Aidan Kelly
@helloaidank.bsky.social
Data Scientist at Nesta in the Discovery team | Passion for clean energy and all things generative AI | Views are my own.
MSc from UvA, PhD in particle physics from UCL.
MSc from UvA, PhD in particle physics from UCL.
Honestly, it is scary seeing fascism on the march across many different western countries. There are Portuguese elections today where Ventura (the Farage of Portugal) is likely to reach the runoff showdown.
We must resist!
We must resist!
January 18, 2026 at 3:45 PM
Honestly, it is scary seeing fascism on the march across many different western countries. There are Portuguese elections today where Ventura (the Farage of Portugal) is likely to reach the runoff showdown.
We must resist!
We must resist!
Reposted by Aidan Kelly
I've had a first stab at making a list of GenAI ppl on here but it's proving harder than anticipated! X is still the place to be for that content.
Let me know if you're interested in being added or know anyone who would be a good addition😎
bsky.app/profile/did:...
Let me know if you're interested in being added or know anyone who would be a good addition😎
bsky.app/profile/did:...
April 25, 2025 at 9:05 PM
I've had a first stab at making a list of GenAI ppl on here but it's proving harder than anticipated! X is still the place to be for that content.
Let me know if you're interested in being added or know anyone who would be a good addition😎
bsky.app/profile/did:...
Let me know if you're interested in being added or know anyone who would be a good addition😎
bsky.app/profile/did:...
Gentle intro to on-policy distillation for language models from Thinking Machines. It defines a lot of key terms used in post-training which I found useful.
Seems like there is scope using this technique to post-train LMs without need for huge compute.
thinkingmachines.ai/blog/on-poli...
Seems like there is scope using this technique to post-train LMs without need for huge compute.
thinkingmachines.ai/blog/on-poli...
On-Policy Distillation
On-policy, dense supervision is a useful tool for distillation
thinkingmachines.ai
January 13, 2026 at 11:11 AM
Gentle intro to on-policy distillation for language models from Thinking Machines. It defines a lot of key terms used in post-training which I found useful.
Seems like there is scope using this technique to post-train LMs without need for huge compute.
thinkingmachines.ai/blog/on-poli...
Seems like there is scope using this technique to post-train LMs without need for huge compute.
thinkingmachines.ai/blog/on-poli...
Reposted by Aidan Kelly
Really interesting piece in the FT with Yann LeCun (@yann-lecun.bsky.social). Insights into the dynamics at Meta and how the winds of AI have changed recently.
Interested in learning more about the world models he has been working on.
www.ft.com/content/e3c4...
Interested in learning more about the world models he has been working on.
www.ft.com/content/e3c4...
Computer scientist Yann LeCun: ‘Intelligence really is about learning’
The AI pioneer on stepping down from Meta, the limits of large language models — and the launch of his new start-up
www.ft.com
January 2, 2026 at 6:04 PM
Really interesting piece in the FT with Yann LeCun (@yann-lecun.bsky.social). Insights into the dynamics at Meta and how the winds of AI have changed recently.
Interested in learning more about the world models he has been working on.
www.ft.com/content/e3c4...
Interested in learning more about the world models he has been working on.
www.ft.com/content/e3c4...
Reposted by Aidan Kelly
Introducing DroPE: Extending Context by Dropping Positional Embeddings
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
January 12, 2026 at 4:07 AM
Introducing DroPE: Extending Context by Dropping Positional Embeddings
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
Listening to the news agent podcast earlier with Mehdi Hasan, and talk of Trump being re-elected in 2020 rather than 2024 may have been the lesser of two evils. On the surface, it seems relatively convincing.
January 11, 2026 at 9:00 PM
Listening to the news agent podcast earlier with Mehdi Hasan, and talk of Trump being re-elected in 2020 rather than 2024 may have been the lesser of two evils. On the surface, it seems relatively convincing.
Really interesting piece in the FT with Yann LeCun (@yann-lecun.bsky.social). Insights into the dynamics at Meta and how the winds of AI have changed recently.
Interested in learning more about the world models he has been working on.
www.ft.com/content/e3c4...
Interested in learning more about the world models he has been working on.
www.ft.com/content/e3c4...
Computer scientist Yann LeCun: ‘Intelligence really is about learning’
The AI pioneer on stepping down from Meta, the limits of large language models — and the launch of his new start-up
www.ft.com
January 2, 2026 at 6:04 PM
Really interesting piece in the FT with Yann LeCun (@yann-lecun.bsky.social). Insights into the dynamics at Meta and how the winds of AI have changed recently.
Interested in learning more about the world models he has been working on.
www.ft.com/content/e3c4...
Interested in learning more about the world models he has been working on.
www.ft.com/content/e3c4...
