God I absolutely love reading about these borrrrrring infrastructure bugs they’re so real they’re so thorny they’re so interesting I LOVE IT www.anthropic.com/engineering/...

> Although the inference server itself can be claimed to be "deterministic", the story is different for an individual user. From the perspective of an individual user, the other concurrent users are not an "input" to the system but rather a nondeterministic property of the system.

There are just under 12 hours left to vote in this year's Tiny Awards! Small, fun, beautiful and occasionally-pointless websites, and a pleasing rebuttal to anyone who thinks everything online is rubbish in 2025: tinyawards.net/vote/

What sort of black magic is this

google estimate that gemini models use 0.24 Wh and 0.26 ml of water per prompt

I'll be talking about @climatepolicyradar.bsky.social's work on building knowledge graphs for policy research at @nestauk.bsky.social's Policy Live event on 11 September

I’m on my way to Vienna for @aclmeeting.bsky.social, where I’ll be giving a keynote for the @climate-nlp.bsky.social workshop on how NLP is being used to address the climate crisis

So much of the recent AI doomerism (environmental, economic, existential, etc) has felt dumb to me... The criticisms seem reactionary, sensationalist, poorly researched, wilfully blind. Anyone with direct experience knows that the reality of the technology is much more mundane than the hype

Some thoughts on how "LLM grooming" could be used to fill information voids and thus further degrade information found on the web. Happy times! dcorney.com/thoughts/202...

Sam Altman just gave ChatGPT's cost-per-query of 0.34 watt-hours: the first time a number has been given in terms of recent LLM power usage and is obviously much lower than the 3 watts still cited by detractors, but there's a lot of asterisks in how that number might be calculated.

The most well-known statistic about AI water use is a lie. This makes it frustrating to talk about AI and the environment, and this is a long deep dive on that specific point. www.verysane.ai/p/the-bigges...

the tone's provocative, but there are loads of sensible, software-development-centric takes about LLM hype vs reality in here

I’ve had such a brilliant weekend at @pydatalondon.bsky.social. So many great presentations, so many great chats with pals old and new. Really feeling the value of the data science community which has grown and cohered here over the last decade. Massive, massive thanks to the organisers 💖

oh my GOD i love pydata events so much. so happy to be back in these spaces with these people

I wrote what turned into a small book with all my big takes on AI and the ways I'd like to see the debate improve andymasley.substack.com/p/all-the-wa...

Deepmind have been doing amaaaaazing things ✨ deepmind.google/discover/blo...

Today we launch our report on the rising threat of misinformation in the UK—featuring expert essays and a new rating system assessing policy, platforms, and progress in tackling false information online. Read it here: buff.ly/Uz6pzTJ

these tarpits are fun/cute/interesting, but also seem quite naive they deliberately poison the well in the hope of preventing a hypothetical, future well-poisoning by another party... but the well still ends up poisoned, with little regard for what the unintended consequences might be

It's just too easy to make yourself insane with the computer

HELLO GOOD MORNING I AM GOING ON THE COMPUTER AGAIN

I'm glad somebody out there is brave enough to push back against the "personal ChatGPT usage is terrible for the environment" message andymasley.substack.com/p/a-cheat-sh...

I am thinking about Reverend Toller again

After a few years of absence, I'll be back at @pydatalondon.bsky.social this year to talk about how we built a knowledge graph for climate policy at @climatepolicyradar.bsky.social

Hell yeah, it sounds like anthropic are really going to push to differentiate themselves from their competitors on the strength of their mechanistic interpretability work

Me and the @climatepolicyradar.bsky.social knowledge graph are coming to @berlinbuzzwords.de this June! I had a great time in Berlin as an attendee last summer, and I'm very excited about returning as a speaker this year - If you're around, come and say hi 👋

Enjoyed this on eval-driven product development from @eugeneyan.com. It chimes with my own experiences building around LLMs and search engines, including the thoughts on automated evaluators. When deconstructed, EDD is just the good old scientific method under a new name

I've spent a lot of this weekend listening to a biography of Alexander Von Humboldt and I'm pleased to report that this dude was COOL AS HELL!!

[Read] LLM-based relevance assessment still can’t replace human relevance assessment arxiv.org/pdf/2412.17156. Really interesting paper, challenging the recent idea that one can use LLLM-based relevance assessments to replace human judgements as gold standard,

🙂

hell yeah hell yeah

enjoyed this on the climate impacts of AI from @simonoxfphys.com opens with a very straightforward premise (how much energy does my chatgpt query use?), but patiently winds its way towards a much more interesting set of conclusions youtu.be/5sFBySzNIX0

brutal www.chalmermagne.com/p/how-not-to...

Alan Turing Institute plans revamp in face of criticism and technological change https://www.ft.com/content/6bfea441-e16c-499a-a887-69f735c29389

i really love the mechanistic interpretability work going on at anthropic, and the way that these transformer circuits papers are published. very exciting / provocative stuff in this one, particularly on planning, and self-awareness of knowledge limitations transformer-circuits.pub/2025/attribu...

If you need a break from SEO, LLMs, and the general emptiness of the modern internet, go here and browse for a bit marginalia-search.com/explore

Megafog

Computer 🙂👍