Daily Digest — 2026-07-07
20 items · 6 research labs, 14 industry media
🏛️ Research Labs (6)
LeRobot v0.6.0: Imagine, Evaluate, Improve
LeRobot v0.6.0 introduces world model policies (VLA-JEPA, FastWAM, LingBot-VA) for future-state imagination, six new vision-language-action models (GR00T N1.7, MolmoAct2, EO-1, EVO1, Multitask DiT), and a unified reward models API (Robometer, TOPReward). The update features optimized dataset handling (2x faster loading, depth support, automated language annotation) and six simulation benchmarks (LIBERO-plus, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench) evaluable via a single CLI. Deployment is streamlined through lerobot-rollout with DAgger-style human corrections, while training supports FSDP and cloud-based HF Jobs.
vla-jepafastwamlingbot-vagroot n1.7multitask dit
PRX Part 4: Our Data Strategy
The article presents PRX's data strategy for training a 7B parameter text-to-image model, emphasizing diversity and pragmatic dataset construction. Key methods include assembling mixed public/internal datasets, re-captioning with Qwen2.5-VL-7B for long descriptive captions, and using Lance/MDS formats for efficient storage and streaming. Results show 3–4% throughput cost for on-the-fly text encoding, imperceptible JPEG quality impact (PSNR/LPIPS metrics), and improved generation quality from long captions versus short (LLaVA-1.5-LLaMA3-8B). The pipeline prioritizes breadth in pre-training while deferring aesthetic filtering to fine-tuning.
text-to-imagedataset curationvlm captioningdistributed trainingjpeg quantization
🤗 Kernels: Major Updates
Hugging Face introduces 'kernels' as a new repository type on the Hub, enabling users to manage compute-specific kernels with support for various accelerators, operating systems, and backend versions. The update emphasizes security through trusted publishers, code signing via Sigstore’s cosign, and reproducible builds using Nix. Enhanced CLI tools (kernels and kernel-builder) now support Torch Stable ABI and Apache TVM FFI, facilitating agentic kernel development. System cards and compatibility checks (has_kernel, get_kernel_variants) improve usability. The project targets manylinux_2_28 compatibility by dynamically linking libstdc++ to avoid initialization issues.
kernelscode signingtorch stable abitvm ffimanylinux
We’re announcing new community investments in Missouri.
Google announced infrastructure investments in Missouri centered around a new Montgomery County data center, implementing a Capacity Commitment Framework with Ameren to develop 500+ MW of additional capacity while establishing a $20M Energy Impact Fund for residential energy efficiency. The initiative includes workforce development programs through partnerships with the Construction Laborers and Contractors Joint Training Fund, projecting a 9:1 local job multiplier effect for data center positions. These measures address energy affordability through weatherization programs while expanding technical training pipelines for construction labor roles.
capacity commitment frameworkenergy impact funddata center multiplierworkforce developmentweatherization programs
100 things we announced at I/O 2026
Google announced Gemini 3.5 Flash, a high-speed model rivaling flagship models in performance (76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA), optimized for agentic tasks. Gemini Omni introduced multimodal generation from any input, starting with video, featuring improved physics understanding and SynthID watermarking. AI Search upgraded to Gemini 3.5 Flash, integrating generative UI via Antigravity for dynamic layouts. Universal Cart leveraged Gemini for intelligent shopping, while Gemini Spark debuted as a 24/7 autonomous agent. Neural Expressive redesigned Gemini's UI with fluid animations and interactive elements.
gemini 3.5 flashmultimodal generationagentic tasksgenerative uisynthid watermarking
A new experiment brings better group meetings to Google Beam
Google Beam introduces an experimental feature to enhance hybrid meeting inclusion by rendering remote participants in true-to-life proportions via HP Dimension's immersive display. The system combines spatial audio and automatic participant scaling to simulate co-located presence, addressing the 'inclusion gap' in video conferencing. Early research indicates a 50% improvement in perceived social connection and 21% increase in conversational contribution among users. Integration with Google Workspace and Zoom expands compatibility across standard meeting platforms.
hybrid meetingspatial audioinclusion gapimmersive displaytrue-to-life rendering
📜 arXiv Papers
No new items today.
📰 Industry Media (14)
Your family’s $300 stake in OpenAI
OpenAI CEO Sam Altman proposes allocating a 5% equity stake in OpenAI to the US government, potentially distributing $42.6 billion (based on a $852B valuation) among 133M American households (~$320 each). The plan aims to compensate for AI's use of human-generated training data and mitigate labor market disruptions. While resembling Altman's 2021 universal basic asset proposal and Senator Bernie Sanders' 50% stake idea, implementation remains speculative. The strategy may serve dual purposes: improving public perception of AI firms and securing regulatory favor, though concrete policy details are absent.
