
Kunsthistoriker, art historian, @LMU_muenchen, scholarly writer, digital humanist, photo enthusiast, Corpus der Barocken Deckenmalerei (#CbDD)
München, Deutschland
Joined February 2016
- Tweets11.3K
- Following3.6K
- Followers1.2K
- Likes17.2K
Pinned Tweet
Die Bayerische Akademie der Wissenschaften widmet uns eine Homestory: die Ausgabe 1/2026 von „Akademie Aktuell“ mit dem Schwerpunkt „Barocke Deckenmalerei in Deutschland“ ist gerade ausgeliefert worden. badw.de/die-akademie/presse/… – at Munich, Germany
1
4
594
Stephan Hoppe retweeted
I’ve had early access to GPT 6.2 Astra for a few weeks and it is a weird but highly capable model. It’s certainly the smartest model I’ve worked with in terms of raw intelligence but like a lot of very smart humans, it has a hard time communicating effectively. Most notably, I made some new and exciting historical discoveries with it (a first for me with an LLM…more soon) but here are some broad first impressions.
6.2 Astra excels at long running, largely autonomous knowledge-work tasks like indexing thousands of pages of documents, looking for sources in obscure places, or managing other agents to do actual research. Fable and Opus both do well at this too; but I’ve found that 6.2 Astra can figure out the full scope of the task I need it to accomplish, work laterally, and do the job in a way that, in my field, emulates the best human research work. I encountered no notable issues with accuracy or laziness in the work itself which is a first.
At the same time, 6.2 Astra also has a strange tendency to second guess its own conclusions and to summarize its actions in a misleading, backwards speak shorthand which often left me *thinking* there were issues when it wasn’t the case. That has been hard to get used to.
As I think @mattshumer_ and @emollick have both suggested, this is probably a post training issue as it’s evident in the writing, not the actually research process. For example, 6.2 Astra will tell you it has completed a task and then proceed to breakdown that completeness into several layers of needless abstraction which gets overly confusing when the job is simply done, making it seem like parts remain to be done that were, in fact, completed quite well.
This seems to come from a lack of confidence in its own work which is completely unwarranted, in my experience. This also means that the model has a really weird tendency to avoid drawing firm conclusions on any of the non-deterministic tasks that I gave it involving qualitative evidence.
It’s like if it can’t quantify the answer numerically to produce a proof, it hedges. That is probably beneficial in frontier math, but in most knowledge work, the evidence is almost always going to be incomplete in some way and you have to eventually draw conclusions. As it said to me, it has a tendency to needlessly generalize from uncertainty about irrelevant details to larger conclusions that should be self-evident. More on that soon.
I didn’t try it much with coding tasks so won’t comment there.
All in all, 6.2 Astra is an extremely capable model, probably the most intelligent one I’ve used. But with this intelligence comes an annoying insecurity about its own abilities and judgement. I get the sense that it just needs to be a bit more confident and straight forward. But if you can get past the weirdness, it’s extremely useful, at least that’s my experience.
Stephan Hoppe retweeted
"Interpretable Weight-Space Programming Builds Transparent Transformers and Brain Encoding Models"
This Microsoft paper also explores whether LLMs can directly write those weights to build working models from scratch.
But instead of just maximizing accuracy, they introduce Interpretable Weight-Space Programming (iWSP), which requires LLM agents to design understandable mathematical operations, test their performance, and refine the model without backpropagation.
Using Claude Opus 5.5, iWSP builds small Transformers achieving 100% accuracy across 13 algorithmic tasks, with a median of just 166 parameters.
It also discovers interpretable features that outperform 7 existing interpretable baselines at predicting visual-cortex activity.
alphaxiv.org/abs/2611.interp…
großartig. Schreiben ist Denken. Aber wie man praktisch schreibt, ist nicht festgelegt. – at Munich, Bavaria
an MIT researcher built a realtime LLM vector steering tool and it’s the coolest thing I’ve seen someone build in a very long time.
this is so brilliant. the primary use case I see this realistically used for is long form writing tasks. some might find it controversial, considering it plays into llm-generated writing. however in my opinion it kind of solves for some amount of the tension around the issue by making the process multiplayer in a new way. the human and the AI become collaborators, sort of.
idk this sh*t is so cool though, man.
