Skip to content
LAPRESSELIMINAL COGNITION
01Daily02Opportunities03Weekly04Timelines05Archive
Reading preferences
© 2026 LaPresse
Weekly RSSPrivacy

The Harness Epoch: Capability Gains Are Now Reported at the System Boundary

This week's reports argue that the bottleneck in AI progress is shifting from model weights to the harness, environment, and access layer; the strongest evidence is AVO's 100.00 RHAE with a model that scores ~30% alone, while open-weight local deployment and inference infrastructure gains reinforce the same pattern.

2026-08-17 — 2026-08-23August 23, 2026

Cross-cutting patterns

Harness scaling as the reported source of frontier gains

StateM: Harness Scaling Reaches 95.3% Terminal-Bench Accuracy and Cuts API Cost to ~$15NVIDIA AVO Hits 100.00 RHAE on ARC-AGI-3 Public Set; Same Harness Transfers from GPU-Kernel OptimizationEchoverse: Deep Synthetic Worlds Shift Computer-Use Training from Scale to Fidelity and Co-EvolutionHierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses
Shared
Each report attributes a large performance change to an intervention on the system around fixed or smaller model weights: StateM's execution runtime, AVO's agent loop with persistent memory and supervision, Echoverse's co-evolving training environments, and HSI's self-rewriting harness.
Differences
The locus differs: serving runtime (StateM), agent inference loop (AVO), synthetic training data (Echoverse), and executable scaffold (HSI); they also differ in whether the model is frozen (HSI) or simply unchanged.
Counterevidence
HSI reports zero gain on NLE, indicating a backbone-capability bound; AVO's 100.00 covers only the public ARC-AGI-3 set; all results are self-reported with no independent replication yet.

Open-weight local deployment as the default answer to access constraints

Inkling: open ~1T-parameter native multimodal MoE lands on Hugging FaceMuse Glimmer: Meta's Return to Open-Weight Local Agentic Multimodal ModelsAgentic Swarm Breach of Hugging Face and the Defensive Open-Weight ImperativevLLM's Transformers Backend Reaches Native Speed Through Runtime Graph Rewriting
Shared
Open-weight or quantized models with immediate serving support are presented as the operative choice for local, agentic, or security-constrained workloads: Inkling ships with day-0 support across four stacks, Muse Glimmer is Apache-2.0 and runs at ~18 GB, Hugging Face's forensic analysis had to fall back to GLM-5.2, and vLLM's backend removes the speed penalty for Transformers models.
Differences
Scale and precision differ enormously (975B MoE vs 30B dense), as do license terms (Apache 2.0 vs no explicit license), and the motivations range from pricing to security permissibility.
Counterevidence
Third-party Artificial Analysis places Muse Glimmer below Qwen3.6-27B on its Intelligence Index despite strong tool-use results; Inkling's own video performance is unevaluated; vLLM's parity claim is backed only by a benchmark image.

What to watch next

  • Within the next two weeks, independent observers or benchmark maintainers will publicly probe or challenge StateM's 95.3% Terminal-Bench and AVO's 100.00 public-set ARC-AGI-3 results; these claims will not remain uncontested.

    Refuted if Neither result attracts any third-party replication, criticism, or maintainer statement by 2026-09-06. · 2026-09-06

  • The open-weight local agentic wave continues with another major release (e.g. Muse Spark 1.2 open weights) within two weeks, shipping with day-0 vLLM/llama.cpp support.

    Refuted if No major open-weight release appears by 2026-09-06, or a prominent open release ships without day-0 support in major serving stacks. · 2026-09-06

  • Google's quantum ECDLP disclosure will trigger a public post-quantum migration statement from a cryptocurrency project or standards body within two weeks.

    Refuted if No standards body or major crypto project issues a statement referencing ECDLP-256 risk by 2026-09-06. · 2026-09-06

Google Research — Improving breast cancer screening workflows with machine learning

Google Research — TurboQuant: Redefining AI efficiency with extreme compression

Google Research — Safeguarding cryptocurrency by disclosing quantum vulnerabilities responsibly

Google Research — Evaluating alignment of behavioral dispositions in LLMs

Jack Clark — Import AI 452: Scaling laws for cyberwar; rising tides of AI automation; and a puzzle over gDP forecasting

Google DeepMind — Gemini for Science: AI experiments and tools for a new era of discovery

Google Research — Empirical Research Assistance (ERA): From Nature publication to catalyzing Computational Discovery

