Mary Branscombe

trivially easy to verify one proof; how about a few 100? now do the pipeline for getting those from verified to part of the discipline of mathematics, that's the bit we need to work on (in every sphere, not just maths)

Mathematics in the age of AI arxiv.org

now that Microsoft has pivoted MAIA to accelerate its MAI models rather than the OpenAI ones, OpenAI had to design its own custom accelerator

openai.com

it occurs to me that all our debates about whether LLMs 'think' or 'reason' are really a proxy for how we come to terms with having something other than us in the world that 'writes' and 'uses language' but we end up dragging along a lot of (IMO irrelevant) baggage about volition and consciousness

kind of vile Microsoft feels forced to add Grok as a model option to Office (been in Copilot Studio for a while); I just assume it's because the manchild will have a hissy fit if they don't but Satya is out there on X saying model diversity. admins should set policies to block Grok in their orgs

so delightful to have a publisher claim *85*% share of a co-written book in the Anthropic claim and to have to go waste time looking for a contract from over a decade ago; a publisher who manages to get ALL the ongoing ebook royalties by dunning *us* for unpaid copies

this will make @ironyuppie.com laugh: I'm reading an IBM blog about how AI fails without real time data and only they can fix it and there is NO DATE on the blog post (but it does link to what it calls a recent study that's from 2022, so there's that)

Resisting the urge to reply to a GitHub comment where someone has clearly pasted the thread into ChatGPT and asked it to parse it and then regurgitated that into the thread complete with the usual “gotcha” telltales to ask if they can’t just get ChatGPT to write the code they want instead

also can anyone remember where to find the huge long list of things reinforcement learning systems get horrifically, terrifyingly wrong that has the video clips of one trying to learn to walk by banging its head on the floor to push itself along? because we're rediscovering how hard reward design is

one interesting thing about this (as you'd expect) thorough and step by step analysis of how 'useful' AI is in mathematics is that it's identifying the same bottleneck in consumption of AI results that may or may not be useful and correct as we see in coding; approval, integration, into 'production'

I have fought SQL and won, and fought Copilot and won at the same time! I wanted to make a virtual collection for URLs in my clipboard manager (the vintage but fabulous clipmate) and virtual collections are written in SQL and I don't SQL but Copilot does, right? only it's DBISAM not SQL

wait is OpenAI on the verge of being in lawsuits with both Microsoft (sued by) AND Apple (suing) as well as being in court with Musk right now? huh

another record month for Patch Tuesday that's bigger than the last two (record) months and if you look at the age of the products affected, it's not new AI-coded bugs, it's almost certainly AI tools helping researchers find bugs coded long ago second order effects: AI means orgs need to patch fast