@jack-kelly.com
Building open source ML to reduce CO₂ emissions from the electricity grid • Python, climate, forecasting weather & energy, AI, open data & science, Rust • Now: Co-founder @openclimatefix.org • Prev: DeepMind, NESO, Imperial • https://github.com/jackkelly
Over the weekend I wrote some code to publish & plot my solar data. It's been a fun project, where I got to try out some new tech, like running Python in the browser by exporting the Marimo notebook as web assembly
Live visualisation: jack-kelly.com/solar/
Blog post: jack-kelly.com/blog/2026-01...
The India Energy Stack looks pretty awesome. It will provide:
• Unique IDs for assets, consumers, & transactions
• Open & standardised APIs
• Ecosystem innovation
• Scalable, plug-n-play architecture that grid operators can adopt
I wish the GB grid had unique IDs 😀
pulseenergy.io/blog/deep-di...
London. I love you. But, wow, you really need to put more thought into how to keep trains and busses cool in heat waves. It doesn't necessarily need AC. Just some windows designed to move outside air through the vehicle.
Many electricity grid operators estimate total renewable generation every half hour. e.g. in Great Britain we have Sheffield Solar's wonderful PVLive. But these estimates are uncertain because many generators don't report their live generation. So, shouldn't estimated totals be *probabilistic*?
The Trump administration's attempts to cut the overhead rates of grants awarded to universities from NIH, NSF, DOE, DoD, etc. have gone largely unnoticed outside of academia, but this move is a crucial aspect of the assault on American research universities. @andrewdessler.com wrote about it. 🧪
A HISTORIC MILESTONE: For the first time, solar power generated more electricity globally than nuclear, making it the world’s 4th largest power source.
In just 5 years, solar output nearly tripled—from 79 TWh (Apr 2020) to 233 TWh (Apr 2025). The energy transition is speeding up.
@ember-energy.org
“It was renewables that did it!” That was the headline after Spain’s blackout on April 28.
But the official report published today points to other factors: chain reaction of overvoltages led to cascading generation disconnections. Human error played a big role too.
media.licdn.com/dms/document...
Webinar on the past and future of IPCC scenarios
TODAY: Tuesday, 17 June 2025, 15:00 – 16:00 CEST
Here is a figure of the first three generation of 'baseline' scenarios used in IPCC assessment reports (SA90, IS92, SRES).
@climatediamond.bsky.social
climate-diamond.eu/diamond-webi...
There's an ML person who will try to convince you that papers don't matter, that only "shipping" impactful research does. That if it's not in prod it's not real. This is a flawed model of how long-term progress, which builds on prior work, happens.
Awesome new podcast geeking out on weather forecasting. From our friends at dynamical.org (including @mrshll.com).
dynamical.org/podcast/
So, in some ways I think the narrative around LLMs should be "wow, isn't it amazing how much knowledge *humans* have accumulated (and is now queriable via LLMs). Well done humans! Let's keep it up!"
I don't think we should consider LLMs to be some alien form of intelligence. Instead, I think of LLMs as a lossily compressed representation of a large portion of all *human* knowledge, that happens to be easy to query using natural language. LLMs don't know anything outside of human knowledge.
All this evidence will be interpreted using hierarchical Bayesian models, to try to understand the interconnected relationships between causes and effects. Everything will be published as an online "living" book. The community will be invited to contribute to every step.
We will then refine these priors by running hundreds of machine learning experiments ourselves, and - crucially - inviting the community to contribute forecasting solutions to long-running challenges scored against standardised open datasets.
For each question, we will distill evidence from the literature into probabilistic estimates. In the language of Bayesian analysis, these estimates will be our "prior beliefs".
The aim is to clarify which forecasting approaches work well, and in which situations, and to quantify the differences. The idea is to first collate a list of questions we'd like answers to, such as in which contexts satellite data improves skill.
This lack of coherence is a pernicious problem: it dramatically slows down and masks progress in research; and it prevents practitioners from building the best forecasts for their users. The ultimate aim of this project is to attempt to fix this problem!
Unfortunately, such recommendations aren't easy to deduce from the literature because each forecasting paper measures skill using a different dataset and different metrics. The end result is that the literature feels like an ocean of unstructured information, rather than a coherent story.
As someone who researches and builds energy forecasts, I'd like recommendations for techniques to improve forecasting skill in each of the myriad different forecasting tasks encountered in the wild.
For years, I've been keen to help the energy forecasting community to accelerate progress. Better forecasts = cheaper electricity & lower CO₂ emissions
I'm super-excited that I'm finally in a position to help. This doc sketches out my plans. I'd love feedback!
docs.google.com/document/d/1...
I'm looking forward to attending this evening's lecture on the "Future of Energy Lecture" by Fintan Slye, the CEO of the GB National Energy System Operator (NESO). Maybe see some of you there?!
www.biee.org/meeting/futu...
Great job opening: Vice President of Engineering at a VC-Backed Startup "on a mission to decarbonise computing power":
docs.google.com/forms/d/e/1F...
I'm updating the UK solar power dataset that Open Climate Fix publishes on Hugging Face. I have some proposals for how to modernise the dataset. Please let me know if you have any concerns! My proposals are described here: huggingface.co/datasets/ope...
Imagine if all you had to do to serve geopspatial data over OGC APIs was to ingest @zarr.dev datasets into cloud object storage...and the platform handled the rest. Scaling, logging, metrics, auth, etc. That's Flux! Join our webinar if you want to learn more.
Great new job vacancy with some of my old colleagues at Google DeepMind: job-boards.greenhouse.io/deepmind/job...
Hot take: We get upset by Trump's inability to tell the truth. But lying is endemic in the business community. It's all about telling a good story. Truth be damned. Theranos was a particularly infamous example because the dangers of fibbing were so obvious. But similar levels of fibbing are common.
We’re building a new static type checker for Python, from scratch, in Rust.
From a technical perspective, it’s probably our most ambitious project yet. We’re about 800 PRs deep!
Criteria for deciding which electric vehicle to buy:
• In 2020: What's the range? How expensive is it?
• In 2025: Will my purchase help fund the rise of fascism in the West? Will the manufacturer decide to use the vehicle as a weapon if they decide they don't like me / my family / my country?
Even fascists have to obey the laws of physics. Weld the fucking doors closed.
Chuck Schumer and those in Congress either need to go to NOAA HQ with federal law enforcement officers to arrest non-government people who are committing crimes by accessing government systems.
Or, they need to go on-air right now to explain why they're not!