DeepMind Core Scientists Continue to Leave: AlphaFold Lead John Jumper Joins Anthropic
AlphaFold team lead and 2024 Nobel Prize in Chemistry winner John Jumper announced he is leaving Google DeepMind to join Anthropic. This career move comes nearly nine years after Hassabis entrusted him with leading AlphaFold just six months after his PhD. Days earlier, Transformer paper author and MoE architecture proposer Noam Shazeer also left Google to join OpenAI. Several Chinese bloggers tracked this wave of top AI research talent movements and noted that DeepMind is experiencing a continuous exodus of core scientists.
Source:
- @JohnJumperSci: https://x.com/JohnJumperSci/status/2068001285173834106
- @op7418: https://x.com/op7418/status/2068190452554281412
- @MaxForAI: https://x.com/MaxForAI/status/2068191506758779261
Codex Launches Cross-Device Task Migration Feature Handoff
Codex has introduced the Handoff feature, allowing users to migrate ongoing development tasks along with the complete Git state (including uncommitted changes and current branch) to a remote host via natural language instructions, and pull them back upon return. Many bloggers believe this feature solves the practical pain point of ‘having to leave while AI is modifying code halfway,’ but also note that the setup steps are numerous—requiring preparatory work in SSH connections, remote Codex installation, same-account login, and Git repository cloning.
Source:
- @guinnesschen: https://x.com/guinnesschen/status/2068062280345162047
- @dotey: https://x.com/dotey/status/2068183780938985827
- @thsottiaux: https://x.com/thsottiaux/status/2068120572673077274
WeChat AI Assistant ‘Xiaowei’ Begins Beta Testing
WeChat’s AI Agent ‘Xiaowei’ beta version has begun beta testing, with an interface similar to a conversational AI assistant. It can answer questions, invoke mini-programs within WeChat, and even generate mini-programs with a single sentence. Some bloggers confirmed that the main model uses WeLM, with DeepSeek as a fallback for some answers. Many industry bloggers believe this marks the beginning of an AI Agent evolution wave for IM software, but some voices point out that Xiaowei’s capability boundaries within the WeChat ecosystem may limit its scalability.
Source:
- @MaxForAI: https://x.com/MaxForAI/status/2068001148691509691
- @Fenng: https://x.com/Fenng/status/2068220272306754016
- @LufzzLiz: https://x.com/LufzzLiz/status/2068341744513823119
Using Built-in Browser + Canvas Tool to Achieve Infinite Canvas Image Generation with Codex GPT Image 2
Multiple bloggers discovered and verified a practical tip: opening a canvas-type web tool in Codex’s built-in browser provides an infinite canvas for GPT Image 2, enabling continuous annotation and image generation. This method requires no additional plugins, and Pro users can leverage GPT’s unlimited image generation permissions (no queue). Some bloggers mentioned that this feature is particularly convenient for annotation images and are planning to open-source related plugins.
Source:
- @cellinlab: https://x.com/cellinlab/status/2068193900570046502
- @zhongerxin: https://x.com/zhongerxin/status/2068027614300893383
- @vista8: https://x.com/vista8/status/2068333691902833072
Nano Banana Upscales GPT-Image-2 Images to Improve Generation Quality
A blogger shared a method to upscale GPT-Image-2 generated images using Nano Banana: GPT-generated non-photorealistic images often have broken textures and rough edges. Inputting them into Nano Banana for redrawing removes messy details, adds meaningful details, and sharpens text and edges. Comparison results show a significant visual improvement—a practical tip to enhance image generation quality without additional API calls.
Source:
Systematic Discussion on Agent Coding Engineering Practices
Multiple bloggers engaged in in-depth discussions on quality management of Agent-generated code. Baoyu shared a complete methodology upon request, covering requirements analysis (providing sufficient context to the Agent), system design (producing a plan first before coding), code review (human fallback), automated testing (CI integration), canary releases and feature flags, and using structured logs to help the Agent locate production issues. Another blogger shared a method of using prompt injection to detect PR submitters who do not review code. Meanwhile, the discussion on ‘whether all rules should be written into AGENTS.md’ also attracted attention, with the view that root causes in testing or processes should be solved with testing and processes, not crammed into AGENTS.md.
Source:
- @dotey: https://x.com/dotey/status/2068363092904276316
- @dotey: https://x.com/dotey/status/2068231396015890449
- @dotey: https://x.com/dotey/status/2068016587953643922
- @Gorden_Sun: https://x.com/Gorden_Sun/status/2068341601173483965
Cloudflare Launches Agent Temporary Account Feature
Cloudflare has introduced a temporary account feature, allowing Agents to directly deploy Worker scripts and databases to Cloudflare without human intervention. After verification and modification, humans can bind the resources to their own Cloudflare account via the claim link returned by the Agent. This design significantly reduces the manual operation cost of AI automated deployment.
Source:
- @Gorden_Sun: https://x.com/Gorden_Sun/status/2068334204337942814
Statistics: Number of scanned timelines=240, Number of hit bloggers=36, Total hit tweets=126, Weighted tweet score=106.95, Original tweets=59, RT tweets=14, Crawl attempts=1, Boundary coverage status=tail_confidently_crossed_target_boundary