Martin Elstner
@martin.elstner.dev
AI and ML engineering. Search and recommendation. Founder of Elstner Analytics. Helping companies solve their data issues
https://elstner.dev
https://elstner.dev
Pinned
Martin Elstner
@martin.elstner.dev
· Jun 25
If your company implements stringent rules regarding the use of Generative AI tools, there's a high probability employees will find ways to circumvent them.
Classrooms as cluster-mixing machines because of poor air? Color me surprised.
This Nat. Comm. study found: In schools, it is not only direct contact, mainly the “shared classroom exposure.” Poor air quality increases the risk exponentially.
www.nature.com/articles/s41...
This Nat. Comm. study found: In schools, it is not only direct contact, mainly the “shared classroom exposure.” Poor air quality increases the risk exponentially.
www.nature.com/articles/s41...
The relative contribution of close-proximity contacts, shared classroom exposure and indoor air quality to respiratory virus transmission in schools - Nature Communications
The relative importance of close-proximity interactions, shared space and air quality to the transmission of respiratory viruses is not well understood. Here, the authors investigate this question by ...
www.nature.com
December 1, 2025 at 12:39 PM
Classrooms as cluster-mixing machines because of poor air? Color me surprised.
This Nat. Comm. study found: In schools, it is not only direct contact, mainly the “shared classroom exposure.” Poor air quality increases the risk exponentially.
www.nature.com/articles/s41...
This Nat. Comm. study found: In schools, it is not only direct contact, mainly the “shared classroom exposure.” Poor air quality increases the risk exponentially.
www.nature.com/articles/s41...
Teile alle genannten Punkte, glaube aber, das viele der aktuellen Probleme eher auf “you are holding it wrong” fußen. Beispiel Bildungswesen: copy-pasta von KI-Antworten ist eine dumme Idee (und das ist so offensichtlich, das ich mich frage wieso man das diskutiert).
Heute vor exakt 3 Jahren wurde ChatGPT veröffentlicht.
Statt der Utopie haben wir:
- Damit überforderte Bildungssysteme
- Propagandamaschinen turned to 11
- Scientific Fraud
- GDP Steigerung nur wegen Kreislaufgeschäften der großen Tech Firmen
- Dead Internet
- "Ich habe GPT gefragt ..."
Cool.
Statt der Utopie haben wir:
- Damit überforderte Bildungssysteme
- Propagandamaschinen turned to 11
- Scientific Fraud
- GDP Steigerung nur wegen Kreislaufgeschäften der großen Tech Firmen
- Dead Internet
- "Ich habe GPT gefragt ..."
Cool.
December 1, 2025 at 12:27 PM
Teile alle genannten Punkte, glaube aber, das viele der aktuellen Probleme eher auf “you are holding it wrong” fußen. Beispiel Bildungswesen: copy-pasta von KI-Antworten ist eine dumme Idee (und das ist so offensichtlich, das ich mich frage wieso man das diskutiert).
Beautiful typographic map
#30daymapchallenge · Jour 30 · Relooking
🌍 Europe · Europa · Ευρώπη · Європа
🔍 Zoomez pour découvrir les villes d'Europe dans leur langue.
🔽 L'ancienne carte, en commentaire, utilisait un texte de substitution, réalisée pour le #30daymapchallenge 2024.
#cartographie #map #europe #europa
🌍 Europe · Europa · Ευρώπη · Європа
🔍 Zoomez pour découvrir les villes d'Europe dans leur langue.
🔽 L'ancienne carte, en commentaire, utilisait un texte de substitution, réalisée pour le #30daymapchallenge 2024.
#cartographie #map #europe #europa
December 1, 2025 at 12:10 PM
Beautiful typographic map
Creative work/parenting 2025 festive edition: ChatGPT creates the Mandala for K3 and you need Gemini nano banana to get the spelling right.
November 30, 2025 at 9:58 AM
Creative work/parenting 2025 festive edition: ChatGPT creates the Mandala for K3 and you need Gemini nano banana to get the spelling right.
Really nice how you can get music notation for your kids to play along for the holidays:
November 27, 2025 at 1:41 PM
Really nice how you can get music notation for your kids to play along for the holidays:
Finetuning for coding tasks probably isn’t worth the effort. Your own model can be great for classification, dense vector search or other small tasks. For complex stuff like ‘coding’ in a broader sense, it’s not holding up to the promises.
