Sam Maddrell-Mander (@smaddrellmander) header
Sam Maddrell-Mander (@smaddrellmander) profile picture

Member of Technical Staff at Dayhoff Labs ex. AI Engineer @graphcoreai ex. CERN

Bristol
Joined September 2010
Photo by Sam Maddrell-Mander (@smaddrellmander)
(Yet) another tale of Rise and Fall: o1 claims extraordinary strong performance, scoring high on olympiad level math & coding problems. Can it handle simple AIW problems, which reveal generalization & basic reasoning deficits in SOTA LLMs? (arxiv.org/abs/2406.02061) 🧵1/n
17
50
14
409
118,704
Sam Maddrell-Mander retweeted
Even after 4 years working in the semiconductor industry, I've learned a lot from reading this Economist's overview. The end of Moore's law, Dennard's scaling, new ideas to reach a trillion transistors, specialized chips for AI... This overview is superb.
AI has returned chipmaking to the heart of computer technology. But it is time for some new ideas. Read our latest Technology Quarterly to learn how advances in chipmaking, both incremental and radical, can keep the exponential engine humming econ.st/4gIs1Ib 👇
1
1
356
Sam Maddrell-Mander retweeted
Sure matplotlib is cool, but what if I want to load my loss curves into the 2006 hit Flash game LineRider?
50
778
92
6,242
438,714
Exciting news for Graphcore, wishing all my ex-colleagues luck with this next chapter. Looking forward to seeing the next generation of chips!
Exciting news: Graphcore joins @SoftBank_Group to build next generation of AI compute. graphcore.ai/posts/graphcore…
2
209
Sam Maddrell-Mander retweeted
17
159
32
609
74,530
Sam Maddrell-Mander retweeted
📢 In our new blogpost w/ @mmbronstein @haitao_mao_ @AndyJiananZhao @zhu_zhaocheng we discuss foundation models in Graph & Geometric DL: from the core theoretical and data challenges to the most recent models that you can try already today! towardsdatascience.com/found…
2
40
5
120
33,310
Sam Maddrell-Mander retweeted
Is Attention all you need? Mamba 🐍, a novel AI model based on State Space Models, emerges as a alternative to the widely used Transformer models 🤖 Read more in our latest article -> thegradient.pub/mamba-explai…
52
6
214
43,791
Sam Maddrell-Mander retweeted
The Mamba paper was rejected? 🤡
12
22
8
466
131,312
Sam Maddrell-Mander retweeted
1/n While everybody’s been busy packing for #NeurIPS2023, our team at @graphcoreai has been busy with this beauty. Let me introduce: ✨SparQ Attention✨ TL;DR This is a plug-and-play inference Attention block for pre-trained LLMs, which evaporates the KV cache bandwidth 🧵
1
10
3
11
3,102
Love to see this work by some excellent ex-colleagues at @graphcore highlighted by @_akhaliq. The paper is excellent, code, results and theory all laid out perfectly. We all want to improve our LLM inference efficiency!
SparQ Attention: Bandwidth-Efficient LLM Inference paper page: huggingface.co/papers/2312.0… Generative large language models (LLMs) have opened up numerous novel possibilities, but due to their significant computational requirements their ubiquitous use remains challenging. Some of the most useful applications require processing large numbers of samples at a time and using long contexts, both significantly increasing the memory communication load of the models. We introduce SparQ Attention, a technique for increasing the inference throughput of LLMs by reducing the memory bandwidth requirements within the attention blocks through selective fetching of the cached history. Our proposed technique can be applied directly to off-the-shelf LLMs during inference, without requiring any modification to the pre-training setup or additional fine-tuning. We show how SparQ Attention can decrease the attention memory bandwidth requirements up to eight times without any loss in accuracy by evaluating Llama 2 and Pythia models on a wide range of downstream tasks.
4
231
Sam Maddrell-Mander retweeted
T-24 hours until #AdventOfCode 2023!!!
10
93
17
639
70,458
It’s the most wonderful time of the year 🎵
For the ninth(!) year in a row, I will be releasing 25 programming puzzles over the first 25 days of December over on adventofcode.com! It's completely free thanks to AoC++ supporters and thanks to sponsors. If you want to check it out early, all past puzzles are still up!
