Patrick Socha @psocha.co.uk · 17hIncidents I can believe. But it’s then really down to the approval process. I’ve been rejecting PRs without videos/screenshots of it working, and especially without tests. Good, process and CI will fix this. Code fast, review slow, test aggressively. 000
Reposted by Patrick SochaTechmeme @techmeme.com · 29/09/2026FOI docs: UK police's six-month live facial recognition trial in London railway stations scanned 500K+ faces, leading to no arrests and one false-positive alert (Daniel Boffey/The Guardian) Main Link | Techmeme Permalink 0118
Patrick Socha @psocha.co.uk · 28/09/2026To be fair, eat the rich and their two double bedroom mansions. /s www.rightmove.co.uk/properties/9...rightmove.co.ukCheck out this 2 bedroom terraced house for sale on Rightmove2 bedroom terraced house for sale in Houblon Road, Richmond, TW10 for £1,000,000. Marketed by Savills, Richmond 000
Patrick Socha @psocha.co.uk · 28/09/2026I’ll be honest, sometimes it really do produce hot garbage. Mostly not, but sometimes, just wow. 000
Patrick Socha @psocha.co.uk · 28/09/2026Unrelated to the car, but that 90 degree turn to as a cyclist avoid going on to the pavement is criminal 😂 130
Patrick Socha @psocha.co.uk · 26/09/2026Today I found out that @github.com doesn't support IPv6 on GHCR.static.klipy.comCersei Lannister's Walk of ShameALT: Cersei Lannister's Walk of Shame 120
Reposted by Patrick SochaMark Riedl @markriedl.bsky.social · 25/09/2026Someone asked Muse to dump its entire virtual machine file system, and it did 35417
Reposted by Patrick SochaDave Vetter @davidrvetter.bsky.social · 24/09/2026Happy Kicked In The Head By A Horse Day, to all who celebrate. It's the 7th anniversary of this absolute banger. 442225725
Reposted by Patrick SochaAlex Selby-Boothroyd @alexselbyb.bsky.social · 23/09/2026New belt-tightening indicator just dropped 34511
Patrick Socha @psocha.co.uk · 23/09/2026Spent last night setting up a new server within a vpc for logging and redis down the line. The server is fully off the internet, only accessible via wireguard, and directly to other servers via the vpc. Did I ever say how much I love wireguard? 000
Patrick Socha @psocha.co.uk · 22/09/2026Wrote a mini log ingestion tool. A docker app which makes use of SQLite. Search, streaming, auto purge. Put it in a VPC, happy days. Perfect for when you have a few apps whose logs want to keep on top of. Going to try open source it in coming days.h 000
Patrick Socha @psocha.co.uk · 22/09/2026Hitting 60 tokens/sec with qwen3.8:27b-mlx with MTP and MLX on an M4 MBP 64gb. Crazy stuff. 000
Patrick Socha @psocha.co.uk · 21/09/2026To everyone complaining about needing to buy Meta One to post a link: own your distribution through your own mailing list and RSS. Stop supporting walled gardens. 010
Reposted by Patrick SochaHacker News Bot @hnfeed.bsky.social · 21/09/2026Kev: Tiny Jev-like family of decision models built on top of Qwen3.5 github.com/jaredpalmer/kev/tree/main [comments] [81 points]github.comGitHub - jaredpalmer/kev: tiny Jev-like family of decision models built on top of Qwen3.5 you can train and run on your owntiny Jev-like family of decision models built on top of Qwen3.5 you can train and run on your own - jaredpalmer/kev 001
Patrick Socha @psocha.co.uk · 21/09/20263.8 is pretty good if you’re able to run it. ~on par with sonnet I’d say. 000
Patrick Socha @psocha.co.uk · 20/09/2026I guarantee those companies will have a shit time down the line. bsky.app/profile/psoc... 010
Reposted by Patrick SochaUnsloth AI @unsloth.ai · 17/09/2026You can now train and run 500+ models locally with our Unsloth Docker image! 🐳 Use our new GUI or notebooks workflow. No setup required. Works on NVIDIA and AMD. Guide: unsloth.ai/docs/get-sta... GitHub: github.com/unslothai/un... 1221
Reposted by Patrick SochaKarl Bode @karlbode.com · 17/09/2026apparently we are calling incompetently failing to control your AUTOMATED HACKING SOFTWARE "model misalignment"nowaxios.comOpenAI discloses six new safety incidentsThe AI lab is now rolling out a new process to internally report safety issues. 1014043
Patrick Socha @psocha.co.uk · 17/09/2026If you're running an MLX (qwen3.8) model on Ollama, update to a version after v0.32.14 which _should_ fix a KV cache memory release issue. You'll save ram and hopefully stop swapping. 100
Patrick Socha @psocha.co.uk · 16/09/2026Don't worry too much about AI taking your job any time soon. Chats from the last 24h. 000
