TLDR AI 2025-08-18

OpenAI is rolling out a friendlier version of GPT-5 with subtle changes after users complained the model felt too formal and that they missed 4o's personality, though internal testing shows no increase in sycophancy.

Anthropic granted its most advanced models the ability to end conversations as a last resort against persistently harmful or abusive interactions, developed partly as an exploration into AI welfare and moral status.

AI startups are growing like nothing we've ever seen - some hit $100M ARR in their first year. Bessemer studied 20 companies and found two types: Supernovas that explode to $125M by year two but have 25% gross margins and fragile retention, and Shooting Stars that quadruple yearly with 60% margins. The new benchmark isn't T2D3 anymore, it's Q2T3 (quadruple, quadruple, triple, triple, triple).

Archon is a copilot for Mac and Windows that takes natural language instructions and then performs tasks by taking screenshots, using GPT-5 to create a plan, and then using a custom fine-tuned model to execute clicks and keystrokes. The copilot uses GPT-5's powerful reasoning, combined with tiny fine-tuned models, to control any interface through natural language. This article looks at how Archon was built.

This ranking evaluates 19 Chinese labs, not on the raw capabilities of their models, but on their contributions to AI research. Deepseek reigns as a top lab, despite recent quiet periods, with newer companies like Moonshot AI and Zhipu closing the gap after their recent releases of K2 and GLM-4.5, respectively.

The impact of advanced AI far outpaces our understanding of advanced AI—a disconnect that will only widen as we race toward AGI. This sprawling list tackles the fundamental philosophical questions about these systems: Why do teenagers master driving in weeks while Waymo burns through millions of hours? How can LLMs store vastly more facts than human brains despite having fewer connections? Does the 7-month doubling of AI task complexity represent something deeper about cognition itself?

Training frontier models? Turing remains the independent choice for the world's top AI teams — from multimodal training to RL gyms. 1000s of domain-trained experts across 60+ languages, 700+ benchmark tasks built from real-world business and STEM use cases, 50+ frontier projects supported successfully. Ask about free migration paths

OpenAI shared a cheat sheet outlining best practices for using GPT-5 in coding scenarios, including prompt design, function calling, and tool use.

A solo dev built a complete web search engine from the ground up in just two months, leveraging 200 GPUs to generate 3 billion neural embeddings across 280 million indexed pages. The project uses transformer-based embeddings, a dramatic improvement over traditional keyword matching, as it supports much more complex queries.

GPT-5, when accessed via the OpenAI API, has its own hidden system prompt that is independent from the system prompt users specify in their API calls. It is very resistant to sharing the details of this secret system prompt. The prompt seems to specify a level of 'oververbosity' where the assistant prioritizes direct answers with minimal extra content and keeps explanations short unless users ask for more.

Since launching in 2023, ChatGPT's mobile app has reached $2B in consumer spending, with an average return of $2.91 per install, vastly outpacing other AI mobile apps.

A Triton implementation of GPT-OSS was included in the release, but the kernels are forward-only - this developer vibe-coded their own kernel to use for training.

Customers of open weight model providers will have to be careful to make the right choice.

Gemini's Projects feature enables the assistant to run Research tasks across uploaded files.