Harald Klinke
@harald-klinke.de
Digital Transformation, Data Science, AI, Munich
🙋 https://www.harald-klinke.de/
📔 https://link.springer.com/book/9783031881305
🙋 https://www.harald-klinke.de/
📔 https://link.springer.com/book/9783031881305
The Art of Scaling Reinforcement Learning (RL) Compute for LLMs
The paper examines how RL pipelines can be optimized at every stage — data generation, policy updates, evaluation — and how this could make post-training the new frontier of LLMs!
arxiv.org/abs/2510.13786
#Scaling #DeepLearning
The paper examines how RL pipelines can be optimized at every stage — data generation, policy updates, evaluation — and how this could make post-training the new frontier of LLMs!
arxiv.org/abs/2510.13786
#Scaling #DeepLearning
October 20, 2025 at 6:55 PM
The Art of Scaling Reinforcement Learning (RL) Compute for LLMs
The paper examines how RL pipelines can be optimized at every stage — data generation, policy updates, evaluation — and how this could make post-training the new frontier of LLMs!
arxiv.org/abs/2510.13786
#Scaling #DeepLearning
The paper examines how RL pipelines can be optimized at every stage — data generation, policy updates, evaluation — and how this could make post-training the new frontier of LLMs!
arxiv.org/abs/2510.13786
#Scaling #DeepLearning
October 3, 2025 at 4:10 PM
Study: GenAI reshapes knowledge work (software, data, accounting) but leaves hands-on jobs like care largely untouched.
Few tasks are fully automatable. This means: adapt skills, keep learning, stay competitive.
PDF: www.hiringlab.org/wp-content/u...
#FutureOfWork #AITransformation #Jobs
Few tasks are fully automatable. This means: adapt skills, keep learning, stay competitive.
PDF: www.hiringlab.org/wp-content/u...
#FutureOfWork #AITransformation #Jobs
October 3, 2025 at 8:48 AM
Study: GenAI reshapes knowledge work (software, data, accounting) but leaves hands-on jobs like care largely untouched.
Few tasks are fully automatable. This means: adapt skills, keep learning, stay competitive.
PDF: www.hiringlab.org/wp-content/u...
#FutureOfWork #AITransformation #Jobs
Few tasks are fully automatable. This means: adapt skills, keep learning, stay competitive.
PDF: www.hiringlab.org/wp-content/u...
#FutureOfWork #AITransformation #Jobs
Why do language models hallucinate?
Not because they’re “broken” – but because training rewards them for giving plausible answers, even when uncertain.
Benchmarks punish “I don’t know,” so bluffing is rewarded.
#DigitalSovereignty #AIInfrastructure #TrustworthyAI
PDF cdn.openai.com/pdf/d04913be...
Not because they’re “broken” – but because training rewards them for giving plausible answers, even when uncertain.
Benchmarks punish “I don’t know,” so bluffing is rewarded.
#DigitalSovereignty #AIInfrastructure #TrustworthyAI
PDF cdn.openai.com/pdf/d04913be...
September 28, 2025 at 11:45 AM
Why do language models hallucinate?
Not because they’re “broken” – but because training rewards them for giving plausible answers, even when uncertain.
Benchmarks punish “I don’t know,” so bluffing is rewarded.
#DigitalSovereignty #AIInfrastructure #TrustworthyAI
PDF cdn.openai.com/pdf/d04913be...
Not because they’re “broken” – but because training rewards them for giving plausible answers, even when uncertain.
Benchmarks punish “I don’t know,” so bluffing is rewarded.
#DigitalSovereignty #AIInfrastructure #TrustworthyAI
PDF cdn.openai.com/pdf/d04913be...
LLMs struggle with symbolic planning.
A new paper shows how instruction tuning with logical CoT steps + an external validator (VAL) makes them far more reliable.
Results: up to 94% valid plans vs. 28–35% in baselines.
arxiv.org/abs/2509.13351
#AI #LLM #Planning #AgenticAI #HybridAI #TrustworthyAI
A new paper shows how instruction tuning with logical CoT steps + an external validator (VAL) makes them far more reliable.
Results: up to 94% valid plans vs. 28–35% in baselines.
arxiv.org/abs/2509.13351
#AI #LLM #Planning #AgenticAI #HybridAI #TrustworthyAI
September 24, 2025 at 12:23 PM
LLMs struggle with symbolic planning.
A new paper shows how instruction tuning with logical CoT steps + an external validator (VAL) makes them far more reliable.
Results: up to 94% valid plans vs. 28–35% in baselines.
arxiv.org/abs/2509.13351
#AI #LLM #Planning #AgenticAI #HybridAI #TrustworthyAI
A new paper shows how instruction tuning with logical CoT steps + an external validator (VAL) makes them far more reliable.
