Last week OpenAI launched five new models: GPT-4.1 (including mini and nano variants), o3 and o4-mini, together with a new command-line tool called Codex CLI. You now have six models to choose from in ChatGPT: GPT-4o, GPT-4o with scheduled tasks, GPT-4.5, o3, o4-mini, and o4-mini-high. If you use the API you can add GPT-4.1, GPT-4.1-mini and GPT-4.1-nano to that list. So which one should you use? In my own experience and based on what I have read on multiple forums, GPT-4o is still the best model for everyday productivity. Ask it about documents, work with texts, get feedback, and use…

Read more…

Last week Meta launched their new Llama 4 models that scored amazingly well on benchmarks. Yann LeCun, Chief AI Scientist at Meta, wrote “BOOM! The Llama-4 brood is out”, and Ahmad Al-Dahle, Head of GenAI at Meta, wrote “Llama 4 Behemoth outperforms GPT4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro”. Boom indeed, because when users got access to the models they performed among the worst of all recent models. In my newsletter last week I wrote “Maybe they published the wrong models online, and somehow released half-baked early internal releases”. Official representatives at Meta quickly denied that they trained their…

Read more…

Last week was one of the most important AI weeks this year for software developers. Microsoft rolled out their all-new Visual Studio Code and GitHub Copilot Pro with massive amounts of new features focused on autonomous AI-based agent coding. They even launched it with a short movie of Satya Nadella “vibe coding” a computer simulator in less than 10 minutes without writing a single line of code. The March 2025 updates of Visual Studio Code and GitHub Copilot are two of the most important updates to Visual Studio Code ever, and not only did they add an autonomous agent mode…

Read more…

Last week OpenAI updated their GPT-4o model with image creation capabilities. If you had previously used ChatGPT to create images using their model “DALL-E 3” you know it was quite a poor performer, so maybe you skipped this news due to low expectations. But the image generation capabilities of GPT-4o is completely unlike anything you have seen before. First, like with Gemini 2.0 Flash, GPT-4o now processes text and images through a unified system, not as separate tasks. This means that the model uses the same neural pathways for understanding both language and visual content, and can access its entire…

Read more…

The wildfires in Los Angeles this year are predicted to have cost over $250 billion, becoming one of the worst natural disasters in U.S. history. Current satellite systems like VIIRS and MODIS can detect fires quickly (within a few hours) but have problems identifying smaller fires (<100m²). Other systems like Sentinel-2 have details down to 10m² but with less frequent intervals (days rather than hours). Last week Google Research in partnership with Earth Fire Alliance, Muon Space, and the Gordon & Betty Moore Foundation, launched the first FireSat satellite, designed to detect wildfires as small as 5×5 meters in less…

Read more…

Will you use Google Search in 2026 the same way you have been using it over the past two decades? I am quite sure the answer is no. I am a firm believer in agentic search, and I believe it will change everything about how we store and publish information online. Last week Andrej Karpathy wrote on X: “It’s 2025 and most content is still written for humans instead of LLMs. 99.9% of attention is about to be LLM attention, not human attention”. This should be your focus going forward with everything you create. When you design a web page,…

Read more…

Most companies I speak with today are rolling out generative AI in two main areas: customer support and digitizing documents for AI access. When it comes to customer support we have finally reached the point where AI agents can now handle most tasks fully autonomously. If you have a good pipeline setup by someone who knows what they are doing the agent will properly escalate and involve humans when needed. For making PDFs and paper documents structured and accessible, the number of affordable solutions have been limited. Traditional PDF-to-markdown tools are effective but expensive and complex. Last week’s launch of Mistral OCR changed…

Read more…

Last week OpenAI finally launched their successor of GPT-4o: GPT-4.5, and the entire Internet quickly became full of (mostly) disappointed old guys claiming AI-development has stalled. GPT-4.5 does not significantly improve performance in test benchmarks, and since it was initially only available for $200 /month Pro subscribers, very few could actually test it. I seem to be one of the few people that appreciate what OpenAI did with GPT-4.5. Instead of focusing strongly on a few select artificial benchmarks, they took the strengths of GPT-4o and made it better. GPT-4.5 is more creative, more nuanced, and more knowledgeable than GPT-4o,…

Read more…

OpenAI just crossed 400 million weekly active users, which is close to 5% of the world’s population. This means that they are growing at an astounding 50 million weekly active users per month, just 2 months ago OpenAI announced that they have 300 million weekly active users. The growth is accelerating, and OpenAI predicts they will reach 1 billion weekly active users by the end of the year. I find myself using ChatGPT, Claude and Perplexity more and more myself as the weeks go by, and as the tools themselves get better and better they really start to feel like…

Read more…

In the coming few weeks we should see a massive influx of new AI models being released. ChatGPT 4.5 is due in a few weeks, Grok 3 is to be released today (!) and is described as “scary smart” by Elon Musk, and Claude 4 is on it’s way too. These models are trained using next-generation hardware and should all show a substantial leap in performance, in particular related to coding and agentic workflows. For those less familiar with AI hardware, this next part might get a bit technical, but it’s worth understanding as it signals a major shift in…

Read more…