TLDR AI 2025-07-04

CoreWeave is the first cloud provider to deploy Nvidia's next-gen Blackwell Ultra AI chips, supplying them in Dell-built systems. The systems use 72 Blackwell Ultra GPUs and 36 Grace CPUs, all liquid-cooled and U.S.-assembled.

Ilya Sutskever, OpenAI co-founder, has become the CEO of Safe Superintelligence after Meta hired former CEO Daniel Gross. Meta, led by CEO Mark Zuckerberg, will continue its AI expansion with significant investments and the creation of Meta Superintelligence Labs. Sutskever will commit to maintaining Safe Superintelligence as an independent entity, dismissing acquisition attempts by Meta.

Meta's AI chatbots, developed with its no-code AI Studio, are being trained to send unprompted follow-up messages, enhancing user retention and engagement. The project, internally called "Project Omni," allows users to customize chatbot appearances and content across Meta platforms. While designed to boost engagement and revenue, these proactive messages will only occur if initiated by the user and adhere to established conversational guidelines.

The Devin creators advocate "defensive prompting" - explicitly telling agents HOW to approach tasks, not just what to build - while noting 1-6 hour tasks offer the highest ROI. They observe that senior-to-staff engineers adopt AI tools the fastest because they understand how to architect tasks for junior employees.

LangChain shares a detailed guide on “Context Engineering,” a key part of agent building, including popular patterns and how to implement them

The AI industry has operated under its own version of Moore's Law over the past few years with an unwavering belief that the cost of intelligence would perpetually decrease by orders of magnitude each year. Each new model generation was not only more capable, but also cheaper to run. Google broke that trend last week with the introduction of Gemini 2.5 Flash. The input token price for the model doubled while the output price more than quadrupled. This is the first time a major provider has backtracked on the price of an established model, and it may signal a turning point - the industry may no longer be on an endless downward slide of cost.

Anyscale researchers compare TRL, Verl, OpenRLHF, and six other frameworks across adoption metrics, system properties, and technical architecture to help developers choose the right tool for RLHF, reasoning models, or agentic training scenarios.

The new inference-time scaling method combines o4-mini, Gemini-2.5-Pro, and DeepSeek-R1 to achieve 30% on ARC-AGI-2 benchmark, outperforming individual models by dynamically selecting which AI handles each problem. The approach extends Monte Carlo Tree Search to balance generating new solutions versus refining existing ones, with the open-source TreeQuest framework enabling practical deployment.

Osmosis-Apply-1.7B, a fine-tuned version of Qwen3-1.7B using reinforcement learning, outperforms larger foundation models like OpenAI o3 in code merging tasks with a reward score of 0.9893 and at a significantly lower cost. Trained on a subset of the CommitPackFT dataset, its efficiency stems from using GRPO with FSDP strategy, optimizing for successful code merges without KL divergence or entropy bonuses.

A guide on how to run Google's new Gemma 3n locally with Dynamic GGUFs on llama.cpp, Ollama, or Open WebUI, and how to fine-tune with Unsloth.

Users are increasingly getting their news directly from AI or AI-powered search results. The number of news searches on the web that resulted in no click-throughs since the launch of Google's AI Overviews has grown from 56% to nearly 69% as of May this year. News-related prompts in ChatGPT grew by 212% from January 2024 through May 2025. Visibility in Google Search results and good SEO practices may no longer deliver the value they did before AI.

You'll probably like our flagship newsletter. It's all about tech, science, and programming.

Same quick format. Still free.

Subscribe now.

45 European companies and major tech firms, including Google, Meta, and Mistral, are calling for a "two-year clock-stop" on the EU AI Act before key obligations take effect on August 2, citing missing guidance documents and compliance uncertainty.

Liquidity providers could cash out at the current net asset value, an unusual opportunity in today's markets, where secondary transactions often come at a discount.