TLDR AI 2025-06-19

U.S. testers can talk to Google Search via the new Live icon, receiving AI‑generated spoken answers and follow‑up suggestions along with on‑screen web links.

Midjourney's V1 turns any image into a 5-second video though it's Discord-only at launch. The company's long-term goal is to generate interactive 3D environments in real time.

Meta is trying to bring on board former GitHub Chief Executive Nat Friedman and Daniel Gross, Friedman's partner in the AI-focused venture capital fund, NFDG. Friedman helped establish GitHub as a pioneer in AI with the launch of GitHub Copilot. He departed GitHub to focus on investments at NFDG. NFDG focuses on identifying hot AI startups and investing anywhere between $1 million to $100 million in seed and growth rounds. It has backed notable companies over the years, like Stripe, Character Technologies Inc., and Perplexity AI.

Google's Release Notes podcast features Connie Fan and Danny Tarlow discussing the design goals behind Gemini's code generation, the emergence of “vibe coding,” and how AI may reshape programming languages.

OpenAI's researchers identified an internal activation pattern linked to misaligned personas in language models and showed it can be detected and dampened through targeted retraining, hinting at an early‑warning system for alignment drift.

Blog posts and technical papers are increasingly plagued by the "synthetic flavor" of LLM writing—empty summary sentences, persistent vagueness, and dull sentence structure. The greatest value of AI is for outlining, polishing early drafts, and targeted editing. Decent writing is cheaper than ever to produce - knowing what's worth writing, or not, still requires human judgment.

How many times have you thought: “I wish we had a GUI for that [automation / admin action / dashboard]”? With Refine AI, any internal tool is one prompt away. Describe what you need, and Refine will guide you through project setup, generate CRUD pages, add authentication, and handle secure deployment. Build your first tool for free

Get a glimpse of Refine AI in action.

Sakana AI researchers built a system that can instantly customize large language models using just a text description, no training data or lengthy fine-tuning required. Their T2L model works by compressing hundreds of LoRA adapters (small, efficient modifications to AI models) into a single network that generates new customizations on demand.

The performance of AI agents on longer-duration tasks can be explained by an extremely simple mathematical model - a constant rate of failing during each minute a human would take to do the task. This implies an exponentially declining success rate with the length of the task. The regularity allows for the success rate for an agent at different task lengths to be estimated. The underlying cause of failure on longer tasks is likely due to the increasingly large set of subtasks - failing any single one fails the whole task.

Claude Code now has support for remote MCP servers. Developers can now connect tools and data sources to personal their coding experience without needing to manage local servers. Claude Code can access both tools and resources exposed by MCP servers. It can also be integrated with any remote MCP server. The growing ecosystem of servers means that new capabilities are constantly coming online.

A new investigative report compiles whistleblower testimonies revealing OpenAI's transformation from a nonprofit to a $300 billion company, including claims that senior staff warned the board Sam Altman was "psychologically abusive" and shouldn't lead them to AGI.

The latest DeepNVMe release broadens support to model checkpointing and inference workloads, adds PCIe Gen5 NVMe scaling, and introduces CPU‑only and offset‑based I/O options to boost data‑bound training speeds in DeepSpeed 0.17.1 and above.

Hear analyst insights, market research, and strategic guidance from 451 Research and Glean on June 25. Register now.

A demo from OpenAI shows AI agents replacing manual QA workflows through Playwright.

OpenAI outlined a prevention‑focused plan, including red teaming, expert partnerships, and security controls, to balance accelerated biotech research with controls that curb potential biothreat misuse.

Moonshot AI's Kimi-Dev-72B achieves 60.4% on SWE-bench Verified—the new state-of-the-art among open models—by using large-scale reinforcement learning that patches real repositories in Docker and only rewards solutions when the entire test suite passes.

OpenAI is cutting ties with Scale AI after Meta's multi-billion dollar investment in the data labeling startup brought CEO Alexandr Wang onto Meta's team, with Google also reportedly planning to drop Scale as a provider.