Hans-Peter Zorn

How do you handle context loss when your AI agent works with large specifications? Tulla is my experimental open-source implementation of Semantic SDD. It is research, beware. But I am open for feedback. no warranties. github.com/hpzorn/tulla 6/6

SDD: hand the construction crew a bunch of textfiles and hope they remember page 42. Semantic SDD: feed them verified instructions for the specific brick they're holding. 5/6

Spec-Driven Development is the current trend: write a detailed spec, let the AI code. But there the fundamental flaw. 1/6

In retrospect, I probably spent too much time grumbling over shifting standards for what counts as "AGI" and not enough time focusing on the massive tidal wave of AI coming straight at us.

When robots take the night shift shopping spree! 🛍️ Robots navigate through dm-drogerie markt Deutschland stores at night to create a digital replica of the store's layout, known as a "digital twin." Developed Ubica Robotics GmbH, these autonomous robots scan shelves to provide real-time information about item positions, pricing, stock gaps, and store layouts. 🏪 This data serves multiple purposes, such as improving staff routes, enhancing inventory management, and informing the creation of planograms for more efficient store layouts. It combines digital twin with robotics and it's really cool use case. What are your thoughts? ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com

ok done

How do LLMs pick the next word? They don’t choose words directly: they only output word probabilities. 📊 Greedy decoding, top-k, top-p, min-p are methods that turn these probabilities into actual text. In this video, we break down each method and show how the same model can sound dull, brilliant, or unhinged – just by changing how it samples.

@growing_daniel unlike claude code, junior devs learn from their mistakes instead of having to be corrected for the same things over and over and over again until the sun burns out

Can an AI model predict perfectly and still have a terrible world model? What would that even mean? Our new ICML paper formalizes these questions One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵

Thanks for sharing! 🙂Here another abstract for the visual people. 😀

This is really BAD news of LLM's coding skill. ☹️ The best Frontier LLM models achieve 0% on hard real-life Programming Contest problems, domains where expert humans still excel. LiveCodeBench Pro, a benchmark composed of problems from Codeforces, ICPC, and IOI (“International Olympiad in Informatics”) that are continuously updated to reduce the likelihood of data contamination.

Ever wondered how linear RNNs like #mLSTM (#xLSTM) or #Mamba can be extended to multiple dimensions? Check out "pLSTM: parallelizable Linear Source Transition Mark networks". #pLSTM works on sequences, images, (directed acyclic) graphs. Paper link: arxiv.org/abs/2506.11997

After ~6 years of building these types of architectures (starting with BERT, eg see Baleen), I think calling these multi-agent systems is a distraction. This is just software. Happens to be AI software. It doesn’t seem so complicated once you internalize it’s just a program.

Bro how was the show Silicon Valley so consistently 10 years ahead of its time 🤣

The data science revolution is getting closer. TabPFN v2 is published in Nature: nature.com/articles/s4158… On tabular classification with up to 10k data points & 500 features, in 2.8s TabPFN on average outperforms all other methods, even when tuning them for up to 4 hours🧵1/19