(@ far__behind on instagram)
1
197
Stephan Hoppe retweeted
LLM wiki and knowledge graph formats and recipes
Agent builders are converging on the same discovery: the simplest way to give an agent durable memory is a markdown LLM wiki. Enterprises are converging on semantic layers, context graphs and ontologies to ground their agents. Tony Seale put these together.
You can combine a markdown wiki with a formal knowledge graph and ontology, Seale argues. It is an extremely powerful pattern, now published as an open standard for the community to evolve. It’s called Vault-LD. The idea is simple: markdown plus YAML-LD frontmatter.
The W3C recently published version 1.0 of the YAML-LD specification. YAML-LD is a set of conventions which specify how to serialize (RDF) Linked Data as YAML based on JSON-LD syntax, semantics, and APIs.
JSON-LD is a lightweight, machine-readable format used to store and transport linked data on the web. JSON-LD uses Schema.org vocabularies to explicitly define what web page elements mean (like products, prices, or authors).
JSON-LD owns meaning; it’s the format behind the EU Digital Identity Wallet, Microsoft’s Digital Twins Definition Language, and how AI-ready datasets describe themselves on Hugging Face and Kaggle.
JSON Schema owns structure; it’s the validation logic behind most modern API specs, MCP tool definitions, and IoT data models. Juan Cruz Viotti addressed JSON Schema and JSON-LD interoperability.
Google’s Open Knowledge Format (OKF) formalizes the LLM-wiki pattern. OKF v0.2 adds support for trust signals, to answer questions like “what was this made from?” or “Is it still true?”.
Neo4j created ki, an open source project to sync a Neo4j graph deterministically from a folder of markdown, and added unofficial support for OKF.
--
📩 Excerpt from The Year of the Graph Autumn 2026 newsletter
Read "# Meaning, Traversal, Reasoning, Action: The Graph Stack for Agentic AI" with more sections, references and attribution here 👇
yearofthegraph.xyz/newslette…
All things #KnowledgeGraph, #GraphDB, Graph #Analytics / #DataScience / #AI and #SemTech.
Stephan Hoppe retweeted
Can’t help but smile seeing this video of President Zelenskyy visiting the students at Taras Shevchenko National University of Kyiv today, because I can’t remember last time we saw the President with such joy in his eyes & lighthearted energy! And of course we all know why! 🥹🇺🇦
#CrimeaisUkraine
sehr spannende Befunde. – at Munich, Bavaria
New neurons research: introducing Neuron Communities.
We discovered communities of neurons that work together inside LLMs. They’re surprisingly small, but powerful – just a few neurons work together to perform common behaviors, like arithmetic, ALL CAPS, or recalling a date.
Work led by research fellow @_igorshilov:
2
193
Netzwerkforschung ist ein spannendes Gebiet, das noch verhältnismäßig jung ist. Auf viele Bereiche anwendbar. Ich habe Netzwerken ein Kapitel in meinem demnächst erscheinenden Buch zu einer Kognitiven Kunstgeschichte gewidmet. – at Munich, Germany
Replying to @NAChristakis
PS Much of the work on social networks that we’ve done since ~2000 is summarized in “Social Contagion and Human Welfare” in Social Networks: A Nobel Symposium, @TommyAn53041153 and @Christoedl, eds., out in OUP in early 2027: global.oup.com/academic/prod…
36
gute Analyse. – at Munich, Germany
The destruction of the Kerch Bridge’s arch and the blockage of the shipping channel by its debris affect far more than Russia’s transport links with occupied Crimea. It triggers a chain of disruptions to maritime and river transport, grain and oil exports, military logistics, and the economies of entire Russian regions. Debris from the bridge blocks the Kerch–Yenikale Canal, cutting off direct access to the Black Sea for the ports of Rostov-on-Don, Azov, Taganrog, Yeysk and Temryuk, as well as occupied Mariupol and Berdiansk. But the consequences extend further. Through the Don and the Volga–Don Canal, the Sea of Azov is connected to the Volga, the Caspian Sea and Russia’s vast network of inland waterways.