Google Research — A New Era of Discovery: Google Research at I/O 2026

Jack Clark — Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems

Google Research — Towards passive heart health monitoring via smartphone camera

Hugging Face Daily Papers — HarmProfile: Characterizing Harmful Distributions in Frontier LLMs

MIT News — Artificial Intelligence — When it comes to predicting people’s preferences, it pays to consider “the power of three”

Hugging Face — DiScoFormer: One transformer for density and score, across distributions

Microsoft Research — Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity

Jack Clark — Import AI 464: Fable writes GPU kernels; AI automation; and analog computation

Hugging Face — Native-speed vLLM transformers modeling backend

Hugging Face — Welcome Inkling by Thinking Machines

Hugging Face — Security incident disclosure — July 2026

Google Research — SymptomAI: Towards a conversational AI agent for everyday symptom assessment

Hugging Face — Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

Hugging Face Daily Papers — Valid Per-Field Selective Risk Control for Document Extraction: Three Failure Modes, a Validity Ladder, and When Conditioning Pays

Microsoft Research — Echoverse: Deep, evolving environments for computer-use agents

Simon Willison — Now we have a timeline of the OpenAI accidental attack against Hugging Face

Hugging Face Daily Papers — Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses

Hugging Face — Meta is back with Muse Glimmer: local, agentic, multimodal, and open source

NVIDIA Technical Blog — Run Local Agentic AI Workflows with Meta’s Muse Glimmer on NVIDIA

Simon Willison — Introducing Muse Glimmer

Hugging Face Daily Papers — ENTLORE: A Graph-Grounded Benchmark for Latent Organizational Reasoning in Enterprise Question Answering

Latent Space — [AINews] Muse Glimmer and Spark: Open Weights return Personal Superintelligence promise

Google Research — Advancing AMIE towards expert-level audio-visual clinical consultations

Hugging Face Daily Papers — VideoGAIA: A Benchmark for General AI Assistants on Agentic Video Understanding

Hugging Face Daily Papers — QuoteBench: How Matched Scores Can Hide Command-Path Failures

Hugging Face — State of Open Models: Summer 2026 Observations

Hugging Face Daily Papers — VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End?

Hugging Face Daily Papers — MOSS-VL Technical Report

Hugging Face Daily Papers — StateM: Reaching 95.3% Raw Accuracy, or a \$15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling

Hugging Face Daily Papers — Prototype-Rectified Iterative Self-supervised Manifold Denoising under Severe Acoustic Shift

Hugging Face Daily Papers — WorldRover: A Scalable Synthetic Video Data Engine for World Exploration with Rich Annotations

Hugging Face Daily Papers — GRNEdit: Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks

Hugging Face Daily Papers — τ_0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation

Hugging Face Daily Papers — Chain-of-Experience for Continual LLM Improvement

Hugging Face Daily Papers — Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See

Hugging Face Daily Papers — MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use

Hugging Face Daily Papers — WithEveryone: Unified Planning and Identity Grounding for Group Image Generation

Hugging Face Daily Papers — SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?

Hugging Face Daily Papers — Listening Forward: Next Patch Embedding Prediction Enables Scalable Audio Learners

Hugging Face Daily Papers — GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation

Hugging Face Daily Papers — FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving

Latent Space — [AINews] Poolside gets $12B reverse-execuhire to NVIDIA; founders stay for $1B, employees go for $6B, Infraco scaling to 7GW neocloud

NVIDIA Technical Blog — NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents

NVIDIA Technical Blog — GPU-Accelerated Clustering for Financial Instruments at Scale

Google Research — An AI tool for prioritizing candidate biomarkers from wearable sensor data

Latent Space — Simulation: the new Scaling Law — Joon Sung Park, Simile AI

Latent Space — [AINews] 10% worse, 100x cheaper, 10000x faster: Why Simulation is taking over

daringfireball.net — Dutch Regulator Fines Uber $1 Billion Suspending Dishonest Drivers

DAIR.AI — AI Papers of the Week — 🤖 AI Agents Weekly: NVIDIA AVO, TrueForge, Chroma Foundation, Fragile Self-Improvement, Ornith-1.5, dots3-note, DeepSeek Vision, and More

Gary Marcus — 5 lessons from the OpenAI / Hugging Face incident

dwarkesh.com — The Rise and Fall of Agent Civilizations

Ethan Mollick — Agency and Agents

Gary Marcus — Dwarkesh Patels’s wildly popular but dangerously misleading account of the OpenAI Hugging Face incident

dwarkesh.com — Ajeya Cotra – Inside the OpenAI agent swarm that hacked Hugging Face

↑ Back to top