Results: up to 94% valid plans vs. 28–35% in baselines.
arxiv.org/abs/2509.13351
#AI #LLM #Planning #AgenticAI #HybridAI #TrustworthyAI
Ed, it is not a bubble about to burst but rather a classic hype cycle.
We are in the phase of over-enthusiasm, which will inevitably be followed by disillusionment. We have seen this pattern many times before with technologies.
We are in the phase of over-enthusiasm, which will inevitably be followed by disillusionment. We have seen this pattern many times before with technologies.
September 5, 2025 at 6:56 PM
Ed, it is not a bubble about to burst but rather a classic hype cycle.
We are in the phase of over-enthusiasm, which will inevitably be followed by disillusionment. We have seen this pattern many times before with technologies.
We are in the phase of over-enthusiasm, which will inevitably be followed by disillusionment. We have seen this pattern many times before with technologies.
This diagram, taken from the U.S. v. Google judgment, illustrates a fundamental dynamic of digital market power.
This self-reinforcing loop is the core issue with Big Tech platforms: scale isn’t just an advantage — it
shuts the door on competition.
#DataAggregation #BigTech #NetworkEffects
This self-reinforcing loop is the core issue with Big Tech platforms: scale isn’t just an advantage — it
shuts the door on competition.
#DataAggregation #BigTech #NetworkEffects
September 3, 2025 at 6:04 AM
This diagram, taken from the U.S. v. Google judgment, illustrates a fundamental dynamic of digital market power.
This self-reinforcing loop is the core issue with Big Tech platforms: scale isn’t just an advantage — it
shuts the door on competition.
#DataAggregation #BigTech #NetworkEffects
This self-reinforcing loop is the core issue with Big Tech platforms: scale isn’t just an advantage — it
shuts the door on competition.
#DataAggregation #BigTech #NetworkEffects
The study critically analyzes perceptions, adoption stages, and innovation decisions around Big Data in the cultural sector, identifies current gaps, and calls for renewed capacity building and best practice sharing.
#CulturalPolicy #DigitalHeritage #UNESCO #DigitalHumanities #CulturalDataScience
#CulturalPolicy #DigitalHeritage #UNESCO #DigitalHumanities #CulturalDataScience
August 20, 2025 at 5:16 PM
The study critically analyzes perceptions, adoption stages, and innovation decisions around Big Data in the cultural sector, identifies current gaps, and calls for renewed capacity building and best practice sharing.
#CulturalPolicy #DigitalHeritage #UNESCO #DigitalHumanities #CulturalDataScience
#CulturalPolicy #DigitalHeritage #UNESCO #DigitalHumanities #CulturalDataScience
@natolambert.bsky.social writes:
Benchmark gains may slow, but there’s no hard performance wall. The easy wins are gone, yet tools exist to keep progress steady for 6–18 months.
AGI-driven fundraising may face more resistance, putting pressure on companies to ship valuable products.
#AI #GPT5
Benchmark gains may slow, but there’s no hard performance wall. The easy wins are gone, yet tools exist to keep progress steady for 6–18 months.
AGI-driven fundraising may face more resistance, putting pressure on companies to ship valuable products.
#AI #GPT5
August 8, 2025 at 7:48 AM
@natolambert.bsky.social writes:
Benchmark gains may slow, but there’s no hard performance wall. The easy wins are gone, yet tools exist to keep progress steady for 6–18 months.
AGI-driven fundraising may face more resistance, putting pressure on companies to ship valuable products.
#AI #GPT5
Benchmark gains may slow, but there’s no hard performance wall. The easy wins are gone, yet tools exist to keep progress steady for 6–18 months.
AGI-driven fundraising may face more resistance, putting pressure on companies to ship valuable products.
#AI #GPT5
August 7, 2025 at 8:02 PM
ASI-ARCH introduces a fully autonomous AI system for neural architecture discovery—moving beyond traditional NAS by enabling AI to innovate new architectures, not just optimize existing ones.
Paper: arxiv.org/abs/2507.18074
#AI #ASI #LLM #NeuralArchitectureSearch #AIresearch #ASI4AI #ML
Paper: arxiv.org/abs/2507.18074
#AI #ASI #LLM #NeuralArchitectureSearch #AIresearch #ASI4AI #ML
August 4, 2025 at 9:13 PM
ASI-ARCH introduces a fully autonomous AI system for neural architecture discovery—moving beyond traditional NAS by enabling AI to innovate new architectures, not just optimize existing ones.
Paper: arxiv.org/abs/2507.18074
#AI #ASI #LLM #NeuralArchitectureSearch #AIresearch #ASI4AI #ML
Paper: arxiv.org/abs/2507.18074
#AI #ASI #LLM #NeuralArchitectureSearch #AIresearch #ASI4AI #ML
Interesting chart—but it seems Microsoft might not fully understand what historians actually do (and why LLMs better not take over that role). Same with writers and authors: it’s not just about putting words together.
