James Requeima (@jamesrequeima) header
James Requeima (@jamesrequeima) profile picture

Postdoctoral fellow at the University of Toronto/Vector Institute. Former PhD student in Machine Learning at the University of Cambridge CBL lab.

Joined July 2011
Photo by James Requeima (@jamesrequeima)Photo by James Requeima (@jamesrequeima)
Replying to @DavidDuvenaud
@DavidDuvenaud and I have a new microscope to look at the beliefs LLMs have about all sorts of quantities. What should we point it at?
LLMs have complex joint beliefs about all sorts of quantities. And my postdoc @jamesrequeima visualized them! In this thread we show LLM predictive distributions conditioned on data and free-form text. LLMs pick up on all kinds of subtle and unusual structure: 🧵
1
1
13
764
James Requeima retweeted
🌡️ Breakthrough: Canadian AI revolutionizes weather forecasting Vector postdoc @jamesrequeima co-developed Aardvark Weather, transforming climate prediction globally. This AI-powered system delivers forecasts 10x faster while using 1000x less computing power than traditional methods. Game-changing features: + Runs on a desktop computer (no supercomputer needed) + Processes satellite, weather station, ship, & plane data simultaneously + Outperforms US national forecasting systems + Can provide 10-day advance predictions with higher accuracy Real-world impact: 🌲 Wildfire prevention in BC 🌊 Flash flood warnings in Toronto 🌾 Agricultural planning ⚡️ Renewable energy optimization Most importantly? This innovation democratizes access to precise weather forecasting, especially benefitting remote communities and developing nations. Read our full story on how this innovation is shaping climate science: vectorinstitute.ai/aardvark-…
2
9
1,286
James Requeima retweeted
Context is Key🗝️ is accepted at ICML 2025! 📈 Let's catch up if you'll be at ICML 🛬 See the poster and tweet thread below for a preview of CiK 👇 teapawt.net/arjunashok37/status/18… And stay tuned for new results ;)
(New paper alert!) Forecasting models typically rely on numerical historical data. However, in many cases, numerical data is insufficient and context is key. E.g., In the series below, would you have predicted the drop? Even the best models do not (forecast in blue).
4
11
57
4,740
James Requeima retweeted
.@jamesrequeima gave the third invited talk of the day at #AABI2025! **LLM Processes**
2
9
512
James Requeima retweeted
Understanding causes is key to science. Finite observational data alone isn't enough. While Bayes offers a framework to deal with this, the calculations are often intractable. We introduce a method to accurately approximate the posterior over causal structures. #ICLR2025 🧵1/15
3
29
2
147
10,607
James Requeima retweeted
(New paper alert!) Forecasting models typically rely on numerical historical data. However, in many cases, numerical data is insufficient and context is key. E.g., In the series below, would you have predicted the drop? Even the best models do not (forecast in blue).
2
31
5
64
17,154
James Requeima retweeted
Really excited about this! We've put effort into making the model accessible with code that fits in a Tweet. 😊 pip install microsoft-aurora from aurora import Aurora model = Aurora() model.load_checkpoint("microsoft/aurora", "aurora-0.25-finetuned.ckpt")
1/5 Excited to announce the public code release of Aurora - a foundation model for atmospheric forecasting! 🌎⛅️ Code: github.com/microsoft/aurora Docs: microsoft.github.io/aurora Paper: arxiv.org/abs/2405.13063
7
42
3,613
📢 Excited to announce 'Environmental sensor placement with convolutional Gaussian neural processes', out now in @EnvDataScience Paper: cambridge.org/core/services/… Talk: youtu.be/v0pmqh09u1Y Code: github.com/tom-andersson/dee… @BAS_News @Cambridge_Uni @turinginst 🧵 1/n
3
16
4
60
9,895
Where should we measure the environment to minimise uncertainty about weather and climate? 🎲⛅ I'm excited to share that my second journal paper, which grapples with this fascinating problem, will be published soon! Interested? Check out my 15-min talk: youtu.be/v0pmqh09u1Y
1
9
1
41
4,284
QUIZ: One of these plots shows temperature around Antarctica. The others are samples from a model conditioned on obs at the black circles. Which one is the true data? Find it tough? That's the power of the AR ConvCNP. Answer below & check out @ikwess's🧵on our #ICLR2023 paper👇
Happy to announce our #ICLR2023 paper Autoregressive Conditional Neural Processes! w/ @stratosmarkou, @jamesrequeima, @AndrewFoongYK, @tom_r_andersson, Anna Vaughan, Anthony Buonomo, @scotthosking, Rich Turner Paper: arxiv.org/abs/2303.14468 Code: github.com/wesselb/neuralpro… 🧵1/4
1
3
8
1,574
James Requeima retweeted
Happy to announce our #ICLR2023 paper Autoregressive Conditional Neural Processes! w/ @stratosmarkou, @jamesrequeima, @AndrewFoongYK, @tom_r_andersson, Anna Vaughan, Anthony Buonomo, @scotthosking, Rich Turner Paper: arxiv.org/abs/2303.14468 Code: github.com/wesselb/neuralpro… 🧵1/4
2
12
2
64
9,842
James Requeima retweeted
Great presentation from @tom_r_andersson on day 2 @Climformatics 🎉 Active learning improve how we understand the Antarctic. Impressive work. See the paper here: arxiv.org/abs/2211.10381
1
9
1
26
4,306
James Requeima retweeted
I'm very excited to announce the release of Plum 2.0! With this release, Plum is fully powered by Beartype, which means that Plum now supports any type supported by Beartype. And that's not all: Plum is now proudly a member of the Beartype org! github.com/beartype/plum
2
4
1
15
2,748
I should have announced this before, but a year ago I switched my research focus to AI existential risk reduction and governance. I think the risk of bad outcomes for humanity due to AGI is substantial, and that coordinating a slowdown in AGI development is probably a good idea.
29
112
20
787
210,316
📢Thrilled to be presenting 'Environmental Sensor Placement with Convolutional Gaussian Neural Processes' at a #NeurIPS22 workshop 🗞️Poster gp-seminar-series.github.io/… 📰Paper arxiv.org/abs/2211.10381 🤝Collaboration betw @BAS_News & @CambridgeMLG 💲Funded by @turinginst 👇More below
1
8
2
22
James Requeima retweeted
Replying to @ylecun
Yes ok, but why do they have to shout so much while they're learning?
9
8
2
307
James Requeima retweeted
Still using latent variables to get correlations out of your Neural Process? Then consider Gaussian Neural Processes (GNP)! ✓ Correlated predictions ✓ Tractable likelihood @stratosmarkou, @jamesrequeima, @ikwess, Anna Vaughan & Rich Turner #ICLR2022 arxiv.org/abs/2203.08775
5
33
1
190
Ever wondered why variational inference + natural gradients + neural networks might be a good idea? Are you curious to follow the details of the mathematical derivations? I wrote a blog post on the new @CambridgeMLG blog: mlg-blog.com/2021/04/13/ngvi… 1/3
1
40
5
205
James Requeima retweeted
Replying to @CambridgeMLG
@CambridgeMLG is launching a blog, featuring a first two-part post about what keeps a Bayesian awake at night by Richard E. Turner and me. 🧵 mlg-blog.com/
2
36
14
172