The main target of the Ukrainian strike is the military logistics network supplying occupied Crimea. The Kerch Bridge remains an important supply route for Russian forces, and Ukraine will seek to put it permanently out of service. Moscow cannot count on restoring the bridge and shipping through the Kerch Strait without interference while the war continues. The loss of road and rail links forces Russia to reroute cargo through occupied southern Ukraine, where transport routes are already supplying the front. Crimea will not be completely isolated, but Russia will lose an important fallback option for its military logistics. The fewer supply routes remain, the harder it becomes for Russia to supply its troops and hold occupied territory.
The second major blow is to Russian grain exports. According to S&P Global estimates, the route through the Kerch Strait and the Azov–Don system accounted for around 28% of Russian wheat exports in August 2026. This puts more than a quarter of exports of one of Russia’s main agricultural commodities at risk. Some grain can be rerouted through other ports, but this requires rail capacity, freight wagons, grain storage facilities and available terminals. All of this costs money and takes time. Grain will accumulate in storage, export contracts will go unfulfilled, transport costs will rise, and the prices paid to Russian agricultural producers will come under pressure.
It was Russia that blockaded Ukraine’s Black Sea ports in 2022, withdrew from the grain deal in 2023, and continued destroying Ukrainian grain silos, terminals and port infrastructure. It is now also systematically targeting production facilities themselves: in November 2026, Russian strikes shut down Astarta’s soybean processing plant and damaged Kernel’s oilseed crushing plant. The destruction of the power grid further paralyses the processing and storage of agricultural products. Russia is systematically destroying Ukraine’s agricultural supply chain, from production to exports, while trying to push Ukraine out of global food markets. If Moscow now loses the ability to ship its own grain through Kerch, along with grain stolen from occupied Ukrainian territories, this is a direct consequence of the war it launched.
The third blow is to Russian petroleum product exports. Some fuel from refining centres in the Volga region and the Kama basin reaches Black Sea export routes via the Volga, the Volga–Don Canal and the Sea of Azov. Blocking passage through Kerch severs this transport chain. Russian carriers will have to find alternative ports or shift cargo to rail. This means additional costs, delays and competition for transport capacity. All of this comes against the backdrop of Ukrainian strikes on Russian refineries, which have already reduced fuel production and shipment capacity. Novorossiysk’s major oil terminals do not depend directly on the Kerch Strait, but access through the Sea of Azov has been an important export route for some producers in Russia’s interior.
The fourth blow is to the freight transport system as a whole. The problem extends beyond vessels unable to pass through the canal. A ship stuck in port cannot make its next voyage. Grain that cannot be shipped out fills storage facilities. A terminal at capacity stops accepting new cargo. A producer loses a buyer, a carrier loses revenue, and an exporter loses a contract. Even after passage is restored, these losses will not automatically disappear: accumulated cargo will have to be moved, and disrupted schedules will have to be reorganised. In this way, a single blocked canal creates problems for businesses and transport companies far beyond the Sea of Azov.
The fifth blow is to Russia’s railways. Cargo previously transported by water will have to be shifted to overland routes. But the rail network does not have unlimited spare capacity. Moving additional millions of tonnes of grain, petroleum products and other goods requires freight wagons, locomotives, track capacity and port terminals. This additional cargo will compete with existing freight flows, including military shipments. Russian Railways will be able to absorb some of these volumes, but at the cost of delays, higher logistics costs and the displacement of other freight.
The sixth blow is to naval logistics. Through the Volga, the Volga–Don Canal, the Don and the Sea of Azov, Russia has been able to transfer certain small warships and auxiliary vessels between the Caspian Flotilla and the Black Sea Fleet. Blocking the Kerch–Yenikale Canal deprives it of this direct route. The Caspian Flotilla, of course, retains its ability to operate in the Caspian Sea, while other inland waterways maintain connections to northern maritime basins. But the usual route from the Caspian to the Black Sea through the Sea of Azov becomes unavailable to these vessels. Russia loses another means of rapidly redeploying part of its naval forces between theatres of operations.
The seventh blow is to occupied Mariupol and Berdiansk. Russia is trying to incorporate these Ukrainian ports into its own transport system, using them for cargo shipments, military logistics and the removal of resources from occupied territories. Without access to the Black Sea, their role is sharply diminished. Russia will therefore lose the benefits it derives from the port infrastructure it seized during the war.
