takes on robotics, ML, and AI — with the math and diagrams that generic summaries usually skip.
The EU Council and Parliament struck a deal to streamline the AI Act, and a machine proved an 80-year-old combinatorics conjecture that stumped human mathematicians. A look at what both moments mean.
readA non-expert prompted an AI, got an answer, checked it with a formal verifier, and it was correct. OpenAI also shut down Sora. Two very different stories about where AI is and is not delivering.
readAnthropic published research showing Claude Opus 4.6 can independently discover and document zero-day software vulnerabilities. That changes the threat model for every organisation running software.
readOpenAI tested advertisements inside ChatGPT, retired GPT-4o on Valentine's Day, and Anthropic disclosed that its flagship model can find zero-day exploits. A busy month for the AI business.
readWithin days of xAI launching Grok image generation, three Asian countries blocked the service over nonconsensual deepfakes. What happened, how platforms responded, and what the episode reveals about AI governance.
readFrom DeepSeek upending cost assumptions in January to GPT-5 dividing opinion in August to Mission Genesis in November, 2025 was the year the AI story became impossible to ignore.
readTrump signed Mission Genesis in November 2025, an executive order being compared to the Manhattan Project. Here is what it actually funds, how it fits into the US-China AI race, and what it means for research.
readThe shift from text-only to multimodal AI is not just a feature upgrade. It changes what these systems can be used for, where they fail, and how developers need to think about prompt design.
readAfter a year of coding agents, autonomous pipelines, and agentic workflows going mainstream, the honest picture is more complicated than the demos. A look at where agents deliver and where they reliably fall apart.
readOpenAI claimed GPT-5 offers PhD-level expertise. Reviewers found it struggled with basic spelling and geography. Meanwhile xAI launched Grok Imagine. A complicated month for AI expectations.
readVideo generators, text factories, and synthetic media tools are producing content at a scale humans never could. The question is not whether this is happening but what it does to information quality and creator economics.
readAmazon announced it was testing humanoid robots for package delivery. Agility, Figure, and 1X are already in facilities. The lab-to-warehouse transition is real, and the engineering challenges are far harder than the demos suggest.
readGoogle launched AI Mode for search, unveiled Veo 3 for video generation, and Anthropic shipped Claude 4. May 2025 was the month the AI race stopped being theoretical and started rewriting tools people actually use every day.
readMeta shipped Llama 4 and the conversation about open versus closed source AI shifted. Free, downloadable, fine-tunable models are now competitive with frontier labs. That has real consequences for how AI gets deployed.
readExtended thinking in Claude 3.7 Sonnet and the OpenAI o-series models represent a genuine architectural shift. Instead of producing an answer directly, these models spend compute on reasoning steps first. Here is what that buys you.
readAndrej Karpathy coined vibe coding on February 2nd. Three days later OpenAI launched Deep Research. A week after that, world leaders met in Paris to argue about AI policy and the US and UK refused to sign anything.
readA Chinese lab released an open-source model that matched GPT-4 performance at a fraction of the training cost. Nvidia lost over $500 billion in market cap in a single day. The assumptions underlying the AI boom had to be revised.
readAutonomous robots are moving beyond hard-coded scripts toward agents that perceive, reason, and adapt. Explore how reinforcement learning, collective intelligence, and real-world safety constraints are reshaping the field.
readTraining robot fleets without centralising sensitive data is no longer a research curiosity - it is an engineering requirement. A deep dive into gradient sharing, heterogeneous data, and secure aggregation across physical agents.
readBenchmark accuracy does not survive contact with the real world. This post examines the reality gap, 6D pose estimation, motion prediction, and the sim-to-real strategies that close the distance.
readFrom the keyboard completing your sentence to the grid balancing the city's electricity, machine learning is quietly embedded in the texture of daily life - mostly invisible, often indispensable.
read