strutsafe.org/ what a great service

Odd that a press helicopter is allowed to film a bomb disposal robot from above. 💣💥

“I will put that [social care] plan to the public at a general election” PM Andy Burnham says he wants to introduce universal social care in the “next parliament” because promises in Labour's last manifesto "make it hard" before then #BBCLauraK bbc.co.uk/news/live/cj4zwkp2…

codex reverse engineered m4 mesa driver: codyho.dev/blog/gpu-driver/

Stonehenge Autumn Equinox Celebrations.

Every Dutch train, one ordinary Tuesday, from first departure to last.

Oh boy you missed the bottom? Thought you nailed all tops and bottoms ? 🤣

If HMRC can't or won't challenge, or if a challenge fails, the government should close the loophole and change the law. Otherwise, we're going to see lots of high-end properties being sold with five teeny flats.

Pi beats claude code and codex You can cut your agent API costs by 50% right now with literally 0% drop in benchmark performance. Anthropic wants you to use Claude Code. OpenAI wants you to use Codex CLI. But a new UC Berkeley study just exposed the "In House Myth": Frontier models actually perform better on minimal, open source frameworks 75% of the time. Stop paying the Big Tech "Harness Tax." Here is the data # UC Berkeley and Arena just dropped the "Harness Tax" paper. They benchmarked 21 model agent pairs on SWE-bench Lite and Terminal-Bench 2.0. They tested SOTA models (Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, Sonnet 4.6, Kimi K3) across 3 agent frameworks: Claude Code, Codex CLI, and a barebones open source harness, Pi. The results prove we are all bleeding API credits for literally zero reason. # Look at the chart: - GPT-5.6 Sol scored 83.3% using the OS Pi harness. On OpenAI's own Codex CLI? It dropped to 78.9%. and cost nearly 2x as much ($0.42 vs $0.76). - Claude Fable 5 hits 97.8% on Claude Code for $1.33/rollout. On Pi? It hits 96.7% for just $0.67. You are paying a 100% markup for a 1% gain. # Why is this happening? First Call Bloat. Proprietary coding agents load up massive system instructions, heavy guardrails, and verbose JSON tool schemas. Claude Code sends 10x the initial context of Pi. You are paying for their dead weight context on every single turn. (check out the replies for more data on this) Pi gives the model exactly 4 tools: read, write, edit, bash. That's it. The takeaway: As frontier LLMs get smarter, heavy agent scaffolding actually holds them back. If your agent API bill is too high, do NOT downgrade your model. Just ditch the bloated harness. I've dropped more data and link to the study in the replies. What agent framework are you guys running in your dev loop right now? Let me know

Cash in the Attic?

Just a reminder that GLM 5.3 Flash, DeepSeek V4.1 Flash, Qwen 3.8 Next Flash, and even Qwen 3.8 27B are all outperforming (in both intelligence and capabilities) every model that was considered "frontier intelligence" in Xmas 2025 (just 10 months ago) Opensource AI is on fire

There is a per constituency spending limit per election so Reform wouldn't be able to spend the 72M

This is how serious a fire incident can look inside an aircraft while cruising thousands of feet above the ground. Flames and smoke are seen coming from the overhead luggage compartment as passengers scream, panic spreads, and people scramble to find anything that can help extinguish the fire. The suspected cause? A carry-on lithium battery. Lithium-ion batteries can enter thermal runaway when damaged, overheated or short-circuited, releasing intense heat and potentially reigniting even after the flames appear to be out. And we recently saw how serious this can become. On September 17, KLM flight KL735 had to turn back over the Atlantic after a power bank caught fire onboard. The crew extinguished the fire, but the aircraft was damaged. At 35,000 feet, there’s no fire truck coming to the rescue. The crew has to contain the situation inside a pressurized cabin while keeping hundreds of passengers safe. A small lithium battery can create a very big problem in an aircraft. 🔥✈️

Divide the property into flats.

Jev is INSANE. I built this in an evening and morning. A real-time trading bot ingesting onchain+offchain data to make rapid decisions about trades. Fully autonomous. So far it has lost me $31,680.

pleased to share the entire arXiv site as a dataset on HuggingFace huggingface.co/datasets/sece… 3,148,796 papers, every version, in LaTeX, PDFs, PostScript, HTML, 16 TB in total