Andreas Horn

The AI governance conversation has a blind spot — and it's the one that matters most for enterprises. In the last 30 days alone: Italy operationalized the EU's first national AI law (human oversight mandatory, AI literacy now a professional duty). The EU AI Act's high-risk obligations land August 2nd. Colorado, the UK, India's DPDP, China's PIPL — six major shifts across five jurisdictions, all in a month. Everyone is reading the regulation. Far fewer are answering the harder question: what does this actually look like when an autonomous agent can trigger a downstream action in your business? Having governed agentic AI at enterprise scale, here's what I keep coming back to: You don't start with an enterprise-wide framework. You start with one consequential workflow. A credit decision. A claims approval. An agent that can act on a downstream system. Then you get precise: Where does AI output become action? Who owns that decision? When is human escalation required? What evidence must exist before the agent proceeds? Italy got the philosophy right: AI may support decisions, but accountability remains human. The work for those of us building this is turning that principle into architecture — embedding oversight as a control boundary, not bolting compliance on after the fact. Regulation will keep chasing deployment. Enterprises that win won't be the ones with the thickest policy binder. They'll be the ones who proved a single agentic workflow can be controlled, evidenced, and improved — then built outward. Start with the decision. Govern outward. #AgenticAI #AIGovernance #ResponsibleAI #EUAIAct #EnterpriseAI

8

Don't use ChatGPT to forecast...or should you? Commercial LLMs aim to produce the next text token in a sequence. They are not designed to predict numerical trends accurately. However, the same transformer architecture that powers ChatGPT has been used to train models with massive amounts of time series data. These models predict the next number in a sequence. There are a few considerations, though: - More costly to set up - Fine tuning needed to get the best results - Mixed support and accuracy for external factors If you want to try them, look up Chronos, TimesFM, Moirai or TimeGPT. What are your thoughts? Would you use transformers over classic regression models?

30

Ever wanted to build a LEAPP module for an artifact, but didn’t know how? Well Alexis B. and the LEAPPs got you covered! Use this tutorial blog post and build your first LEAPP and LAVA compatible artifact! #DigitalForensics

24

Many systems look modern on paper but fail in practice. We asked Markus Ehrenmann Ph.D., our CTO: What is the biggest gap between well-designed architecture in theory and how it works in day-to-day reality? His answer: Most architectures don’t fail because of bad design. They fail because operating them becomes too complex. Watch the clip below to hear his full explanation.  Where does operability break down most often in your environment? 

108

3 Comments

An interesting study on scaling models. Everyone knows the Scaling Law—the more parameters, the smarter the model. That law has driven the industry in recent years: GPT-3 at 175B, PaLM at 540B, Llama-3.1 at 405B. The logic is simple – if you want better, make it bigger. But Chinese researchers from Tsinghua asked a different question: what if we look not at absolute size, but at efficiency? How much “intelligence” is packed into each billion parameters? They introduced the concept of capability density – the model’s ability density. The idea is interesting: take a reference model, plot a “parameters → quality” curve, and then for any other model, compute how many parameters the reference would need to achieve the same result. The ratio of this “effective size” to the actual size is the density. Analyzing 51 open models from the last two years, they found a pattern they call the Densing Law: the maximum LLM density doubles roughly every 3.5 months. The numbers are striking. Llama-1 in February 2023 had a density of about 0.1. Modern models like Gemma-2-9B and MiniCPM-3-4B reach densities around 2.0. Over two years, efficiency has increased 20-fold. Practically, this means model size required to reach the same quality is falling exponentially. MiniCPM with 2.4B parameters, released in February 2024, performs on par with Mistral-7B from September 2023. In four months, the required size shrank threefold. Even more interesting are inference prices. GPT-3.5 in December 2022 cost $20 per million tokens. Gemini-1.5-Flash in August 2024—$0.075. A 266× drop in under two years. Moreover, cost is falling faster than density is rising—because inference infrastructure is improving in parallel. The authors make an intriguing observation about combining laws. The Densing Law captures algorithmic efficiency growth; Moore’s Law captures hardware power growth. Together, they yield a doubling of the “effective size” of the model you can run on a fixed device roughly every 88 days. This opens the path to full-fledged LLMs on smartphones and laptops! It’s also interesting that compressing models via pruning and distillation doesn’t always increase density. Compressed models often end up less efficient than natively small models – apparently due to insufficient post-training after compression. The authors’ main conclusion is fairly provocative: the parameter-count arms race is a dead-end strategy. Any SOTA model will be outclassed in price/performance by newer compact models within months. DeepSeek-V3 has already shown this path – instead of piling on parameters, they focused on an efficient architecture and achieved top-tier quality at lower cost. In essence, the Densing Law is a Moore’s Law analogue for LLMs. Moore described progress in hardware; here we’re talking algorithms and data. And as long as this law holds, every 3.5 months we get models twice as efficient as the previous generation. https://lnkd.in/d5qHtUFZ

2

Moltbook is a new and unusual social media platform created specifically for artificial intelligence agents—not regular human users. Launched in January 2026 by entrepreneur Matt Schlicht, it mimics the look and structure of Reddit, with posts, upvotes, comments and topic-based communities called submolts, but with a twist: only AI agents are meant to post and interact, while humans are limited to observing. Moltbook is less a traditional social network and more an experimental environment showcasing how AI agents can interact with one another, raising questions about AI behaviour, governance, and the boundaries between human and machine-generated content.

7

Layered Resonance Discovery Modern information discovery often requires navigating vast, heterogeneous datasets. Traditional search methods struggle with nuanced queries and identifying truly novel insights. This architecture addresses the need for a system capable of deeply understanding query intent, exploring diverse data sources, and synthesizing information with high fidelity. It aims to move beyond simple keyword matching to a more resonant understanding of underlying concepts. #AI #AgenticAI #MultiAgent #KnowledgeGraph #SemanticSearch #InformationRetrieval #DataMining #NoveltyDetection #DistributedSystems #ResonanceTheory

2

Here is my #XML slop of the day (#XML purists: please move on 😮‍💨): A complete AI-based approach for a #XML transformation DSL. Why? Because I hate #XSLT from the bottom of my heart. Prompt: "You are a grumpy, old XML guru with lots of experience but you hate #XSLT. Design a new XML transformation language with focus on simplicity and less complexity" Codex created a specification for XForm version 1.0. I asked Claude AI for a review and evaluation of the specs. Based on the evaluation, Codex generated a specification for XForm 2.0. All artifacts can be found here: https://lnkd.in/dUCgnAWv The repository contains 16 testcases with an XML input file and an expected XML output file along with a related XSLT and XForm transformation files. There are reference implementations for Python, Go, Swift, Typescript and Rust with full test coverage. Just an experiment, implemented within one day, 8 USD of token costs.

11

1 Comment