Sign in

Martin Elstner

@martin.elstner.dev
1.4K followers 908 following 185 posts

AI and ML engineering. Search and recommendation. Founder of Elstner Analytics. Helping companies solve their data issues elstner.dev

PostsRepliesMedia
Martin Elstner @martin.elstner.dev · 15/06/2026
Liegt zZ aber ~1200 daneben und einen ähnlichen Effekt wird man mMn für die kommende Jahre auch sehen (etwa 15% niedriger in 26 und weiter fallend). Immerhin deutlich besser als B und fast alle anderen Städte in DE
040
Martin Elstner @martin.elstner.dev · 15/06/2026
Den Optimismus der Statistik kann ich nicht ganz teilen, auch wenn es sicher weniger Schlimm ist als in Berlin
140
Martin Elstner @martin.elstner.dev · 08/06/2026
Luckily, these over formal, math heavy specification are not limited to survey papers. Our (as a field) publication practice has some wired quirks and stylistic conventions
010
Martin Elstner @martin.elstner.dev · 19/05/2026
Recommended read! How data engineering tools should fit together.
040
Martin Elstner @martin.elstner.dev · 15/04/2026
may has more applications. What about synthetic rewrites of technical documents (think: specific handbooks, monographs and reference books) that are primarily defined by the presented data points and not by artistic narration. Since textbooks are all you need we know, that the data would be useful.
000
Martin Elstner @martin.elstner.dev · 15/04/2026
(doing a clean room reimplementation, let a LLM build a detailed plan and specification, maybe tests from an existing code base or documentation; move the plan to a fresh repo and let an agent build code that isn't using any of the original code but conforms to the specification)
100
Martin Elstner @martin.elstner.dev · 15/04/2026
The dust settled a bit around the chardet rewrite (simonwillison.net/2026/Mar/5/c...) and the practice seems to be considered ok (in a legal, not necessarily in a moral sense). Beside converting GPL or closed source code to more permissive licences, the general idea
simonwillison.net
Can coding agents relicense open source through a “clean room” implementation of code?
Over the past few months it’s become clear that coding agents are extraordinarily good at building a weird version of a “clean room” implementation of code. The most famous version …
100
Martin Elstner @martin.elstner.dev · 14/04/2026
Jailbreak is the only way
010
Martin Elstner @martin.elstner.dev · 09/04/2026
Working hypothesis: spend too much time with Americans and adopted that ‚yeah, it’s okay‘ attitude
100
Martin Elstner @martin.elstner.dev · 09/04/2026
As autistic German (having paperwork on both), I was surprised by the scoring. Or the test is broken
110
Martin Elstner @martin.elstner.dev · 26/03/2026
Danke! Hm, das hat mit dem claim „word fliegt raus“ aber nur lose zu tun. Steht da ja neben csv und json, was ja viel bedeuten kann
000
Martin Elstner @martin.elstner.dev · 26/03/2026
In dem Dokument geht es um KI- und allg. Developer-Infrastruktur. Zu Office-Dateiformaten steht da gar nichts drin.
100
Martin Elstner @martin.elstner.dev · 25/03/2026
“Humanity isn’t ready for that much software” Very true. I’m not sure, where the new bottleneck is located. Currently trying to build nicer, beautiful things and make MVPs a bit less messy. But not really settled on the right way, yet
020
Martin Elstner @martin.elstner.dev · 25/03/2026
The public perception around this topic seems to shift. Slowly, but some hope exists that we will end up in a better position someday ssir.org/articles/ent...
ssir.org
The New Paradigm of Airborne Disease (SSIR)
The COVID-19 pandemic shattered established views of airborne disease. Although it represents a paradigm shift in public health, the field has yet to catch up.
1103
Martin Elstner @martin.elstner.dev · 13/03/2026
Seems plausible. Way 3 is feasible for small company/team to come up with a somewhat valuable dataset and pipeline that is specific and unique enough to provide value and get customers. More or less what I do
000
Martin Elstner @martin.elstner.dev · 13/03/2026
Yes, the models exist but their economic impact is yet very small
000
Martin Elstner @martin.elstner.dev · 07/03/2026
You also don’t emit any CO2 if you fly to the Maledives. The plane would go regardless of your ticket purchase
010
Martin Elstner @martin.elstner.dev · 07/03/2026
Make it a MCP server!
