It's just gone 5am here, and with just under 19 hours to go in the #24hrlemans it's time to put some flour, yeast, salt, and water to prove so after I go for a nap I can put a loaf in the oven for tomorrow!

Been struggling with the amount of contact form #spam a couple of sites I look after has been getting even with CAPTCHA solutions (the non-profit I work with has been getting maybe 30-40 a week, which is tough as the recipients aren't hugely computer literate so can get flustered). Switched to IPQS.com for IP and #email validation and seeing a big drop-off in crap (and no reported issues from legit contacts). Referral link if you want to help support the non-profit https://ipqualityscore.com/?r=869497

Doing some #ai took testing today and I have watched both #Qwen3.6 and #Gemma4 get into a loop while trying to hallucinate the code needed to solve the problem I was using to compare them. I wonder how many #tokens both burnt through by not being able to recognize they were in a loop and self correct. Luckily both local so no real costs, just time wasted and negative scores for both. LLMs still need to address a lot of the fundamental architecture issues and optimisations to be reliable tools

So many #ai #llm scenarios seem to rely on a single massive model in the cloud or a hamstrung local model. People work best when they collaborate, so why don't LLMs? For instance, when coding use a local model to build a knowledge graph and handle appropriate tasks but if it struggles or has a problem completing it only then offload to a cloud model. Would help keep the budget down and encourage local models to get smarter vs spending billions on vanity data centers with sketchy ROIs

Is there any reason for the continued push in the #AI #LLM world purely for bigger models requiring more compute/RAM? At this point the smart decision would seem to be an optimisation phase - go for a 2x, or even 10x improvement on how it can run on smaller, more generalised hardware. With the economics of #tokenmaxxing suddenly hitting home the way to get ahead of the bubble would be to take a current frontier model and optimise that to make it run on consumer hardware then build from there

I don't think I've made #bread since I was at school. Remarkably simple process (flour, yeast, salt, water) and then patience while it rises, cooks, and cools!

Any #Azure, specifically #AzureSQL product peeps on here? I would *really* like a realtime layer on top of AzureSQL without having to roll my own with SQL Change Feed/Functions/SignalR etc. I want Supabase simplicity without having to spin up a whole bunch of stuff.

I know that #AI is a constantly evolving feast of buzzwords, but are there any recommendations for actual useful products and use cases to help a small business provide better service to their customers (I stress not by replacing humans but augmenting them so people can do less drudge and more fulfilling and creative things). Even better if they are cost effective and have good tutorials!