Dawid Ostrowski

The German AI Federal Association welcomes EU plans to establish "AI factories" where young companies can train and test their "artificial intelligences" and language models on supercomputers. This is in response to a vote in the European Parliament on the "EuroHPC Startup Initiative to Strengthen Europe's Leadership in Trustworthy Artificial Intelligence". The Commission's President Ursula von der Leyen believes that Europeans can develop a leading role with a focus on "trustworthy" AI and wants to facilitate access to suitable supercomputers for AI startups. As the EU continues to invest in its AI infrastructure and support startups, enterprises should closely monitor developments and explore ways to leverage (talent and supercomputing resources) the emerging opportunities in the European AI landscape. #ai #enterpriseai #euAI #trustworthyai https://lnkd.in/eNDxMBPp

5

Finally - my review of Stephan 🚀 Schmidt's book. It is a short book with many good thoughts on the whole spectrum of a CTO's job, all general enough to immediately think about how those apply to one's situation - basically like a meditation.

7

1 Kommentar

At the ATX Mega Meetup on AI two weeks ago, one of the panelists - Ian - used a metaphor that has stayed with me since. He said that LLMs are like new material for us and we have to use them to get a tactile feel for what they can do. Think of a time when you handled a new material or object or tool for the first time. You encountered the unfamiliar. It felt different in your hand - maybe soft or rough or slippery. It looked or smelled or sounded different. It was novel. Using #GenerativeAI is a new material for us and we are still feeling our way to familiarity. We are sculpting with words. I experience this in how I frame questions in #Perplexity. In how I prompt and re-prompt in #ChatGPT. Each time I try, I am learning what shape of words works better to get the shape of result I want. This is a new skill and I feel like such a beginner. It reminds me of my struggles to learn the saxophone. Reading music notation on the page so that I place my fingers on the right keys and use the right muscles in my mouth and enough air in my lungs to create as series of sounds that create music. Sometimes it all comes together and I am elated - I am playing effortlessly. And sometimes, it does not and I play a phrase over and over and over, struggling with the notation or the rhythm or the sound. Ezra Klein recently interviewed Ethan Mollick, a professor at Wharton and his advice was to use a GenAI tool for at least 10 hours to gain familiarity (link to the show “How Should I be Using AI Right Now?” in the comments). Professor Mollick admits that it is an unscientific threshold but reasonable to learn these new tools. He says: “The key is to use it in an area where you have expertise, so you can understand what it’s good or bad at, learn the shape of its capabilities.” In my day job as a #productmanager- I am not the one writing the prompts or building the software system. But it still helps that I am familiar with this form of cognitive sculpting. My contribution to the creation of our #GenAI system is providing context for what problems we want to solve for customers and where it would be useful in their work. Think of a time when you encountered a new material or a new skill. What do you remember about the tactile unfamiliarity or the cognitive novelty? What steps did you take to become more familiar, to scaffold your way to mastery? What are you doing to use GenAI in your daily work? Where are you finding it helpful and more importantly unhelpful? What are you learning about this new tool that you can share with others? #futureofproduct 

38

5 Kommentare

I ran across this post in the Changelog newsletter: https://lnkd.in/eCdjVekq - I like the name slop for unwanted AI-generated content. I love LLMs. They are part of my daily workflow, and I've been exploring them more seriously while on my sabbatical. I wouldn't misrepresent generated content without critical review as my own, but the web seems to be filled with more generated garbage (or 'slop')—content that looks fine at first but isn't accurate or helpful. LLMs (and other generative AI) should help us to make things of higher quality, not just churn out slop more quickly.

11

1 Kommentar

Amusing though it is, the massive negative feedback that Google is getting on its new "#AI Overview" feature seems exaggerated. In "normal" searches you'd be bound to find crackpot results, too - and would very likely find them funny, or would just dismiss them out of hand. Yet when a #statistical model is trained on those results, and then regurgitates the crackpot info as part of its response (along with lots of more useful bits), that somehow is bad. It's simply a reminder that #LLMs have no moral compass, no magical guide on what in their training data is right or what is wrong. Data is data, and they happily play it back to us in ways that often sound intelligent and insightful - largely because no human has a chance of digesting this much data, and the unexpected correlation surprises us. Which in turn, of course, is one of the great opportunities in the technology when used appropriately (and with the right set of expectations).

