Rasmus Aagaard's Avatar

Rasmus Aagaard

@rasgaard.com

Industrial PhD Student @ Laerdal.com & DTU.dk, research in efficient machine learning, edge deployment and model compression

326
Followers
540
Following
175
Posts
15.11.2024
Joined
Posts Following

Latest posts by Rasmus Aagaard @rasgaard.com

Post image Post image

Claude Code Copenhagen meetup ๐Ÿป๐Ÿš€

26.02.2026 16:03 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0
Preview
Open source er en strategisk nรธdvendighed i AI: Sรฅdan kommer vi med i kaplรธbet | DataTech Pรฅ tvรฆrs af it-branchen spiller open source software en fundamental rolle for bรฅde offentlige og private lรธsninger. Dette gรธr sig ogsรฅ gรฆldende i AI-verdenen.Transformer-arkitekturen, grundstenen i nu...

Together with fellow ddsc.io members I helped a bit with this piece on @ing.dk :)

25.02.2026 10:33 ๐Ÿ‘ 4 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Post image

First blogpost of my cofounder post-Delhi on the deafening silence of public/political sphere as we come closer to another industrial revolution. anastasiastasenko.substack.com/p/the-countr...

23.02.2026 19:08 ๐Ÿ‘ 73 ๐Ÿ” 18 ๐Ÿ’ฌ 3 ๐Ÿ“Œ 8
The Danoliterate Generative Large Language Model Benchmark

Strongly agree! Sounds like a fantastic idea for a special course :) Maybe you can take inspiration from Danoliterate danoliterate.compute.dtu.dk which was a master's project at DTU

23.02.2026 14:01 ๐Ÿ‘ 3 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

Helt klart. Der er stadig tonsvis af quirks. Jeg er mest begejstret for tendensen der bliver vist. At software og hardware begge udvikler sig i retning af at lokal inference bliver konkurrencedygtigt ift. cloud.

20.02.2026 12:01 ๐Ÿ‘ 0 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Post image Post image

LLM cloud inference dominates usage, but should it? Local models and accelerators have improved massively over recent years.

Perfect routing to best local model "reduce energy consumption by 80.4%, compute by 77.3%, and cost by 73.8% versus cloud-only deployment"

arxiv.org/pdf/2511.07885

20.02.2026 10:27 ๐Ÿ‘ 9 ๐Ÿ” 3 ๐Ÿ’ฌ 3 ๐Ÿ“Œ 1
Video thumbnail

Introducing โœจTiny Ayaโœจ, a family of massively multilingual small language models built to run where people actually are.

Tiny Aya delivers strong multilingual performance in 70+ global languages in a 3.35B parameter model, efficient enough to run locally, even on a phone.

17.02.2026 15:15 ๐Ÿ‘ 97 ๐Ÿ” 15 ๐Ÿ’ฌ 2 ๐Ÿ“Œ 5
Post image

Excited to be in fantastic company amongst the speakers for IDA Driving AI 2026 :)

ida.dk/driving-ai/t...

18.02.2026 10:03 ๐Ÿ‘ 5 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0
Post image

The paper suggests that poor reasoning abilities is one failure mode that comes from removing deeper layers. But yeah, wonder what else unexpectedly takes a hit

17.02.2026 08:09 ๐Ÿ‘ 2 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Post image

Due to residual structures in Transformer-models it's possible that many layers contribute very little to the downstream performance of the network, allowing for removal those redundant layers with little impact.

openreview.net/pdf/72eeca23...

16.02.2026 12:15 ๐Ÿ‘ 12 ๐Ÿ” 2 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 2
Video thumbnail

Local AI inference through the browser is great. Transformers.js makes it feasible for someone like me who has little web dev experience :)

Here's a simple transcription tool using Whisper-large-v3-turbo, making use of your local hardware and ensuring privacy.
rasgaard.com/webai-stuff/...

10.02.2026 08:30 ๐Ÿ‘ 3 ๐Ÿ” 1 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

ofc there's already a python plotting library for this sort of thing

github.com/jndean/LossR...