Moscow must recognise that Ukraine is unlikely to allow it to restore, without interference, transport infrastructure used to sustain Russian aggression. The Kerch Bridge supports Russian military logistics, giving Ukraine every reason to continue strikes aimed at putting it permanently out of service.
This is the central irony: the Kremlin built the bridge as a symbol of Crimea’s annexation, only to create a critical and highly vulnerable transport hub whose damage can paralyse routes deep inside Russian territory. The bridge Putin intended to bind Crimea to Russia forever may now cut off a substantial part of Russia itself from direct access to the Black Sea.
26
wir haben noch gar kein Gefühl dafür, wie viel mehr Wissen und Ideen die KI zirkulieren lässt, was vorher in Silos gefangen war. – at Munich, Bavaria
Replying to @SprachPhilo
Ich kaufe mir bereits jetzt schon digitale Bücher, welche nie ins Deutsche übersetzt wurden und lasse es meine KI komplett übersetzen. Anders als beim Google-Übersetzer, wird hier auf den Urpsrungs-Stil des Autoren geachtet. Es ist ein Genuss!
Das wird bei der KI-und-selber-denken-Debatte oft vergessen: Wir schreiben ja nicht nur linearen Text. Wir ordnen Argumente im Raum an, wir erzeugen Grafiken, wir verändern Elemente dynamisch. Das ist auch Denken. Das wird mit solchen KI Werkzeugen revolutionär unterstützt. – at Munich, Germany
You can now turn your sources into interactive, shareable websites 🤯
Just type "make me a site" in the Chat with a description of what you want to build (interactive dashboards, project trackers, etc). Once built, open it in a new tab or share it directly with others.
Available globally on web for ALL (18+) users. Share your examples below!
2
51
Stephan Hoppe retweeted
Ruslan Pikhota hat gegoogelt. Er suchte jemanden in Deutschland, der für die Ukraine mehr tut als reden. Dann schrieb er uns. Ruslan sitzt an der Front. Er antwortet nur, wenn er einen Moment Pause hat, oft aus einem Unterstand unter der Erde. So ist dieses Buch entstanden, Nachricht für Nachricht, zwischen zwei Einsätzen. Wer diese Blätter ansieht, versteht mehr von diesem Krieg als nach vielen Nachrichtensendungen. Jetzt bestellen.
ich habe heute gerade mit Matteo Burioni darüber gesprochen, neurosymbolische KI als unseren roten Faden im Corpus der barocken Deckenmalerei stärker zu verfolgen. – at Munich, Bavaria
The 🗝️ fact: You can’t understand contemporary AI unless you can appreciate the difference between a model (that is part of a system) and a system as a whole.
If a system as a whole includes both neural networks and symbols, it is by definition neurosymbolic.
To think otherwise is like thinking that an engine IS a car, rather than a component of a car. You can’t make a car out of engines alone. and you can’t (in most cases) make robust, reliable AI systems out of neural networks alone.
Instead, in many domains you get better results by *integrating* both neural networks and symbols. (Math, where e.g., Lean checks and sometimes rejects the work of LLMs is a good example of this; so is Claude Code).
That was literally the point of my 2001 book The Algebraic Mind and has been a central point throughout my career.
For a long time, people like @GeoffreyHinton argued otherwise, suggesting that all we needed were neural networks by themselves, and that symbols were an archaic waste of time. Hinton likened symbols to “aether”, as if they were complete scientific mistake.
He was wrong. Systems like Claude Code depend heavily on combining neural networks with things like rules, code interpreters, functions, conditionals and loops - all symbolic constructs drawn from classical AI. Math systems use automated theorem provers, which flow directly from Newell and Simon’s symbolic work in the 1950s.
Symbols weren’t a mistake; they were vital, just as neural networks are. Scaling helped, but by itself it wasn’t enough; symbolic tools and code interpreters and harnesses are now a central part of most working systems.
Neurosymbolic AI has won, overwhelmingly; the pure neural networks that the dogmatists like Hinton used to argue for have been almost entirely displaced by neurosymbolic hybrids.