24

5 Kommentare

How many models is too many? The current situation reminds me of early search engine days when everyone was investing and none had a clear business model until some company came and found a decent way to scale the mission of organising the world's information. Relevance was the key back then and relevance is the key today. What do you think? https://lnkd.in/dhkpjnyk

39

6 Kommentare

The high degree nonlinearity exists within networks, almost everywhere. It is cumulative over training time and hidden layer depth, which was demonstrated by James Zou 7 years ago. A key to this is to figure out how to measure its nonlinearity. I'm not too sure about the way in which nonlinearity was measured in the "Your Transformer is Secretly Linear" paper. Manifold can only handle low degree nonlinearity (smooth), not high degree nonlinearity (fragile and sharp), something I am working on. https://lnkd.in/gxeAJCS5

8

1 Kommentar

⚙ 3D modeling - Raycasting - with #Perspective ⚙ https://lnkd.in/ed9ndTAE Fun with #wasm, #browsertech ➡ no server component, all ray-casting done entirely in #Perspective https://lnkd.in/eFQTZju2 https://lnkd.in/eD3fb4RF

5

A wonderful demonstration of the practical utility of Knowledge Graphs deployed using Linked Data principles! Imagine a clickable image where each link takes you directly to what it denotes. The thinking behind a Semantic Web is so powerful, obvious, and game-changing—once you give it a chance. BTW, UniProt publishes the largest knowledge graph for protein-related information in the massive Linked Open Data (LOD) Cloud collective, representing a whopping 156,318,054,494 entity relationships as RDF triples, deployed using Linked Data principles. Check it out here: https://lnkd.in/giH8rPxM. This Knowledge Graph is also super useful for building #AI solutions that are loosely coupled via predefined SPARQL query templates. Learn more: https://lnkd.in/ezp5JYhM. #VirtuosoRDBMS #UseCase #KnowledgeGraph #SemanticWeb #Bioinformatics #MolecularBiology #Proteins #RAG #GenAI #SPARQL

9

1 Kommentar

Recently Google has pulled back on some of the AI capabilities it rolled out in support of search.    While I’m confident that the skilled engineers at Google will take this practical user testing and turn it into an even better product, it does remind us of how critical robust training data – and a well-defined use case – is in AI applications.    We’ve been talking a lot lately about AI for ag, and the ways in which it can enhance agronomic knowledge when done right.    Crop Science CIO Amanda McClerren and I will be recording another Inside Innovation conversation together in a month, this time focused on AI applications in ag.    What questions do you have for how we’re using artificial intelligence and large language models to help farmers? Drop them below and we’ll try to address them in the recording! #AI #LLM #DigitalFarming

102

3 Kommentare

Very thoughtful analysis. Worth a read. My 2 cents: it seems that the 2 companies (NVIDIA & Apple) that have the lion share of TSMC most advanced capabilities might indeed have a small moat for now

1

I'd like to share a very exciting new AI product release from my former team in the area of AI for Science: in this case applying generative AI techniques to chemical process prediction, recipe creation and synthesis automation in a product called #RXNforChemistry. AI promises to help scientists, especially materials scientists and chemists, make discoveries and address societal challenges a lot faster and more effectively (and is becoming a key focus in #AIAlliance). RXN is a nice step toward that long term promise! Those who know me also appreciate my deep interest in product and business incubation, a big focus when I led the creation of the IBM Quantum business and our Accelerated Discovery platform. Like others, my view is that you need to bring groundbreaking products to market early, ship often, and have a focused and compelling way for users to get started that includes a business/pricing model relevant to the product and fast feedback paths to improve it. That's often a lot harder in a large company where the new product does not fit into existing lines of business than it is in a startup, but in IBM Research we have been figuring out how to do it. 😀 #IBM #IBMResearch #RXNforChemistry #AIforScience Teodoro Laino John R. Smith Srikanth Koneru Alessandro Curioni Karl Wehden Lauren McHugh Olende Jonathan Miller and many others! https://lnkd.in/dTEeH_jP

