Thoughts on engineering, automation, AI agents, and building systems that last.
The claim that the Pentagon used Grok AI to fire 2,000 missiles at Iran misses what the reporting actually says. The disclosure emerged in a Justice Department filing tied to an xAI pollution case, and the documented figure refers to more than 2,000 munitions against 2,000 targets in 96 hours, not autonomous missile launches by Grok.
Read more →
Meta has struck a three-part nuclear power package with Vistra, TerraPower and Oklo to support its AI data centres, including the Prometheus site in Ohio. The deal reaches up to 6.6 gigawatts by 2035, but the much-quoted ‘up to eight’ figure refers to TerraPower options rather than eight guaranteed new reactors.
Read more →
Anthropic is making two arguments at once. It says Claude is already doing a striking share of the company’s coding work and speeding up AI development, but it also says that same trend is dangerous enough to justify tighter model limits and even the option to slow frontier AI work. That tension matters because it looks like safety policy, competitive defence, and product strategy all converging in public.
Read more →
Claude Fable 5 and Claude Mythos 5 are not two entirely separate models. They are Anthropic’s product split for the same underlying system, with Fable 5 aimed at broader use and Mythos 5 reserved for a small group of vetted cybersecurity and biology partners. Here is what changed, what it costs, where the limits sit, and why the fallback to Opus 4.8 matters.
Read more →
There is no separate trick for ‘AI indexing’. For Google’s AI search features, your pages need to be indexed and eligible for a normal search snippet first. That means getting the basics right: crawlability, internal links, text-based content, canonical signals, structured data that matches the page, and sensible monitoring in Search Console. For Bing, IndexNow can speed up discovery of new URLs, but it is still only a prompt, not a guarantee.
Read more →
Claude Code looks simple at first glance: a terminal assistant with sensible defaults, approval gates, and a short list of commands. The deeper story is more interesting. Official Anthropic documentation covers settings, hooks, subagents, status lines, CLI flows, and security, but close reading of the public source has revealed extra hook behaviours and configuration patterns that advanced users can exploit carefully. This article maps the documented core, explains what sits beyond it, and shows why serious users should treat source-discovered features as powerful but version-specific tools.
Read more →
Google is giving its homepage search box a serious AI overhaul, with support for text, images, files, videos and even Chrome tabs. Some of that shift is already live, including AI Mode and conversational follow-ups from AI Overviews, while other features such as search agents, business calling and custom mini-app dashboards are still rolling out on a staggered timetable.
Read more →
Most Americans do not trust AI governance, and the numbers suggest that is the real story. Pew found major concern about job loss, misinformation and weak regulation, while Gallup shows even heavy AI users in Gen Z feel anxious. The public mood is not technophobic so much as unconvinced that the people building and policing AI deserve confidence.
Read more →
Malta has announced a national AI programme with OpenAI that offers eligible citizens and residents a free online literacy course and, after completion, a year of ChatGPT Plus access. The design matters as much as the giveaway: training comes first, access is managed centrally, and officials say the initiative will scale as participation grows.
Read more →
Impeccable teaches AI design by loading a seven-part reference stack before generation starts, then exposing 23 commands that turn typography, color, motion, layout, and live browser iteration into deliberate actions. The result is a more designer-directed way to work with AI.
Read more →
Figure AI is making one of the boldest consumer robotics pitches in years: a humanoid robot designed for laundry, dishes, and cleaning inside messy, unpredictable homes. That is a much bigger ambition than a Roomba-style vacuum. The question is not whether the vision is exciting. It is whether Figure 03 and Helix are close to replacing the dependable, purpose-built cleaning robots already rolling around our floors. The evidence suggests a fascinating long-term direction, not an imminent handoff.
Read more →
Hermes Agent is Nous Research's self-improving AI agent runtime built for people who want more than a one-off chatbot. Its persistent memory, on-demand skills, MCP tool connectivity, multi-platform messaging gateway, and cron-based automation make it a practical system for real work instead of a novelty interface.
Read more →
Open Design lands as a timely answer to a changing design stack: local-first, open-source, artifact-driven, and built to work with the coding-agent CLIs people already run. Its release matters because AI design is no longer experimental, and the workflow between code and canvas is quickly becoming two-way.
Read more →
For most auto-post agents, the best temperature is a low one. Lower values produce more focused and consistent output, while higher values increase variety and creativity. The practical default for repeatable, on-brand posting is usually 0.0 to 0.2, with higher settings reserved for deliberate experimentation.
Read more →
A practical account of letting Codex update a site, discovering the change introduced a regression, and using review, branch isolation, and revert tooling to recover cleanly.
Read more →
The next phase of AI isn't about better prompts, but about agents that inhabit our environment. From Vibe Coding to Personalised Synthetic Clones, I look at the underrated breakthroughs of 2026.
Read more →
DeepSeek-R1 and Llama 3.3 have shifted community focus from claustrophobic free-tier APIs to local 'sovereign' engineering. This post explores why owning your hardware is the only way to truly get AI for free in 2026.
Read more →
Watching an agentive swarm move cards across a Kanban board is the definitive bottleneck of 2026. I explore why Hermes Agent's stability is being wasted in sequential workflows and why we need to move toward DAG-based orchestration.
Read more →
90+ security incidents in early 2026 have killed the 'cowboy' era of AI agents. I dive into why we're shifting toward the Microsoft Governance Toolkit and Claude Code.
Read more →
Images 2.0 is killing the 'plastic' look. I dive into the community's reaction to the realism leap, the search for jailbreak vectors, and why taste is replacing keywords.
Read more →
MiniMax 2.7 is hitting 86.2% on PinchBench and clearing filters GPT-4o misses. I dive into why this GPT-4 class model is the new efficiency standard for production engineering.
Read more →No posts yet. Check back soon.