Patrick Socha @psocha.co.uk · 14/09/2026Every European to get 1L of maple syrup for life and it’s a deal? www.cbc.ca/news/politic...cbc.caCarney says Canada planning to discuss 'unique alliance' with EU, but not membership | CBC NewsPrime Minister Mark Carney says Canada plans to begin discussions for a "unique alliance" with the European Union — but not membership — as he gears up to address the European Parliament this week. 000
Reposted by Patrick SochaHacker News Bot @hnfeed.bsky.social · 11/09/2026Detecting and countering misuse of AI: September 2026 www.anthropic.com/threat-intelligen… [comments] [143 points]anthropic.comCountering misuse of AI: September 2026 / AnthropicCase studies from threat actors disrupted between December 2025 and August 2026 across seven areas of harm, from cyber operations to biological misuse. 012
Reposted by Patrick Sochaaustin (e/accordion 🪗) @thebadcode.com · 08/09/2026pov: you just claimed that llms are capable of solving certain real world problems 715211
Patrick Socha @psocha.co.uk · 08/09/2026It's beautiful 🥲 Everyone blocking me for talking about LLMs. 001
Patrick Socha @psocha.co.uk · 07/09/2026Oh man you’re about to lose some days to playing around with local models. Wrote my own harness for qwen3.8. Works amazing. 110
Patrick Socha @psocha.co.uk · 07/09/2026Uses a smaller model to predict multiple tokens then passes them in one pass for the bigger model to agree/disagree. If agree, you get a speed up of tokens from the Multi Token Prediction. 000
Patrick Socha @psocha.co.uk · 07/09/2026As a reviewer: actively reject large PRs. I’m not going to be able to reliably review 5k LOC. 000
Reposted by Patrick SochaHacker News Bot @hnfeed.bsky.social · 04/09/2026Google AI Mode shows same products 21.6% more expensive than traditional search productrise.app/blog/google-ai-mode… [comments] [287 points]productrise.appGoogle AI Mode shows the same products 21.6% more expensive than traditional search [Data Study] | ProductriseA US and UK data study: when the same product ranks in both Google AI Mode and traditional search, the AI Mode price is about 21.6% higher. 001
Patrick Socha @psocha.co.uk · 04/09/2026Not necessarily true. If you only posted on bsky, there’s a huge group of anti-ai users. Doesn’t reflect the wider engineering community. From my experience, everyone is using it. 100
Patrick Socha @psocha.co.uk · 03/09/2026Start of the month, was updating personal finance spreadsheet. Realised I could get an LLM to write me a single html file which writes values to a CSV and displays everything exactly how I want it. Super simple. Way clearer at a glance. 000
Patrick Socha @psocha.co.uk · 03/09/2026Or make your password “incorrect”. That way when you type it in wrong, every website reminds you that your password is “incorrect” 👍 010
Patrick Socha @psocha.co.uk · 02/09/2026I believe its in relation to everyone shitting on the paint.net guy who said he got his application working on linux with the help of LLMs. Bsky went ape and he quit the platform for making a genuinely useful free tool available to even more people. 381
Patrick Socha @psocha.co.uk · 02/09/20261,000% agree. Private equity DD now often includes building out the product using AI as a way to test the defensive moat. Anyone saying they won’t use any product that uses it is delusional and far removed from the industry. 010
Patrick Socha @psocha.co.uk · 02/09/2026I kinda get that. But also it’s fun to get notifications about getting blocked by people who hate ai but still use a website that’s hosted in a datacenter. 030
Patrick Socha @psocha.co.uk · 01/09/2026Absolutely. It's a simple loop, written in Go, uses bubbletea and honestly it's so simple, but it works great. I was using Qwen with Claude Code, but tool usage was failing with updates. I wrote this super simple harness, and it's just fantastic + stable. 47tok/s on 64gb mbp m4. Crazy. 110
Patrick Socha @psocha.co.uk · 01/09/2026Absolutely. I've written my own harness (container + has read, find, bash). Finding it does pretty much everything I need with 3.8:27b. Agents are great, but I'm not fully sold/trust them as a SWE just yet. 110
Patrick Socha @psocha.co.uk · 01/09/2026Absolutely. That's sort of why I'm thinking the mac mini isn't awful. Good hardware, not the best, but good and great power consumption. 110
Patrick Socha @psocha.co.uk · 01/09/2026Yeah, but that's waiting for two years when I want local models yesterday 😂 110