OpenAI launches gpt-oss open-weight LLMs
Two sizes: 120B parameters for powerful hardware, 20B for desktops/laptops.
Enable agentic tasks with chain-of-thought reasoning, tool use (web search, Python execution).
Fully customizable with fine-tuning and adjustable reasoning effort.
Provide full chain-of-thought outputs for transparency and debugging.
Apache 2.0 licensed for commercial use without patent or copyleft risks.
Performance close to OpenAI’s proprietary models on benchmarks like MMLU and AIME.
Extensive safety testing with external reviews marks progress for open model safety.
Developer-friendly playground and broad vendor collaboration enhance accessibility.
Community excited about local frontier-quality LLMs but highlight performance trade-offs vs. other open models.
AI tools don’t make engineers 10x productive—here’s why
AI coding assistants excel at boilerplate and small scripts but struggle with large codebases, complex contexts, and nuanced language.
Software delivery involves many steps beyond coding (ideation, review, testing, deployment) that AI has not notably shortened.
“10x engineer” productivity often comes from reducing unnecessary work, something AI does not replicate.
Many 10x productivity claims are hype or management-driven pressure rather than measurable gains.
Emphasizes maintaining coding joy and mastery over speed, urging realistic expectations about AI’s impact.
Advises managers to foster trust and avoid unrealistic productivity demands fueled by AI hype.
DeepMind unveils Genie 3: scalable real-time 3D world model
Generates diverse, immersive 3D environments at 720p/24fps without explicit 3D representations like NeRFs.
Supports text-prompted creation of dynamic, interactive worlds including natural, historical, and fantastical settings.
Simulates natural phenomena (water, lighting) and complex environment interactions.
Enables text-driven user interactions and powers embodied AI agents (e.g., SIMA) for navigation and task pursuit.
Demonstrates emergent long-term consistency over minutes but limited multi-agent social interaction and geographical accuracy.
Released as a controlled research preview emphasizing safety in open-ended world generation.
Use cases include education, AI training, robotics simulation, and generative media.
Prompts community reflection on neural world models versus traditional 3D engines and prospects of robots “learning in their dreams.”
US pressures TSMC to invest $400B and buy 49% stake in Intel for tariff relief
US ties tariff relief on Taiwan to TSMC acquiring a large Intel stake and massive US semiconductor investments.
Intel faces revenue decline ($79B in 2021 to $53B in 2024), production delays, and strategic uncertainty despite federal grants.
The $400B investment plus forced acquisition is financially and politically controversial.
Industry doubts feasibility and critiques the approach as extortionate, likely inflating consumer costs.
Seen as a geopolitical move to bolster US semiconductor independence and tie Taiwan semiconductor capability to US defense commitments.
Alternative partnership suggestions exist, including collaborations with Apple or Nvidia.
Highlights the complex interplay of trade policy, national security, and global chip supply chains.
uBlock Origin Lite: minimal, declarative content blocker for Apple devices
Lightweight, free content blocker for iPhone, iPad, Mac, and Apple Vision, available via Mac App Store.
Uses declarative filtering leveraging browser-native CSS/JS injection—no persistent background service.
Integrates popular filter lists (EasyList, EasyPrivacy, Peter Lowe’s Ad servers).
Minimal CPU/memory footprint; service worker activates only during UI interactions.
Compatible with iOS 18+, macOS 15+, visionOS 2.0+.
No user data collection ensured by a detailed privacy policy.
Appeals to privacy-conscious users wanting streamlined ad blocking without extension bloat or performance overhead.
View all episodes
