teach sand to think, give sand a telescope 🔭 PhD @cambridge_uni; building agentic research systems for physics & astronomy. prev @guardian data science.
Manchester
Joined October 2026
- Tweets178
- Following198
- Followers37
- Likes1.9K
Pinned Tweet
I built a scale 3D map of the universe you can fly through, from Earth to the edge of the observable universe, including 192 million stars from Gaia, 15 million galaxies from DESI, 7,000 exoplanets, and more, stitched together.
Runs in your browser with @threejs!
zoom-the-universe.gamccarthy…
4
5
287
Since yesterday: AlexNet training is at 4x up from 3x on Vulkan-Honeykrips on @omarchymac compared to CPU. Full pytorch operation coverage with 226/641 operations on PyTorch OpInfo GPU accelerated.
Check it out and help reach full covereage: github.com/georgeamccarthy/o…
1
14
Turns out I did not need to buy an old iPhone Pro from Giffgaff for Lidar
Testing 3d scanning with this ~$50 lidar sensor (ST VL53L9CX) attached to an iPhone.
Works pretty well for quick indoor scans using just ARKit poses.
PCB is open source from RemiFM: github.com/RemiFM/USB-LiDAR-…
2
50
Running pytorch on @OmarchyMac? Join my new project github.com/georgeamccarthy/o…, a PyTorch backend for the Honeykrisp Vulkan GPU driver to accelerate deep learning on Apple Silicon Linux.
2
1
77
Currently bare minimum support (I'm adding operation support as I use it) please start doing it and do the same, agent PRs welcome!
1
1
13
Inspired by @joshuaswarren's @omarchy-mlx, but for PyTorch.
1
2
28
I'm surprised the AI productivity multiplier reported by Anthropic staff is only 2x. I would have thought it would be a lot higher, given they are also quoted saying the experiment implementation time is ~0. Does ideation and compute quickly become the limiting factor for RSI?
Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development.
We are now publishing a report summarizing the results of this risk assessment.
2
3
58
This viral post has been widely misinterpreted because the claims in the post and main tweet are heavily caveated in the details of the post.
"natively read neural network weights" sounds like Astra reads neuralease, but it just recognises shapes in the first weights layer.
Look at the image below, astra is told that the player is in the left column, the moving obj in the middle column etc. and it figures these correspond to pong (row 2), pacman (row 3) etc. I asked Opus 5.5 about this image: claude.ai/share/28c8163f-464… and it sees it too, many humans would also guess pong for row 2 and could take a decent guess at the others.
Astra also did not improve qwen's mathematical ability via direct weight read and write, it applied an out of the box mech interp technique to remove a bug where the model was getting stuck in an output loop, causing qwen to score 0 when it in some cases knew how to solve the problem. It did this by running an experiment to find a weights vector to control the behaviour rather than reading and editting the weights directly.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
1
53
Really inspiring event the KCL Tech Summit. Entirely organised by students at King's University, showcasing some incredible founders and companies in the UK. I went to three talks about UK startups and feel really optimistic about the UK as a place for tech innovation
A highlight was this lock picking robot by naude.dev. A lot of lock pickers do not actually know how to open doors they just remove the lock, something like this could make it so if you are locked out of your flat you can actually get it picked.
1
38
There is a lot of wisdom in this blog post I learnt a lot from it.
I know many PhD students are questioning what their paths mean as AI accelerates scientific discovery. That anxiety deserves to be taken seriously.
I wrote some reflections in a recent blog post. My students found them helpful, so I’m sharing them more widely.
I hope it offers a little light as we navigate this strange time. Please share it with someone who might need it.
pages.cs.wisc.edu/~sharonli/…
2
66
I wondered whether there is a safe route for a roadtrip from London to China so I vibe coded this web site that has scraped travel advice from gov.uk and shows it on a globe
It's quite useful for general road trip / backpacking planning
You can try it out at travel-advice-globe.gamccart… but the data will be out of data after today
2
2
77