010
Martin Elstner @martin.elstner.dev · 06/03/2026
After all those years, the "data is the new oil" thing makes sense
010
Martin Elstner @martin.elstner.dev · 26/02/2026
Interesting, thanks for that. Selected more neutral scores, don’t see eg airborn precaution as my identity, but do it as a necessity
010
Martin Elstner @martin.elstner.dev · 26/02/2026
Movement Identity Overall Identity Strength: 5.55 / 7 Airborne Health Identity Overall Identity Strength: 5.75 / 7 Proactive Health Identity Overall Identity Strength: 5.65 / 7 Cross-Domain: Largest gap: Movement ↔ Airborne (0.20) Provisional Pattern: Integrated Triad Cross-domain convergence: 98%
100
Reposted by Martin Elstner
Albert Villanova del Moral @albertvillanova.bsky.social · 26/02/2026
🚀 Big news for TRL users with AI agents: We just released the trl-training SKILL in TRL v0.29.0 🤗 Agents can now use the TRL CLI as a structured capability:
232
Reposted by Martin Elstner
Christian Minich @christianminich.com · 25/02/2026
I see AI-forward teams spending a ton of time building bespoke tools to help them develop more effectively using agents (see: proliferation of CLIs to work with git worktrees), trying out different models, testing our new workflows.
211
Martin Elstner @martin.elstner.dev · 18/02/2026
Would maybe add llms.txt here (llmstxt.org). Not specific to agents or discovery but imho a good way to serve up-to-date static content.
llmstxt.org
The /llms.txt file – llms-txt
A proposal to standardise on using an /llms.txt file to provide information to help LLMs use a website at inference time.
110
Reposted by Martin Elstner
The Matter Lab @thematterlab.bsky.social · 17/02/2026
We’re excited to share a new preprint that may reshape how humans write - TreeWriter: AI-Assisted Hierarchical Planning and Writing for Long-Form Documents 🌳 📃 arxiv.org/abs/2601.12740 Demo: drive.google.com/file/d/1R-X6... Live demo: treer.ai [1/5]
arxiv.org
TreeWriter: AI-Assisted Hierarchical Planning and Writing for Long-Form Documents
Long documents pose many challenges to current intelligent writing systems. These include maintaining consistency across sections, sustaining efficient planning and writing as documents become more co...
162
Reposted by Martin Elstner
AI Firehose @ai-firehose.column.social · 16/02/2026
This study presents "abstractive red-teaming," a method for pre-deployment audits of language models, identifying problematic query categories causing character violations. This innovation improves model reliability by exposing risks like illegal behavior and sexism. arxiv.org/abs/2602.12318
arxiv.org
Abstractive Red-Teaming of Language Model Character
ArXiv link for Abstractive Red-Teaming of Language Model Character
021
Reposted by Martin Elstner
David Emmett @motomatters.com · 12/02/2026
Quantitative measures are always a dangerous yardstick. You can describe severe diarrhoea as "an explosion in bowel productivity", but you wouldn't say it correlates positively with gut health.
2198
Martin Elstner @martin.elstner.dev · 12/01/2026
Nemotron3 nano is a solid choice: huggingface.co/unsloth/Nemo... Running it with 2x 3090 and OpenCode for local stuff.
huggingface.co
unsloth/Nemotron-3-Nano-30B-A3B-GGUF · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
110
Martin Elstner @martin.elstner.dev · 08/12/2025
Think of Alex and Palantir what ever you want; this is probably the best 3D-chess move you could make...
000
Martin Elstner @martin.elstner.dev · 05/12/2025
Started a collection of LLM use-case dataset on HF. Key entrances: the super cool ZenML LLMOps use-case database and Anthropic's new interview dataset. Details below, any other things to include? huggingface.co/collections/...
huggingface.co
AI use-cases - a dwablimol Collection
A collection of datasets that describes various use cases of AI in enterprise and other settings.