Custom training a model on a specific codebase mostly doesn't work from what I've seen
November 26, 2025 at 4:35 PM
Finetuning for coding tasks probably isn’t worth the effort. Your own model can be great for classification, dense vector search or other small tasks. For complex stuff like ‘coding’ in a broader sense, it’s not holding up to the promises.
Skimming the comments, acceptance is still far away
Is this platform still massively against AI or has it moved more towards acceptance?
November 26, 2025 at 11:06 AM
Skimming the comments, acceptance is still far away
Reposted by Martin Elstner
The water thing is mostly overblown andymasley.substack.com/p/the-ai-wat...
The electricity and CO2 impact is real and deserves serious attention, which is why it's annoying that the water issue steals so much of that attention
The electricity and CO2 impact is real and deserves serious attention, which is why it's annoying that the water issue steals so much of that attention
The AI water issue is fake
On the national, local, and personal level
andymasley.substack.com
November 26, 2025 at 7:09 AM
The water thing is mostly overblown andymasley.substack.com/p/the-ai-wat...
The electricity and CO2 impact is real and deserves serious attention, which is why it's annoying that the water issue steals so much of that attention
The electricity and CO2 impact is real and deserves serious attention, which is why it's annoying that the water issue steals so much of that attention
Reposted by Martin Elstner
Building datasets to train smaller, task-focused models used to be incredibly time-consuming.
Very excited to see SAM3 massively lower that barrier. Describe the class you want to detect and get annotated datasets automatically!
Try it yourself: huggingface.co/datasets/uv-...!
Very excited to see SAM3 massively lower that barrier. Describe the class you want to detect and get annotated datasets automatically!
Try it yourself: huggingface.co/datasets/uv-...!
November 21, 2025 at 1:30 PM
Building datasets to train smaller, task-focused models used to be incredibly time-consuming.
Very excited to see SAM3 massively lower that barrier. Describe the class you want to detect and get annotated datasets automatically!
Try it yourself: huggingface.co/datasets/uv-...!
Very excited to see SAM3 massively lower that barrier. Describe the class you want to detect and get annotated datasets automatically!
Try it yourself: huggingface.co/datasets/uv-...!
Yes, LLMs can automate parts of consumer research. But if you tell ChatGPT your SaaS idea and ask for feedback, don’t be surprised when it pats you on the back. The real question is: how do you get beyond the fluff?
October 16, 2025 at 10:40 PM
Yes, LLMs can automate parts of consumer research. But if you tell ChatGPT your SaaS idea and ask for feedback, don’t be surprised when it pats you on the back. The real question is: how do you get beyond the fluff?
Interesting point (still problematic tasks) but wrong conclusions. I’m again and again puzzled how intelligent people neglect the utility of LLMs/VLMs. They are far from flawless but can do so much already. If their abilities don’t overlap with your problem space, don’t use them there.
Can generative AI be trusted to draw chemical structures? Not yet, according to two chemists who want to see the community take a tough stance against its use.
The chemistry community should ban drawing chemical structures with generative AI, chemists warn
AIs like Microsoft's Copilot, Google's Gemini and OpenAI's ChatGPT still make serious errors rendering structural formulae
www.chemistryworld.com
October 15, 2025 at 3:05 PM
Interesting point (still problematic tasks) but wrong conclusions. I’m again and again puzzled how intelligent people neglect the utility of LLMs/VLMs. They are far from flawless but can do so much already. If their abilities don’t overlap with your problem space, don’t use them there.
Have to dig deeper and maybe running a few experiments on this. But the abstract reads very promising:
arxiv.org/abs/2509.24372
arxiv.org/abs/2509.24372
Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
Fine-tuning pre-trained large language models (LLMs) for down-stream tasks is a critical step in the AI deployment pipeline. Reinforcement learning (RL) is arguably the most prominent fine-tuning meth...
arxiv.org
October 7, 2025 at 3:26 PM
Have to dig deeper and maybe running a few experiments on this. But the abstract reads very promising:
arxiv.org/abs/2509.24372
arxiv.org/abs/2509.24372
As a data guy who analyzes all sorts of effects on and from C19 since early 2020, I second this
Let me be clear. As someone who has been professionally involved with the biodefense / biosecurity sector for nearly 20 years, and has worked on biological risk analysis for Covid-19 since early 2020, reducing one's risk of being infected is not health anxiety. It's just smart.
