Harrison Pim
@harrisonpim.com
I'm working on search, machine learning, and knowledge graphs at climatepolicyradar.org | harrisonpim.com
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/...
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
September 26, 2025 at 10:13 PM
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/...
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.
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
September 21, 2025 at 12:05 PM
> 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.
Reposted by Harrison Pim
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
September 1, 2025 at 11:21 AM
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/
Reposted by Harrison Pim
What sort of black magic is this
August 19, 2025 at 9:04 PM
What sort of black magic is this
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
August 21, 2025 at 3:53 PM
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
Policy Live 2025 - Homepage
www.policylive.org
August 20, 2025 at 11:18 AM
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
July 30, 2025 at 10:05 AM
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
June 21, 2025 at 12:47 PM
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
Reposted by Harrison Pim
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...
dcorney.com/thoughts/202...
What happens when "LLM grooming" fills an "information void"?
dcorney.com
June 18, 2025 at 3:00 PM
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...
dcorney.com/thoughts/202...
Reposted by Harrison Pim
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.
June 10, 2025 at 10:06 PM
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.
Reposted by Harrison Pim
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...
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
June 8, 2025 at 11:54 PM
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...
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
June 9, 2025 at 7:02 AM
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 💖
June 8, 2025 at 9:52 PM
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
June 7, 2025 at 5:28 PM
oh my GOD i love pydata events so much. so happy to be back in these spaces with these people
Reposted by Harrison Pim
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...
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
May 26, 2025 at 5:04 AM
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...
andymasley.substack.com/p/all-the-wa...
Reposted by Harrison Pim
Deepmind have been doing amaaaaazing things ✨
deepmind.google/discover/blo...
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
May 14, 2025 at 9:39 PM
Deepmind have been doing amaaaaazing things ✨
deepmind.google/discover/blo...
deepmind.google/discover/blo...
Reposted by Harrison Pim
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
Read it here: buff.ly/Uz6pzTJ
May 13, 2025 at 9:50 AM
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
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
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
Basically traps AI crawlers and sends them down an "infinite maze" of static files with no exit links, where they get stuck and thrash around for months. Once trapped, the crawlers can be fed gibberish data, aka Markov babble, which is designed to poison AI models arstechnica.com/tech-policy/...
AI haters build tarpits to trap and trick AI scrapers that ignore robots.txt
Attackers explain how an anti-spam defense became an AI weapon.
arstechnica.com
May 10, 2025 at 11:44 AM
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
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
Reposted by Harrison Pim
It's just too easy to make yourself insane with the computer
May 9, 2025 at 4:57 PM
It's just too easy to make yourself insane with the computer
HELLO GOOD MORNING I AM GOING ON THE COMPUTER AGAIN
May 1, 2025 at 9:31 AM
HELLO GOOD MORNING I AM GOING ON THE COMPUTER AGAIN
Reposted by Harrison Pim
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...
A cheat sheet for why using ChatGPT is not bad for the environment
Arm yourself with knowledge
andymasley.substack.com
April 29, 2025 at 3:30 PM
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
April 28, 2025 at 9:09 PM
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
April 28, 2025 at 4:05 PM
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