TLDR AI 2025-04-22

Wish AI understood your context and could take actions for you? Concierge is the first connected AI Assistant that can read & write to your apps in real-time. Simply connect in 2 clicks (Gmail, Slack, Notion, Jira, HubSpot, Linear, Asana, Confluence, Airtable, Attio, Salesforce, etc) and ask Concierge anything.

Reply to emails, update CRMs, create new tickets, summarize customer feedback, draft a PRD, do research across knowledge sources, and more.

Try Concierge for free or use the code TLDR3 to get Concierge Pro for 3 months.

Meta is using AI to identify teens on Instagram and automatically place them in restricted teen account settings, which now include even stronger default protections and require parental approval for changes.

Anthropic unveiled a multidimensional framework to assess and mitigate AI harms, addressing both catastrophic and everyday risks. The approach evaluates physical, psychological, economic, societal, and autonomy impacts, guiding policy, testing, and enforcement strategies. This initiative, aiming for adaptable, proportionate safeguards as AI capabilities evolve, complements its Responsible Scaling Policy.

This post explores the challenge of the new crop of reasoning models and presents evidence to suggest that OpenAI is over-optimizing on certain objectives that make its models brittle and prone to hallucinations.

Writing quality reward models (WQRM) are a way to evaluate the quality of creative writing. They can be used to train creative writing models. WQRMs are a new development for RL-trained models with verifiable rewards. This thread shows a case where WQRM correlated strongly with overall writing quality.

Recent advancements in LLMs like OpenAI's o3 and o4-mini have made them effective for search-based tasks, avoiding past issues with hallucinations. Unlike earlier models, these LLMs integrate search results into their reasoning process, providing reliable real-time insights. This progression diminishes reliance on traditional search engines and suggests shifts in the Web's economic model.

Happening live this Thursday! Join Erika Rice Scherpelz, VP of Engineering at Sourcegraph and Kristina Hager, Principal Software Engineer at Indeed, for a discussion on accelerating developer iteration (inner loop), automating deployment and delivery (outer loop), and the role of AI-driven engineering practices throughout the entire SDLC. Join the conversation

LOO-StabCP boosts the speed of conformal prediction by using leave-one-out stability, providing scalable uncertainty estimation without sacrificing accuracy.

This study presents a simulated self driving agent that went 2 years between collisions. It was trained fully with self-play and is a substantial improvement over the state-of-the-art trained on Gigaflow.

DeepMind has released the next version of its geometry model, which answers 84% of questions correctly, up from 54% by prior methods. The main improvements come from the Gemini language model and better search.

The MAGI 1 model is a new autoregressive video generation model that can generate long form and consistent videos. It performs on par with Wan video generation while being slightly below some closed models.

TLDR is hiring an AI curator to help write our daily newsletter, read by over 700,000 subscribers.

We're looking for someone who works full time in AI (at a frontier lab, AI startup, or major tech company) and regularly keeps up with the latest AI news and research.

Time commitment is ~5-10 hours/week paid at a rate of $100/hr.

To apply please send your LinkedIn or resume to jobs@tldr.tech along with a few sentences on how you currently keep up with AI news!

Stackblitz, the creators of Bolt, is hosting the world's largest hackathon for non-devs and vibe coders on 5/30 for anyone around the world to participate.

Offered by the Cloud Security Alliance (CSA), this free self-paced course delivers a foundational understanding of how generative AI models work, and how to create and refine prompts to improve output. Register for free

Mercor launched a $50,000 fellowship for PhD students and postdocs in STEM focused on identifying raw talent based on ideas, not pedigree or connections.

New code strings in ChatGPT's web bundle reference “buy_now” buttons and a “shopify_checkout_url,” suggesting OpenAI is building a native purchase flow into the assistant.

Google is inviting AI startups focused on U.S. infrastructure to apply for its six-month accelerator, offering mentorship, technical support, and strategic guidance.