September 26, 2025 at 11:33 AM
As a data guy who analyzes all sorts of effects on and from C19 since early 2020, I second this
It‘s unbelievable. Hunderts of “yeah, me too” replies and quotes.
I mean, there are points to be concerned about the LLM/AI space. But this 🙈 tactic
I mean, there are points to be concerned about the LLM/AI space. But this 🙈 tactic
I have *never* used ChatGPT. Not once.
I’m not saying I’m better than you… oh fuck it, I’m better than you.
I’m not saying I’m better than you… oh fuck it, I’m better than you.
September 18, 2025 at 3:59 PM
It‘s unbelievable. Hunderts of “yeah, me too” replies and quotes.
I mean, there are points to be concerned about the LLM/AI space. But this 🙈 tactic
I mean, there are points to be concerned about the LLM/AI space. But this 🙈 tactic
Probably the most interesting podcast on AI I listened to in the last months
www.youtube.com/watch?v=2Ngd...
www.youtube.com/watch?v=2Ngd...
Why Everyone Is Wrong About AI (Including You) | Benedict Evans
YouTube video by The Knowledge Project Podcast
www.youtube.com
September 3, 2025 at 6:09 PM
Probably the most interesting podcast on AI I listened to in the last months
www.youtube.com/watch?v=2Ngd...
www.youtube.com/watch?v=2Ngd...
This channel has some nuggets! Relatively new and only a few videos so far, but I read and watched the dbt and DuckDB related ones and they get much deeper than the usual 'hey look, cool new tools' fluff.
MLOps: A/B Testing with MLflow, Kafka, and DuckLake - Learn how to implement an end-to-end machine learning workflow, from data ingestion, to A/B testing and monitoring, using MLflow, Kafka and DuckLake.
MLOps: A/B Testing with MLflow, Kafka, and DuckLake
Learn how to implement an end-to-end machine learning workflow, from data ingestion, to A/B testing and monitoring, using MLflow, Kafka and DuckLake.
datalabtechtv.com
August 28, 2025 at 6:16 AM
This channel has some nuggets! Relatively new and only a few videos so far, but I read and watched the dbt and DuckDB related ones and they get much deeper than the usual 'hey look, cool new tools' fluff.
And in Germany you can directly map the medium of the back and forth to an industry:
Fax: healthcare
email: public
Slack: startups
Teams: DAX
Whatever the google tool is called this week: SME
Discord: 🙃
Fax: healthcare
email: public
Slack: startups
Teams: DAX
Whatever the google tool is called this week: SME
Discord: 🙃
What Data Management looks like in the real world 😊
Researcher: Here’s some data, can you wrangle it?
DM: Sure! Do I have everything I need?
Researcher: Yep!
The next 6 months: A series of back and forth emails about missing files, unclear values, and why things are the way that they are.
Researcher: Here’s some data, can you wrangle it?
DM: Sure! Do I have everything I need?
Researcher: Yep!
The next 6 months: A series of back and forth emails about missing files, unclear values, and why things are the way that they are.
August 27, 2025 at 8:34 PM
And in Germany you can directly map the medium of the back and forth to an industry:
Fax: healthcare
email: public
Slack: startups
Teams: DAX
Whatever the google tool is called this week: SME
Discord: 🙃
Fax: healthcare
email: public
Slack: startups
Teams: DAX
Whatever the google tool is called this week: SME
Discord: 🙃
Reposted by Martin Elstner
Today is the new semester for @db.cs.cmu.edu's Intro to Database Systems! We're going harder into material than before. More challenging projects but you can use LLMs to help. We also have 10min talks each Wed from leading DB companies: 15445.courses.cs.cmu.edu/fall2025
CMU 15-445/645 :: Intro to Database Systems (Fall 2025)
You want to know whether this is the premier course at Carnegie Mellon University on the design and implementation of database management systems? Well, it is. This course rips through data models (re...
