NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-06-08ENGLISH EDITION
This issue
—
All time
—

AI Blog

2 stories
01

Launches Economic Research Exchange

OpenAI has launched the Economic Research Exchange, a collaborative initiative designed to research and analyze the impact of artificial intelligence technologies on labor markets, economic growth, and global productivity. The program is currently accepting applications for select research projects, offering a platform and potential resources to academic and professional researchers investigating these socio-economic transitions. The initiative aims to deepen collective understanding of how AI integration reshapes jobs and productivity patterns across the broader global economy. (source: https://openai.com/index/economic-research-exchange)

OpenAIEconomic Research ExchangeAI economic impactlabor market researchproductivity studyArtificial IntelligenceIndustry NewsResearch Progress
02

Unitree Positions to Lead Global Humanoid Robotics Market

Chinese robotics manufacturer Unitree is preparing for an initial public offering (IPO) as it approaches the shipment of its 10,000th humanoid robot. Over the past 12 to 18 months, Unitree reduced the pre-tax price of its flagship G1 humanoid from over $50,000 to $27,300, while maintaining a 67% gross margin. Backed by approximately $300 million in planned AI research and development spending and triple-digit year-over-year revenue growth, the company leverages a QDD actuator design for structural cost advantages. Currently, over 250 of its humanoids are deployed in real-world labor settings. (source: https://newsletter.semianalysis.com/p/chinas-unitree-will-dominate-global)

UnitreeG1 HumanoidRoboticsIPOQDD ActuatorRoboticsIndustry NewsTech News

Hacker News

5 stories
01

Siri AI

Apple officially introduced Apple Intelligence, a deeply integrated personal intelligence system built for iPhone, iPad, and Mac. The framework leverages on-device generative AI models combined with user personal context to deliver more natural and context-aware assistance through Siri. It is supported by Private Cloud Compute, processing data on-device and on secure Apple silicon servers to ensure user privacy. Apple's WWDC 2026 event was also covered in other reports highlighting iOS 27 and a strategic architecture integrated with Google Gemini models. (source: https://www.apple.com/apple-intelligence/)

Apple IntelligenceSiriOn-device AIPrivate Cloud ComputeGenerative AIArtificial IntelligenceGenerative AIAI Agent
02

MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second

The MiMo-v2.5-Pro-UltraSpeed artificial intelligence model architecture has been released, scaling up to one trillion parameters. The model addresses latency and computational bottlenecks in deploying ultra-large-scale neural networks, achieving an inference speed of 1000 tokens per second. It leverages advanced inference optimization, proprietary hardware-software co-design, and parallelization strategies to run massive generative models with real-time response rates suitable for enterprise applications and interactive agents. The release establishes a new benchmark for high-throughput inference performance on massive models. (source: https://mimo.xiaomi.com/blog/mimo-tilert-1000tps)

Trillion-Parameter ModelInference OptimizationHigh-Throughput GenerationTokens Per SecondHardware-Software Co-designArtificial IntelligenceLarge Language ModelDeep Learning
03

AI Is Slowing Down

An emerging industry consensus suggests that progress in training frontier large language models is facing a period of diminishing returns, a phenomenon known as the flattening of scaling laws. Despite multi-billion dollar investments in larger data centers, the performance gains between successive model generations are becoming less pronounced than the transition from GPT-3 to GPT-4. Contributing factors include a dwindling supply of high-quality human text data alongside high power and financial costs, shifting corporate focus toward algorithmic efficiency and sustainable engineering. (source: https://www.wheresyoured.at/ai-is-slowing-down/)

Scaling LawsGenerative AIModel TrainingLarge Language ModelsCompute EfficiencyArtificial IntelligenceLarge Language ModelGenerative AI
04

Launch HN: Intuned (YC S22) – Build and run reliable browser automations as code

Intuned, a Y Combinator S22 startup, has launched a platform to build, deploy, and maintain browser automations as code. Designed for web targets that lack APIs, the platform employs an autonomous AI agent to handle data scraping, report retrieval, and form submissions. The system captures comprehensive execution metadata, which allows the AI agent to automatically self-heal and debug the underlying code when target websites undergo structural or design updates. This approach aims to eliminate manual maintenance overhead. (source: https://intunedhq.com)

Browser AutomationWeb ScrapingSelf-Healing CodeAI AgentDeveloper ToolsArtificial IntelligenceAI AgentOthers
05

Replies to comments on my "LLMs are eroding my career" post

The author published a follow-up analysis addressing public feedback regarding how large language models are altering the software engineering profession. The post synthesizes recurring reader perspectives on human-AI collaboration, shifting expectations of developer productivity, and the devaluation of deep technical expertise. It explores collective industry anxieties about automated code generation, the threat of code homogenization, and the transition of the developer's role from active, creative problem-solving to passive system monitoring and validation. (source: https://human-in-the-loop.bearblog.dev/replies-to-comments-on-my-llms-are-eroding-my-career-post/)

Large Language ModelsDeveloper ProductivityAI AutomationSoftware Engineering CareerHuman-AI CollaborationArtificial IntelligenceLarge Language ModelGenerative AI

