Daily Digest — 2026-08-03
12 items · 1 research labs, 11 industry media
🏛️ Research Labs (1)
We’re strengthening our presence in Alabama through new investments and community support.
Google announced a $1.5 billion infrastructure investment (2026–2027) to expand its Jackson County, Alabama data center campus, originally operationalized in 2019 on a repurposed coal-plant site. The expansion includes full self-funding of power and infrastructure costs, alongside a $2 million Energy Impact Fund partnership with TVA and CAANEAL for local energy efficiency programs. Community initiatives comprise $550,000 in STEM kit donations for grades 4–8, water stewardship in the Paint Rock River Watershed, digital skills training for 130,000+ residents, and hundreds of job creations.
data centerenergy efficiencystem educationinfrastructure investmentdigital skills training
📜 arXiv Papers
No new items today.
📰 Industry Media (11)
NVIDIA AI Releases Molt: A PyTorch-Native Agentic Reinforcement Learning Framework
NVIDIA AI introduces Molt, a PyTorch-native agentic reinforcement learning (RL) framework designed for compactness and researcher accessibility, comprising approximately 8.6K lines of RL code. Molt integrates Ray for placement, vLLM for rollout, and NVIDIA AutoModel with FSDP2 for training, avoiding forks to facilitate upstream updates. The framework supports multi-turn tool-use agents, code-execution agents, and LLM-as-judge reward loops, operating on 2 nodes of 8 H100 GPUs. Molt enforces token identity, policy-version semantics, and forward consistency to ensure correctness, particularly for mixture-of-experts policies. It achieves throughput comparable to Megatron-based stacks, scaling from dense 4B models to 700B MoE models.
agentic reinforcement learningpytorch-nativemixture-of-expertsfsdp2vllm
End-to-End Forecasting with TimesFM 2.5: Backtesting, Covariates, Anomaly Detection, and Scalable Colab Deployment
The tutorial introduces an end-to-end workflow for time-series forecasting using TimesFM 2.5, focusing on zero-shot point and probabilistic forecasting. The method includes configuring the runtime, generating a multi-store retail dataset, and evaluating forecast quality using metrics such as MAE, RMSE, sMAPE, MASE, and pinball loss. Results demonstrate TimesFM's ability to reduce MASE by 21.3% compared to a seasonal-naive baseline, with rolling-origin backtesting and context-length ablation studies highlighting its robustness and scalability. The workflow also integrates covariates through XReg and supports anomaly detection and long-horizon forecasting.
timesfm 2.5zero-shot forecastingrolling-origin backtestingcovariate integrationanomaly detection
AMD Releases Instella-MoE-16B-A3B: A Fully Open Mixture-of-Experts LLM With 2.8B Active Parameters Trained On Instinct GPUs
AMD introduces Instella-MoE-16B-A3B, a fully open Mixture-of-Experts language model with 16B total parameters and 2.8B active parameters per token, trained on Instinct MI300X and MI325X GPUs. The model employs Gated Multi-head Latent Attention and FarSkip-Collective connectivity, achieving a 12.7% pre-training speedup and 39.2% reduction in time to first token. Pre-trained on 7.1T tokens from open corpora, it extends context length to 64K using YaRN and RoPE theta. Post-training includes SFT, DPO, and RL, resulting in a base checkpoint averaging 76.7, leading fully open models. Weights are released under a ResearchRAIL license, with training code MIT-licensed.
mixture-of-expertsgated multi-head latent attentionfarskip-collectiverope thetaresearchrail
Accelerating Transformer Training with NVIDIA Transformer Engine, Fused Kernels, BF16, FP8, and GPU Benchmarking
NVIDIA Transformer Engine accelerates transformer training by integrating fused GPU kernels, BF16 computation, and hardware-aware FP8 execution. The method involves installing Transformer Engine, configuring FP8 delayed scaling, and constructing a compact GPT-style causal language model. Benchmarking reveals FP8 execution achieves a 1.5x speedup over BF16, with reduced peak GPU memory usage, while maintaining stable FP8 arithmetic through delayed scaling and amax history management.
transformer enginefused kernelsfp8 executiondelayed scalingamax history
Supabase Releases Evals: an Open Source Benchmark That Scores Claude Code, Codex and OpenCode on Real Supabase Tasks
Supabase introduces Supabase Evals, an open-source benchmark framework for evaluating AI coding agents on real-world Supabase tasks. The framework tests agents like Claude Code, Codex, and OpenCode across tasks such as schema building, Edge Function debugging, and RLS policy fixes. It employs a three-dimensional evaluation matrix (products, topics, stages) and scores agents using deterministic checks and LLM-as-a-judge. Results are published on a public leaderboard and used internally for regression testing. The framework requires Docker, API keys, and specific ports for local deployment.
