XKCD Bot#3254 - Detector
| GitHub / Twitter | Fishrock123 |
| Website | https://www.jeremiah-senkpiel.com |
| Location | North Okanagan BC |
Jeremiah Senkpiel
@Fishrock
- 213 Followers
- 166 Following
- 1.8K Posts
Staff Rust-lang developer. Backcountry powersports nerd. Adventure motorcyclist. Dirtbiker. Snowmobiler. Ski tourer. Trad rock climber. Former dungeon master.
Jacob PrattI am available (immediately) for Rust work! Whether it is finding nontrivial performance gains, reviewing code for correctness and soundness, or writing implementations from scratch, I am a knowledgeable engineer with experience at companies large and small from various industries.
If your company would like to discuss setting something up, contact me via the email on my GitHub profile (https://github.com/jhpratt).
WaltteriIt occurs to me now that even JEPA based AI models are missing a key feature of human communication - the ability to separate ideas fully.
It feels as though a hierarchical system needs to have the ability to distinguish if there are multiple ideas in a given input and be able to put those into their own separate high dimensional representations at lower levels.
So if you tell Claude code to “not delete any files”, it is literally incapable of actually respecting that.
It seems unclear to some people, so let’s be clear: telling an LLM what to do does not “tell it what to do”. They cannot truely process instructions. All they can do is give you a median centered normal distribution of text that would follow from such text as provided. Yes, they can maintain context and focus, but they cannot plan by instruction.
Jed BrownRE: https://neuromatch.social/@jonny/116667022863059853
Excellent thread on what "LLM-assisted" testing actually does, with context from rsync replacing its test suite with a poorly vibe-ported Python version (as regressions roll in)
More generally, the primary product of LLMs is deception: the unfounded belief by the insufficiently-critical observer that the task has been performed. From vibe-coders to user communities, managers, and investors, there are many targets of the deception. Many are playing along for their own self-interested reasons (e.g, executives need investor confidence more than they need a product). And many feel hit when peers and communities see through it.
David GerardIf you previously donated to the Wikimedia Foundation (e.g. via the beg notices on Wikipedia), it is still good to donate to the local chapters if there's one in your country. They're chronically underfunded (e.g. by the WMF) and generally are tax deductible charities too.
Florian GilcherA massively multiplayer discussion game: Naïve Online
Annika Backstrom@chris Learning is forever but a Claude Code subscription bills monthly.