A 27B model that runs on a phone is not something I expected to see in 2026
Bonsai 27B ships GPT-3.5-class intelligence on mobile hardware, needing no cloud, API keys and no data leaving the device.
I tested local models on a laptop last month and the gap between promising and actually useful was considerable. Will phone-sized models be any different?
@TikTokSupport I canโt follow anyone. My account is 6 days old and I havenโt been able to follow anyone. Have tried restarting, clearing cache etc but nothing has resolved it
Taking the CB650R up the A9 to Cullen this weekend.
This is my third annual run, with a couple of mates from the veterans' group.
The scenery through the Scottish Highlands genuinely clears the head and really is great for the mental health.
We've got ties to the community centre in Cullen, where we are staying, and the Cullen Skink is hard to turn down.
It's a lot of riding for a bowl of fish soup and a few beers.
Worth it though ๐
Anthropic now requires government ID and a selfie to access certain Claude features. Third-party verification through Persona.
The safety case is straightforward. Prevent abuse, meet compliance obligations and so on.
What's less straightforward is where this leads. ID checks for AI tools won't stay contained to one company. Every provider gets measured against the new bar.
Every API integration you build today sits on a platform that might ask for your passport tomorrow.
That โlocal models are good nowโ post is getting shared everywhere, so I tested Qwen 7B on a real task: building a SQL formatter for my Astro site.
It understood the framework, reused my CSS variables and followed the component structure.
The code didnโt run.
Deprecated API. Typos in function names.
Local models have crossed a threshold: theyโre now good enough to produce code youโll spend ten minutes reviewing before realising itโs broken.
Progress, technically.
Microsoft just had to pull 73 of their own github repos!
Someone injected password-stealing malware into them, specifically targeting AI coding tools like claude code, VS code, cursor etc.
I clone repos all the time and rarely check what's actually in them.
If Microsoft can't keep their own repos clean, I probably shouldn't assume that I can spot a compromised one.
Next task for today: Revoke my tokens and generated fresh ones.
the "open source AI must win" manifesto is doing the rounds. 600+ points on HN.
I'm not an open source evangelist but after this week, having alternatives to proprietary models feels like a sensible choice.
I think it's time to keep at least one workflow that runs without a cloud API, just so i have options.
The US government just forced Anthropic to pull Fable 5 and Mythos 5.
Not sure what that means long term but if your entire workflow depends on one vendor's API key, this is a wake-up call.
Today's meant to be wet and windy.
Genuinely not bothered. Last October I rode 300 miles through Storm Amy to the National Memorial Arboretum for Ride to the Wall which kind of set the bar for bad riding weather
After that, a bit of Scottish drizzle doesn't register. (Famous last words ๐)
AI isn't making my skills less valuable.
But to stay ahead of the curve I have to adapt.
So here's my approach:
Let AI handle speed: mockups, scaffolding, rubber-ducking problems.
Keep control of the stuff that matters: business logic, critical paths and decisions.
I'm not trying to out-code the model. Just staying one layer above it.
I've been playing Texas Chainsaw Massacre over Christmas. Have uploaded some gameplay of me getting schooled by the family ๐
https://t.co/ayQfC0I6Cw
@WindyWilson88 A waterspout tornado is one over water. One over land is just called a tornado. That was over land as far as I can tell, given that the sea is about 10 miles east of it