Building AI knowledge system @ElstnerAnalytic | Product-focused | Passionate about LLMs, search systems & data science | Knowledge for LLMs
Germany
Joined November 2014
- Tweets847
- Following2.2K
- Followers517
- Likes11K
Pinned Tweet
The Unintended Consequences of Strict AI Policies
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.
Our initial survey data from a large, mid-sized German technology firm supports this. It indicates that a limited number of approved AI offerings often leads to unregulated usage through other channels. Approximately one-third of respondents admitted to using unauthorized LLM chatbots via their personal phones, unapproved websites, or both. While this data point represents a single company with 3,409 survey respondents, it provides valuable insight. Furthermore, a quarter of all respondents (789, or 23%) reported not using LLMs at all.
We conducted an anonymous survey across all employees, comprising 12 questions on GenAI usage and 7 demographic classification questions. The company itself is focused on industrial goods production, rather than IT, though about 20% of the workforce is in software roles and another 20% in hardware engineering.
All employees have access to an internally developed web UI that integrates OpenAI models hosted in Azure—essentially, an isolated, internal ChatGPT clone. Both GPT-4o and GPT-4o-mini are available, and the use of internal, confidential data within this tool is explicitly permitted (e.g., copying confidential email content into the chat window for context).
Despite this internal tool, company regulations prohibit the use of any other GenAI tools for work-related purposes. Access to certain tools like ChatGPT or DeepSeek is also blocked on the internal network.
Many users express a desire for more features than the basic in-house chat UI provides. Prominently mentioned wishes include improved chat history and search functionalities, access to advanced reasoning models, document upload capabilities, and image generation. The availability of these features in free or $20/month external tools significantly contributes to what we term "gray usage." Specifically, 21% of employees use LLM chatbots on their private phones for work-related queries, 10% use unauthorized web-based LLM chatbots (not all of which are blocked), and 3% engage in both.
While not all unauthorized usage may pose high risks (e.g., many programming questions can be asked without exposing company intellectual property), data exfiltration via chat UIs remains a genuine concern, particularly given the activity of certain dubious entities in this domain.
Martin Elstner retweeted
Ich wiederhole mich. Wir können uns nicht vorstellen, wie übel es in Europa werden wird, wenn die Amerikaner wirtschaftlich davonziehen.
Martin Elstner retweeted
We can’t complain about Europe lagging in AI while dismissing every effort to catch up.
Surprise: no single initiative will fix everything. Progress has to start somewhere.
I'm increasingly bothered by people dismissing and joking about every effort as “not enough”. If you do that without contributing yourself or suggesting an improvement, then you're part of the problem.
Start supporting the people who are making an effort if you want something to change.
Martin Elstner retweeted
Einer der besten Kommentare, die ich zu diesem Thema gelesen habe. Vielen Dank. Aus europäischer Sicht stellt sich für mich die Situation zusätzlich noch auf eine andere Art und Weise dar: Je mehr wir bremsen, statt gestalten, je mehr kreative Energie wir auf Compliance lenken, desto weniger spielen unsere Überlegungen eine Rolle. Das Gedachte kann nicht zurückgenommen werden, und die Welt ändert sich schneller denn je. Die Zukunft wird passieren, und wenn Europas Beitrag nur der erhobene Zeigefinger ist, dann müssen wir eine Welt akzeptieren, die andere gebaut haben.
Not a single cause, quite sure. SM and diffused attention isn’t helping. But we (as a society) should have a look at what is directly in front of us.
This is exactly what one would expect to see with covid completely uncontrolled now. No one is safe from the cognitive effects of covid infections, not even kids:
theconversation.com/mounting…
Sicher alles nicht monokausal (SM und unser schlechter Umgang mit schwachen Schülern hilft auch nicht), aber ein Aspekt der in der offiziellen Diskussion mMn zu kurz kommt.
Recommended read:
I ran Boltz-2 100 million times 🤯
A cell is shaped by billions of interactions between proteins, RNA, DNA and small molecules. Co-folding can now simulate some of these interactions at scale.
So what can we do with all this data?
Could it open a bottom-up path to Virtual Cells?
Can it help us understand the limitations of co-folding models? And learn when they actually translate to biology?
🧵
Great work and a nice example/tutorial if anyone wants to finetune their own model.
Always amazed by how well the BM25 baseline still performs. I mean, all the effort we put into embedding models and still not too easy to beat the baseline
Martin Elstner retweeted
Really cool new dataset on @huggingface: PubCheF pairs ~2.5 million molecules with their biological functions, extracted from the literature🧬
This kind of dataset is exactly what open science needs 🤗 thank you for making it available!
Datasets can now be found on HF!
huggingface.co/collections/k…
Martin Elstner retweeted
One part of this project I'm very excited about is the website we made to explore the PubCheF-1 predictions over 13.2 million molecules! 🧵
Martin Elstner retweeted
The LLM Hackathon for Applications in Materials and Chemistry is now the Open Scientific Intelligence Hackathon, and our next event will be Oct 21–22, 2026! As always, the event is free, hybrid, and worldwide.
