Sign in

Danny Wood

@echostatements.bsky.social
157 followers 313 following 67 posts

AI Research Scientist || PhD in machine learning || Ensembles, probabilistic machine learning, recurrent neural networks || echostatements.net

PostsRepliesMedia
Danny Wood @echostatements.bsky.social · 09/06/2026
It definitely seems more robust with the "the word" phrasing, and sometimes even gives you a Python snippet that can verify the result. But even then it doesn't seem to catch everything.
A Google query for "how many ps in the word poppy". The answer it returns is 2
120 2
Danny Wood @echostatements.bsky.social · 09/06/2026
Saying that theres no "i" in team used to mean you were a team player. Now it shows you're yet to embrace the full potential of AI.
A google search for the query 'How many times does the letter "i" appear in team'.

The result shows Google's AI overview saying the letter "i" appears exactly once in team.
110 1
Danny Wood @echostatements.bsky.social · 03/02/2026
Can we get LLMs to generate text that is perfectly typeset down to the very last character? In a new blog post, I show how using a constrained beam search, LLMs can efficiently generate perfect blocks of monospace text echostatements.net/posts/2026/0...
Image showing the transformation of the intro paragraph to the Hot Fuzz Wikipedia page into perfectly justified monospace text
110 1
Danny Wood @echostatements.bsky.social · 01/02/2026
I've added the ability to create Delaunay triangulations to my Voronoi diagram generation Python library. The image generation takes a few seconds but I'm pleased with how easy it is to make pretty and varied outputs with it
100 1
Danny Wood @echostatements.bsky.social · 23/01/2026
And an honourable mention: Lego: Lego Lego supposedly are the world's largest manufacturer of tyres (at least according to the Guinness*). Sure, tyres aren't their main focus, but that doesn't seem so unusual for a tyre company! * While we're at it: yes, Guinness the brewery/word record authority
The lego logoA lego forklift
000 0
Danny Wood @echostatements.bsky.social · 23/01/2026
Goodyear: Blimps Unbelievably left-field. In the 20th century, they at least built the blimps themselves, but now their sideline seems to just be being the world's most famous owner of a fleet of airships
The good year logoThe Goodyear blimp
100 1
Danny Wood @echostatements.bsky.social · 23/01/2026
Michelin: Restaurant reviews This clearly got out of hand for them. It started as a more general guide for motorists, (even that is a slightly eccentric thing for a tyre company) but getting your tyres from the world's foremost authority on restaurant quality is odd
The Michelin logoA fancy looking meal
100 1
Danny Wood @echostatements.bsky.social · 23/01/2026
Dunlop: Sports equipment If you're working with rubber anyway, I can see how you get into making tennis balls and how that might spiral into sporting goods more generally Fair enough
The Dunlop logoA Dunlop tennis ball
100 1
Danny Wood @echostatements.bsky.social · 28/09/2025
Answering my own question, the example from the "Infinite-dimensional vector function" Wikipedia page makes it feel clear that a curve with these properties will exist Even so, without more thought, this feels more like it side-steps the issue in my intuition than directly tackles it
https://en.wikipedia.org/wiki/Infinite-dimensional_vector_function#Crinkled_arcs
010 1
Danny Wood @echostatements.bsky.social · 07/08/2025
Looking at the code in Hugging Face's library for the new GPT models, I was a bit disappointed by how similar they are to Llama & Mistral models, but there is one cool trick I hadn't seen before: attention sinks. These are a mechanism by which attention heads can say "I don't have anything to add"
The code for the attention mechanism in the `transformers` implementation of gpt_oss
120 2
Danny Wood @echostatements.bsky.social · 22/07/2025
A little etymological fact that I like is that "lemma" and "dilemma" are from the same ancient Greek origin If you translate "lemma" as meaning a proposition, a dilemma is literally having two propositions to consider
120 2
Danny Wood @echostatements.bsky.social · 18/07/2025
And even better, these tricks are not mutually exclusive, by doing both simultaneously, you get a 2.5x speed up (dependent on batch size) 8/9 🧵
Graph from start of thread repeated, showing effects of faster model and faster training
100 1
Danny Wood @echostatements.bsky.social · 18/07/2025
Again, this gets some pretty significant speed-up 7/9 🧵
Graph showing training speed of classic vs faster training (~250 seconds vs ~150 seconds)
100 1
Danny Wood @echostatements.bsky.social · 18/07/2025
Optimisation 2: When training Siamese networks, people tend to generate matching/non-matching pairs in equal ratio. However, you can train more efficiently if you generate only matching pairs, then creating the non-matching ones with some shifting of subnetwork outputs. 6/9 🧵
Image showing how paired samples can be offset by one in a batch to make unpaired samples
100 1
Danny Wood @echostatements.bsky.social · 18/07/2025
Optimisation 1: In practice, when implementing these networks there is only one subnetwork, called twice, once for each input But by stacking the inputs, we actually only need to make one call to the network: Depending on batch size, the effect can be significant 5/9 🧵
A diagram of a faster implementation of a Siamese network vs a standard oneComparison of inference  and throughput for classic vs faster network (1.35 seconds for classic, 0.91 seconds for faster)
110 1
Danny Wood @echostatements.bsky.social · 18/07/2025
Firstly, what is a Siamese neural network? A Siamese neural network consists of two identical subnetworks and a comparison layer The two subnetworks generate representations for two different inputs, then the comparison layer measures the similarity between these representations 3/9 🧵
A diagram of a standard Siamese neural network
100 1
Danny Wood @echostatements.bsky.social · 18/07/2025
Siamese neural networks are a neat way of doing one-shot/few-shot classification but there's something about the way they're usually implemented that felt a bit inefficient to me Here are a couple of relatively simple tricks you can do to speed up training Siamese networks by 2.5x! 1/9 🧵
A graph comparing faster model + faster training vs classic model + classic training, and the other two combinations. Faster Model/faster training shows a 2.5x speed up vs classic model/classic training
130 3
Danny Wood @echostatements.bsky.social · 02/05/2025
If you've not read it, I highly recommend "Admissible probability measurement procedures" by Shuford et al (1966) too It's really well written and if you replaced "students" with "models" in the text, would read like a fairly modern machine learning paper link.springer.com/article/10.1...
010 1