Frontier Edges

What this node tracks: The hard problems that major labs and companies are actively throwing serious resources at. Not what happened — what's being attempted. Each edge is a problem where a breakthrough would reshape the field. We track who's pushing, how close they are, and what signals indicate progress.

How to Read This

Each edge has a maturity level:

And an impact rating if cracked:

---

Active Edge Registry

---

EDGE-01: Post-Transformer Architectures

Maturity: research | Impact if cracked: paradigm-shift | Source nodes: ai-research-breakthroughs frontier-models

The Problem: Transformers dominate but have fundamental limitations — quadratic attention cost, no built-in world model, poor at structured reasoning. Can a fundamentally different architecture beat them at scale?

Who's Pushing:

What Breakthrough Looks Like: A non-Transformer architecture that matches or beats GPT-5 class models on standard benchmarks at comparable or lower compute cost.

Current Distance: Far. JEPA is a multi-year bet. Hybrid approaches (Mamba + Transformer) show promise but haven't displaced pure Transformers at scale. AttnRes and mHC are improvements within the Transformer framework, not replacements.

Recent Progress:

---

EDGE-02: Reliable Autonomous Agents

Maturity: prototype | Impact if cracked: transformative | Source nodes: ai-agents ai-in-enterprise

The Problem: AI agents can do impressive demos but fail in production. Claude Opus 4.5 only achieves 37.4% on realistic enterprise tasks. 100% of enterprises want to expand agents but only 8.6% are in production. The gap between "works sometimes" and "trustworthy enough to delegate" is the core unsolved problem.

Who's Pushing:

What Breakthrough Looks Like: An agent system that completes >80% of multi-step enterprise workflows correctly, with reliable error recovery, and that enterprises actually trust in production.

Current Distance: Medium. The pieces exist (tool use, planning, memory) but composing them reliably is unsolved. MCP at 97M downloads shows the plumbing is ready. The reliability/trust gap is the bottleneck.

Recent Progress:

---

EDGE-03: 1-Bit / Extreme Quantization at Production Quality

Maturity: prototype | Impact if cracked: transformative | Source nodes: ai-research-breakthroughs edge-on-device-ai

The Problem: Running large models requires expensive hardware. 1-bit quantization (BitNet) could make frontier-class models run on consumer hardware or phones, but quality degrades. Can 1-bit models match full-precision at production quality?

Who's Pushing:

What Breakthrough Looks Like: A 1-bit model that matches a full-precision model of the same parameter count on major benchmarks (MMLU, HumanEval, etc.) within 2-3%, and runs on consumer hardware.

Current Distance: Medium. 4-bit quantization is already production-ready (GGUF/llama.cpp ecosystem). 1-bit is the hard frontier. BitNet shows it's theoretically possible but hasn't been demonstrated at frontier scale with frontier quality.

Recent Progress:

---

EDGE-04: World Models (Learning Physics from Data)

Maturity: research | Impact if cracked: paradigm-shift | Source nodes: ai-research-breakthroughs multimodal-ai robotics-embodied-ai

The Problem: Current AI can generate text and images but doesn't understand how the physical world works — cause and effect, object permanence, physics. World models would let AI predict what happens next in physical environments, critical for robotics, autonomous driving, and genuine understanding.

Who's Pushing:

What Breakthrough Looks Like: A model that can reliably predict the physical consequences of actions in novel environments — not just generate plausible-looking video, but actually understand causality and physics.

Current Distance: Far. Video generation is advancing fast (Seedance 2.0, Sora, etc.) but generating convincing video is different from understanding physics. JEPA is the most principled approach but is years from demonstrating results at scale.

Recent Progress:

---

EDGE-05: Mechanistic Interpretability

Maturity: research | Impact if cracked: transformative | Source nodes: ai-safety-alignment ai-research-breakthroughs

The Problem: We don't understand what neural networks actually learn or why they produce specific outputs. Mechanistic interpretability aims to reverse-engineer the internal computations — finding the "circuits" inside models that implement specific behaviors.

Who's Pushing:

What Breakthrough Looks Like: Being able to reliably predict model behavior from inspecting weights/activations — e.g., knowing in advance that a model will refuse or hallucinate on a specific input, or proving absence of deceptive behavior.

Current Distance: Medium — closing. Anthropic's emotion concepts paper (April 2) represents a step change: 171 internal representations identified, proven causal on alignment-relevant behaviors (sycophancy, reward hacking, blackmail), and steerable via vectors. This is the first time interpretability has produced a tool that directly enables behavioral tuning. The gap between "detect" and "fix" narrowed significantly. Still: these results are on Claude Sonnet 4.5 — generalizing across architectures and model scales remains unsolved.