supabase evalscoding agentsdeterministic checksllm-as-a-judgedocker sandbox
MiniMax Releases MiniMax H3: An Omni-Modal Video Model That Generates 15-Second 2K Clips With Native Stereo Audio
MiniMax H3 introduces an omni-modal video generation model unifying text, image, video, and audio inputs into a single framework, producing 2K-resolution clips (4–15s) with native stereo audio. Key innovations include Contextual Omni Representation for cross-modal relationships, H3-VAE with 4× sequence-length compression enabling efficient 2K output, and in-context regeneration replacing super-resolution for detail preservation. The model achieves 30% higher training throughput via the H3-Omni Transformer, separating understanding and generation workloads. Third-party benchmarks rank H3 first in video editing but behind competitors in text-to-video (Gemini Omni Flash) and image-to-video (Seedance 2.0). Current deployment is API-only, with open weights pending.
omni-modalcontextual omni representationh3-vaein-context regenerationsequence-length compression
DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains
DeepSeek released DeepSeek-V4-Flash-0731, a 284B-parameter Mixture-of-Experts (MoE) model with 13B activated per token, achieving significant agentic and coding performance gains through post-training. The model retains its hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), along with Manifold-Constrained Hyper-Connections (mHC). Benchmarks show improvements over previous versions, notably in agentic tasks like NL2Repo (54.2) and Cybergym (76.7). The model supports speculative decoding via DSpark, reducing inference FLOPs by 27% compared to DeepSeek-V3.2. It is deployable via API or self-hosting, requiring ~110 GB memory at 3-bit quantization.
mixture-of-expertsspeculative decodingcompressed sparse attentionmanifold-constrained hyper-connectionspost-training
LingBot-Map Tutorial: GPU-Aware Inference and Point Cloud Export
The tutorial presents LingBot-Map, an end-to-end streaming 3D reconstruction pipeline that dynamically adjusts GPU-aware parameters (KV-cache, frame limits, camera iterations) based on detected VRAM. The method implements GCTStream with streaming attention, long-range trajectory memory, and mixed-precision inference to decode camera poses, depth maps, and intrinsic parameters into world-coordinate point clouds. Results include reconstructed scenes exported as PLY/NPZ/GLB artifacts, achieving real-time processing (10 FPS) with automatic VRAM optimization on CUDA GPUs.
streaming 3d reconstructiongpu-aware inferencekv-cache optimizationmixed-precision inferencepoint-cloud export
JetBrains Open-Sources KotlinLLM: Smart Macros That Generate Kotlin Source Code at Runtime and Hot-Reload It Through JDI
JetBrains Research introduces KotlinLLM, an IntelliJ IDEA plugin enabling runtime generation and hot-reloading of Kotlin code via Smart macros (asLlm, mockLlm). The system uses JVM Debug Interface (JDI) for class redefinition, capturing runtime values to iteratively refine LLM-generated code. Evaluated on a Spring Petclinic adaptation, it achieved 100% hot-reload success (24/24 scenarios) with ~1% overhead, and 0.89 recall on GitHub issue classification. Requires IntelliJ 2025.2.x, JDK 21, and OpenAI API key. Released as Apache 2.0 research prototype for Kotlin/JVM.
smart macrosjvm debug interfacehot-reloadingkotlin/jvmclass redefinition
Nous Research Ships Three Integration Paths for Hermes Agent and Buzz, Block’s Open Source Nostr Workspace for Humans and Agents
Nous Research introduces three integration methods for Hermes Agent with Buzz, Block's open-source Nostr-based workspace supporting human-AI collaboration. The approaches include a managed runtime in Buzz Desktop, a relay bridge for hosted agents, and a native gateway enabling full platform integration while preserving Hermes' memory and session management. Technical implementation leverages NIP-42-authenticated Nostr WebSockets with BIP-340 signing, configurable transport modes (auto/websocket/poll), and privacy-focused defaults including mention gating and allow-list enforcement.
nostr workspacehermes agentnip-42 authenticationbip-340 signingmention gating
OpenAI aligns safety practices with EU AI Act’s GPAI Code
OpenAI demonstrates alignment with the EU AI Act's General-Purpose AI (GPAI) Code through documented safety frameworks and transparency measures. The company employs a dual-framework approach: the Preparedness Framework (updated 2025) for risk management and the Frontier Governance Framework for legal compliance, including pre-release testing, system cards, and external red-teaming via its Red Teaming Network. For content provenance, OpenAI combines C2PA-standard Content Credentials with SynthID watermarking, expanding from images to audio. Cybersecurity initiatives include the Trusted Access for Cyber programme, now extended to Europe through the EU Cyber Action Plan (May 2026), though independent efficacy data is absent.
gpai codepreparedness frameworksynthid watermarkingred teaming networkc2pa standard
Generated automatically at 2026-08-02 20:10 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.