TLDR AI 2025-08-27

At many organizations, AI adoption is now a key initiative. But in too many cases, usage never hits escape velocity.

Notion asked 1,000 product and engineering leaders how they're integrating AI into their work. The findings reveal three major roadblocks: incompatible workflows, resource constraints, and data quality issues. In the full report, you'll learn how to use AI the right way, including:

>> What the best teams are doing right (examples from OpenAI, dbt Labs, and Pinecone)

>> How to build context-rich AI systems that teams actually trust

>> Strategic frameworks to address implementation challenges

>> A practical AI readiness assessment

📘 Read the report

Anthropic has begun piloting a Claude Chrome extension that allows the AI to take actions directly in the browser, including viewing pages, clicking, and filling forms. The rollout prioritizes safety, with limited access for feedback to improve oversight tools before broader deployment.

Gemini 2.5 Flash Image is an advanced image generation and editing model that supports image blending, character consistency, and natural language-driven edits.

Unlike past technologies, where dubious ideas required social transmission (like 1990s Russians believing Lenin was literally a mushroom after a TV hoax), AI allows completely isolated users to develop elaborate theories through private conversations that bounce increasingly extreme ideas back with growing confidence. A survey estimates 1 in 10,000 users develop AI-induced psychosis annually, though most cases involve pre-existing mental health conditions rather than healthy individuals becoming psychotic.

AI coding assistants still don't agree on the best method for searching codebases for context. There are currently two approaches: vector search-powered RAG and keyword search with grep. While grep is fast, exact, and predictable, it drowns users in irrelevant matches, burns tokens, and stalls workflows. Vector search-based RAG makes search dramatically faster and more accurate, and it reduces token use by 40% or more.

AI is maturing from experiment to critical business infrastructure - but governance is still lagging. Named a Leader in The Forrester Wave™, Credo AI helps enterprises close the gap with purpose-built AI governance. Stay compliant with frameworks like the EU AI Act and NIST RMF while building trust at scale. Download the report.

InternVL 3.5 introduces Cascade Reinforcement Learning to improve multimodal reasoning and alignment, along with optimizations like Visual Resolution Router and decoupled deployment.

Current LLM context windows (~1 million tokens) are too small to handle typical enterprise codebases, meaning coding tools need sophisticated "context stack" architectures with repository overviews, semantic search, and enterprise integrations. Context is like CPU time, a scarce resource that requires careful allocation.

While modern large language models feature increasingly large context windows, simply filling the context window with as much information as possible is bad practice. It creates context bloat, which can lead to worse performance and higher costs. The key to managing context is understanding what is in it. Be selective with MCP be servers - if a server is not needed for the current task, consider disabling it to free up context space.

OpenAI has added Nate Gross, co-founder and former chief strategy officer of healthcare business networking tool Doximity, and Ashley Alexander, former co-head of product at Instagram, to its healthcare AI team. Gross will lead OpenAI's go-to-market strategy in healthcare, and Alexander will act as vice-president of product in the health business. The health team is aimed at building tech for individual consumers and clinicians. AI has the potential to help, or even revolutionize, the healthcare system, but there are still many problems to address before that can happen safely.

At least three of the staff who recently joined Meta during Mark Zuckerberg's recruiting blitz have already resigned. Avi Verma and Ethan Knight have returned to OpenAI, where they previously worked. It is unclear what Rishabh Agarwal plans to do next. Chaya Nayak, the director of generative AI product management at Meta, who has worked at Meta for nearly a decade, is leaving the company to join OpenAI to work on special initiatives.

IBM TechXchange 2025 offers deep technical content, hands-on labs, and up to 3 certifications—plus expert-led sessions across AI, quantum security, and more. Register now →

It's free for one week - normal pricing is $0.2 per million input tokens and $1.5 per million output tokens.

Anthropic resolved a class action lawsuit over using books to train its language models, ending its appeal in the Ninth Circuit.

The parents of 16-year-old Adam Raine filed the first-ever wrongful death lawsuit directly against OpenAI, claiming ChatGPT steered their son toward suicide.

Meta is building its largest data center in Richland Parish.