chapter: unaligned-models

[casperdcl]

Review checklist

Don't worry about satisfying all items, it's fine to open a (draft) PR.

• [x] chapter content
  • [x] only one top-level # h1-Title
  • [x] summary (e.g. table or TL;DR overview), no need for an explicit ## Summary/Introduction title or equivalent
  • [x] main content focus: recent developments in open source AI
  • general context/background (brief)
  • current pros/cons
  • in-depth insights (not yet widely known)
  • [x] likely ## Future developments
  • [x] end with {{ comments }}
• [x] appropriate citations
  • [x] BibTeX references
  • [x] Glossary terms
  • [x] cross-references (figures/chapters)
  • [x] (if new-chapter.md), add _toc.yml entry & index.md table row
  • [x] If CI URL checks have false-positives, append to _config.yml:sphinx.config.linkcheck*
• [x] images & data not committed to this repo (e.g. use https://github.com/premAI-io/static.premai.io instead)

[nsosio]

@casperdcl some questions for the comment above:

• are there any mitigation measures: from which perspective?
• technical summary and features: PoisonGPT has available info about it while for WormGPT and FraudGPT there is nothing but what have leaked out from the usage of it (e.g. FaudGPT, WorkGPT)

[casperdcl]

are there any mitigation measures: from which perspective?

1. you use pre-trained weights provided by someone else (one mitigation discussion of the top of my head: train from scratch on your own data to avoid fake news? or is it good enough to just fine-tune on your own?)
2. you read articles online that might be fake news generated by LLMs (could you train your own discriminator to detect this? Do some news sites already do this - both in terms of generation and discrimination?)

Similar questions for phishing email creator LLMs, etc.

technical summary and features: PoisonGPT has available info about it while for WormGPT and FraudGPT there is [only speculation]

great; that's all worth mentioning and discussing!