iiinova — where your digital self keeps evolving

Rasmus Aagaard @rasgaard.com · 22/09/2026

As I'm approaching the end of my 1st year as a PhD student I'm struggling a bit with how coding agents affect my learning. In that light I found Tim Dettmers' new post very good and inspiring! "Let go of how you work. Not who you are." timdettmers.com/2026/09/21/d...

010

Rasmus Aagaard @rasgaard.com · 16/08/2026

Was at @ijcai.org yesterday for the GLOW workshop glow-ijcai-2026.github.io/glow-ijcai-2... :)

160

Rasmus Aagaard @rasgaard.com · 02/08/2026

I guess my account is now dedicated to posting my latest dyno #bouldering

130

Rasmus Aagaard @rasgaard.com · 31/07/2026

That first move took about a million attempts #bouldering

040

Rasmus Aagaard @rasgaard.com · 20/07/2026

Probably my hardest climb so far. Super stoked:) Guess the grade? 🧐 #bouldering

020

Rasmus Aagaard @rasgaard.com · 18/06/2026

Got an extended abstract paper accepted at the GLOW workshop at IJCAI :) Turns out you can just delete 6 layers from the Whisper encoder and it doesn’t hurt transcription performance all that much (but you have to be pretty specific about which layers)

040

Rasmus Aagaard @rasgaard.com · 15/04/2026

Researchers identifies "Super Weights" which are at most a handful of weights (amongst billions!). Models end up generating complete gibberish if these weights are removed, underlining the brittleness of LLMs and importance and careful consideration of outlier features. arxiv.org/pdf/2411.07191

2172

Rasmus Aagaard @rasgaard.com · 26/02/2026

Claude Code Copenhagen meetup 🍻🚀

110

Rasmus Aagaard @rasgaard.com · 20/02/2026

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

393

Rasmus Aagaard @rasgaard.com · 18/02/2026

Excited to be in fantastic company amongst the speakers for IDA Driving AI 2026 :) ida.dk/driving-ai/t...

150

Rasmus Aagaard @rasgaard.com · 17/02/2026

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

020

Rasmus Aagaard @rasgaard.com · 16/02/2026

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...

0122

Rasmus Aagaard @rasgaard.com · 10/02/2026

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/...

031

Rasmus Aagaard @rasgaard.com · 03/02/2026

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

111

Rasmus Aagaard @rasgaard.com · 23/01/2026

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

131

Rasmus Aagaard @rasgaard.com · 20/01/2026

Perplexity is often used as a measure for how well a language model fits data. However, it can give you a false sense of security when performing model compression such as quantization and pruning. arxiv.org/pdf/2310.01382

080

Rasmus Aagaard @rasgaard.com · 16/01/2026

LiteASR (arxiv.org/pdf/2502.20583) finds that latency caused by the Whisper encoder and decoder varies significantly at different settings. Compressing the decoder is common (distill/turbo variants) but the encoder is largely ignored. The authors look into encoder compression - super cool work!

160

Rasmus Aagaard @rasgaard.com · 15/01/2026

From the recent Ministral 3 paper (arxiv.org/pdf/2601.08584) Love these deceptively simple pruning strategies

020

Rasmus Aagaard @rasgaard.com · 09/12/2025

skaftenicki.github.io/dtu_mlops/

150

Rasmus Aagaard @rasgaard.com · 05/11/2025

At Digital Tech Summit today and tomorrow :)

040

Rasmus Aagaard @rasgaard.com · 15/10/2025

Testing out offline language model (SmolLM2-360M) along with in-browser database (pglite.dev with pgvector) for a completely local RAG system that runs on a phone. Thanks @xenova.bsky.social for building so many demos and examples to build on top of! :)

130

Rasmus Aagaard @rasgaard.com · 07/10/2025

pytorch.org/blog/when-qu... Sparsity + quantization is a super exciting direction

000

Rasmus Aagaard @rasgaard.com · 02/10/2025

DHH's fireside chat at DTU was so great. My main takeaway is that it's an amazing time to be optimistic and ambitious.

000

Rasmus Aagaard @rasgaard.com · 01/10/2025

First official day starting my new position as industrial PhD student at DTU.dk and Laerdal.com 🎉 So great to be back on campus. Be sure to reach out if you're interested in model compression and edge deployment of deep neural networks 🔥🚀

150

Rasmus Aagaard @rasgaard.com · 18/09/2025

Times were simpler when BERT was considered enormous proceedings.neurips.cc/paper/2020/f...

010

Rasmus Aagaard @rasgaard.com · 02/09/2025

Haven't seen this before but it sounds super interesting. If your LLM application outputs a limited set of tokens why not prune away all the other unused output connections? arxiv.org/pdf/2411.17713

050

Rasmus Aagaard @rasgaard.com · 02/09/2025

Tried to naively use Nano Banana to make a "Where's Waldo"-type image. Can you spot the monkey?

