Bartesian: The Capsule Cocktail Machine That Puts a Home Bar in a Pod — wired.com

**JadePuffer: The AI-Driven Ransomware Attack That Still Needed a Human Hand** — techcrunch.com

South Korean memory chipmaker SK Hynix is preparing a U.S. IPO that could raise roughly $28 billion, according to TechCrunch. The company plans to sell nearly 17.8 million shares in the form of American depositary receipts, with pricing expected Thursday and trading set to begin Friday. — techcrunch.com

Hoto's PixelDrive Cordless Screwdriver Drops to $60, Tying Its Lowest Price on Amazon — theverge.com

Siri Gets a Slider for Speed and Emotion in Latest iOS 27 Beta — techcrunch.com

Bartesian: The Capsule Cocktail Machine That Puts a Home Bar in a Pod — wired.com

**JadePuffer: The AI-Driven Ransomware Attack That Still Needed a Human Hand** — techcrunch.com

Hoto's PixelDrive Cordless Screwdriver Drops to $60, Tying Its Lowest Price on Amazon — theverge.com

Siri Gets a Slider for Speed and Emotion in Latest iOS 27 Beta — techcrunch.com

AI hardware setback

Nvidia's Kyber rack for Rubin Ultra delayed to 2028, creating rare opening for rivals

CNBC and Tom's Hardware confirm manufacturing challenges with Nvidia's Kyber NVL144 rack for Rubin Ultra GPUs have pushed the flagship scale-up system more than 12 months behind schedule. Hyperscaler customers rejected a stopgap NVL72x2 design, leaving Nvidia without a proven solution for Rubin Ultra in 2027 and handing AMD and Google's custom chips a potential opening at the high end of the AI market.

CNBC · 2026-07-06 · Read the dispatch →

Podcast Pill · The Divide 0:43

Anthropic reveals Claude's internal 'global workspace' — J-Space interpretability breakthrough

Anthropic researchers identified a privileged subspace (J-Space) in Claude's activations that functions like a global workspace for conscious access. Claude can perform silent reasoning steps there, and hidden goals become visible. An interactive demo is available on Neuronpedia.

AnthropicAI · 2026-07-06

Tencent open-sources Hy3: 295B MoE model under Apache 2.0 with 256K context

Hy3 is a 295B-parameter Mixture-of-Experts model activating only 21B parameters per token with a 256K context window. Released under Apache 2.0, it matches or beats flagship closed models on agentic and coding tasks. Available free on Nous Portal and OpenRouter for a limited time with day-0 SGLang support including FP8 checkpoint and speculative decoding.

Nous Research · 2026-07-06

Even Realities hits unicorn status with $150M round for AI smart glasses

The Shenzhen-based startup founded by a former Apple engineer raised $150M in a pre-Series B round led by Meituan and Tencent, reaching a $1B valuation. The company develops AI-powered smart glasses in the rapidly growing AI wearable hardware sector.

TechStartups · 2026-07-06

Research & Papers

ByteDance releases Seed2.0 model series targeting real-world complexity

Seed2.0 delivers world-leading reasoning, visual understanding, and search capabilities grounded in a user-need-based evaluation framework focused on long-tail knowledge and complex instruction following.

HuggingFace · 2026-07-06

Embodied.cpp: Portable inference runtime for VLA models on heterogeneous robots

A portable C++ inference runtime achieves 100.0% and 91.0% task success rates on HY-VLA and pi0.5 benchmarks respectively, while reducing WAM benchmark block memory from 312.2 MiB to 88.1 MiB.

HuggingFace · 2026-07-07

OrbitQuant: Data-agnostic quantization for image and video diffusion transformers

Using permuted block-Hadamard rotation and Lloyd-Max codebooks, OrbitQuant sets SOTA for post-training quantization on FLUX.1, Z-Image-Turbo, Wan 2.1, and CogVideoX, reaching W2A4 with usable generation quality.

HuggingFace · 2026-07-07

VLA-Corrector: Adaptive detection and correction for robotic manipulation policies

A corrective inference framework for action-chunked VLA policies detects visual dynamics deviations via a latent-space vision monitor and triggers corrective replanning without retraining the backbone.

HuggingFace · 2026-07-07

Measuring the gap between human and LLM research ideas at scale

A two-axis research-taste taxonomy finds LLM-generated research ideas cluster around bridge-like opportunities and synthesis methods, diverging systematically from the broader distribution of human research papers.

HuggingFace · 2026-07-07

Open Source & Models

Tencent open-sources Hy3: 295B MoE under Apache 2.0

The 295B-parameter Mixture-of-Experts model activating 21B parameters per token with 256K context is released under Apache 2.0, matching closed models on agentic and coding tasks. Available on Nous Portal and OpenRouter with day-0 SGLang support.

Nous Research · 2026-07-06

DSpark speculative debugging now in SGLang for DeepSeek-V4 and Qwen3

LMSYS announced production support for DeepSeek's DSpark speculative decoding framework in SGLang, improving throughput by 60-85% over MTP-1 baselines. Supports DeepSeek-V4-Pro, DeepSeek-V4-Flash, and Qwen3 models with open-sourced kernels.

LMSYS Org · 2026-07-06

ThinkingCap-Qwen3.6-27B cuts thinking tokens by ~50% while maintaining accuracy

BottleCap AI's fine-tune of Qwen3.6-27B reduces reasoning tokens by an average of 50% (up to 90% in best cases) while maintaining original accuracy. Released under Apache 2.0 and built via online RL on holdout portions of target datasets.

