2026-06-21 (Sunday) Journal

Day summary: Today’s RSS activity: 2 item(s) from infosec.exchange including “Training an LLM on a heavily cleaned, de-identified corpus can be like correcting every grammatical mistake in a large collection of texts: the result may look cleaner, but it can also lose the context, variation, and imperfections that reflect real-world language and behaviour.A corpus scrubbed of every sensitive detail and irregularity can become a polished imitation of reality. Privacy protection could be necessary, but a model trained mostly on synthetic or over-sanitised data risks producing equally synthetic answers: coherent on the surface, yet disconnected from the messy context in which real life happens.So what do you expect from an LLM following the regulation? and over-cleaning the input data? #llm #ai Ref: https://ora.ox.ac.uk/objects/uuid:fa1155e9-c2ff-436a-8391-455b622f4e64/files/r3b5919575 “AI models collapse when trained on recursively generated data””; “« Once an organisation accepts that the difficult software will be bought elsewhere, internal teams slowly lose the habit of building. Procurement becomes a substitute for strategy. Legal review becomes a substitute for leadership. Risk management becomes a substitute for execution. »https://foo.be/2026/06/Sovereignty-Is-Engineered-Not-Procured.html#sovereignty #europe #opensource”.

Hourly activity:

hour  00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
pulse  .  .  .  .  .  .  .  .  .  .  .  .  .  @  .  .  .  .  .  .  .  .  .  @
count 00 00 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00 00 00 00 00 00 00 01

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

adulau pushed gcve-enriched-dumps

Training an LLM on a heavily cleaned, de-identified corpus can be like correcting every grammatical mistake in a large collection of texts: the result may look cleaner, but it can also lose the context, variation, and imperfections that reflect real-world language and behaviour.A corpus scrubbed of every sensitive detail and irregularity can become a polished imitation of reality. Privacy protection could be necessary, but a model trained mostly on synthetic or over-sanitised data risks producing equally synthetic answers: coherent on the surface, yet disconnected from the messy context in which real life happens.So what do you expect from an LLM following the regulation? and over-cleaning the input data? #llm #ai Ref: https://ora.ox.ac.uk/objects/uuid:fa1155e9-c2ff-436a-8391-455b622f4e64/files/r3b5919575 “AI models collapse when trained on recursively generated data”