Lightnews — Scholar-powered news

Harrison Pim @harrisonpim.com · 15d

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/...

A postmortem of three recent issues

This is a technical report on three bugs that intermittently degraded responses from Claude. Below we explain what happened, why it took time to fix, and what we're changing.

www.anthropic.com

1

Harrison Pim @harrisonpim.com · 21d

> 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.

Defeating Nondeterminism in LLM Inference

Reproducibility is a bedrock of scientific progress. However, it’s remarkably difficult to get reproducible results out of large language models. For example, you might observe that asking ChatGPT the...

thinkingmachines.ai

1

Reposted by Harrison Pim

Matt Muir @mattmuir.bsky.social · Sep 1

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/

Tiny Awards

This is the home of the Tiny Awards, which, since 2023, has celebrated the best of the small, poetic, creative, handmade web.

tinyawards.net

12 31

Reposted by Harrison Pim

Grace @gracekind.net · Aug 19

What sort of black magic is this

5 6 72

Harrison Pim @harrisonpim.com · Aug 21

for some extra context, this excellent piece by @andymasley.bsky.social makes a very convincing argument that the environmental impact of LLMs is overstated, all while assuming that LLMs use 3 Wh per query (roughly 10x higher than the actual numbers!)

Using ChatGPT is not bad for the environment

And a plea to think seriously about climate change without getting distracted

andymasley.substack.com

2

Harrison Pim @harrisonpim.com · Aug 21

those numbers are roughly in line with sam altman's claim that the average chatgpt request uses 0.34 Wh.

all of these values are also much, much lower than those assumed in a lot of the discourse around LLMs!

The Gentle Singularity

We are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence, and at least so far it’s much less weird than it seems like it should be. Robots...

blog.samaltman.com

1 2

Harrison Pim @harrisonpim.com · Aug 21

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

Measuring the environmental impact of AI inference | Google Cloud Blog

A methodology for measuring the energy, emissions, and water impact of Gemini prompts shines a light on the environmental impact of AI inference.

cloud.google.com

1 4

Harrison Pim @harrisonpim.com · Aug 20

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

Policy Live 2025 - Homepage

www.policylive.org

1

Harrison Pim @harrisonpim.com · Jul 30

This will be my first time at ACL, and I’m very excited about the opportunity to meet and catch up with other folks doing work at the frontiers of NLP research :) drop me a message if you’re around!

Harrison Pim @harrisonpim.com · Jul 30

I’ll be talking about the knowledge graph that we’re building at @climatepolicyradar.bsky.social - By weaving together the connections between policy documents from all over the world, we’re hoping to make them easier to explore, understand, and explain

1 2

Harrison Pim @harrisonpim.com · Jul 30

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

1 1

Harrison Pim @harrisonpim.com · Jun 21

But in this, @aworkinglibrary.com gets straight at the bit which really does worry me.

"AI", not as a technology, but as an ideology, is being used as a prop in the revival of scientific racism, and old, eugenicist hierarchies of intelligence.

It's a very good essay

Toolmen

Even the best weapon is an unhappy tool.

aworkinglibrary.com

1 2

Harrison Pim @harrisonpim.com · Jun 21

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

1 1

Reposted by Harrison Pim

David Corney @dcorney.com · Jun 18

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...

What happens when "LLM grooming" fills an "information void"?

dcorney.com

1 1

Reposted by Harrison Pim

Max Woolf @minimaxir.bsky.social · Jun 10

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.

As datacenter production gets automated, the cost of intelligence should eventually converge to near the cost of electricity. (People are often curious about how much energy a ChatGPT query uses; the average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.)

3 9 28

Reposted by Harrison Pim

Very Sane AI Newsletter @verysane.ai · Jun 8

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 Biggest Statistic About AI Water Use Is A Lie

How did it become the main story?

www.verysane.ai

1 6 44

Harrison Pim @harrisonpim.com · Jun 9

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

Thomas Ptacek @sockpuppet.org · Jun 2

I regret nothing yet! fly.io/blog/youre-a...

My AI Skeptic Friends Are All Nuts

My smartest friends have bananas arguments about LLM coding.

fly.io

1

Harrison Pim @harrisonpim.com · Jun 8

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 💖

1 3

Harrison Pim @harrisonpim.com · Jun 7

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

Reposted by Harrison Pim

Andy Masley @andymasley.bsky.social · May 26

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...

All the ways I want the AI debate to be better

A lot of rules and ideas for talking about AI

andymasley.substack.com

2 7 53

Reposted by Harrison Pim

Slideshare Gems @goodslides.bsky.social · May 17

2 57 190

Harrison Pim @harrisonpim.com · May 14

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

AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms

New AI agent evolves algorithms for math and practical applications in computing by combining the creativity of large language models with automated evaluators

deepmind.google

Reposted by Harrison Pim

Full Fact @fullfact.org · May 13

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

Access to accurate information is not a luxury, it is the foundation of our democracy. We cannot let large online platforms which wield so much influence over our daily lives walk away from commitments to make our online world a safer place.

Government and regulators must hold them to account, to the full extent of the law. This is no time for half measures.

1 23 37

Harrison Pim @harrisonpim.com · May 10

IMO the more interesting question is whether this sort of tarpit could be used to maliciously steer LLM training, beyond just adding noise. given a large/subtle enough maze, can a more pointed set of disinformation be baked into LLMs' circuits?

Harrison Pim @harrisonpim.com · May 10

not sure one can effectively fight slop by generating even-sloppier-slop

1