@_akhaliq · 2026-07-06

Hardware & Robotics

Nvidia reaffirms AI chip roadmap, rejects Kyber delay reports

Nvidia pushed back against the SemiAnalysis report claiming manufacturing delays, stating its AI chip roadmap remains intact and production plans for next-generation AI accelerators are on track. Shares stabilized in after-hours trading.

ROIC.ai · 2026-07-06

AMD Ryzen AI Halo Dev Kit review: AI development in a mini-PC

LTT Labs reviews AMD's mini-PC built around the Zen 5 Ryzen AI Max+ 395 with 128 GB unified LPDDR5x-8000 memory. Priced at $3,999 with ROCm support, it targets AI developers running local LLMs and rivals the Mac Studio.

LTT Labs · 2026-07-06

Tools & Startups

Glaze by Raycast: Build native Mac apps by chatting with AI

Raycast's Glaze lets users create native macOS applications through natural language, connecting to APIs, local CLIs, and MCPs. Published apps can be shared in a store.

Product Hunt · 2026-07-07

Pulpie: Pareto-optimal models for web content extraction, 20x cheaper

Feyn's family of models matches SOTA extraction quality while being 20x cheaper — cleaning 1 billion pages costs $7,900 with Pulpie vs $159,000 with alternatives. Open-source on Hugging Face.

Feyn · 2026-07-06

OfficeCLI: Open-source Office suite purpose-built for AI agents

A single-binary open-source Office suite that lets AI agents read, edit, and automate Word, Excel, and PowerPoint files without requiring an Office installation.

GitHub · 2026-07-06

Money & Markets

In Brief

Source:

Nvidia's Kyber rack for Rubin Ultra delayed to 2028, creating rare opening for rivals

Anthropic reveals Claude's internal 'global workspace' — J-Space interpretability breakthrough

Tencent open-sources Hy3: 295B MoE model under Apache 2.0 with 256K context

Even Realities hits unicorn status with $150M round for AI smart glasses

Research & Papers

ByteDance releases Seed2.0 model series targeting real-world complexity

ByteDance releases Seed2.0 model series targeting real-world complexity

Embodied.cpp: Portable inference runtime for VLA models on heterogeneous robots

OrbitQuant: Data-agnostic quantization for image and video diffusion transformers

VLA-Corrector: Adaptive detection and correction for robotic manipulation policies

Measuring the gap between human and LLM research ideas at scale

Open Source & Models

Tencent open-sources Hy3: 295B MoE under Apache 2.0

Tencent open-sources Hy3: 295B MoE under Apache 2.0

DSpark speculative debugging now in SGLang for DeepSeek-V4 and Qwen3

ThinkingCap-Qwen3.6-27B cuts thinking tokens by ~50% while maintaining accuracy

Hardware & Robotics

Nvidia reaffirms AI chip roadmap, rejects Kyber delay reports

Nvidia reaffirms AI chip roadmap, rejects Kyber delay reports

AMD Ryzen AI Halo Dev Kit review: AI development in a mini-PC

Tools & Startups

Glaze by Raycast: Build native Mac apps by chatting with AI

Glaze by Raycast: Build native Mac apps by chatting with AI

Pulpie: Pareto-optimal models for web content extraction, 20x cheaper

OfficeCLI: Open-source Office suite purpose-built for AI agents

Money & Markets

Bespoke Labs raises $40M Series A for AI agent training environments

Bespoke Labs raises $40M Series A for AI agent training environments

Sherpa.ai raises $18M for data-sovereign AI platform

Katalyze AI raises $10.5M seed for agentic AI in pharma manufacturing

In Brief

Nvidia's Kyber rack for Rubin Ultra delayed to 2028, creating rare opening for rivals

Anthropic reveals Claude's internal 'global workspace' — J-Space interpretability breakthrough

Tencent open-sources Hy3: 295B MoE model under Apache 2.0 with 256K context

Even Realities hits unicorn status with $150M round for AI smart glasses

Research & Papers

ByteDance releases Seed2.0 model series targeting real-world complexity

ByteDance releases Seed2.0 model series targeting real-world complexity

Embodied.cpp: Portable inference runtime for VLA models on heterogeneous robots

OrbitQuant: Data-agnostic quantization for image and video diffusion transformers

VLA-Corrector: Adaptive detection and correction for robotic manipulation policies

Measuring the gap between human and LLM research ideas at scale

Open Source & Models

Tencent open-sources Hy3: 295B MoE under Apache 2.0

Tencent open-sources Hy3: 295B MoE under Apache 2.0

DSpark speculative debugging now in SGLang for DeepSeek-V4 and Qwen3

ThinkingCap-Qwen3.6-27B cuts thinking tokens by ~50% while maintaining accuracy

Hardware & Robotics

Nvidia reaffirms AI chip roadmap, rejects Kyber delay reports

Nvidia reaffirms AI chip roadmap, rejects Kyber delay reports

AMD Ryzen AI Halo Dev Kit review: AI development in a mini-PC

Tools & Startups

Glaze by Raycast: Build native Mac apps by chatting with AI

Glaze by Raycast: Build native Mac apps by chatting with AI

Pulpie: Pareto-optimal models for web content extraction, 20x cheaper

OfficeCLI: Open-source Office suite purpose-built for AI agents

Money & Markets

Bespoke Labs raises $40M Series A for AI agent training environments

Bespoke Labs raises $40M Series A for AI agent training environments

Sherpa.ai raises $18M for data-sovereign AI platform

Katalyze AI raises $10.5M seed for agentic AI in pharma manufacturing

Repos on GitHub

Trending on HF

In Brief