73

1 Kommentar

Brain Scans of Jazz Musicians Reveal How to Reach a Creative ‘Flow State’ Both expertise and the ability to release one’s focus can help people enter a state of effortless attention Ludwig van Beethoven’s notebooks show that he spent countless hours laboriously developing and revising the musical ideas on which his great compositions were based. It was a torturous and all-consuming process. Beethoven was also the most gifted improviser of his time. He would sit at the piano and create, on demand, fleeting compositions so beautiful and imposing that they would reduce people to tears. Beethoven illustrates two modes of creativity that can be used at different times by the same person. Most people are familiar with the arduous type—the creative struggle—from personal experience. Generating a stream of high-quality creative ideas is difficult. But the latter kind—the flow state, or the experience of being “in the zone”—is more elusive. Since psychological scientist Mihály Csíkszentmihályi first identified flow and systematically studied it in the 1970s, this state of “effortless attention” has gained widespread interest. It is believed to enhance innovation, productivity, meaning making and joy. But until recently, all that was known about flow came from introspection and behavioral research. That body of work revealed important characteristics of the flow state but left basic, though important, questions unanswered. In particular, researchers knew little about its inner mechanisms, which hindered the development of techniques for training or inducing flow to boost one’s creative production. https://lnkd.in/gy693pni

1

1 Kommentar

AI agents, that will represent you, talk to other agents and find a match. This is just a digital version of Arranged marriages … my parents have been replace by AI bots :) Miles Walker thanks for the news article.

24

4 Kommentare

It is the last day at Berlin Buzzwords and the show is definitely living up to its hype. Following yesterday’s post, I am happy to share a link to my session that explains how the limits of search can be shattered using domain specific computing. Interested in learning about the architecture of domain specific computing? Or how it achieves hyper candidate generation, better ranking efficiency and fast menory access? Tune in to this talk to learn more. https://lnkd.in/dD4GKZdK #vectorsearch #keywordsratch #search #elasticsearch #bbuzz

9

It is the last day at Berlin Buzzwords and the show is definitely living up to its hype. Following yesterday’s post, I am happy to share a link to my session that explains how the limits of search can be shattered using domain specific computing. Interested in learning about the architecture of domain specific computing? Or how it achieves hyper candidate generation, better ranking efficiency and fast menory access? Tune in to this talk to learn more. https://lnkd.in/dD4GKZdK

48

I've encountered this countless times: you receive some experiment code and assume you can simply run it on your computer, server, or in the cloud. After all, it worked before, right? The harsh reality? You spend countless hours wrestling with annoying issues: trivial dependency mismatches, major operating system incompatibilities, data located in unexpected places, or the lack of an adequately powerful GPU for training a model. Perhaps no one documented where the data is stored, and the one person who knew left the company two years ago. What follows is a wave of shame and frustration. "Why is this so difficult?" "Is it because I'm not technical enough?" "If only I were more organized!" "I'm sure [name of colleague you secretly envy] wouldn't struggle like this!!" The truth is that we are dealing with a crisis of epic proportions ... A 2024 study showed that 45% of all Jupyter notebooks that were associated with actual publications cannot be reproduced. Think about it. "Standing on the shoulders of giants"? Over 50% of the work researchers cannot reproduce their own (!) work. Know that you're not alone. This isn't just your problem, nor is it solely your organization's (though some handle it better than others). There are simply too many rapidly changing variables—frequent package updates, new hardware and tools, diverse storage configurations, and security settings tailored to specific environments. I've made a film about this struggle. Do you recognize any of this?

10

Better to plan for success and put the hard work in early with a comprehensive brand framework... than risk falling victim to the PLG trap. "For ambitious managers of software companies, the lesson is not to avoid taking a PLG approach. Instead, embrace PLG as an initial strategy, but commit to winning the enterprise over the long term. Plan ahead and you can avoid the PLG Trap." "Seamlessly transitioning from a PLG to an enterprise company, by avoiding the PLG Trap, requires deliberate planning from an early stage." #brandvalue #PLG

5

This is the code repository for Microsoft's #VSCode, the dominating code editor of today, with well over 10 million monthly active users: 2026 contributors 120,000+ commits (code changes) 167,000+ closed issues (fixed bugs) Some issues are solved within minutes. I filed an issue number 210624, referencing similar symptoms from a 9-month old issue. Some take 3 years (the 114425 in 'resources'). There is no point I'm trying to make. Just a little appreciation for the scale and pace of work on major software products.

27

2 Kommentare