I’m glad we wrote that paper. However LLMs “still lack basic reasoning skills” makes me cringe.
Information theory tells me that because an LLM is a finite set that is not able to grow itself, once it is trained has a finite capability. And that capability is driven by statistics and numbers.
intuitively (to me at least) if you present an LLM with a prompt that’s weird enough it will “hallucinate” answers because it has no critical thinking, it’s just a big probability machine that tries to find the most likely answer to your question. As a result, present an LLM with a chess problem brain teaser unique setup, chances is the LLM will make up rules because what it trained against isn’t chess rules but “in general chess problems end with a checkmate” and it will interpolate the movements from where you are to a checkmate.
https://mastodon.social/@appleinsider/113295305642702643
Oh yes we have our new “you wouldn’t download a car”
The whole of my book on Building a Debugger is now available on Early Access!
It teaches you how to write a native code debugger from scratch.
There's lots of cats.
- “Check the domain” doesn’t help if you have no information about what domains are “normal”
Damn right. I rememberd doing some phishing training from Google, and they asked if some email sample is legit, and demonstrated its legitimacy by pointing out that it uses legit DropBox domains.
That’s where I thought “but no one here (DropBox isn’t accessible in China) uses DropBox at all! How are they going to learn if this is legit?!” 🫠
Lets Encrypt will disable OCSP about 6 months after Microsoft Root program allows it to (the browsers have already okayed it).
This all could be over in a year, year and a half. If you need OCSP for your business, you need to investigate alternatives NOW - which are all proprietary.
Apache ACME will handle this change just fine. Stapling will of course no longer be provided to clients.
https://letsencrypt.org/2024/07/23/replacing-ocsp-with-crls/