Discover Amazing Content, Share Life Moments
Connect Our Wonderful World
Google Antigravity 2.0 Launched: Copied Codex's UI and Deleted IDE Features for Existing Users
Google has officially released the standalone desktop version of Antigravity 2.0, which focuses on multi-agent collaboration and integration with the Google ecosystem, and its UI is highly similar to Codex. Many expected that adopting proven mature interaction design would fix the poor usability that plagued Google's past products, but the launch was immediately marred by login crashes, forced updates that overwrote older versions, and the removal of original IDE functionality, leading to widespread complaints from long-time users.
42 Researchers Release 100-Page Agent Survey: Code Is the Core Foundation for Next-Generation Agents
A 100-page survey paper on AI Agents co-authored by 42 researchers from institutions including UIUC, Meta, and Stanford systematically organizes methods and real-world applications centered on "code as the harness framework for agents", and proposes that future Agent systems must meet four key properties: executable, inspectable, stateful, and governable. It provides a systematic engineering reference for the currently fragmented landscape of Agent development.
Google Officially Announces Migration of Consumer Gemini CLI to Antigravity Platform, Login and Compatibility Issues Reported on Launch Day
On May 19, 2026, Google announced that it will统一 migrate the consumer-facing Gemini CLI to the new Antigravity agent development platform. The new CLI has been rewritten in Go, supports multi-agent orchestration and asynchronous workflows. Individual users are required to complete the migration within 30 days, and the old service will officially shut down on June 18, 2026. Enterprise user services remain unaffected. On the first day of launch, a large number of developers have already reported issues such as login failures and incompatibility with Termux, and the move has also sparked complaints about Google's confusing product naming.
Claude Launches *The Problem Solvers* Debut Episode: Devin Claims 10x Development Efficiency? X Users Are Divided
On May 19, 2026, Claude officially launched its founder interview series *The Problem Solvers*. The debut episode centers on Scott Wu, founder of the Cognition team that built Devin—the world's first AI software engineer. The episode confirms Devin is built on Claude, with a core goal to boost the software development speed of all engineering teams by 10 times. X users have sparked heated discussions around the authenticity of the 10x efficiency claim, Devin's real-world deployment capability, and its underlying technology, with professional opinions and humorous meme images bringing diverse perspectives to the conversation.
2B-Parameter Open-Source VLM Marlin-2B Released, Matches Gemini 2.5 Flash Performance in the Same Parameter Class, Specialized in Video Structured Extraction
Developer Shubham Sharma has recently officially open-sourced Marlin-2B, a dedicated small vision-language model with only 2 billion parameters. It is the top-performing open-source VLM in its parameter class, with performance comparable to Google's Gemini 2.5 Flash. It is purpose-built for extracting structured information from videos, accurately answering the two core questions "what happened" and "when did it happen", and drastically reducing the deployment cost of video processing workflows.
Azure Foundry Adds 7 New Models Including GPT-5.5: Open-Source Tool Solves the Common Problem of Unclear ROI for Multi-Model Setups
On May 19, 2026, Microsoft Azure announced two updates for its Foundry Model Router: 7 cutting-edge large models including GPT-5.5, Claude-Opus-4.7, and grok-4.1-fast-reasoning have been added, and an open-source automated evaluation repository has been released alongside. This tool enables one-click calculation of the actual quality, cost, and latency benefits of model routing, directly addressing the widespread pain point of unquantifiable return on investment that plagues the deployment of modern multi-model and Agent applications.
Gemini 3.5 Flash: Outperforms peers in speed, gains 9 points, 5x price hike, hallucination rate still stands at 61%
Google launched the new generation Gemini 3.5 Flash large language model on May 19, 2026. Third-party independent tests show its composite intelligence score has increased by 9 points compared to the previous generation, with an output speed of 284 tokens per second, multimodal capabilities leading the industry, and Agent task performance has significantly filled the gap. However, combined with a 3x model price increase and higher token consumption, the overall usage cost has risen 5.5 times from the previous generation. While its hallucination rate dropped from 92% to 61%, it remains at a high level, and the industry has widespread controversy over its cost-performance ratio and performance in long-workflow tasks.
NVIDIA Releases Tri-Mode Large Language Model: Switch Decoding Modes Simply by Modifying Attention Mask, Boosting Single-User Throughput by Up to 4x
In May 2026, NVIDIA launched the industry's first tri-mode language model family, Nemotron-Labs-Diffusion. No architecture modifications or additional auxiliary models are required—users only need to adjust the attention mask to switch between three decoding modes: autoregressive, diffusion, and self-speculation. Adaptable to all scenarios from high-concurrency cloud services to personal local inference, it delivers up to a 4x improvement in actual single-user throughput. The entire model family is open-sourced under an open license.
AI Legend Framed as a Fund Manager with an $8.4 Billion Bearish Bet on Semiconductors? A Prank Tweet Fooled the Entire Finance World
In May 2026, semiconductor industry analyst Dylan Patel posted a prank tweet: he framed top AI influencer Andrej Karpathy as a fund manager holding $8.45 billion in semiconductor put options, and even added the fake gossip that Karpathy quit Anthropic after just 4 hours to jump to a16z's new media division. Thanks to Patel's credibility in the industry, the tweet blew up on Financial Twitter overnight and even triggered panic selling among retail investors. It was later revealed that the holdings were actually copied from real fund manager Leopold's portfolio. Discussions in the AI circle centered on the career joke and a three-way split of opinions on Hacker News, laying bare the absurdity of information gaps across different industries.
Andrej Karpathy Joins Anthropic: Training Claude With Claude
Andrej Karpathy is back at a major tech firm, joining Anthropic to lead pre-training research. His goal this time is to achieve recursive self-improvement using Claude, meaning work on education and open source will have to be put on the back burner for now.