Watching what's happening over on grok in terms of image manipulation is really quite worrying and a clear violation of privacy. It needs to be regulated with a big stick, however I fear this is only the tip of the iceberg of what's out there on the web.
January 2, 2026 at 12:17 PM
Watching what's happening over on grok in terms of image manipulation is really quite worrying and a clear violation of privacy. It needs to be regulated with a big stick, however I fear this is only the tip of the iceberg of what's out there on the web.
Reposted by Aidan Kelly
There's been a lot of discourse lately about how coding agents have become overwhelming, and that SWEs who don't use them yet just won't be able to catch up. I don't believe this is true, but we do need to know what problems coding agents are solving to approach them with an engineering mindset.
Coding Agents Are Easy, Actually
Software engineering is here to stay
karashiiro.leaflet.pub
January 1, 2026 at 8:45 PM
There's been a lot of discourse lately about how coding agents have become overwhelming, and that SWEs who don't use them yet just won't be able to catch up. I don't believe this is true, but we do need to know what problems coding agents are solving to approach them with an engineering mindset.
Reposted by Aidan Kelly
I didn't see enough news about this crazy/cool paper from DeepSeek - you take text, and convert it to imagery, and you get BETTER performance. Oh and some really neat ideas about contextual memory too.
www.distributedthoughts.org/a-picture-is...
www.distributedthoughts.org/a-picture-is...
A Picture Is Worth Ten Thousand Tokens
DeepSeek's new OCR model achieves 97% accuracy while using one-tenth the tokens by rendering text as images. The computationally "heavy" modality turns out to be the efficient one, which should make u...
www.distributedthoughts.org
December 29, 2025 at 7:27 PM
I didn't see enough news about this crazy/cool paper from DeepSeek - you take text, and convert it to imagery, and you get BETTER performance. Oh and some really neat ideas about contextual memory too.
www.distributedthoughts.org/a-picture-is...
www.distributedthoughts.org/a-picture-is...
Can't stop listening to this. Mixture of synth, bass and guitar alongside this genre must have felt like it was from another planet in 1980.
www.youtube.com/watch?v=qh-j...
www.youtube.com/watch?v=qh-j...
A Forest
YouTube video by The Cure - Topic
www.youtube.com
December 29, 2025 at 6:29 PM
Can't stop listening to this. Mixture of synth, bass and guitar alongside this genre must have felt like it was from another planet in 1980.
www.youtube.com/watch?v=qh-j...
www.youtube.com/watch?v=qh-j...
Reposted by Aidan Kelly
Exploring TabPFN: A Foundation Model Built for Tabular Data
Understanding the architecture, training pipeline and implementing TabPFN in practice
Telegram AI Digest
#ai #foundationmodel #news
Understanding the architecture, training pipeline and implementing TabPFN in practice
Telegram AI Digest
#ai #foundationmodel #news
Exploring TabPFN: A Foundation Model Built for Tabular Data
Understanding the architecture, training pipeline and implementing TabPFN in practice
towardsdatascience.com
December 27, 2025 at 9:20 PM
Exploring TabPFN: A Foundation Model Built for Tabular Data
Understanding the architecture, training pipeline and implementing TabPFN in practice
Telegram AI Digest
#ai #foundationmodel #news
Understanding the architecture, training pipeline and implementing TabPFN in practice
Telegram AI Digest
#ai #foundationmodel #news
Trying to understand some code in Rust and it's giving me vietnam style flashbacks to C++
December 25, 2025 at 8:46 PM
Trying to understand some code in Rust and it's giving me vietnam style flashbacks to C++
Merry Christmas! 🎄
December 25, 2025 at 12:55 PM
Merry Christmas! 🎄
Thoughtful piece by my current favourite podcaster @dwarkesh.bsky.social on where we're at with AI progress and reflections around intelligence itself: substack.com/home/post/p-...
Thoughts on AI progress (Dec 2025)
Why I'm moderately bearish in the short term, and explosively bullish in the long term
substack.com
December 6, 2025 at 11:03 AM
Thoughtful piece by my current favourite podcaster @dwarkesh.bsky.social on where we're at with AI progress and reflections around intelligence itself: substack.com/home/post/p-...
Yann LeCun being let go by Meta might have some unexpected benefits 🍿
November 15, 2025 at 5:37 PM
Yann LeCun being let go by Meta might have some unexpected benefits 🍿
This may be very old man of me, but I feel like there is something about the bluesky algorithm/my feed that is not sucking me in enough. I know, I know, it's sort of the point but still, I need to have some pull to check out the content and it isn't quite itching that scratch.
November 4, 2025 at 9:49 PM
This may be very old man of me, but I feel like there is something about the bluesky algorithm/my feed that is not sucking me in enough. I know, I know, it's sort of the point but still, I need to have some pull to check out the content and it isn't quite itching that scratch.
Reposted by Aidan Kelly
With the AWS outage, now‘s as good a time as any to post this old strip.
