Tijmen Blankevoort (@TiRune) / X

Tijmen Blankevoort

512 posts

Tijmen Blankevoort

@TiRune

Deep Learning Researcher Nvidia - Efficiency/Numerics

Amsterdam, The Netherlands

Joined May 2009

Tijmen Blankevoort
@TiRune
Jul 13, 2025
I recently made the news because of a doc I wrote in Meta’s GenAI organization. ‘The Information’ wrote about it as if I did a big raging ‘mic drop’ before leaving the company. Nothing could be further from the truth - so setting the record straight here. open.substack.com/pub/blankevoor…
18K
Tijmen Blankevoort
@TiRune
May 1, 2020
Our new paper is up on Arxiv! We found out that rounding-to-nearest in #neuralnetwork #quantization is actually really really bad. Our new method AdaRound requires no training, little data, and pushes networks to 4 bit weights with high accuracy! arxiv.org/abs/2004.10568
Tijmen Blankevoort
@TiRune
May 15, 2020
We put our paper "Bayesian Bits: Unifying Quantization and Pruning" on Arxiv! lnkd.in/dHezcH7 In this work, we train a network to automatically choose between different bit-widths for each layer, and structured-prune the network too. All in one principled Bayesian method.
Tijmen Blankevoort
@TiRune
Apr 15, 2020
So im doing a webinar on neural network quantization! Please join if you're interested in running your deep learning models on device :)
Qualcomm Research & Technologies
@QCOMResearch
Apr 14, 2020
Interested in shrinking #AI models so that they run faster and consume less power? Join our webinar to learn about the latest quantization techniques from @Qualcomm AI Research. event.on24.com/wcc/r/2232517/…
Tijmen Blankevoort
@TiRune
Jun 3, 2024
Our new Bitune method is out - improving results on instruction-tuning! We add bidirectional attention to a decoder-only architecture, and concoct an IT-specific PEFT method for Q&A and instruction following settings
2.1K
Tijmen Blankevoort
@TiRune
Aug 5, 2025
Replying to @tydsh
Damn, even Evan Miller’s blog post got the citation - but not our original work on it: arxiv.org/abs/2306.12929 We should’ve done better PR 😂
arxiv.org
Quantizable Transformers: Removing Outliers by Helping Attention...
Transformer models have been widely adopted in various domains over the last years, and especially large language models have advanced the field of AI significantly. Due to their size, the...
612
Tijmen Blankevoort
@TiRune
Dec 5, 2023
Talking in this Dutch podcast, about how a general AI, something that can do many tasks just like a human (and perhaps better) might not be as far away as you might think. 😃 Specifically RL+LLMs has the potential ability to supercharge the current model performance.
De Technoloog
@detechnoloog
Dec 5, 2023
Na het interne conflict bij OpenAI is het gesprek over ‘artificial general intelligence’ (AGI) relevanter dan ooit. @hmblank en @ikbenechtben zoeken samen met @TiRune uit wat het precies is, en hoe slim AI nu echt kan worden. bnr.nl/podcast/de-tec…
442
Tijmen Blankevoort
@TiRune
Sep 18, 2024
Replying to @tydsh
I fully agree on the ‘2-layer MLP can fit anything’ argument. These theoretical fitting capacity papers are not super useful. I wonder what happens if we regularize the in-between CoT outputs to be shorter. Humans might be forced to strong abstractions as we are ‘stupid’
27K
Tijmen Blankevoort
@TiRune
May 2, 2020
Replying to @honualx and @gsautiere
Not yet, but we released our model efficiency toolkit open source today! Adaround code might find its way in there too! github.com/quic/aimet
Tijmen Blankevoort
@TiRune
Jul 16, 2025
Replying to @rosstaylor90
I 100% agree on this. Inside Meta - there are so many insanely talented people. I don’t think any attracted talent is 10-100x more impactful than the people they already had, or lost to other companies in the process.
246
Tijmen Blankevoort
@TiRune
Jul 21, 2024
I worked on this in Qualcomm for a while. You can take a phone that has a great ISP, capture the raw bayer input and pretty output and train with that. Works very well, but it’s a bit expensive energy/silicon-wise. Also very nice if you e.g. have dead pixels, can still work 😄
78
Tijmen Blankevoort
@TiRune
Jun 16, 2021
Everything you need to know about neural network quantization, right here! :)
Guillaume Sautière
@gsautiere
Jun 16, 2021
My colleagues @mnagel87 @mfournarakis @TiRune et al from our model efficiency team released a white paper on neural network quantization, with all there is to know to write your own Post-Training Quant / Quant-Aware pipelines. Lot of gold to dig through arxiv.org/abs/2106.08295
Tijmen Blankevoort
@TiRune
Jul 22, 2024
Replying to @3scorciav @y_m_asano and 2 others
Yup! Just the .raw bayer or nonacell inputs. I’ve seen many camera algorithms that work on that. A neural network can learn all the ISP stuff directly. Even 4 bit quantized networks can do this 😄
65
Tijmen Blankevoort
@TiRune
Oct 19, 2018
Replying to @TiRune @SethManfield and 2 others
I'll add 2 promo flooded strands. Got both of them in two goodie bags from a gp.