A change this year as the scope now includes the physical sciences and mathematics.
Register here!
llmhackathon.github.io
Leonie’s recommendation may be not completely unbiased, but mine is. Excellent resource for fine tuning SLMs
The @liquidai cookbook is such an underrated developer resource:
Curious about fine-tuning text, vision, audio, or encoder models?
Curious about fine-tuning with CPT, SFT, DPO, or GRPO?
Curious about fine-tuning LFMs with Unsloth or TRL?
It has it all.
I just did a little cleanup. Enjoy!
github.com/Liquid4All/cookbo…
Classic
Och, menno, jetzt hat man das neue schicke 🇩🇪 #LLM schon beim Schummeln erwischt. Dass die Architektur eine NVIDIA-Nematron-3-Clone ist - geschenkt, dafür sind diese Nematrons ja da. 1/3
Martin Elstner retweeted
I joined @Stripe in January to build a new philanthropic initiative with @nanransohoff. Today we’re announcing Intercept, a $500m innovation fund focused on radically reducing the burden of respiratory viruses.
I’ve spent the past decade investing in vaccines and drugs for infectious disease, and this is the most exciting thing I’ve ever been a part of. Thrilled to be working with @Stripe, @AnthropicAI, @TheFluLab, @FoundationOAI, @coeff_giving and individuals from Jane Street.
We've raised the capital and now it’s time to get to work.
Today we're launching Intercept: a $500M philanthropic initiative to make respiratory infections, like the common cold and flu, a thing of the past.
We treat respiratory infections as a minor nuisance, but that’s really not the case. Most of us will spend 5% of our lives (!) sick from these viruses, they kill 1M people a year, cost $600B annually in productivity, and periodically threaten civilization through pandemics.
So, if they’re such a big problem, why haven’t we dealt with them yet? Last year we convened ~40 leading scientists, pharma R&D leaders, biotech investors, and regulatory experts to better understand that.
We heard two main reasons:
(1) First, it’s just technically very challenging: respiratory viruses represent hundreds of distinct, mutating strains across several families. Fortunately, recent breakthroughs make this newly possible.
(2) Second is a lack of funding: broad-spectrum solutions have historically been underfunded, in part because they’re not a great fit for most philanthropic or commercial funding (and while COVID generated a burst of activity around preventing and understanding respiratory infections through an influx of new funding, that hasn't been sustained).
We think that with enough focus and funding, this might be solvable. Intercept is a $500 million philanthropic initiative that will take advantage of new tools to catalyze the development and deployment of two types of products: broad-spectrum preventatives and air cleaning technologies.
This problem is undoubtedly difficult. But it’s more tractable now than it’s ever been. We think we should give it our best shot.
We’re enormously grateful to our anchor funders: @stripe, @AnthropicAI, @TheFluLab, @FoundationOAI and individuals from Jane Street.
And, I’m very excited to be building this with @incredutility and the rest of the team.
Martin Elstner retweeted
Ist das der Anfang einer Ära, in der Atemwegserkrankungen endlich ernst genommen und als Business Case behandelt werden? An dem neuen, 500 Millionen Dollar schweren Fonds intercept.com beteiligen sich US-Techkonzerne wie Stripe, Anthropic ... 1/ technologyreview.com/2026/06…
Gibt da Schätzungen zu das man in heißen Gegenden rund 1°C mehr kumuliert.
agupubs.onlinelibrary.wiley.…
Meine für Vegas auch mal 2-3° gelesen zu haben, finde da aber grad nichts zu.
Spannend wäre mittelfristig Verbrennerautos aus den Städten zu bekommen, die liefern mit ~15kW thermischen Beitrag neben den anderen Schadstoffen lokal recht viel Wärmeabfall.
@miriam_vollmer, sehen Sie das als meine Bewerbung für den inneren Zirkel.
Yep, just keep an open training pipeline / recipes and a few reference and base models. GPUs will sell themselves as a nice side effect
. @NVIDIAAI should lead an effort for middle power AI sovereignty via as good as possible open source AI
Martin Elstner retweeted
My latest in the @FT today. I make four points:
1) Capitalism's great achievement was moving activity out of the household and into the market — turning domestic production into paid specialisation, creating jobs, and making output visible to the national accounts. AI-enabled self-service might quietly reverse that centuries-long trend.
2) When a technology automates tasks within a service, it can trigger a Jevons paradox: the service gets cheaper, demand expands, and employment grows. When it lets people do the work themselves, demand for the service collapses.
3) AI extends this even to the manual trades, the supposed safe haven of the AI age. If a homeowner can ask a chatbot why their boiler keeps losing pressure, heating engineers lose call-outs.
4) When work shifts to the consumer, it vanishes from the economy statisticians measure. Replace a billing department with a chatbot and a firm records lower costs and higher output per worker; the national accounts register a productivity gain. But the hours patients spend decoding their own test results appear nowhere — not in labour statistics, not in GDP. As self-service spreads into professional domains, that blind spot will grow.