Twitter

5 stories
01

OpenEnv Governance Shifts To Multi-Stakeholder Committee

OpenEnv has transitioned to a decentralized governance model overseen by a multi-stakeholder committee to direct its open agentic Reinforcement Learning (RL) stack. The newly established committee is comprised of major AI organizations and industry leaders including Meta-PyTorch, Hugging Face, Nvidia, Unsloth, Reflection, Modal, Prime Intellect, Mercor, and Fleet AI. This collaborative effort aims to accelerate open-source agentic RL architecture development by combining shared engineering resources and collective strategic oversight. (source: https://x.com/Thom_Wolf/status/2064021681723547793)

OpenEnvReinforcement LearningOpen SourceAgentic AIAI CollaborationArtificial IntelligenceAI AgentOpen Source
02

Reflecting on One Year of Claude Code and Future Agentic Coding Directions

Boris Cherny and Cat Wu have released a technical retrospective celebrating the one-year anniversary of Claude Code's general availability. The update outlines the architectural motivations behind the system's auto mode, highlighting key verification best practices and robust iterative execution loops that stabilize autonomous programming workflows. Looking forward, the development roadmap focuses heavily on increasing the overall reliability and direct task autonomy of coding agents within automated software engineering pipelines. (source: https://x.com/ClaudeDevs/status/2064032814392352816)

Claude CodeAgentic CodingSoftware EngineeringAutonomous AgentsDevelopment ToolsArtificial IntelligenceAI AgentProduct Launch
03

Why AI Coding Progress Outpaces Biology Infrastructure Challenges

Anthropic has published a new science blog explaining why biological research progress lags behind software engineering in AI model adoption. The analysis demonstrates that contemporary biological databases are akin to legacy municipal infrastructure and are not structured for automated agents. Anthropic emphasizes the critical engineering need to overhaul and optimize these biological data structures for agentic frameworks to accelerate scientific discovery in life sciences. (source: https://x.com/AnthropicAI/status/2064054837294354677)

AI AgentsBiologyData InfrastructureScientific ResearchAnthropicArtificial IntelligenceAI AgentResearch Progress
04

Runway Introduces AI-Powered Universal Video Reformatting Tool

Runway has launched an AI-powered universal video reformatting tool designed to automate content adaptation across popular social media platforms. Leveraging generative AI, the utility lets creators upload a single video source and automatically reframe it to different aspect ratios suited for TikTok, Instagram Reels, and YouTube Shorts. This tool aims to streamline content creation workflows and maintain high visual fidelity without traditional, labor-intensive manual editing. (source: https://x.com/c_valenzuelab/status/2064043562816119046)

RunwayVideo EditingGenerative AIContent CreationAspect RatioGenerative AIComputer VisionProduct Launch
05

Wharton Research Indicates AI Needs Massive Productivity Gains to Sustain Growth

A research paper published by the Wharton School indicates that generative artificial intelligence technologies must deliver a 2.7-fold productivity increase across target sectors to justify current market valuations and corporate capital expenditures. The findings emphasize that these massive operational efficiency gains need to be achieved rapidly to prevent a severe mismatch between infrastructure spending and financial output. This study has sparked notable industry debate. (source: https://x.com/ylecun/status/2064041550527508785)

WhartonProductivityArtificial IntelligenceTech InvestmentEconomic ImpactArtificial IntelligenceIndustry NewsResearch Progress

huggingface

8 stories
01

Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills

Researchers have introduced Socratic-SWE, a closed-loop self-evolution framework designed to train software engineering agents using their own historical solving traces. Unlike traditional synthetic data generation, Socratic-SWE distills execution traces into structured skills detailing recurring failures and repair patterns, which are then used to generate targeted repair tasks. Validated via execution-based tests and scored with a solver-gradient alignment reward, the framework achieved 50.40% on SWE-bench Verified after three training iterations. Evaluated on SWE-bench Verified, SWE-bench Lite, SWE-bench Pro, and Terminal-Bench 2.0, Socratic-SWE consistently outperforms self-evolving baselines. (source: https://huggingface.co/papers/2606.07412)

self-evolutionsoftware engineering agentssolving tracesagent skillsSWE-benchArtificial IntelligenceLarge Language ModelAI Agent
02

SIA: Self Improving AI with Harness & Weight Updates

Researchers proposed SIA, a unified self-improving loop where a Feedback-Agent simultaneously updates both the execution scaffold (harness) and the underlying weights of a task-specific agent. While previous methods isolated harness rewrites from model fine-tuning, SIA integrates both mechanisms to enhance agentic planning and domain-specific intuition. Tested across Chinese legal charge classification (LawBench), GPU kernel optimization, and single-cell RNA denoising, the approach significantly outperformed scaffold iteration alone. The framework demonstrated performance gains of 56.6% on LawBench, a 91.9% runtime reduction on GPU kernels, and a 502% improvement on denoising. (source: https://huggingface.co/papers/2605.27276)