08.02.2026 16:33 ๐Ÿ‘ 7 ๐Ÿ” 4 ๐Ÿ’ฌ 2 ๐Ÿ“Œ 0
This bullshit has been going on for generations. A 1996 Simpsons episode tried to teach us.
This bullshit has been going on for generations. A 1996 Simpsons episode tried to teach us. YouTube video by Gangstagrass

This is crazy:

youtu.be/v81b0XllvgI

05.02.2026 17:41 ๐Ÿ‘ 375 ๐Ÿ” 130 ๐Ÿ’ฌ 14 ๐Ÿ“Œ 15
AI-utopierne: En morgen i Elon Musks drรธmmeverden
AI-utopierne: En morgen i Elon Musks drรธmmeverden YouTube video by Verden forsรธgt forklaret

www.youtube.com/watch?v=b64v...

06.02.2026 12:05 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
A photograph of sunny Copenhagen in the summer!

A photograph of sunny Copenhagen in the summer!

๐Ÿ“ข I am hiring a highly-motivated Ph.D student at the University of Copenhagen to work on tokenization-free NLP.

Read our previous work in this topic: aclanthology.org/2025.emnlp-m...
aclanthology.org/2023.emnlp-m...
openreview.net/forum?id=FkS...

Apply by March 8: employment.ku.dk/phd/?show=1563

04.02.2026 10:40 ๐Ÿ‘ 19 ๐Ÿ” 9 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Post image

Sometimes a single line changes the reproducibility-game completely:))

The paper is really cool though: Can the transformer+conv Mimi (NAC) decoder be replaced with a purely transformer-based one?

Answer is yes, and it's 10x faster! Benchmarked on actual mobile hardware.

arxiv.org/pdf/2601.20094

03.02.2026 09:24 ๐Ÿ‘ 1 ๐Ÿ” 1 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0
Post image

WE ARE BACK! ๐Ÿš€๐Ÿš€๐Ÿš€

๐Ÿ“ข #databeers #copenhagen x Absalon is back on March 10th, SAVE THE DATE!

Many thanks to our sponsors, the Danish Data Science Academy and the Danmarks Tekniske Universitet

Speaker announcement and tickets out soon, stay tuned!

02.02.2026 16:15 ๐Ÿ‘ 15 ๐Ÿ” 11 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 1

Slowly but surely getting a better intuition for how and when compression methods such as quantization is beneficial and, perhaps more crucially, detrimental to overall model efficiency

02.02.2026 11:13 ๐Ÿ‘ 2 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Preview
Understanding Efficiency: Quantization, Batching, and Serving Strategies in LLM Energy Use Large Language Models (LLMs) are increasingly deployed in production, contributing towards shifting the burden in terms of computational resources and energy demands from training to inference. While ...

"numerical precision reduction yields the most benefit in the prefill phase of large models, where compute dominates. In contrast, the decode phase remains memory-limited, and aggressive quantization (e.g., int8 or int4) may incur overheads that outweigh theoretical savings"

02.02.2026 11:11 ๐Ÿ‘ 16 ๐Ÿ” 0 ๐Ÿ’ฌ 2 ๐Ÿ“Œ 2
Post image

It took me weeks, but finally it's there: an overlong blogpost on synthetic pretraining. vintagedata.org/blog/posts/s...

01.02.2026 17:53 ๐Ÿ‘ 86 ๐Ÿ” 21 ๐Ÿ’ฌ 3 ๐Ÿ“Œ 2
Preview
Ny open source-baseret platform kan fjerne de store tech-virksomheders greb om danske skoler ยท Dataetisk Tรฆnkehandletank Photo: publiccode.eu Open source-vรฆrktรธjer i skoler og kommuner vil ikke kun spare samfundet penge og...

"Nรฅr 32 % af alle danske skoler bruger OS2 Skole, gรฅr det i nul. Omkostningerne ved OS2 Skole vil ikke vรฆre de licenser, der betales til Big Tech, men penge, der betales til de [skattebetalende] personer, der ansรฆttes til at vedligeholde og udvikle systemet."
dataethics.eu/da/new-open-...

