Nice Reads

Some nice reads and initiatives.

Collection of nice initiatives

The School of Life - on the skills everyone should be taught, but no school actually does (on Instagram, LinkedIn, YouTube)
About how to use your working life in a purposeful way and many general life advices: 80,000 hours carrier guide
GiveWell - Highly recommended on effective donations

If in Austria: can donate to effektiv-spenden.org s.t. the donation is automatically deducted from tax

Why renewables became so cheap so fast - article
"The world is awful. The world is much better. The world can be much better." - blog post

The School for Moral Ambition - Interesting initiative on making moral ambition part of your work life

CrashCourse Philosophy YouTube Series
exurb1a: Nice philosophical mind games on YouTube
"HUMAN" movie series by Yann Arthus-Bertrand: portraits of people around the world about life - YouTube movies

Google paper on the importance of data before model work
"The oldest monetization engine on the internet is coming to save the newest. From VHS to broadband to streaming, and now to LLMs. When in doubt, sell dopamine 🤷‍♂️. [...] Because every tech revolution starts with ideals, then ends with monetization." LinkedIn opinion
Highly recommended: "AI Should Challenge, Not Obey": preprint opening the door for an new view on human-centered AI-based content generation. With AI as a provocateur encouraging critical thinking of humans instead of simply creating a final response/output based on a prompt. paper
When models fail to reason beyond known benchmarks paper
Interesting benchmark on real business tasks: OpenAI GDPval
Benchmark on tough expert questions: Humanities Last Exam (website)
"State of AI in Business 2025" -> the famous 95% of Ai projects fail MIT report
DeepSeek OCR release paper and high-level explanation
OpenAI: Why models hallucinate. On how gpt-5 has a huge reduction in false positive answers (hallucinations) - blog post
"Emergent misalignment": Fine-tuning gpt-4o on writing insecure code lead to major misalignment of the model in many other fields (creating deceptive, harmful content unrelated to code generation) - paper
Open Models of OpenAI
Rinat Abdullin on **Schema Guided Reasoning** utilizing structured outputs (constrained decoding of LLMs)

How the TIMETOACT Enterprise RAG Challenge was won: Lots of insights on building a working RAG solution for extracting KPIs from annual company reports - blog post
Subliminal learning: "During distillation on model-generated outputs, students exhibit subliminal learning, acquiring their teachers’ traits even when the training data is unrelated to those traits." and "Subliminal learning occurs for different traits (including misalignment), data modalities (number sequences, code, chain of thought), and for closed- and open-weight models." - paper
On which tasks and jobs have high AI applicability (i.e. likely of being strongly affected by AI) - paper
Google Agent2Agent Protocol (here)
First large scale transformer-Mamba fusion LLM: Hunyuan T1
Facts & Data about AI trends: https://epoch.ai/
Tracing the thoughts of a large language model - Anthropic blog post
Anthropic Model Context Protocol (MCP) - blog post
FunSearch by Google - evolutionary agent to come up with new (code) solutions for complex problems (blog)
DeepSeek R1 release - challenging OpenAIs o1 model

Thoughts: Compute will be one of the most precious resources in the future (after oil and data)