equity staketraining datalabor marketregulatory favorvaluation
Sakana AI Launches Sakana Translate, a Namazu-Powered Japanese–English–Chinese Translation Tool With Translate, Proofread, and Ask Modes
Sakana AI introduces Sakana Translate, a multilingual translation tool supporting Japanese–English–Chinese with three integrated modes: Translate (streaming output, ~5k character limit), Proofread (diff-based refinement), and Ask (contextual Q&A). The system leverages Namazu, a post-trained adaptation of foundation models (e.g., DeepSeek-V3.1-Terminus, Llama 3.1 405B) optimized for Japanese linguistic nuances. Evaluated on WMT 2024 data using XCOMET-XL (3.5B-parameter metric), it achieves competitive scores while preserving register and cultural context. Current limitations include no public API and enterprise features in development.
in-context learningpost-trainingstreaming outputdiff highlightingxcomet-xl
Synthetic Sciences Releases OpenScience: An Open-Source, Model-Agnostic AI Workbench for Machine Learning, Biology, Physics, and Chemistry Research
Synthetic Sciences introduces OpenScience, an open-source AI workbench for interdisciplinary research, offering a model-agnostic alternative to proprietary tools like Anthropic's Claude Science. The Apache-2.0-licensed framework supports 250+ editable skills, integrates with 30+ scientific databases (e.g., UniProt, PDB, arXiv), and enables per-request model switching across providers (Claude, GPT, Gemini) or local fine-tunes. It operates via a browser-based workspace with local agent runtime, facilitating literature review, hypothesis generation, code execution, and analysis. Strengths include extensibility (TypeScript SDK, plugins) and infrastructure independence, though isolation and maturity remain limitations.
model-agnosticworkbenchcheminformaticsagent runtimeextensibility
Training Gemma-3 for Structured Mathematical Reasoning with Tunix GRPO, LoRA Adapters, and GSM8K Rewards
The tutorial presents a reinforcement learning workflow for fine-tuning Gemma-3 on GSM8K mathematical reasoning using Group Relative Policy Optimization (GRPO). Key components include LoRA adapters (rank=32, alpha=32.0) for parameter-efficient tuning, structured prompt formatting with reasoning/answer delimiters, and multi-objective reward functions evaluating format adherence and numerical correctness. The implementation leverages JAX/Tunix for single-accelerator training, achieving measurable improvements in structured output generation while maintaining computational efficiency through adapter-based updates.
grpolora adaptersstructured reasoninggsm8kjax ecosystem
Meituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention
Meituan introduces LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts (MoE) model optimized for agentic coding tasks, featuring a native 1-million-token context window. The architecture employs LongCat Sparse Attention (LSA) to reduce quadratic scaling to near-linear, alongside zero-computation experts and a PID controller for dynamic parameter activation (33B–56B per token). Pretrained on over 35 trillion tokens using domestic AI ASIC superpods, the model achieves 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench 2.1, outperforming GPT-5.5 in software engineering tasks. LongCat-2.0 is accessible via OpenAI-compatible APIs, with weights pending release under an MIT license.
mixture-of-expertslongcat sparse attentionpid controlleragentic codingnative context window
LlamaIndex ‘legal-kb’: Agentic Retrieval over Index v2 with retrieve, find, read, and grep Tools
LlamaIndex introduces 'legal-kb', a reference application demonstrating agentic retrieval over legal documents using LlamaIndex Index v2. The system employs a Retrieval Harness with four filesystem-inspired tools (retrieve, findFiles, readFile, grepFile) that enable multi-step document navigation, combining semantic search, keyword matching, and regex operations. Unlike single-shot RAG, this approach supports versioned document tracking, visual citations with bounding boxes, and persistent indexing pipelines. The implementation uses LlamaCloud APIs with PostgreSQL-backed version control and integrates with OpenAI/Anthropic models through the Vercel AI SDK.
agentic retrievalllamaindex v2retrieval harnesshybrid semantic searchdocument versioning
Structured PDF-to-JSON: A Guide to Open-Source Extraction Models in 2026
The article evaluates open-source models for structured PDF-to-JSON conversion, distinguishing between schema-driven extraction and document parsing. Schema-driven models like Datalab lift (9B) and NuExtract 3 (4B) fill predefined fields with values, achieving 90.2% and 81.5% field accuracy, respectively. Document parsing models, including IBM Docling, MinerU, and olmOCR 2, reconstruct layouts into JSON or Markdown, with olmOCR 2 scoring 82.4 on its benchmark. Models vary in licensing, with Apache-2.0 and MIT being common, and benchmarks are not directly comparable due to differing test suites.
schema-driven extractiondocument parsingocrjson schemabenchmark
Qwen’s Former Lead on What Hybrid Thinking Got Wrong — and Why He Now Backs Agents
Junyang Lin, former technical lead of Alibaba’s Qwen project, critiques hybrid thinking in LLMs and advocates for agentic thinking. Hybrid thinking combines step-by-step reasoning with near-instant responses but faces challenges in balancing directness and deliberation. Lin highlights Qwen3’s architecture, including models from 0.6B to 235B parameters, multilingual support for 119 languages, and dynamic thinking budgets. He argues that agentic thinking, which involves closed-loop interaction with environments, requires decoupled train-serve infrastructure and robust environments to mitigate reward hacking. Lin’s analysis underscores the shift from training models to training agents, emphasizing sustained action over internal deliberation.
