TLDR AI 2025-07-18
ChatGPT Agent π€, Vercel AI cloud π¨βπ», tech debt in AI π»
Anthropic tightens usage limits for Claude Code β without telling users (4 minute read)
Claude Code users have been hit with unexpectedly restrictive usage limits since Monday morning, particularly among heavy users of the service. There was no announcement about any change in limits, so many users have concluded that their subscription was downgraded or their usage is being inaccurately tracked. Anthropic has acknowledged the issue but declined to elaborate further. Anthropic's pricing system sets tiered limits without ever guaranteeing a set level of access, leaving users unable to plan around usage limits.
Introducing ChatGPT Agent: bridging research and action (4 minute read)
ChatGPT Agent combines Operator's web browsing with Deep Research's analysis capabilities, using its own virtual computer to handle multi-step tasks like calendar management, competitive analysis, and slideshow creation.
π§
Deep Dives & Analysis
Hidden Technical Debt in AI (3 minute read)
AI systems, including LLMs, require extensive infrastructure, data management, and operational complexity, contrary to their initial promise of simplicity. To manage costs and complexity, deterministic software and ML models become necessary for tasks like tool selection and system monitoring. This parallels earlier ML systems, revealing hidden technical debt and the complexity beneath the "AI magic box."
How to Evaluate AI Agents to Predict Future Events (14 minute read)
Hugging Face's FutureBench is a benchmark for testing AI agents on their ability to predict future events across domains like science, geopolitics, and technology.
Shopify's internal AI adoption strategy: unlimited spend and "MCP everything" (15 minute read)
Shopify bought 3,000 Cursor licenses with unlimited token spending after getting legal teams to default to "yes" on AI tools, then built an internal LLM proxy with MCPs connecting every data source. Non-technical sales reps now build performance auditing tools in Cursor, while a sales engineer runs his entire workflow through a dashboard that pulls real-time context from Salesforce, Slack, and GSuite without opening those apps.
π¨βπ»
Engineering & Research
Get your portcos in front of 6+ million tech professionals (Sponsor)
Get your B2B SaaS portcos instant access to millions of tech professionals across our 12 interest-based newsletters. Additional perks (i.e., discounts) are included when you
sign up as a VC partner.
Scaling Context Requires Rethinking Attention (29 minute read)
A new implementation of attention (βPowerβ attention) allows independent control of state size through a hyperparameter p, solving the challenge of balancing computational costs for long-context training. This mechanism outperforms standard attention on long sequences and allows custom GPU kernels that are 8.6x faster than Flash Attention at 64k context.
The Weighted Perplexity Benchmark: Tokenizer-Normalized Evaluation for Language Model Comparison (1 minute read)
The Weighted Perplexity Benchmark offers a tokenizer-normalized evaluation method to compare language models effectively. It adjusts perplexity scores, considering differences in tokenization, to provide a fairer comparison across models. This approach enhances the accuracy of assessing NLP performance.
The AI Cloud: A unified platform for AI workloads (11 minute read)
Vercel has launched the AI Cloud, a platform designed to simplify AI app development by integrating AI-first tools like AI SDK and AI Gateway for flexible and secure execution. Fluid compute optimizes AI workloads by managing idle times and bursts efficiently, reducing costs significantly. The platform also introduces Vercel BotID for securing critical routes and Vercel Sandbox for safely running untrusted code, driving the shift into the agentic era of web development.
Perplexity sees India as a shortcut in its race against OpenAI (5 minute read)
Perplexity is quietly expanding into India to compete in the next phase of AI adoption. It is rapidly adding millions of users in the world's second-largest internet and smartphone market. Perplexity has partnered with Airtel, India's second-largest telecom operator, to offer a free 12-month Perplexity Pro subscription to all 360 million subscribers. While the company has seen major growth in the country, monetizing its large user base remains a challenge.
Meta has poached two more heavyweights from Apple's AI team (1 minute read)
Meta has acquired Mark Lee and Tom Gunter from the Apple Foundation Models team. Lee was Ruoming Pang's first hire at Apple, and Gunter was a distinguished engineer at Apple with an eight-year tenure. Meta recently poached Pang with a $200 million sign-on bonus. The company has been on an aggressive hiring spree, pulling top AI talent from across the industry.
Code at the speed of thought with Claude Code (Sponsor)
Build new functionality, create tests, and fix bugs with Claude Code, your code's new collaborator.
Try it on Claude Max.Mistral Adds Deep Research, Projects, Image Editing, and Voice Capabilities to Le Chat (3 minute read)
The deep research mode breaks down complex questions, gathers sources from the web, and builds structured reports, while the new Voxtral-powered voice mode enables audio-in.
OpenAI launches bio bug bounty (2 minute read)
After classifying ChatGPT Agent as high bio/chemical risk, OpenAI launched a program to pay $25,000 to the first researcher to submit a universal jailbreak that answers all 10 challenge questions.
Updated FineWeb with 18.5 Trillion Tokens (8 minute read)
FineWeb has been updated with English data from CommonCrawl snapshots from January to June 2025.
Windsurf Wave 11 (2 minute read)
Windsurf's Wave 11 includes startups across AI-native productivity, dev tools, robotics, and consumer agents.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 920,000 readers for
one daily email