AI News Highlights from 30th of May, 2025

AI News Highlights from 30th 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


🚀 Headlines & Launches

Perplexity Labs Enables AI-Built Reports, Dashboards, and Apps Perplexity has launched Perplexity Labs, a new workspace for Pro users to turn prompts into end-to-end deliverables—like reports, spreadsheets, apps, and dashboards—using tools such as code execution and web browsing. 🔗 https://www.perplexity.ai/hub/blog/introducing-perplexity-labs

Black Forest Labs Debuts FLUX.1 Kontext for AI-Driven Image Editing The new FLUX.1 Kontext models allow users to manipulate and generate images in context using both text and visual input, powered by flow-matching techniques. 🔗 https://bfl.ai/announcements/flux-1-kontext

Anthropic Open-Sources Circuit Tracing for LLMs Anthropic has released tools to visualize how large language models make decisions by generating “attribution graphs.” These diagrams trace internal model logic and work with open-weight models via a frontend called Neuronpedia. 🔗 https://www.anthropic.com/research/open-source-circuit-tracing


🔍 Deep Dives & Analysis

DeepSeek R1 Reaches Gemini 2.5 Pro Intelligence Levels DeepSeek R1’s latest update scored 68 on the AAI Index, matching Google’s Gemini 2.5 Pro and outperforming Grok 3 mini, Qwen 3 253, and Meta's Llama 4 Maverick. The improvement came without architectural changes, signaling major open-source progress. 🔗 https://threadreaderapp.com/thread/1928071179115581671.html

Why Slower Thinking Yields Smarter AI Lilian Weng draws parallels between AI and human cognition, suggesting that models given more time (i.e., test-time compute) behave more like “System 2” thinkers—logical and accurate. She also dives into RL training, reward hacking, and how this affects models like o1 and R1. 🔗 https://lilianweng.github.io/posts/2025-05-01-thinking/

Chatterbox: Open-Source TTS with Emotional Control Resemble AI released Chatterbox, a new text-to-speech model that beats ElevenLabs in benchmark tests and lets users modulate emotion intensity. 🔗 https://github.com/resemble-ai/chatterbox

RenderFormer: Photorealistic Neural Renderer Microsoft unveiled RenderFormer, a zero-shot neural renderer that generates photorealistic scenes with global illumination from triangle-based inputs—no per-scene tuning needed. 🔗 https://microsoft.github.io/renderformer

Web Bench Sets New Standard for Browser Agent Testing Web Bench offers over 5,700 tasks across 450+ websites to evaluate AI web agents. Anthropic’s Sonnet 3.7 currently leads the benchmark. 🔗 https://blog.skyvern.com/web-bench-a-new-way-to-compare-ai-browser-agents/

Meta’s Zero-Shot Grafting Slashes Vision-Language Training Costs Meta researchers introduced a technique that uses shallow LLM layers to train vision encoders with smaller “surrogate” models, cutting costs by 45% with no performance loss. 🔗 https://github.com/facebookresearch/zero


🧠 Trends & Commentary

‘Max-Performance Domains’ Reward Hyper-Specialization An insightful thread explores “max-performance domains”—fields where being world-class at one narrow skill outweighs mediocrity elsewhere. These roles reward output, not breadth. 🔗 https://threadreaderapp.com/thread/1928174505148698909.html

Nvidia: We're Building 'AI Factories' Now Nvidia's earnings reveal unprecedented demand, with Microsoft generating 5× more tokens and hyperscalers buying 72,000 GPUs weekly. Simpler AI is out; complex reasoning is in. 🔗 https://tomtunguz.com/nvda-2025-05-29/


⚡ Quick Links


To view or add a comment, sign in

Others also viewed

Explore topics