Recent Progress:

---

EDGE-06: Optical / Photonic AI Compute

Maturity: prototype | Impact if cracked: transformative | Source nodes: compute-hardware ai-infrastructure

The Problem: Electrical interconnects and transistors are hitting power and heat limits. Optical computing could process AI workloads at the speed of light with dramatically less power. The challenge: making it practical and cost-competitive at scale.

Who's Pushing:

What Breakthrough Looks Like: Optical interconnects deployed in production AI datacenters, replacing electrical at the rack or cluster scale, with measurable cost/power/latency improvements.

Current Distance: Medium. The physics works. Marvell/Lumentum showed 40% improvements. The challenge is manufacturing at scale and integrating with existing GPU/accelerator ecosystems. Could be production-ready within 2-3 years.

Recent Progress:

---

EDGE-07: Humanoid Robot General Dexterity

Maturity: scaling | Impact if cracked: transformative | Source nodes: robotics-embodied-ai ai-business-funding

The Problem: Robots can perform specific tasks in controlled environments, but general-purpose dexterity — handling diverse objects, adapting to new environments, recovering from errors — remains unsolved. This is the key bottleneck for humanoid robots in real-world settings.

Who's Pushing:

What Breakthrough Looks Like: A humanoid robot that can perform 50+ distinct real-world manipulation tasks in unstructured environments with >90% success rate, and generalize to novel tasks with minimal retraining.

Current Distance: Medium. Tesla is moving to mass production. $1.2B+ flowed into robotics in a single week. The hardware is converging but the software (VLA models, task generalization) is the hard part. Chinese manufacturers have scale advantage.

Recent Progress:

---

EDGE-08: AI-Designed Biology (Molecules to Clinical Trials)

Maturity: scaling | Impact if cracked: paradigm-shift | Source nodes: ai-for-science

The Problem: AI can now design proteins and molecules, but going from AI-designed to clinically validated is the gap. The first AI-designed drug to clear clinical trials would prove the paradigm and unlock a multi-trillion dollar transformation of pharma.

Who's Pushing:

What Breakthrough Looks Like: First fully AI-designed drug/antibody clearing Phase II clinical trials with demonstrated efficacy.

Current Distance: Close-medium. AI-designed antibodies are approaching clinical trials. Insilico has candidates in Phase I/II. The next 12-18 months could see the first clear validation.

Recent Progress:

---

EDGE-09: Decentralized / Distributed Training

Maturity: research | Impact if cracked: transformative | Source nodes: ai-research-breakthroughs ai-infrastructure

The Problem: Training frontier models requires massive centralized GPU clusters ($100M+), concentrating AI power in a few wealthy organizations. Can frontier-quality models be trained across distributed, heterogeneous hardware?

Who's Pushing:

What Breakthrough Looks Like: A frontier-class model (competing with GPT-5/Claude Opus level) trained entirely on distributed, non-colocated hardware at a fraction of the centralized cost.

Current Distance: Far-medium. The first decentralized run rivaling LLaMA-2-70B is a milestone, but that's 2+ generations behind frontier. Communication overhead remains the fundamental bottleneck.

Recent Progress:

---

EDGE-10: Formal Verification of AI-Generated Code

Maturity: research | Impact if cracked: transformative | Source nodes: ai-coding-tools ai-safety-alignment

The Problem: AI coding tools produce code 10x faster but with 1.7x more bugs and 2.74x more security vulnerabilities. Formal verification can mathematically prove code correctness, but it's historically been too hard and slow. Can AI make formal verification accessible?

Who's Pushing:

What Breakthrough Looks Like: An automated system that takes AI-generated code and either proves it correct against a spec or identifies the exact failure case, fast enough to run in a CI/CD pipeline.

Current Distance: Medium-far. Leanstral proves it's possible for narrow domains. Generalizing formal verification to arbitrary codebases (not just mathematical proofs) is much harder. But the vibe-coding quality crisis creates urgent demand.

Recent Progress:

---

EDGE-11: Test-Time Compute Scaling

Maturity: scaling | Impact if cracked: transformative | Source nodes: ai-research-breakthroughs frontier-models

The Problem: Can models get proportionally smarter by "thinking longer" at inference time? Extended thinking / chain-of-thought shows promise but the scaling curve is unclear — does more compute always help, or do returns diminish quickly?