15445.courses.cs.cmu.edu
August 25, 2025 at 2:29 PM
Today is the new semester for @db.cs.cmu.edu's Intro to Database Systems! We're going harder into material than before. More challenging projects but you can use LLMs to help. We also have 10min talks each Wed from leading DB companies: 15445.courses.cs.cmu.edu/fall2025
Reposted by Martin Elstner
Happy Friday 🌍! We are giving away one copy of 'Spatial Linear Models for Environmental Data'. To enter, simply follow me and re-post this by the end of Sunday, August 17. #spatial #environment # #data #StatsSky routledge.com/9780367183349 not associated with bsky
August 15, 2025 at 9:56 AM
Happy Friday 🌍! We are giving away one copy of 'Spatial Linear Models for Environmental Data'. To enter, simply follow me and re-post this by the end of Sunday, August 17. #spatial #environment # #data #StatsSky routledge.com/9780367183349 not associated with bsky
You can’t park here
August 14, 2025 at 7:19 AM
You can’t park here
The Open Model release from OpenAI is a good reason to check out local AI models on your device.
Is it the best model out there? Definitely not, even in it's parameter class, it may not be THE best, benchmark wise (aside: OpenAI build sizes that are uncommon, so direct comparison is harder), but
Is it the best model out there? Definitely not, even in it's parameter class, it may not be THE best, benchmark wise (aside: OpenAI build sizes that are uncommon, so direct comparison is harder), but
August 6, 2025 at 7:33 AM
The Open Model release from OpenAI is a good reason to check out local AI models on your device.
Is it the best model out there? Definitely not, even in it's parameter class, it may not be THE best, benchmark wise (aside: OpenAI build sizes that are uncommon, so direct comparison is harder), but
Is it the best model out there? Definitely not, even in it's parameter class, it may not be THE best, benchmark wise (aside: OpenAI build sizes that are uncommon, so direct comparison is harder), but
I honestly think, we need to adapt our notion of reality. The tools to reshape it are here already. We can try to regulate them, we can mourn for our loss or we can adapt. I would see it as an exercise in media literacy...
bsky.app/profile/mart...
bsky.app/profile/mart...
Fake footage of rabbits on trampolines misleads millions of viewers
Footage of rabbits on a trampoline has been viewed millions of times, but turns out to be fake. Numerous versions of these fake images are circulating on social media.
www.nu.nl/318817/vid...
Footage of rabbits on a trampoline has been viewed millions of times, but turns out to be fake. Numerous versions of these fake images are circulating on social media.
www.nu.nl/318817/vid...
August 4, 2025 at 10:42 AM
I honestly think, we need to adapt our notion of reality. The tools to reshape it are here already. We can try to regulate them, we can mourn for our loss or we can adapt. I would see it as an exercise in media literacy...
bsky.app/profile/mart...
bsky.app/profile/mart...
Dropping generated videos is a good test if you want to know what other platforms your followers use.
Bluesky has a large user base of AI sceptics and seems generally isolated from the hot topics on the video platforms (not judging, but the latter is probably a good thing).
Bluesky has a large user base of AI sceptics and seems generally isolated from the hot topics on the video platforms (not judging, but the latter is probably a good thing).
August 1, 2025 at 9:16 AM
Dropping generated videos is a good test if you want to know what other platforms your followers use.
Bluesky has a large user base of AI sceptics and seems generally isolated from the hot topics on the video platforms (not judging, but the latter is probably a good thing).
Bluesky has a large user base of AI sceptics and seems generally isolated from the hot topics on the video platforms (not judging, but the latter is probably a good thing).
Need to read more carefully. Was watching Sabine's recent video and was expecting to learn about "Bohemian mechanics". You can't imagine my disappointment
www.youtube.com/watch?v=DvpS...
www.youtube.com/watch?v=DvpS...
New Experiment Rules out Bohmian Mechanics. It’s Serious.
YouTube video by Sabine Hossenfelder
www.youtube.com
July 20, 2025 at 6:52 PM
Need to read more carefully. Was watching Sabine's recent video and was expecting to learn about "Bohemian mechanics". You can't imagine my disappointment
www.youtube.com/watch?v=DvpS...
www.youtube.com/watch?v=DvpS...
DHH on the Lex Fridman Podcast (2h in): probably you can see a productivity drop this week from all the techies wading through it...
July 17, 2025 at 8:38 PM
DHH on the Lex Fridman Podcast (2h in): probably you can see a productivity drop this week from all the techies wading through it...