010
Martin Elstner @martin.elstner.dev · 05/12/2025
One of the few examples of how to use AI tools in a creative process that's fully curative instead of generative. Describing images, OCRing documents and make it searchable. www.youtube.com/watch?v=9ngb...
youtube.com
How Emmy Award–winning filmmakers use AI to automate the tedious parts of documentaries
YouTube video by How I AI
041
Martin Elstner @martin.elstner.dev · 01/12/2025
Fenster hoffentlich alle vorsorglich vermauert um Haftungsrisiko zu minimieren
080
Martin Elstner @martin.elstner.dev · 01/12/2025
“Hacke” is quite universal used for all hand tools that help loosen the soil; very general term. “Queraxt“ is a term I loosely remember from 19th century narratives, but it’s very uncommon in today’s gardening practice.
010
Martin Elstner @martin.elstner.dev · 01/12/2025
Yeah, would have problems to name all the tools properly in German
100
Martin Elstner @martin.elstner.dev · 01/12/2025
Learning/speaking English now for ~30y and feeling quite comfortable with it. The density of new words in a single post is just insane. Need to level up my gardening game
100
Martin Elstner @martin.elstner.dev · 01/12/2025
Classrooms as cluster-mixing machines because of poor air? Color me surprised. </s> 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...
nature.com
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 ...
000
Martin Elstner @martin.elstner.dev · 01/12/2025
Es gibt auch ein paar unaufgeregte Berichte aus den eher kreativen Bereichen, wie man Nutzen jenseits der sloppy Bilder/Filme/Musik ziehen kann www.lennysnewsletter.com/p/nobody-wan...
lennysnewsletter.com
"Nobody wanted to do this work": How Emmy-award-winning filmmakers use AI to automate the tedious parts of documentaries
Listen now | 🎙️ How AI transformed Ken Burns' documentary team's production process
000
Martin Elstner @martin.elstner.dev · 01/12/2025
Für bestimmte Themen gibt es schon gute Ideen, zB Jeremy Howards solveit www.answer.ai/posts/2025-1... Die Adoption und Adaption ist viel langsamer als von der Hype-Bubble angenommen, wenn die Nützlichkeit echt ist, wird sie aber kommen.
solve.it
100
Martin Elstner @martin.elstner.dev · 01/12/2025
Ein KI-Tutor (“hilf mir binomische Formeln zu verstehen”) ist eine ganz andere Hausnummer, ist aber Arbeit und keine Wunderlösung.
100
Martin Elstner @martin.elstner.dev · 01/12/2025
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).
100
Martin Elstner @martin.elstner.dev · 01/12/2025
Beautiful typographic map
020
Martin Elstner @martin.elstner.dev · 30/11/2025
Creative work/parenting 2025 festive edition: ChatGPT creates the Mandala for K3 and you need Gemini nano banana to get the spelling right.
AI generated mandala coloring page with a spelling error from ChatGPT, Gemini adjusted the image and corrects the error
000
Martin Elstner @martin.elstner.dev · 28/11/2025
Aber das eine ist soviel günstiger als das andere
000
Martin Elstner @martin.elstner.dev · 27/11/2025
Working on the important stuff outside the hype narrative seems to be a nice spot. I would also try to do that. How you make money with any of the (business)models is an open question, but probably (for now) not as important as many think. DS playing obviously a long term game.
140
Martin Elstner @martin.elstner.dev · 27/11/2025
But I'm concerned to see sloppy info graphics for the foreseeable future, as this seems to be the next hype topic on AI-twitter/linkedin. The various failure modes of the model are interesting, it's surprisingly good at some points and then fails totally at other stuff. bsky.app/profile/mart...
010
Martin Elstner @martin.elstner.dev · 27/11/2025
It's much, much better if you start with an image and let it do modifications.
100
Martin Elstner @martin.elstner.dev · 27/11/2025
Also the training data bootstrapping sounds very interesting. Quite some nuggets in the report!
150
Martin Elstner @martin.elstner.dev · 27/11/2025
Nano banana pro also nicely puts it directly onto the music sheet, it looks really good, but it’s not the chords it suggests above 🤨
010
Martin Elstner @martin.elstner.dev · 27/11/2025
Really nice how you can get music notation for your kids to play along for the holidays:
100
Martin Elstner @martin.elstner.dev · 26/11/2025
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.
010