A NOTE FROM THE EDITOR

Welcome to the latest Datakami Bot Bulletin! This month, we cover major releases from OpenAI and Google, a new study on AI's impact on developer productivity, and the latest features for AI-assisted learning. I hope you enjoy the read!

—Alexander

GPT-5 Release

OpenAI has released GPT-5! It shows improved performance across many benchmarks and, according to OpenAI, it is more transparent and hallucinates less. GPT-5 also shows major improvements in safety: instead of completely refusing to respond to possibly harmful questions, GPT-5 is trained to provide the most helpful answer possible without breaking safety guidelines. In ChatGPT, users no longer need to manually pick which model to use. Instead, the user prompt is used to automatically determine whether to use a cheap and fast model or a more capable but expensive reasoning model.

Do AI Tools Actually Increase Development Speed?

A recent study by METR measured the impact of AI coding tools on expert developers fixing issues in real-world open-source repositories. While the developers expected that using AI tools (mainly Cursor with Claude 3.5/3.7 Sonnet) would speed them up by 24%, it actually slowed them down by nearly 20%. While this result is surprising, it should be noted that this was a relatively difficult scenario for the AI, since the developers were extremely familiar with the existing codebase but relatively inexperienced with the AI tools. Check out this blog for a more detailed discussion.

Google DeepMind's Genie 3

Google DeepMind has unveiled Genie 3, its newest “world model”, which converts a text prompt into a navigable, interactive environment rendered in real time. The model can maintain visual consistency for several minutes, even for objects that move out of view. Such realistic simulated environments could be used for training AI agents or robots, allowing them to interact with a wide variety of environments and tasks.

Open-Source GPT Models from OpenAI

For the first time since GPT-2, OpenAI has released the weights of two large language models: GPT-oss-20B and GPT-oss-120B. According to OpenAI, their performance is comparable to the proprietary o3-mini and o4-mini models, respectively. Both models utilize a Mixture-of-Experts (MoE) architecture, which means only a fraction of the parameters are active during inference. They also support chain-of-thought reasoning and tool use. While these models might not show groundbreaking performance, it is great to see OpenAI contribute to open-source again.

Supporting Learning with Large Language Models

Both Gemini and ChatGPT now have a dedicated mode to support learning with LLMs. Gemini’s Guided Learning and ChatGPT’s Study Mode are designed to help users through a problem step-by-step instead of directly providing an answer. This approach aims to engage users actively in order to foster learning. Based on my initial testing, their approaches vary slightly: ChatGPT’s Study Mode often provides clear step-by-step explanations to a question, while Gemini’s Guided Learning tends to ask the user guiding questions rather than immediately offering solutions.

Datakami news

Judith and Yorick were in Zurich to visit startups in the area and attend NixCon 2025. Some highlights:

• Yorick merged a whole bunch of Nix-related Pull Requests
• Judith experienced the Night Jet sleeper train for the first time
• Lunch dates with some of our favorite startups in Zurich
• Meeting friends old and new at the VIP drinks and the "hallway track" next to Lake Zurich