030

Rasmus Aagaard @rasgaard.com · 27/08/2025

Had a really good time at d3aconference.dk :)

010

Rasmus Aagaard @rasgaard.com · 21/08/2025

Today's mission is to put an ML model on this little guy

010

Rasmus Aagaard @rasgaard.com · 12/08/2025

Visited Laerdal's office in Stavanger to share ideas on how my upcoming ph.d. on model compression and edge deployment can potentially contribute to several projects :) Super exciting, and Stavanger is a lovely city to explore

110

Rasmus Aagaard @rasgaard.com · 06/08/2025

I wonder if we're going to see children somehow mask themselves as AI scraping bots to avoid internet age restrictions #showerthought

110

Rasmus Aagaard @rasgaard.com · 05/08/2025

KittenTTS is under 25 megabytes at ~15M parameters. Quality is crazy good for this size. I'm super excited about the opportunities here. www.reddit.com/r/LocalLLaMA...

010

Rasmus Aagaard @rasgaard.com · 26/07/2025

Great day at the @aicentre.dk for the Pre-ACL Workshop with lots of good talks, posters and pastries 🥐

050

Rasmus Aagaard @rasgaard.com · 30/06/2025

"pruning is the equivalent of lobotomizing the LLM" Not sure how I feel about that analogy medium.com/@eaddario/sq...

010

Rasmus Aagaard @rasgaard.com · 24/06/2025

Just completed Quantization Fundamentals with Hugging Face on learn.deeplearning.ai :) Already looking forward to the more advanced course! @deeplearningai.bsky.social

030

Rasmus Aagaard @rasgaard.com · 20/06/2025

CPH NLP Symposium 🤗🔥 cphnlp.github.io

040

Rasmus Aagaard @rasgaard.com · 29/05/2025

Haven't ordered new books in a good while because I have been using the library so much but now that I got a few gift cards for my birthday I checked out these :) #booksky

040

Rasmus Aagaard @rasgaard.com · 23/05/2025

I have been working on bootstrapping AI evals before product launch and came up with something like this. Using QA to inform "problematic on purpose" synthetic data that can validate LLM-as-a-Judge. Would love to chat with people who have experience with this :) #mlsky #buildinpublic

020

Rasmus Aagaard @rasgaard.com · 22/05/2025

I usually don't really like off-the-shelf AI evaluation tools. Like, what do you even mean when you measure "groundedness"? Using AI Toolkit for VS Code for LLM-as-a-Judge alignment evals has been the first time I have had an okay experience with a tool like this.

040

Rasmus Aagaard @rasgaard.com · 20/05/2025

Gemma 3n E2B available through Edge Gallery: github.com/google-ai-ed... Running on my Nothing Phone 1 here. #MLSky

110

Rasmus Aagaard @rasgaard.com · 20/05/2025

Driving AI hos @ida.dk er i dag. Glæder mig især til at høre om: - Region Hs erfaring med AI til håndtering af hudkræft - Generativ AIs miljøpåvirkning af folk fra ddsc.io - Vision Transformers fra Capacit som havde et maks-nørdet oplæg sidste år om DSPy. #dkai #dktech

010

Rasmus Aagaard @rasgaard.com · 19/05/2025

Vil gerne blive bedre til at dele hvad jeg har gang i. Også selvom det måske er en smule upoleret. Her er sidste nyt: Whisper-large-v3-turbo som web app, 100% lokalt så din data aldrig forlader din computer. Virker bedst i Chrome! (pga. webgpu support) rasgaard.com/p/transcribe/

020

Rasmus Aagaard @rasgaard.com · 18/05/2025

Har leget med en idé om annotering som del af servicen: - Simpel Whisper-app der kører lokalt -> gratis med Transformers.js - Fokusér på UI/UX så det er lækkert og nemt at bruge (nok det sværeste) - Gør det muligt at rette fejl i teksterne - Optional: Indsend rettelser + lyd til database

100

Rasmus Aagaard @rasgaard.com · 08/05/2025

Demoen her llm-tool-calling.nico.dev som ser rimelig lovende ud

011

Rasmus Aagaard @rasgaard.com · 07/05/2025

Er imponeret over hvor godt Qwen3-0.6B kører med Transformers.js. Det hele er lokalt i browseren(!) Det her er et minimalt hjemmebrygget eksempel for at forstå lidt af hvad der sker under motorhjelmen. @xenova.bsky.social har lavet en bedre demo :) Prøv den her: huggingface.co/spaces/webml...

121

Rasmus Aagaard @rasgaard.com · 06/05/2025

PyData Copenhagen @ Google Denmark's offices 🐍🔥

020

Rasmus Aagaard @rasgaard.com · 06/05/2025

Resultater på Coral test: CER: 15.93% · WER: 34.30% Sammenlignet med modellerne som er 40x større er jeg egentlig virkelig imponeret. Næste skridt vil være at træne på Common Voice. Men indtil da kan modellen findes her: huggingface.co/rasgaard/whi... :)

000

Rasmus Aagaard @rasgaard.com · 05/05/2025

I have found that using "Show more/less like this" to tune my Discover feed actually works quite well!

000

Rasmus Aagaard @rasgaard.com · 05/05/2025

Træner Whisper-tiny på Coral-datasættet. Er spændt på hvor meget man kan få ud af en så lille (38M) model. Som bonus er det fuldkommen gratis: Det kan lade sig gøre på free-tier GPU compute på Kaggle :)

130