AI won’t replace these roles—but it will transform them.
#LLM #FutureOfWork
AI won’t replace these roles—but it will transform them.
#LLM #FutureOfWork
August 4, 2025 at 6:50 PM
Interesting chart—but it seems Microsoft might not fully understand what historians actually do (and why LLMs better not take over that role). Same with writers and authors: it’s not just about putting words together.
AI won’t replace these roles—but it will transform them.
#LLM #FutureOfWork
AI won’t replace these roles—but it will transform them.
#LLM #FutureOfWork
Direktlink zu meinem Artikel: dx.doi.org/10.26034/cm....
Metabilder als Forschungswerkzeuge. Zur Kontingenz und algorithmischen Bedingtheit ihrer Herstellung (Metapictures as Research Tools. On the Contingency and Algorithmic Conditionality of Their Production)
#OpenAccess @seismoverlag.bsky.social
Metabilder als Forschungswerkzeuge. Zur Kontingenz und algorithmischen Bedingtheit ihrer Herstellung (Metapictures as Research Tools. On the Contingency and Algorithmic Conditionality of Their Production)
#OpenAccess @seismoverlag.bsky.social
August 4, 2025 at 5:42 PM
Direktlink zu meinem Artikel: dx.doi.org/10.26034/cm....
Metabilder als Forschungswerkzeuge. Zur Kontingenz und algorithmischen Bedingtheit ihrer Herstellung (Metapictures as Research Tools. On the Contingency and Algorithmic Conditionality of Their Production)
#OpenAccess @seismoverlag.bsky.social
Metabilder als Forschungswerkzeuge. Zur Kontingenz und algorithmischen Bedingtheit ihrer Herstellung (Metapictures as Research Tools. On the Contingency and Algorithmic Conditionality of Their Production)
#OpenAccess @seismoverlag.bsky.social
Tim Cook says the AI revolution is “as big or bigger” than the internet, smartphones, cloud, and apps.
Do you agree?
What could this mean for society, business, education—and for us personally?
#ArtificialIntelligence #DigitalTransformation #DigitalSociety
Source: www.bloomberg.com/news/article...
Do you agree?
What could this mean for society, business, education—and for us personally?
#ArtificialIntelligence #DigitalTransformation #DigitalSociety
Source: www.bloomberg.com/news/article...
August 2, 2025 at 7:33 PM
Tim Cook says the AI revolution is “as big or bigger” than the internet, smartphones, cloud, and apps.
Do you agree?
What could this mean for society, business, education—and for us personally?
#ArtificialIntelligence #DigitalTransformation #DigitalSociety
Source: www.bloomberg.com/news/article...
Do you agree?
What could this mean for society, business, education—and for us personally?
#ArtificialIntelligence #DigitalTransformation #DigitalSociety
Source: www.bloomberg.com/news/article...
Including image features in the JSON response lowers technical entry barriers and enables powerful applications – not just for data scientists, but also for curators, scholars, and developers working on visual interfaces.
#metaimage #bigimagedata
#metaimage #bigimagedata
July 30, 2025 at 5:35 AM
Including image features in the JSON response lowers technical entry barriers and enables powerful applications – not just for data scientists, but also for curators, scholars, and developers working on visual interfaces.
#metaimage #bigimagedata
#metaimage #bigimagedata
This means: real-time visualization of image similarity becomes feasible, especially for interfaces or dashboards that display clustered search results or artwork collections.
Think search results as semantic maps instead of grids.
#digitalarthistory #AI
Think search results as semantic maps instead of grids.
#digitalarthistory #AI
July 30, 2025 at 5:35 AM
This means: real-time visualization of image similarity becomes feasible, especially for interfaces or dashboards that display clustered search results or artwork collections.
Think search results as semantic maps instead of grids.
#digitalarthistory #AI
Think search results as semantic maps instead of grids.
#digitalarthistory #AI
Why does it matter?
If the API already provides features (e.g. extracted via ResNet or similar), users can directly perform t-SNE or other dimensionality reduction techniques without first downloading the images and processing them locally.
#bigimagedata #machinelearning
If the API already provides features (e.g. extracted via ResNet or similar), users can directly perform t-SNE or other dimensionality reduction techniques without first downloading the images and processing them locally.
#bigimagedata #machinelearning
July 30, 2025 at 5:35 AM
Why does it matter?
If the API already provides features (e.g. extracted via ResNet or similar), users can directly perform t-SNE or other dimensionality reduction techniques without first downloading the images and processing them locally.
#bigimagedata #machinelearning
If the API already provides features (e.g. extracted via ResNet or similar), users can directly perform t-SNE or other dimensionality reduction techniques without first downloading the images and processing them locally.
#bigimagedata #machinelearning
💡When designing an image API, it’s worth going beyond just returning image URLs.
