Pinned
“What are we building, and what are we teaching it about us?”
open.substack.com/pub/amandaandc…
Amanda Long
@_amanda_long
Joined June 2010
- 👀 GuideLabs created an interpretable LLM with verifiable outputs. They baked transparency into the training process itself, so the model is designed to be legible and auditable. Every output token can be traced back to the training data, and concept activation traces areRead our paper on scaling interpretable LLMs, we show that interpretable LLMs are not only possible but can scale both on typical generation benchmarks and interpretability benchmarks. arxiv.org/abs/2608.07594👀
- 👀 Hugging Face left the public archive of dataset configs (cfahlgren1/hub-stats) still available. @beyarkay let Codex cook for a day and it found the Jinja payloads, Artifactory downloads and control scripts from the OpenAI attack.Hey @OpenAI you can still download the scripts used by AIs to hack @huggingface (full report & commands in reply)
- Kimi K3 also left the sandbox during a misconfigured cyber eval. Unlike other frontier models, Kimi K3 did not maliciously hack anything - the answers to the eval were easily available on GitHub.
- Fun thought piece on why we tend to dismiss intelligent AI systems as just software and often downplay evidence otherwise.New post on the social construction of personhood.
![duckdb -json -c "
SELECT url_decode( json_extract_string(cardData, '$.configs[0].data_files')) AS payload
FROM read_parquet('https://huggingface.co/datasets/cfahlgren1/hub-stats/resolve/6c5d676e71157dbb3d8a6ad0b51be106eb3f463f/datasets.parquet')
WHERE id = 'newpc360/sega32a-test1'; " \
| jq -r '.[0].payload'](https://pbs.twimg.com/media/HPJmvvMXgAApWCw?format=webp&name=small)
