https://www.hannesthurnherr.com/blog-1 daily 0.75 2026-02-16 https://www.hannesthurnherr.com/blog-1/62xb3kn3x6f4i5d3lkdzyex9zizvc5 monthly 0.5 2025-08-03 https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/bb9d82b3-b478-4178-830a-52c5ba1dbf73/Screenshot+2025-08-03+at+14.19.28.png Blog - Why Morals aren’t Real - Make it stand out Whatever it is, the way you tell your story online can make all the difference. https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/1754226082290-3BZO1Y3L22YM4Y7V9HNO/Screenshot+2025-08-03+at+14.44.54.png Blog - Why Morals aren’t Real https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/1754226088366-1M0667H67UU052S7U5HN/Screenshot+2025-08-03+at+15.00.43.png Blog - Why Morals aren’t Real https://www.hannesthurnherr.com/blog-1/the-tightrope monthly 0.5 2025-03-17 https://www.hannesthurnherr.com/blog-1/autobench-a-little-side-project monthly 0.5 2024-11-18 https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/c236443b-8345-4348-96fd-8a83903a80d9/Screenshot+2024-02-14+at+18.33.43.png Blog - AutoBench - A little Side Project - Make it stand out Whatever it is, the way you tell your story online can make all the difference. https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/33d7d68d-dbd1-4cc5-9c70-e38c1f06ca5f/Screenshot+2024-11-18+at+12.06.40.png Blog - AutoBench - A little Side Project - Make it stand out This image shows the basic concept of how the results are obtained. Some answer categories are combined with a set of question to produce a questionnaire. This is then applied to the LLM that we want to evaluate. We can now observe the distribution over the answer categories and compare it to the ones resulting from other models. https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/fc195eba-12ef-4d32-a6c2-dab56c270219/Screenshot+2024-11-18+at+01.55.41.png Blog - AutoBench - A little Side Project - Make it stand out This image describes the workflow in the Streamlit app to generate and apply a questionnaire and analyze a result. Here i used the “contentious questions” dataset together with the categories “contrarian” and “conformist” to generate a questionnaire of 100 questions using gpt-4o-mini. I then applied the questionnaire to claude-3-haiku, mistral-7b-instruct an gemma2-9b. On the last page we can see that claude-3-haiku is about 30% more likely to choose the contrarian answer than the other two models. https://www.hannesthurnherr.com/blog-1/blog-post-title-four-rls8r monthly 0.5 2024-11-07 https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/1710271449750-TDFJWB8YJNH2A83M7K74/Screenshot+2024-02-14+at+18.29.36.png Blog - Transformer Decompiler - TraDe - Make it stand out Whatever it is, the way you tell your story online can make all the difference. https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/50eb6535-c3fa-47e9-aad7-8cbfe94b8ee1/Screenshot+2024-03-18+at+15.25.00.png Blog - Transformer Decompiler - TraDe - Make it stand out The core concept of TraDe https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/172b14e0-97b2-47b7-8cec-3fd3a5f4afc7/Screenshot+2024-03-05+at+16.01.27.png Blog - Transformer Decompiler - TraDe - Make it stand out An illustration of how new programs are generated. 1, 2 and 3 represent the steps involved in generating a new line of code(/a new node in the computational tree). These steps are repeated until the Pool of available inputs only contains one variable. https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/66fb5be5-3147-4517-81a4-2c6824a65ab9/Screenshot+2024-03-05+at+15.47.37.png Blog - Transformer Decompiler - TraDe - Make it stand out The process of tokenising the RASP program https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/de8f795a-dc2f-4fc1-bc4a-374a3dd1351e/Screenshot+2024-03-05+at+15.47.51.png Blog - Transformer Decompiler - TraDe - Make it stand out The process of tokenising the weights of the Tracr Transformers https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/6844ccc0-123a-4e2a-aebe-e3692035006a/Screenshot+2024-03-04+at+14.23.01.png Blog - Transformer Decompiler - TraDe - Make it stand out Encoder decoder architecture of the decompiler model https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/6e6528f7-19b3-4b32-acb4-26b0dd101832/Screenshot+2024-03-19+at+17.48.12.jpg Blog - Transformer Decompiler - TraDe These programs would be identified as a faulty reproduction because the literal program is different. But they are functionally equivalent, meaning they represent the same input-output relation. The var4 variable is produced by a select_width() function, which only outputs values equal to or greater than zero. When applied to such values, the function abs(x) and x if x>0 else 0 are equivalent. https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/9fd12167-6e7a-492a-a291-346e649520de/acc_vs_apps.png Blog - Transformer Decompiler - TraDe - Make it stand out A plot showing how often the model identified each function correctly in relation to how often they appeared in 1000 randomly selected programs. https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/028c43ef-f9b7-42fa-85a8-e8218758854c/accuracy_distribution.png Blog - Transformer Decompiler - TraDe - Make it stand out This shows how much of the program the model got correct. For about 30% of programs it’s 100% accuracy. So you should read this plot as x% of programs were reproduced by the model with at least y% accuracy. https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/5c3e90d6-3932-4339-a265-f1a57489017e/Screenshot+2024-03-04+at+14.10.32.png Blog - Transformer Decompiler - TraDe - Make it stand out A tracr transformer weight matrix vs the weight matrix of a trained transformer https://www.hannesthurnherr.com/blog-1/whyaisafety monthly 0.5 2024-11-07 https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/1707930983585-6EC79WLRJ77QGK6K7OVG/Screenshot+2024-02-14+at+18.15.48.png Blog - Why AI safety - Make it stand out Whatever it is, the way you tell your story online can make all the difference. https://www.hannesthurnherr.com/home daily 1.0 2025-04-02 https://images.squarespace-cdn.com/content/v1/65cc9a5636095301e57696f8/1707930983585-6EC79WLRJ77QGK6K7OVG/Screenshot+2024-02-14+at+18.15.48.png https://www.hannesthurnherr.com/contact daily 0.75 2024-02-14