Thoughts on AI, technology, and the future we're building.
New posts every week
Thoughts on AI, technology, and the future we're building.
New posts every week
Archive
Browse every article, guide, product review, and analysis piece in one place.
A plain-English guide to Perplexity AI: what the answer engine does, how Search, Deep Research and Perplexity Computer work, how it compares with Google and ChatGPT, and where its cited answers and agentic workflows still need checking.
A practical, source-backed comparison of OpenAI Codex and Anthropic Claude Code, covering local CLI versus cloud-worker architecture, workflow feel, cost/token efficiency and Codex OAuth advantages for OpenClaw and Hermes.
Anthropic has launched Claude Opus 4.8 with stronger coding and agentic workflow claims, unchanged standard API pricing, cheaper fast mode and new controls for how much effort Claude applies to a task.
Claude Mythos Preview is Anthropic's restricted frontier model for AI-assisted cybersecurity. This explainer breaks down what it is, how Project Glasswing works, what the reported data shows and why the bottleneck is shifting from finding vulnerabilities to fixing them.
Anthropic says Project Glasswing and Claude Mythos Preview have surfaced more than 10,000 high- or critical-severity vulnerabilities across critical software. The signal is serious, but the bottleneck is shifting from finding bugs to validating, disclosing and patching them safely.
OpenAI's ChatGPT personal finance experience is a U.S. Pro preview that connects financial accounts through Plaid, creates a money dashboard and lets users ask ChatGPT questions grounded in their own financial context. The useful part is personalised insight. The hard part is trust, privacy and knowing where AI stops short of professional advice.
Andrej Karpathy says he has joined Anthropic, giving the Claude maker one of the field's best-known AI researchers at a moment when frontier labs are fighting for pre-training talent, research judgement and ways to make model development more efficient.
Google Gemini Omni is a new multimodal model family that starts with video: it can take text, images, audio and video as input, then create or edit video through natural language. The first release, Gemini Omni Flash, is rolling out through Gemini, Google Flow and YouTube Shorts, with API access planned later.
DeepSeek is reportedly hiring a Beijing-based Code Harness team, pointing to a possible Claude Code and Codex rival. The product has not launched, and key details remain unconfirmed.
Google used I/O 2026 to push Gemini beyond a chatbot and into a cross-product agent layer, led by Gemini 3.5 Flash, Gemini Omni Flash, Gemini Spark, Antigravity and Gemini-powered product launches.
Vibe coding is a style of AI-assisted software development where you describe what you want in plain language and an AI tool writes or edits the code. It is useful for prototypes, MVPs and learning, but generated code still needs review, testing and security checks.
An AI agent harness is the runtime and control layer around an AI agent. It connects the agent to tools, context, memory, permissions, workflows, logging, evaluation and human approval so the agent can complete tasks safely and reliably.
Tokenmaxxing is the habit of pushing AI token usage as high as possible, often through long prompts, coding agents, parallel workflows and internal usage leaderboards. Used well, it can encourage serious AI experimentation. Used badly, it turns productivity into an expensive token bonfire.
A transformer model is an attention-based neural network architecture behind many modern LLMs. It helps models weigh relationships between tokens in context, which is why GPT-style systems, BERT-style systems and other AI tools can handle prompts, examples and language patterns so flexibly.
A beginner-friendly explainer of AI model training, datasets, training examples, pattern learning, and why model quality depends so heavily on data quality.
RAG connects an AI model to external documents, databases, files, or knowledge bases so it can retrieve relevant context before generating a more useful answer.
A plain-English guide to AI agents, including how agentic AI can plan steps, use tools, observe results, complete tasks, and differ from a chatbot that mainly responds in conversation.
A beginner-friendly explainer of AI prompts as the instructions, context, examples, constraints, and output format given to tools like ChatGPT, Claude, Gemini, and other AI assistants.
OpenAI has brought Codex remote access to the ChatGPT mobile app in preview, giving users a mobile way to start threads, approve actions, review outputs, and supervise coding-agent work from iOS and Android while execution continues on the connected host.
A clear practical guide to the hierarchy of AI, machine learning, deep learning, and generative AI, with examples, comparisons, limitations, and common misconceptions.