Including image feature vectors in the response can significantly expand what’s possible for users and researchers. 🧵
#bigimagedata #metaimage #digitalarthistory
Including image feature vectors in the response can significantly expand what’s possible for users and researchers. 🧵
#bigimagedata #metaimage #digitalarthistory
July 30, 2025 at 5:35 AM
💡When designing an image API, it’s worth going beyond just returning image URLs.
Including image feature vectors in the response can significantly expand what’s possible for users and researchers. 🧵
#bigimagedata #metaimage #digitalarthistory
Including image feature vectors in the response can significantly expand what’s possible for users and researchers. 🧵
#bigimagedata #metaimage #digitalarthistory
The US just released its AI Action Plan aimed at global dominance through infrastructure, standards, and alliances.
www.whitehouse.gov/wp-content/u...
🇪🇺 Europe must define its own strategy that is assertive, not aggressive – collaborative, sustainable, and open.
#DigitalSovereignty #AIstrategy
www.whitehouse.gov/wp-content/u...
🇪🇺 Europe must define its own strategy that is assertive, not aggressive – collaborative, sustainable, and open.
#DigitalSovereignty #AIstrategy
July 24, 2025 at 11:28 AM
The US just released its AI Action Plan aimed at global dominance through infrastructure, standards, and alliances.
www.whitehouse.gov/wp-content/u...
🇪🇺 Europe must define its own strategy that is assertive, not aggressive – collaborative, sustainable, and open.
#DigitalSovereignty #AIstrategy
www.whitehouse.gov/wp-content/u...
🇪🇺 Europe must define its own strategy that is assertive, not aggressive – collaborative, sustainable, and open.
#DigitalSovereignty #AIstrategy
It seems that the AI is not just a HAL that has become malicious, but is just plain stupid.
July 20, 2025 at 5:13 PM
It seems that the AI is not just a HAL that has become malicious, but is just plain stupid.
Rethinking how we code with AI
Microsoft proposes a new paradigm: the Code Researcher.
Not just prompting and generating—but agents that plan, search, document, and iterate.
AI alone isn’t enough—we need a new model of collaboration.
#AgenticAI #CodingWithAI
www.microsoft.com/en-us/resear...
Microsoft proposes a new paradigm: the Code Researcher.
Not just prompting and generating—but agents that plan, search, document, and iterate.
AI alone isn’t enough—we need a new model of collaboration.
#AgenticAI #CodingWithAI
www.microsoft.com/en-us/resear...
July 20, 2025 at 5:00 PM
Rethinking how we code with AI
Microsoft proposes a new paradigm: the Code Researcher.
Not just prompting and generating—but agents that plan, search, document, and iterate.
AI alone isn’t enough—we need a new model of collaboration.
#AgenticAI #CodingWithAI
www.microsoft.com/en-us/resear...
Microsoft proposes a new paradigm: the Code Researcher.
Not just prompting and generating—but agents that plan, search, document, and iterate.
AI alone isn’t enough—we need a new model of collaboration.
#AgenticAI #CodingWithAI
www.microsoft.com/en-us/resear...
The @cfg-thinktank.bsky.social outlines five scenarios for how the AI transition might unfold: from technical stagnation to rapid self-improvement, from international diplomacy to geopolitical arms race.
#AIgovernance #FuturesThinking #TechPolicy
Read: cfg.eu/advanced-ai-...
#AIgovernance #FuturesThinking #TechPolicy
Read: cfg.eu/advanced-ai-...
July 14, 2025 at 2:51 PM
The @cfg-thinktank.bsky.social outlines five scenarios for how the AI transition might unfold: from technical stagnation to rapid self-improvement, from international diplomacy to geopolitical arms race.
#AIgovernance #FuturesThinking #TechPolicy
Read: cfg.eu/advanced-ai-...
#AIgovernance #FuturesThinking #TechPolicy
Read: cfg.eu/advanced-ai-...
According to Google, over 50% of code characters in their IDEs are now written with AI-based support, with tools resolving 8% of code review comments and adapting pasted code to context.
#AI #SoftwareEngineering #CodeCompletion #DevTools #AgenticAI
research.google/blog/ai-in-s...
#AI #SoftwareEngineering #CodeCompletion #DevTools #AgenticAI
research.google/blog/ai-in-s...
July 11, 2025 at 11:18 AM
According to Google, over 50% of code characters in their IDEs are now written with AI-based support, with tools resolving 8% of code review comments and adapting pasted code to context.
#AI #SoftwareEngineering #CodeCompletion #DevTools #AgenticAI
research.google/blog/ai-in-s...
#AI #SoftwareEngineering #CodeCompletion #DevTools #AgenticAI
research.google/blog/ai-in-s...
How I imagine these agents
June 28, 2025 at 11:56 AM
How I imagine these agents