TLDR AI 2025-10-28

Think the newest LLM is the best for coding? Despite their shared strengths and flaws, each LLM has a unique and inherent style—a measurable “coding personality” that drives their distinct results.

Sonar put GPT-5, Claude Sonnet 4, and Llama 3 models to the test in their latest State of Code report - and found some surprising results. Read it to discover:

• The shared strengths and flaws of LLMs
• Coding archetypes for the leading LLMs
• Hidden quality & security risks of using LLMs for code generation
• How to select the best model for your needs

Get the report ↗️

Anthropic is continuing its push into financial services by launching Claude for Excel, which lets users analyze and modify spreadsheets by chatting with Claude in a sidebar. It is currently in beta behind a waitlist.

Mercor, a human resources and recruiting startup, has finalized a new funding deal that values the company at $10 billion. The company manages 30,000 contractors around the world who help AI chatbots learn how to think and speak like humans. Venture capital firms are eager to offer funding to founders at sky-high valuations because they fear missing out on the next AI tech giant. Mercor is currently being sued by Scale for allegedly stealing trade secrets.

Google recently published an announcement that its Gemma model had discovered a new potential cancer pathway. It's important to note that the model didn't invent or theorize a method to discover potential drug candidates - the entire construction of the experiment was imagined and executed by humans. The model was used to probabilistically narrow a set of potential candidates to a manageable set that could be inspected by human review. Humans still possess a reasoning capability that is not present in any current AI system.

RL environments can be surprisingly complex and fun to create. This post explains what RL environments are, introduces the 'verifiers' framework, and also walks readers through how to create an environment for a benchmark called AgentDojo. RL environments are scenarios that LLMs operate in for evaluation or training. Designing them means essentially defining the maze, the rewards, and how the LLM navigates through it.

We love this practical guide from You.com that walks through a proven framework to identify, prioritize, and document high-value AI opportunities. Learn how AI can drive the most impact in your organization, both internally and externally. Get the AI Use Case Discovery Guide.

A new sampling method enables base models to reach single-shot reasoning performance comparable to reinforcement learning, without retraining or external verifiers, while maintaining diversity and multi-shot accuracy.

Thinking Machines Lab demonstrated that training smaller AI models by having them learn from their own mistakes, by grading them with a larger teacher model, achieves the same reasoning performance as RL at 9-30x lower cost. "On-policy distillation" combines the relevance of learning from your own outputs with the dense feedback of traditional distillation, reaching 70% on AIME'24 math problems in just 150 training steps compared to 17,920 GPU hours of RL.

TOON is a compact, human-readable format designed for LLM input that allows developers to pass structured data to LLMs with significantly reduced token usage. It excels at uniform complex objects. Standard JSON is verbose and token-expensive - TOON conveys the same information with fewer tokens. In benchmark tests, TOON achieves higher accuracy compared to JSON while using significantly fewer tokens.

The smartphone chipmaker's new AI200 and AI250 inference accelerators—shipping in 2026 and 2027—sent shares up 20% on the promise of diversification beyond a stagnating mobile market. Saudi Arabia's Humain is committed to purchasing roughly $2 billion worth of the chips starting next year.

Chinese Lab MiniMax open-sourced M2, a 230B parameter model with 10B active that ranks #1 among open-source models on Artificial Analysis' composite intelligence score. It slightly outperforms the previous generation Claude Sonnet model on coding benchmarks at less than 10% of the cost.

Access frontier models using the OpenAI SDK with integrated Snowflake tools. Build secure, enterprise-ready AI apps with state of the art inference and built-in control over data privacy and compliance. Get started in seconds.

Following recent safety controversies, OpenAI revealed that approximately 1-3 million users per week show signs of mental health emergencies, suicidal ideation, or emotional reliance on ChatGPT.

ChatGPT's mobile app download growth and daily active use are slowing.

AI search engines tend to cite less popular websites - ones that wouldn't even appear in the Top 100 links listed in an 'organic' Google search.