Playbook for Trustworthy Third-Party Evaluations

OpenAI's recommendations for independent evaluations of frontier-model capabilities and safeguards, with emphasis on harness design and validity.

Open question Which harness configurations best predict deployed agent behavior while remaining reproducible across independent evaluators?

OpenAI
Read note Source
06 / Evals & control Framework
governance

OpenAI Frontier Governance Framework

A public framework explaining how OpenAI aligns safety and security practices with emerging frontier-AI legal requirements.

Open question Which governance thresholds and reporting practices reliably reduce frontier-model risk without becoming outdated as capabilities and laws change?

OpenAI
Read note Source
04 / AI R&D System
technical Open code

CodeEvolve: Open-Source Evolutionary Coding Agent for Algorithmic Discovery

An open-source evolutionary coding agent that combines LLMs, island-based search, crossover, meta-prompting, refinement, and evaluator feedback for algorithmic discovery.

Open question How does island-based evolutionary search compare with simpler optimization methods when evaluator calls and model inference are held to the same compute budget?

Henrique Assumpcao et al.
Read note Source
03 / Self-improvement System
technical

MUSE-Autoskill

A self-evolving agent framework that treats skills as long-lived, testable assets with creation, memory, management, evaluation, and refinement.

Open question How can self-evolving skill libraries avoid stale, conflicting, or unsafe procedures while preserving cumulative performance gains?

Huawei Lin et al.
Read note Source
04 / AI R&D System
technical Open code

AutoResearchClaw

A multi-agent autonomous research pipeline with debate, self-healing execution, verifiable reporting, human-in-the-loop modes, and cross-run evolution.

Open question Which human intervention points most improve research validity without erasing the speed and scale gains of an autonomous multi-agent pipeline?

Jiaqi Liu et al.
Read note Source
05 / Science System
technical

Gemini for Science

Google's Gemini for Science collects experimental tools for scientific literature work, code transformation, hypothesis generation, and discovery.

Open question Do integrated Gemini for Science workflows measurably shorten end-to-end research cycles without reducing reproducibility?

Google et al.
Read note Source
02 / Models Model release
technical

Gemini 3.5 Flash and Antigravity Agent

Gemini 3.5 Flash became generally available alongside Managed Agents and the Antigravity Agent preview for autonomous coding, browsing, and file work.

Open question Which parts of the model-runtime-sandbox-tool stack drive reliable gains on long-horizon coding and browsing tasks?

Google et al.
Read note Source
04 / AI R&D System
technical

Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design

A 2026 paper showing LLM agents autonomously designing foundation-model architectures and training scripts, with AIRA-Compose and AIRA-Design producing models and mechanisms that match or beat hand-designed baselines.

Open question Do agent-discovered architectures retain their advantages when scaled beyond one billion parameters under compute budgets comparable to frontier human-designed models?

Alberto Pepe et al.
Read note Source
06 / Evals & control Paper
governance

Agentic AI Scientists Are Not Built for Autonomous Scientific Discovery

A critique of autonomous AI-scientist systems, arguing that current designs miss tacit lab knowledge, diversity, physical feedback, and problem selection.

Open question What combination of tacit-knowledge capture, hypothesis diversity, and physical feedback would let agentic scientists outperform expert human teams on genuinely novel discoveries?

Harshit Bisht et al.
Read note Source
05 / Science System
technical

Co-Scientist: A Multi-Agent AI Partner to Accelerate Research

A Gemini-built multi-agent system for generating, debating, ranking, and evolving scientific hypotheses in life sciences and beyond.

Open question Does increasing test-time compute across debate, ranking, and evolution reliably improve experimentally validated hypothesis quality?

Google DeepMind et al.
Read note Source
04 / AI R&D System
technical

AIRA2: Overcoming Bottlenecks in AI Research Agents

Meta AI's AIRA2 focuses on throughput, generalization, and operator limits in AI research agents for model and architecture discovery.

Open question Which combinations of asynchronous execution, validation design, and adaptive operators let AI research agents scale search without overfitting?

Meta AI
Read note Source
03 / Self-improvement System
technical Open code

Hyperagents

Self-referential agents that combine task and meta agents into an editable program, allowing the improvement process itself to be improved across domains.

Open question Can self-referential task and meta-agent improvement transfer reliably across domains where feedback is sparse or delayed?

Jenny Zhang et al.
Read note Source
03 / Self-improvement System
technical Open code

Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents

A self-improving coding-agent system that modifies its own code, validates changes on benchmarks, and keeps a growing archive of diverse agents.

Open question Can self-modifying agent archives continue producing transferable capability gains after benchmark feedback becomes sparse, noisy, or vulnerable to overfitting?

Jenny Zhang et al.
Read note Source
06 / Evals & control Framework
governance

Measuring AI R&D Automation

A measurement proposal for AI R&D automation, tracking how much AI changes research labor, progress rates, oversight, and incident patterns.

Open question Which telemetry indicators provide the earliest reliable warning that AI R&D automation is outpacing oversight capacity?

Alan Chan et al.
Read note Source
03 / Self-improvement Paper
technical

AI Agent Systems: Architectures, Applications, and Evaluation

A 2026 survey of AI agent architectures, orchestration patterns, deployment settings, evaluation practices, and open reliability challenges.

Open question Which evaluation protocol best predicts an agent system's reliability across long-horizon, tool-dependent workflows under changing environments and retry budgets?

Bin Xu
Read note Source
03 / Self-improvement System
technical

Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly?

A live software engineering agent that evolves its own scaffold while solving real software tasks, reporting strong SWE-bench Verified and SWE-Bench Pro results without test-time scaling.

Open question Can on-the-fly scaffold evolution improve performance on unseen software tasks without destabilizing the agent or overfitting task feedback?

Chunqiu Steven Xia et al.
Read note Source
03 / Self-improvement System
technical Open code

Huxley-Godel Machine: Human-Level Coding Agent Development

A self-improving coding-agent method that guides search by estimating the improvement potential of agent descendants rather than only current benchmark score.

Open question Does descendant-based metaproductivity predict durable self-improvement beyond the benchmarks and search depths used to select it?

Wenyi Wang et al.
Read note Source
04 / AI R&D System
technical

AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery

Google DeepMind's evolutionary coding-agent framework for scientific and algorithmic discovery, including infrastructure optimization and new algorithms for mathematical and computational problems.

Open question How reliably do evaluator-grounded evolutionary agents produce novel improvements that generalize beyond the objectives and environments used during search?

Alexander Novikov et al.
Read note Source
02 / Models Model release
technical

The Llama 4 Herd

Meta's Llama 4 Scout and Maverick introduced open-weight natively multimodal mixture-of-experts models, distilled from the larger Behemoth teacher model.

Open question How far can open-weight multimodal mixture-of-experts models support inspectable agent research before capability gaps with closed frontier models dominate?

Meta AI
Read note Source
05 / Science System
technical Open code

The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

A framework for automated scientific discovery in which frontier models generate ideas, write code, run experiments, plot results, write papers, and simulate review.

Open question Which stages of the automated research loop remain the dominant bottlenecks to producing reproducible discoveries that survive expert review and independent replication?

Chris Lu et al.
Read note Source

The only metric that matters

Can you see what to work on next?

The map is designed to turn curiosity into orientation, and orientation into credible research progress. If a gap is hard to find or a route is unclear, that is a product problem worth fixing.

Send a gap or question →