27.01.2026 12:15 ๐Ÿ‘ 2 ๐Ÿ” 1 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Post image

This reddit post made me laugh

26.01.2026 10:01 ๐Ÿ‘ 19 ๐Ÿ” 4 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Preview
Kan man klare sig uden amerikansk 'big tech'? Her er nogle af alternativerne Der findes masser af alternativer til de store amerikanske platforme og programmer, siger dataetisk rรฅdgiver og forbrugerrรฅd.

Kan man klare sig uden amerikansk 'big tech'? Her er nogle af alternativerne https://www.dr.dk/nyheder/viden/teknologi/kan-man-klare-sig-uden-amerikansk-big-tech-her-er-nogle-af-alternativerne

#dk #nyheder #dkbot

25.01.2026 19:32 ๐Ÿ‘ 22 ๐Ÿ” 6 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0

In Grenoble/France where I work, a steep rise in bicycle commutes (green) happened, mainly because of strong support from the local municipality through infrastructure changes - bike lanes! This coincides with a drop in car usage (red), while walking (yellow) and public transp. (black) stayed flat.

25.01.2026 11:43 ๐Ÿ‘ 52 ๐Ÿ” 6 ๐Ÿ’ฌ 2 ๐Ÿ“Œ 0
Preview
Creating the worldโ€™s best model for Scandinavian languages The Tech Collective offers the worldโ€™s best embedding model for Danish, Swedish, and Norwegian, transforming text into numerical data for AI applications.

Der er en artikel om det her thetechcollective.eu/insights/cre... :)

23.01.2026 18:35 ๐Ÿ‘ 1 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0

The winners were,
- CoRal: huggingface.co/CoRal-project/
- Top Scandinavian text embedding model: huggingface.co/jealk/TTC-L2...
- EuroEval: euroeval.com

23.01.2026 17:59 ๐Ÿ‘ 4 ๐Ÿ” 0 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0
Post image

Had a great time at @danskerhverv.dk today where the Danish Data Science Community (ddsc.io - join the Slack!) revealed the winners of the first Open Source Awards

23.01.2026 17:54 ๐Ÿ‘ 3 ๐Ÿ” 1 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 1
Preview
Building Offline-First AI for Community Health Workers in West Africa Weย build reasoning models for advanced context engineering in the agentic AI

Super excited to see focus on on-device AI for healthcare workers.

There's an enormous technology/impact gap in resource-constrained environments. One that Laerdal also works extensively with (@laerdalgh.bsky.social). Offline-first is essential for bridging this gap.

Great work over at pleias!

22.01.2026 15:01 ๐Ÿ‘ 2 ๐Ÿ” 0 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 0
Day 121 โ€“ Vibe Coding an App Until I Make $1,000,000 | Revenue: $30,790.95
Day 121 โ€“ Vibe Coding an App Until I Make $1,000,000 | Revenue: $30,790.95 YouTube video by BridgeMind

Marketing your app is really hard and getting harder now days. So this guy started live vlogging his vibe-coding.

Day 121 โ€“ Vibe Coding an App Until I Make $1,000,000 | Revenue: $30,790.95

www.youtube.com/watch?v=4Oml...

21.01.2026 11:27 ๐Ÿ‘ 12 ๐Ÿ” 1 ๐Ÿ’ฌ 1 ๐Ÿ“Œ 0
Preview
Microsoft vokser i Danmark trods Trumps trusler: ยปDer er en risiko for spionageยซ | Ingeniรธren Tech-giganten vil bygge tre nye datacentre i Vestjylland, der skal skabe op mod 500 nye arbejdspladser. Eksperter frygter en sikkerhedsrisiko, og at Danmarks afhรฆngighed af amerikansk tech vil vokse.

"Microsoft vokser i Danmark trods Trumps trusler: ยปDer er en risiko for spionageยซ | Ingeniรธren"

Forstรฅ nu alvoren, #dkpol

#betalingsmur

https://ing.dk/artikel/microsoft-vokser-i-danmark-trods-trumps-trusler-der-er-en-risiko-spionage

20.01.2026 06:58 ๐Ÿ‘ 5 ๐Ÿ” 3 ๐Ÿ’ฌ 0 ๐Ÿ“Œ 1