TLDR AI 2025-06-17
OpenAI-Microsoft relationship 👊, Character AI writing evals 📝, OpenAI’s prompt API primitive 👨💻
The Microsoft-OpenAI Relationship Falters Over Windsurf Acquisition (1 minute read)
OpenAI executives have considered accusing Microsoft of anticompetitive behavior as tensions escalate over the AI startup's $3 billion acquisition of Windsurf, whose coding IP OpenAI wants to keep from Microsoft's GitHub Copilot. The standoff highlights OpenAI's struggle to reduce its dependence on Microsoft's cloud services while still needing the tech giant's approval to complete its for-profit conversion.
OpenAI Wins $200 Million U.S. Defense Contract (2 minute read)
The one-year deal marks OpenAI's first public Defense Department contract and follows similar moves by rival lab Anthropic and the weapons-maker Anduril, which recently signed a $100 million contract.
Coding Agents Have Crossed a Chasm (9 minute read)
Powerful AI coding tools are shifting developers from writing code to architecting solutions and crafting precise problem descriptions. Feeding Claude ASCII sequence diagrams of OAuth flows or entire DOM trees from Chrome DevTools turned one developer's 45-minute debugging sessions into 10-minute fixes. However, AI amplifies both strengths and misconceptions—it'll confidently help you implement terrible architecture faster than ever before.
AI Should Be Warm But Not Alive (8 minute read)
The lead of model behavior and policy at OpenAI outlines the company's approach to managing the increasingly emotional connections users form with ChatGPT. While AI consciousness remains scientifically unknowable, OpenAI wants to avoid encouraging users to perceive models as sentient without evidence to support it. Its goal is to create a model that's warm and approachable that doesn't claim to have an inner life.
👨💻
Engineering & Research
Build what's next with AI: Save $700 on tickets to GitHub Universe (Sponsor)
Code Interpreter Reasoning (GitHub Repo)
CoRT post‑trains large reasoning models with hint engineering so they can delegate calculations to external code interpreters.
TreeRL for On‑Policy LLM Training (16 minute read)
TreeRL uses on‑policy tree search and intermediate supervision to train LLMs without a separate reward model, delivering stronger math and code reasoning than ChainRL.
🇵 🇷 🇴 🇲 🇵 🇹 🇸 are an API primitive now! (1 minute read)
Prompts are now an API primitive on OpenAI's services. Developers can now centrally manage, version, and optimize prompts across the Playground, API, Evals, and Stored Completions. Prompts can be preconfigured with tools, models, and messages. The change was designed to help developers refine and reuse prompts better and faster.
Nanonets OCR Small (4 minute read)
Nanonets-OCR-s is a state-of-the-art image-to-markdown OCR model that goes far beyond traditional text extraction. The model can transform documents into structured markdown with intelligent content recognition and semantic tagging. It understands document structure and content context and can deliver intelligently formatted markdown output ready for downstream processing by large language models. Nanonets-OCR-s can recognize LaTeX equations, intelligently describe images, detect and isolate signatures, extract watermarks, handle checkboxes, and extract complex tables.
Call Intelligence that doesn't suck. (Sponsor)
Retellio listens to every sales call, surfaces the key customer insights, and delivers them as a podcast, newsletter, or Slack message. Giving sales leaders super intelligence that is easy and fun to consume.
See how it works.
Models.dev (Website)
This site contains an open-source database of AI models that lists AI model specifications, pricing, and capabilities.
Character AI's Writing Evaluation Framework (2 minute read)
Character AI outlined a new framework that scores conversational storytelling by combining creative writing heuristics with objective metrics to judge pacing, engagement, and fun.
TikTok Launches AI-Generated Product Models and Reviewers (3 minute read)
TikTok's Symphony platform will now generate videos of virtual avatars modeling products and trying on clothes, potentially replacing human influencers.
AstraZeneca Signs $5B AI Research Deal with China's CSPC (3 minute read)
AstraZeneca will pay $110 million upfront for CSPC to use AI-driven research to discover small molecule therapies, including an oral treatment for immunological diseases.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 920,000 readers for
one daily email