Self-Improving AIHarness UpdateWeight UpdateAI AgentsFeedback-AgentArtificial IntelligenceLarge Language ModelAI Agent
03

Streaming Video Generation with Streaming Force Control

A research team has developed StreamForce, a causal streaming video generation framework that enables real-time, physically grounded control using continuous force inputs. Unlike prior models that rely on separate systems for different forces or non-causal processing, StreamForce is a unified model that responds to time-varying local and global force inputs while preserving photometric and dynamic realism. Supported by a newly designed force-controllable distillation pipeline, the system runs at up to 16.6 FPS on a single GPU. It delivers stable generation and outperforms previous baselines in force compliance and motion fidelity. (source: https://huggingface.co/papers/2606.07508)

Streaming Video GenerationStreaming Force ControlPhysically Grounded ControlUnified Force RepresentationMotion RealismComputer VisionGenerative AIDeep Learning
04

Reinforcement Learning from Rich Feedback with Distributional DAgger

Researchers have developed DistIL, a reinforcement learning method that leverages multi-dimensional feedback such as execution traces, tool outputs, and self-evaluations, moving beyond simple binary rewards. Based on a distributional variant of the DAgger imitation learning algorithm, DistIL employs a forward cross-entropy loss that facilitates robust credit assignment by propagating future expert-student mismatches to past decisions. The authors prove that this objective ensures monotonic policy improvement and theoretical regret guarantees. In empirical evaluations, DistIL consistently outperformed standard reinforcement learning from verifiable rewards (RLVR) across scientific reasoning, programming, and mathematics. (source: https://huggingface.co/papers/2606.05152)

Reinforcement LearningDAggerImitation LearningRich FeedbackPolicy ImprovementMachine LearningArtificial IntelligenceLarge Language Model
05

HarnessForge: Joint Harness and Policy Evolution for Adaptive Agent Systems

To address the limitations of fixed agent architectures, researchers proposed HarnessForge, a meta-adaptive framework for the joint evolution of LLM agent systems. HarnessForge structures an agent system as a paired harness and policy, executing co-evolution through fault-guided harness tailoring and harness-conditioned policy alignment. Evaluated across five distinct benchmarks, the framework improved Qwen3-4B and Qwen3-8B base models, exceeding harness-only and policy-only methods by up to 12.0%. The experiments demonstrate that joint optimization of execution structures and internal reasoning parameters significantly enhances overall system performance and rollout efficiency. (source: https://huggingface.co/papers/2606.01779)

meta-adaptationco-evolutionLLM agentsreasoning policiesharness-policy alignmentArtificial IntelligenceLarge Language ModelAI Agent
06

OpenSkill: Open-World Self-Evolution for LLM Agents

Researchers have proposed OpenSkill, an open-world self-evolution framework designed for LLM agents deployed without pre-curated skills, target-task supervision, or human verifiers. OpenSkill enables an agent to autonomously bootstrap its learning cycle by retrieving knowledge and verification anchors from web resources and documentation. It synthesizes these materials into transferable skills and refines them using self-generated virtual practice tasks. Across three benchmarks, OpenSkill achieved the highest automated pass rates while maintaining strict zero-supervision constraints. The generated skills successfully transferred across different language model backbones without model-specific retraining. (source: https://huggingface.co/papers/2606.06741)

self-evolutionLLM agentsopen-worldverification signalsskill synthesisArtificial IntelligenceLarge Language ModelAI Agent
07

Data-Efficient Autoregressive-to-Diffusion Language Models via On-Policy Distillation

A new method called On-Policy Diffusion Language Model (OPDLM) has been introduced to transform autoregressive language models into diffusion language models. To prevent data distribution shifts and loss of pretraining knowledge, OPDLM uses an On-Policy Distillation (OPD) objective where a student model with bidirectional attention generates its own text trajectories while a frozen teacher model provides training logits. This on-policy distillation approach eliminates the typical train-inference mismatch found in traditional diffusion models. OPDLM achieves strong performance across tasks while using 15x to 7,000x fewer training tokens. (source: https://huggingface.co/papers/2606.06712)

Diffusion Language ModelsOn-Policy DistillationAutoregressive ModelsKnowledge RetentionModel TransformationNatural Language ProcessingLarge Language ModelGenerative AI
08

Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation

Researchers studied the post-hoc compression of long chain-of-thought reasoning traces before distilling knowledge to student models. Using 283k traces generated by Qwen3.5-397B-A17B and gpt-oss-120B teachers, instruction-tuned models compressed the sequences to 8.6-21.0% of their original lengths. Evaluation across 48 experimental configurations showed that compressed traces reduce training tokens to 12-30% of raw inputs, boosting training speeds by 2.0-7.6x and shortening inference outputs. However, raw traces maintained the highest accuracy, revealing a clear performance-efficiency trade-off where students retain up to 96% of accuracy while gaining 18x token efficiency. (source: https://huggingface.co/papers/2606.05988)

knowledge distillationreasoning trace compressionlarge language modelschain-of-thoughtefficiency trade-offArtificial IntelligenceNatural Language ProcessingLarge Language Model