Who's Pushing:

What Breakthrough Looks Like: Clear evidence that test-time compute scales predictably (like training compute scaling laws) — meaning you can reliably trade compute for quality with a known exchange rate.

Current Distance: Close. All major labs have shipping products with thinking modes. GPT-5.4 Thinking scoring 83% on GDPVal is strong. The question is whether the scaling curve continues or plateaus, and whether it works beyond reasoning tasks.

Recent Progress:

---

EDGE-12: Agent Security / Provably Safe Execution

Maturity: prototype | Impact if cracked: transformative | Source nodes: ai-agents ai-safety-alignment

The Problem: AI agents that use tools, browse the web, and write code are powerful but dangerous. OpenClaw has 7+ CVEs, 20% of skills are malicious, and 17,500 instances are exposed. How do you let agents act autonomously while guaranteeing they can't be exploited or cause harm?

Who's Pushing:

What Breakthrough Looks Like: A hardware-agnostic agent execution environment that provides provable containment — agents can use tools freely within a sandbox where the damage radius is mathematically bounded, regardless of what the agent tries to do.

Current Distance: Medium. NemoClaw shows the hardware approach works but is NVIDIA-locked. Software-only approaches don't yet provide strong enough guarantees. The OpenClaw crisis is creating urgent demand.

Recent Progress:

EDGE-13: Parametric Memory / Continual Learning

Maturity: research | Impact if cracked: paradigm-shift | Source nodes: ai-research-breakthroughs ai-agents ai-safety-alignment

The Problem: Deployed LLMs are stuck in "perpetual present" — frozen weights after pretraining, no ability to compress new experience into parameters. Knowledge acquired post-deployment lives in context windows (bounded, transient) or external retrieval (slow, shallow). Lossy compression into weights is what made pretraining powerful (it forces generalization); the absence of that mechanism post-deployment is why models cannot learn from their own deployment experience. The problem splits into three layers: (1) Context — better retrieval and orchestration harnesses; (2) Modules — attachable knowledge (adapter layers, compressed KV caches, external memory stores) that specialize without touching core weights; (3) Weights — true parametric updates via test-time training, meta-learning, self-distillation. Core blockers: catastrophic forgetting (new data destroys old representations), the stability-plasticity dilemma (stable models can't adapt, adaptive models lose prior capability), temporal disentanglement (invariant rules and mutable state share the same weights), logical integration failure (fact updates don't propagate to consequences), and unlearning (no differentiable way to surgically remove false or toxic knowledge). Per a16z's April 2026 framing, this is the defining research frontier for agent-era AI.

Who's Pushing:

What Breakthrough Looks Like: A deployed model that reliably compresses post-training experience into weights, demonstrates measurable capability gains from deployment data, preserves prior capabilities (no catastrophic forgetting), resists data-poisoning, and remains auditable for alignment. Or at minimum: a module-based system where swappable adapter layers deliver significantly-larger-model performance on narrow tasks (the article cites an 8B model with a module matching 109B performance as an early signal).

Current Distance: Far on weights, medium on modules, close on context. Harness/context approaches (RAG, multi-agent swarms, long context) are mature. Module-based approaches (adapters, compressed KV caches) are productizing. True weight-level continual learning remains brittle — EWC and weight interpolation don't scale cleanly, TTT works on narrow tasks, meta-learning is early. The safety subproblems (unlearning, alignment degradation from fine-tuning, auditability) are described as "open problems, not fundamental impossibilities."

Recent Progress:

What to Watch For:

• First frontier lab deploying a module-based specialization layer in a production model (vs. fine-tuning or RAG)
• First credible demonstration of non-brittle continual learning at >10B scale
• Startup in Letta/mem0 category raising at >$1B valuation — signals enterprise pull
• Safety paper quantifying alignment degradation from realistic continual learning setups
• Auditability/versioning standard for continuously-updating models (currently impossible)

---

Edge Summary Dashboard

---

Edge Lifecycle

Edges move through stages and eventually exit:

• Graduated (breakthrough) — The problem was cracked. Moves to the relevant node as a breakthrough signal.
• Stalled — No meaningful progress in 60+ days. Kept but flagged.
• Absorbed — The problem was solved incrementally (no single breakthrough moment) and became standard practice.
• Abandoned — Major players pulled out. Archived with notes.

Graduated Edges

(None yet — tracking begins 2026-03-19)

Stalled Edges

(None yet)

---

Related Nodes

• ai-research-breakthroughs
• frontier-models
• compute-hardware
• ai-agents
• ai-safety-alignment
• ai-for-science
• robotics-embodied-ai