October 20, 2025 at 10:18 AM
With the AWS outage, now‘s as good a time as any to post this old strip.
Happy to share my first piece of analysis work with the Discovery team at @nestauk.bsky.social on 𝗠𝗶𝘀𝘀𝗶𝗼𝗻 𝗥𝗮𝗱𝗮𝗿!
The Mission Radar is a briefing series on innovation signals and emerging trends in Nesta’s missions. We use AI to scan for innovation across business, academia and policy. (1/7)
The Mission Radar is a briefing series on innovation signals and emerging trends in Nesta’s missions. We use AI to scan for innovation across business, academia and policy. (1/7)
October 20, 2025 at 11:39 AM
Happy to share my first piece of analysis work with the Discovery team at @nestauk.bsky.social on 𝗠𝗶𝘀𝘀𝗶𝗼𝗻 𝗥𝗮𝗱𝗮𝗿!
The Mission Radar is a briefing series on innovation signals and emerging trends in Nesta’s missions. We use AI to scan for innovation across business, academia and policy. (1/7)
The Mission Radar is a briefing series on innovation signals and emerging trends in Nesta’s missions. We use AI to scan for innovation across business, academia and policy. (1/7)
What a really interesting conversation between two of my favourite people in AI in the form of @karpathy.bsky.social and @dwarkesh.bsky.social. I wish more ppl in this space took the same approach in terms of being open to debate, challenge and overall being humble.
www.youtube.com/watch?v=lXUZ...
www.youtube.com/watch?v=lXUZ...
Andrej Karpathy — “We’re summoning ghosts, not building animals”
YouTube video by Dwarkesh Patel
www.youtube.com
October 19, 2025 at 8:38 PM
What a really interesting conversation between two of my favourite people in AI in the form of @karpathy.bsky.social and @dwarkesh.bsky.social. I wish more ppl in this space took the same approach in terms of being open to debate, challenge and overall being humble.
www.youtube.com/watch?v=lXUZ...
www.youtube.com/watch?v=lXUZ...
So theoretically LLMs can be deterministic if we set their temperature to zero, but in practice they aren't. This comes down to how we do calculations such as matrix multiplication at inference time. It turns out the order you do these calculations in does matter. (1/6)
October 5, 2025 at 3:24 PM
So theoretically LLMs can be deterministic if we set their temperature to zero, but in practice they aren't. This comes down to how we do calculations such as matrix multiplication at inference time. It turns out the order you do these calculations in does matter. (1/6)
Does anyone have cool (or boring) agentic use-cases that they've read about or used?
Doing some horizon scanning work to see what's out there but thought my bluesky network might have ideas!
Doing some horizon scanning work to see what's out there but thought my bluesky network might have ideas!
September 22, 2025 at 10:12 AM
Does anyone have cool (or boring) agentic use-cases that they've read about or used?
Doing some horizon scanning work to see what's out there but thought my bluesky network might have ideas!
Doing some horizon scanning work to see what's out there but thought my bluesky network might have ideas!
What if there was an AI-powered tool that allowed policymakers to rapidly identify the most effective policies to tackle any given social challenge, based on the latest evidence? (1/2)
September 12, 2025 at 10:49 AM
What if there was an AI-powered tool that allowed policymakers to rapidly identify the most effective policies to tackle any given social challenge, based on the latest evidence? (1/2)
This was a fun ML problem to tackle, underlined some of the data quality issues with EPCs (92% of the storey count field was missing) 😅.
Think the conclusion around the distinct heating strategies needed for areas with different rise types is definitely a salient one!
Think the conclusion around the distinct heating strategies needed for areas with different rise types is definitely a salient one!
How are *flats* in England and Wales heated?
It's a surprisingly interesting question, and my Nesta colleagues Roisin Gorman and @helloaidank.bsky.social have been looking into it...
www.nesta.org.uk/project-upda...
It's a surprisingly interesting question, and my Nesta colleagues Roisin Gorman and @helloaidank.bsky.social have been looking into it...
www.nesta.org.uk/project-upda...
August 29, 2025 at 2:53 PM
This was a fun ML problem to tackle, underlined some of the data quality issues with EPCs (92% of the storey count field was missing) 😅.
Think the conclusion around the distinct heating strategies needed for areas with different rise types is definitely a salient one!
Think the conclusion around the distinct heating strategies needed for areas with different rise types is definitely a salient one!
Favourite feature from the GPT-5 release is being able to quickly spin out different UIs (which can be previewed), make easy edits without needing to manually edit JS code myself. It's pretty good at instruction following now too!
August 29, 2025 at 2:42 PM
Favourite feature from the GPT-5 release is being able to quickly spin out different UIs (which can be previewed), make easy edits without needing to manually edit JS code myself. It's pretty good at instruction following now too!