4
133
We think multitask learning is key for the next generation of molecular models - our paper presents datasets and baselines for jointly learning quantum, chemical and biological properties of molecules. Check it out!
Laying the foundations for foundational models in molecular learning: presenting seven novel labelled datasets for supervised learning covering ~100m molecules and 3,000+ sparsely defined tasks - by @graphcoreai and our research partners. hubs.la/Q025Pp_F0
1
4
212
Sam Maddrell-Mander retweeted
There are not many human beings in the entire world that have as much holistic full stack LLM experience as this man here. 👇 Tons of wisdom in these slides here by @hwchung27.
I gave a talk at Seoul National University. I titled the talk “Large Language Models (in 2023)”. This was an ambitious attempt to summarize our exploding field. Video: youtu.be/dbo3kNKPaUA Slides: docs.google.com/presentation… Trying to summarize the field forced me to think about what really matters in the field. While scaling undeniably stands out, its far-reaching implications are more nuanced. I share my thoughts on scaling from three angles: 1) Change in perspective is necessary because some abilities only emerge at a certain scale. Even if some abilities don’t work with the current generation LLMs, we should not claim that it doesn’t work. Rather, we should think it doesn’t work yet. Once larger models are available many conclusions change. This also means that some conclusions from the past are invalidated and we need to constantly unlearn intuitions built on top of such ideas. 2) From first-principles, scaling up the Transformer amounts to efficiently doing matrix multiplications with many, many machines. I see many researchers in the field of LLM who are not familiar with how scaling is actually done. This section is targeted for technical audiences who want to understand what it means to train large models. 3) I talk about what we should think about for further scaling (think 10000x GPT-4 scale). To me scaling isn’t just doing the same thing with more machines. It entails finding the inductive bias that is the bottleneck in further scaling. I believe that the maximum likelihood objective function is the bottleneck in achieving the scale of 10000x GPT-4 level. Learning the objective function with an expressive neural net is the next paradigm that is a lot more scalable. With the compute cost going down exponentially, scalable methods eventually win. Don’t compete with that. In all of these sections, I strive to describe everything from first-principles. In an extremely fast moving field like LLM, no one can keep up. I believe that understanding the core ideas by deriving from first-principles is the only scalable approach.
2
27
181
49,718
This is really cool, their new GNN design allows each node to decide how to propagate information both to and from its neighbours. Properly dynamic message passing as best suited to any given task. Can’t wait to try it out!
bit.ly/3PLHs5s **Cooperative GNNs** with @benfinkelshtein @hxyscott @ismaililkanc In message-passing GNNs, each node is updated based on messages from its neighborhood. In Co-GNNs, every node can choose to either ‘listen’, ‘broadcast’, ‘listen&broadcast’, or ‘isolate’
1
3
126
Sam Maddrell-Mander retweeted
📢 Exciting News! 📢 We're launching a monthly lecture series on Generative AI in Drug Discovery hosted by Prof. Michael Bronstein & Prof. Bruno Correia, where researchers will present their latest work. Our inaugural session is 📅 next Friday 📅, by @befcorreia and @mmbronstein on Harnessing Geometry For Molecular Design 📆 October 6, 11 am ET / 5 pm CET Join us LIVE on Zoom: us02web.zoom.us/j/8349010708… All lectures will be uploaded on Youtube youtube.com/@vant_ai #AIDrugDiscovery #GenerativeAI #GenAI #GeometricDeepLearning #tpd #biotech #computationalchemistry #compchem #compbio #AI #machinelearning #drugdiscovery
36
3
119
24,203
Sam Maddrell-Mander retweeted
Inspired by the wonderful "Computational Discovery on Jupyter" by By Neil J. Calkin, Eunice Y.S. Chan, and @corless_rob and the gallery of bohemianmatrices.com. Each plot represents eigenvalues of the corresponding matrix. Made with #python, @matplotlib and @numpy_team .
24
220
22
1,173
127,462
Sam Maddrell-Mander retweeted
Just added "learning on graphs and other geometries & topologies"! Primary areas should be modifiable until the final paper deadline.
2
8
84
14,048