A beginner-friendly explainer of reinforcement learning from human feedback, how reward models work, and why human preference data can improve AI model behaviour.
A beginner-friendly explainer of AI evals, test cases, grading methods, and the pre-deployment measurement habits teams use to make AI systems more reliable.
A beginner-friendly explainer of how AI image generators turn prompts into images, how diffusion-style generation and editing work, and where text-to-image AI is useful or limited.
GPT stands for Generative Pre-Trained Transformer: a transformer-based language model style that is broadly trained before use and generates text from context. GPT is usually an LLM, but it is not the same as all AI, all generative AI, or every transformer model.
A practical guide to open source AI versus closed AI models, including open weights, proprietary models, transparency, cost, safety, and business control.
AI inference is the process of using a trained model to generate answers, predictions, labels, images, recommendations or other outputs from new input.
A beginner-friendly explainer of machine learning as pattern learning from data, including how ML differs from fixed-rule software, how models train on examples, common use cases, main types, benefits and limitations.
A plain-English guide to multimodal AI, including how models process text, images, audio and video together, where multimodal systems are useful, and why high-stakes outputs still need human review.
A context window is the active token budget an AI model can use for a request or turn. It explains why ChatGPT can seem to forget details in long conversations and why large documents need careful context management.
A vector database stores embeddings as searchable numerical representations so AI apps can retrieve related documents, products, images, or knowledge base chunks by similarity rather than exact keywords.
Temperature in AI controls how much randomness a model uses when generating responses. It affects why ChatGPT outputs vary, how creative or consistent answers feel, and how much risk a workflow accepts.
Prompt engineering is the practical skill of writing clearer AI instructions. This beginner explainer defines the concept, shows examples, explains best practices, and flags common mistakes that lead to weak or risky AI responses.
A beginner-friendly explainer of MCP as the open standard that connects AI assistants to tools, files, data sources, systems and workflows, with practical examples and security caveats.
AI tokens are the small units of text or data that models process. They matter because token counts shape AI pricing, context windows, memory-like working context, output length, latency, and usage limits.
A beginner-friendly explainer of AI grounding, source-backed answers, RAG, citations, and the practical habits that help reduce unsupported AI claims.
A clear beginner-friendly explainer of tool calling in AI, including how models request functions, APIs, apps and external data, plus practical examples and safety guidance.
A plain-English explainer of AI bias, how training data can create unfair outcomes, and why biased AI systems matter for business, search, hiring, finance, and public services.
A beginner-friendly comparison of AI chatbots and AI agents, including reactive chat, goal-oriented planning, tool use, autonomy, examples, risks, and when to choose each one.
A practical explainer of AI hallucinations, why generative AI can produce false or unsupported answers, and how to reduce risk with grounding, citations, verification, and human review.
A plain-English explainer of foundation models as large reusable AI base models, covering how they work, how they are adapted, where they appear in real products, and why their strengths and weaknesses travel downstream.
AI model parameters are learned internal values, often weights and biases. This explainer separates model size from user settings like temperature, top_p and max output tokens.
Generative AI is artificial intelligence that creates new content, including text, images, audio, video and code. It differs from traditional AI because traditional AI usually predicts, classifies, detects, ranks or recommends, while generative AI produces new outputs from learned patterns and prompts.
A beginner-friendly explainer of responsible AI principles, governance, transparency, bias, fairness, and the practical controls teams use to manage AI risk.
A plain-English explainer of large language models, including training data, token prediction, how LLM-powered products work, common use cases, benefits, limitations and practical review habits.
A beginner-friendly explainer of artificial neural networks as layered pattern-learning models, including how neurons, layers, weights, biases, training, examples, benefits, and limitations fit together.
A beginner-friendly explainer of deep learning as machine learning with layered neural networks, including how deep learning works, why layers matter for complex data, real-world examples, benefits, limits and related AI terms.
A complete beginner guide to artificial intelligence, with a plain-English definition, everyday examples, common use cases, key terms, benefits, limitations, and FAQs.
Fine-tuning adapts an existing AI model for a narrower task or domain, but it is only worth using when prompting, RAG, tools, and evals are not enough.