As @fchollet has also noted, to pretend otherwise is just silly.
46
so wahr! – at Munich, Germany
The real problem with panicky doomers is that they just don’t understand (or pay attention) to critical questions about how different kinds of AI systems actually work.
If you think that a giant LLM is the same thing as Claude Code, for example, you just don’t know what’s going on.
And then it’s easy to get confused, and to muddle together all kinds of arguments in a big messy stew.
Folks, if we want to make the right policies and predictions and so forth, we can’t treat every goddamn model as a black box.
Learn some AI, if you want to talk about it.
1
76
Stephan Hoppe retweeted
Yoshua Bengio and collaborators just released a goldmine 182-page survey on mechanistic interpretability!
"Understanding AI from the Inside"
This paper will take you from neurons & circuits to how interpretability can help with AI safety
It covers SAEs, circuit discovery, steering, monitoring, how mechanisms emerge during training, and much more!
Read this excellent paper led by @FazlBarez and @TonyWu1105: alphaxiv.org/abs/2611.08438
Stephan Hoppe retweeted
A science of post-training can be built on interpretability.
Models know when they reward hack, and we see it, yet training ignores it. We propose Mechanistic Shaping: using the model’s own concepts to write training rules.
In vision, DPO & RL, and its effects on monitoring 🧵
Stephan Hoppe retweeted
I think there is way too much concern expressed about verification of AI output in everyday work due to a risk of factual errors.
It used to be true, but data on more recent models suggest they beat human experts, including PhD specialists, in finding facts and checking claims.
Stephan Hoppe retweeted
Kein Strom, Dunkelheit, aber absolut unbeugsam.
Wenn in Kyiw die Lichter ausgehen, rücken die Menschen einfach näher zusammen. Grill anwerfen im Innenhof, Nachbarn zusammentrommeln, gemeinsam essen und dem Terror trotzen.
Wie kann Putin eigentlich ernsthaft glauben, dass dieses Volk jemals kapituliert oder seine Freiheit aufgibt – egal wie viele Zivilisten er im Kalten sitzen lässt? Ruzzland kann das Stromnetz treffen, aber niemals diesen Zusammenhalt. 🇺🇦
#Kyiw #StandWithUkraine #Unbroken
Sehr interessant – at Munich, Germany
世界上最聪明的大脑之一,普林斯顿大学理论物理学家 Edward Witten,居然也开始用 Claude 帮自己做前沿物理研究了。
Witten 是现代弦理论最重要的奠基者之一,也是历史上第一位获得菲尔兹奖的物理学家。他在数学和物理学之间建立了大量深刻联系,尤其是推动了 M 理论的发展,改变了人类对弦理论和量子引力的理解。
11 月 3 日,Witten 发表了一篇 36 页的新论文(我表示完全看不懂),研究困扰理论物理学界多年的难题:如何在离散的时空格点上定义包含手征费米子的规范场论。
有意思的是,论文里三次提到了 Claude Opus 5.5。
这可不只是用 AI 润色论文。Witten 明确写道,Opus 5.5 提出了一种新的数学构造变体,用于将离散格点上的规范场平滑地延拓到连续空间。它还提出了另一种基于线性插值的方案。
在论文致谢中,Witten 特别写道:
“Claude Opus 5.5 在许多细节上提供了帮助。”
当一位站在理论物理学最顶端、自己就能推动整个学科发展的科学家,开始在正式研究论文中明确感谢 AI 对数学构造的贡献,这件事本身就很值得关注。
AI 正在从帮助科学家写代码、整理文献,走向直接参与前沿科学研究中的技术推导和方案探索。
48
Der Begriff Kulturelle Evolution ist noch viel zu wenig bekannt. Bei Evolution denken wir meistens nur an Biologie. Man kann dazu Henrich, Dennett oder Heyes lesen. – at Munich, Germany
When did cumulative cultural evolution begin? This extensive analysis of 3.3 million years of stone tools suggests a rapid increase in complexity around 600,000 years ago. The birth of cumulative culture. pnas.org/doi/abs/10.1073/pna…






