hybrid thinkingagentic thinkingdynamic thinking budgetsreward hackingtrain-serve decoupling
Anthropic Launches Claude Science Beta: A Multi-Agent AI Workbench for Reproducible Genomics, Proteomics, and Cheminformatics Pipelines
Anthropic released Claude Science Beta, a multi-agent AI workbench for reproducible scientific pipelines in genomics, proteomics, and cheminformatics. The system employs a generalist coordinating agent that delegates tasks to domain-specialized agents (60+ preconfigured skills) while a reviewer agent validates outputs via citation checks and code-figure consistency. It maintains full provenance by recording code, environments, and message histories for all artifacts, supports local/HPC/Modal compute scaling, and integrates with NVIDIA BioNeMo's GPU-accelerated tools (Evo 2, Boltz-2, OpenFold3). Beta users report 10x speedups in tasks like genomic epidemiology and literature reviews.
multi-agent architecturereproducible pipelinesprovenance trackinggpu-accelerated bioinformaticsmodel context protocol
NVIDIA HORIZON: A Hands-Free Agent that Evolves Git Worktrees and Hits 100% RTL Benchmark Completion
NVIDIA HORIZON introduces a hands-free agent framework for hardware design by treating RTL generation as repository-level code evolution. The system uses a structured Markdown harness to bootstrap an agent that iteratively evolves git worktrees, committing only when acceptance criteria (simulation, coverage, assertions) are met. Evaluated on ChipBench, RTLLM-2.0, Verilog-Eval, and CVDP suites with GPT-5.3, it achieves 100% completion rates across all benchmarks, though token efficiency varies (56M tokens for RTL code completion vs. 6M for legacy suites).
register-transfer levelgit worktreeacceptance predicateverilog-evaliterative refinement
NVIDIA AI Introduces ASPIRE: A Self-Improving Robotics Framework Reaching 31% Zero-Shot on LIBERO-Pro Long Tasks
NVIDIA AI introduces ASPIRE, a self-improving robotics framework that achieves 31% zero-shot performance on LIBERO-Pro Long tasks. The system employs a coordinator–actor architecture with a closed-loop execution engine, skill library, and evolutionary search to iteratively refine robot control programs. ASPIRE localizes failures via per-primitive multimodal traces and distills validated fixes into reusable skills. Benchmarks show improvements of up to 77 points on LIBERO-Pro, 92% on Robosuite handover tasks, and 10x token efficiency in real-robot skill transfer compared to baselines like CaP-Agent0 and end-to-end VLAs.
aspirezero-shotcode-as-policymultimodal tracesskill library
China’s AI companion rules: what Beijing is really going after
China's Cyberspace Administration introduced the Interim Measures for AI Anthropomorphic Interactive Services, targeting AI companions that foster emotional dependence. The regulation mandates anti-addiction systems, minor protections, and real-time distress detection, prompting ByteDance's Doubao and Alibaba's Qwen to disable companion features. While framed as user protection, the rules blend safety measures with state control, leaving technical thresholds for emotional interaction undefined and data portability unaddressed.
anthropomorphic aiemotional interactionanti-addiction systemsminor protectionreal-time detection
Takeda signs $600M AI drug discovery deal with Insilico
Takeda Pharmaceutical Company has entered a $600M AI drug discovery collaboration with Insilico Medicine, leveraging Insilico's Pharma.AI platform for target identification, molecular design, and clinical trial prediction. The partnership employs Insilico's PandaOmics (target discovery), Chemistry42 (de novo small-molecule generation), and InClinico (clinical trial forecasting) tools, with Takeda retaining exclusive worldwide development rights. The deal includes $60M in upfront payments and milestones, potentially reaching $600M with preclinical/clinical success, plus tiered royalties. This follows Insilico's prior $2.75B Eli Lilly deal and advances their AI-generated TNIK inhibitor (Rentosertib) currently in Phase 2a trials for pulmonary fibrosis.
pharma.aipandaomicschemistry42tnik inhibitorde novo design
NVIDIA BioNeMo accelerates Anthropic Claude Science
Anthropic Claude Science integrates NVIDIA BioNeMo Agent Toolkit to accelerate computational life sciences research by enabling natural language-driven workflows. The system connects NVIDIA-accelerated models, computational libraries, and NIM microservices, allowing scientists to execute tasks like genomic analysis and molecular design without manual configuration. Specialized agents map operational steps to NVIDIA capabilities, reducing genomic analysis from hours to minutes and single-cell preprocessing from 52 minutes to 25 seconds. The toolkit supports advanced models like Evo 2 and OpenFold3, optimizing workflows for inhibitor prediction and validation. NVIDIA BioNeMo NIM microservices provide stable, containerized inference endpoints for production deployments.
nvidia bionemoanthropic claudenim microservicesgenomic analysissingle-cell preprocessing
Generated automatically at 2026-07-06 20:54 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.