AI News Highlights from 29th of May, 2025
Together with JobsAICopilot, The Best AI Job Application Bot. Effortlessly apply to jobs at 400,000+ companies worldwide with the power of AI-driven job search and smart application automation. Try it now
🚀 Launches & Headlines
DeepSeek Quietly Updates R1 Model Ahead of Nvidia Earnings DeepSeek published an ambiguous update to its R1 reasoning model on Hugging Face, with no release notes or changelog. The timing, just before Nvidia’s earnings call, raised eyebrows in the AI community. 🔗 https://techcrunch.com/2025/05/28/deepseek-updates-its-r1-reasoning-ai-model-releases-it-on-hugging-face/
Mistral Introduces Codestral Embed for Code Search and Retrieval Codestral Embed, Mistral’s newest embedding model, is tuned for code. It outperforms competitors on retrieval tasks and allows users to fine-tune dimensions and precision for efficiency. 🔗 https://mistral.ai/news/codestral-embed
Meta’s AI Assistant Hits 1 Billion Users Meta's assistant, now available in a standalone app, has reached one billion monthly users across Facebook, Instagram, and WhatsApp. Monetization strategies—ads or subscriptions—are still on the horizon. 🔗 https://www.cnbc.com/2025/05/28/zuckerberg-meta-ai-one-billion-monthly-users.html
🔍 Insights & Analysis
You Could Have Invented Transformers A new essay reframes transformer architecture as a logical evolution from n-gram models and RNNs. It argues that transformers are essentially enhanced MLPs with tweaks that solved bottlenecks in prior systems. 🔗 https://gwern.net/blog/2025/you-could-have-invented-transformers
Structured JSON Outputs Make Code Agents Smarter Hugging Face shows that agents generating structured code outputs via JSON can outperform traditional methods, offering more reliable performance in code-generation benchmarks. 🔗 https://huggingface.co/blog/structured-codeagent
⚙️ Research & Engineering
Paint With Ember: Directly Manipulate Image Model Concepts Goodfire’s Paint With Ember lets users guide image models by sketching instead of prompting. The tool translates pixel sketches into neural activations using autoencoders trained on Stable Diffusion XL-Turbo. 🔗 https://www.goodfire.ai/blog/painting-with-concepts
Spurious Rewards Match RL on Qwen2.5 Researchers found that even incorrect or random training signals produced 15–24% performance gains on Qwen2.5-Math models—suggesting pretraining plays a major role in reward learning. The effect didn’t generalize to Llama3 or OLMo2. 🔗 https://rethink-rlvr.notion.site/Spurious-Rewards-Rethinking-Training-Signals-in-RLVR-1f4df34dac1880948858f95aeb88872f
PixelFlow: Direct Image Generation in Pixel Space PixelFlow generates high-fidelity images directly in pixel space, bypassing VAEs while maintaining semantic control and competitive quality. 🔗 https://github.com/shoufachen/pixelflow
🧠 Commentary & Trends
Dario Amodei: White-Collar Jobs Face Imminent AI Disruption Anthropic’s CEO warns of a coming “white-collar bloodbath,” projecting 10–20% unemployment within five years as AI automates entry-level roles in law, finance, and tech. 🔗 https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic
China’s AI Race Faces Structural Hurdles Despite bold goals to dominate AI by 2030, China continues to lag the U.S. in semiconductor tech and private funding. Government backing and open-source innovation remain key to their strategy. 🔗 https://www.recordedfuture.com/research/measuring-the-us-china-ai-gap
Opera Launches AI-Driven Browser That Can Code For You Opera Neon is a new browser optimized for task automation and AI workflows, including coding websites and games directly through built-in assistants. 🔗 https://techcrunch.com/2025/05/28/operas-new-browser-can-code-websites-and-games-for-you/
⚡ Quick Bits
The AI Discourse Is a Mess – A candid blog post calling out misinformation from both AI doomsayers and cheerleaders. 🔗 https://steveklabnik.com/writing/i-am-disappointed-in-the-ai-discourse/
Model Transparency in Action – Users can now view model “thought processes” mid-task. 🔗 https://threadreaderapp.com/thread/1927423603940663402.html
Proving Humanity Will Get Harder – The CAPTCHA paradox: better AI makes it harder for humans to prove they're human. 🔗 https://talkingrobot.com/2025/05/the-captcha-paradox/
Shorter Reasoning = Better Accuracy – A Meta study found cutting down reasoning steps boosts accuracy by 34% and reduces compute needs. 🔗 https://venturebeat.com/ai/less-is-more-meta-study-shows-shorter-reasoning-improves-ai-accuracy-by-34/
AMD Buys Enosemi – Boosts AMD’s silicon photonics and AI infrastructure capabilities. 🔗 https://techcrunch.com/2025/05/28/amd-buys-silicon-photonics-startup-enosemi-to-fuel-its-ai-ambitions/