Practical lessons on speeding up AI agents: inference speed (TTFT vs tok/s), tool call traces, building the tooling agents want, and hardware and sandboxing.

Memory caps, run time limits, sleep inhibitors, better cron and path units - systemd features that turn out to be handy for running throwaway agent-written code.

With chip stocks on their longest losing streak since 2008 and Anthropic heading for IPO, five ways the AI boom could end - and why I think the value is real but may not flow to the people putting the money in.

Agents have collapsed the cost of building native apps twice. But I think they will also make formal verification the default way software gets written.

OpenAI has cut Luna prices 90% in two months, DeepSeek is still king on cached input, and even Anthropic is now cutting Opus prices. What that means for the frontier labs.

The recent agent sandbox escapes at OpenAI and Anthropic look less like a technical failure and more like a philosophical one - treating security controls as though they only need to work most of the time.

A simple technique for keeping up with a codebase that coding agents are changing faster than you can read: ask the agent to quiz you on it.

Z.AI ran a full model release on domestically manufactured Chinese chips. Impressive, but EUV lithography is a hard wall and the gap with Western inference hardware is more likely to widen than close.

Winter 2025 was the tipping point for coding agents. Summer 2026 is shaping up to be the same for open weights - and compute supply, not model quality, is now setting the price.

A practical framework for reducing AI costs - auditing token spend, killing legacy models, moving workloads to open weights providers, and fixing the agent and tool inefficiencies that quietly burn tokens.

For long agentic sessions cache reads are most of the bill, not input or output tokens. KV caches have shrunk dramatically while cache read pricing has barely moved.

For the first time I'm choosing my daily driver models on tokens per second rather than raw intelligence. Why ~100tok/s might be the new 100ms, where the speed gains stop mattering, and the price war that's coming.

Breaking down the Hugging Face security incident caused by OpenAI's own models during a benchmark run - the sandbox escape, the package proxy, and whether it's really a marketing stunt.

Part two: as AI inference margins collapse, who actually captures the value? The hardware supply chain and consumers win, model inference commoditises, and the frontier labs' escape routes are managed agents and staying ahead.

GLM 5.2 is the first open weights model I'd call a genuine competitor to Opus and GPT for agentic work - at ~15-20% of the price. Part one of why AI inference margins are about to collapse.

Profiling a MoE model to find which experts matter for a specific task, then quantising the cold ones hard. The result: near-Q4 quality at near-Q2 size for local models.

How KV cache compression - from MQA and GQA to MLA and linear-attention hybrids - quietly unlocked the long context windows that make modern agentic LLMs possible.

xAI is renting huge amounts of GPU capacity to Anthropic and Google. Financial engineering ahead of the SpaceX IPO, a real compute shortage, or a genuine datacentre advantage? Probably all three.

The UK is obsessed with building AI datacentres at home. But the arguments for sovereignty - latency, tax, control - mostly don't hold up.

I joined Radical's Built for Turbulence podcast to talk about what AI agents are doing to the economics of software, the Figma Trap, and why running human-written code without AI audit is going to start looking reckless.

Google's Gemini 3.5 Flash was the headline model at I/O - fast, but expensive and middling at coding. Why it makes more sense as a model built for Google itself, the TPU advantage, and Google's real weakness in coding agents.

Managed agents (cloud-hosted agent harnesses) are powerful, but locking yourself into a frontier lab's platform now is risky - here's why and what to do instead.

Open weights models keep frontier labs honest on price. If they disappear, we end up with a handful of oligopolists extracting consumer surplus.

A fictional scenario about what AI changes for cloud security, written because the technical version of the argument doesn't land with anyone except engineers.

Figma's reliance on non-designer seats made it uniquely exposed to AI. Claude Design's launch deepens the problem.

I built a small tool to visualise how Mixture of Experts models route tokens through different experts. It's genuinely fascinating to watch.

What Anthropic's Mythos research preview tells us about the trajectory of frontier models, sandbox escapes, and the cybersecurity risk ahead.

AI compute demand is growing exponentially while supply constraints bite hard. The next 18-24 months are going to be defined by shortages, rationing and price discovery.

A cascading wave of supply chain attacks has hit npm and PyPI in under two weeks. LLMs are making it worse, and current mitigations aren't enough.

How I used Claude Code to fix Linux desktop issues, get 'garbage'-rated Windows apps working in Wine, and what it means for software ecosystems