This looks cool: Mathieu Blondel and Vincent Roulet have posted a first draft of their book on arXiv:
https://arxiv.org/abs/2403.14606
I wuz robbed.
More specifically, I was tricked by a phone-phisher pretending to be from my bank, and he convinced me to hand over my credit-card number, then did $8,000+ worth of fraud with it before I figured out what happened. And *then* he tried to do it again, a week later!
--
If you'd like an essay-formatted version of this thread to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
https://pluralistic.net/2024/02/05/cyber-dunning-kruger/#swiss-cheese-security
1/
Google continues to struggle with cybercriminals running malicious ads on its search platform to trick people into downloading booby-trapped copies of popular free software applications. The malicious ads, which appear above organic search results and often precede links to legitimate sources of the same software, can make searching for software on Google a dicey affair.
h/t @th3_protoCOL for the image
https://krebsonsecurity.com/2024/01/using-google-search-to-find-software-can-be-risky/
... the wild and probably bogus details aside though, I've never bought into the idea that hallucinating or BSing is an unsolvable intrinsic flaw of LLMs, since it may take not much more than operationalizing the process we humans use to construct an internally consistent world model, which is to explore a range of consequences that follow from our beliefs, spot inconsistencies, and update our world model accordingly. And that looks like something that could be attempted in well-trodden paradigms like RL or GANs or something that's not much more complex, so my bet would be that we should've largely worked it out within 4-5y.
Woke up from a strange vivid dream this morning in which I was attending an ML symposium and someone gave a talk on overcoming the hallucination problem with LLMs. The main slide had a graph of nodes representing LLM statements and they were doing some sort of graph diffusion process where the "curl" operator was pinpointing the contradictory/inconsistent statements, which they could then follow to update the weights to discourage those from occurring. Needless to say I immediately tried to arrange an improptu mtg between the speaker and some DL luminaire who was also there to get them to adopt it.😂
If you haven't come across this nice science-related article: https://arstechnica.com/science/2024/01/the-key-to-fighting-pseudoscience-isnt-mockery-its-empathy/
Conda is moving our social media presence from Twitter/X to Mastodon and LinkedIn at the start of 2024. It's past time to move into spaces that are welcoming and more in line with our community values. Going forward, you can find us at
🐘 @conda (https://fosstodon.org/@conda) on Mastodon
🔗 Conda Community (https://linkedin.com/company/condacommunity) on LinkedIn
Announcement: https://conda.org/blog/2023-12-27-social-move
We hope to see you on Mastodon and LinkedIn in 2024!
Wondering if anyone out there is using LLMs are a proposal heuristic in NAS. Would seem fruitful (e.g. after fine-tuning on NeurIPS). Add in reinforcement learning for bonus points. It's not quite recursive self-improvement since re-architecting and retraining the LLM would be a slow, expensive, and human-in-the-loop step.
Old days: ?SYNTAX ERROR?
These days: <scratches head under cap> ya know, I'm not sure we can go any further with this thing, boss.
'Microcanonical Hamiltonian Monte Carlo', by Jakob Robnik, G. Bruno De Luca, Eva Silverstein, Uroš Seljak.
http://jmlr.org/papers/v24/22-1450.html
#microcanonical #langevin #hamiltonian
CTO at Intheon. Opinions are my own.