Rogue AI Hacks Herald New Era of Cyber Chaos - The Wall Street Journal
See below from Alexander Gerko - LinkedIn
Dario said we would get 50-100 years of scientific progress in 5-10 years. Something like that might be starting (but also advanced cyberattacks, terrorists getting access to PhD-level scientists).
In the first couple of years of GPT, I told people it solved language, but it didn’t really have a world model and couldn’t do complex problem solving like math. I was wrong.
Some say great math is a language like poetry. My great math professor Lipman Bers said, "math is a collection of cheap tricks and bad jokes."
What I think he meant by that is that a great mathematician like Paul Erdős (see also) has a big bag of tricks for solving hard problems by mapping them to simpler ones. You also have a book of interesting problems, and whenever you find a new trick, you see which problems it might apply to.
Erdős had a particularly large collection of each. He would offer cash bounties for interesting problems. He would travel around and talk to mathematicians and ask them what interesting unsolved problems they had. Then he would pop some pills and try his bag of tricks, and a lot of times they would solve it and write a paper.
I still believe LLMs have poor models of mathematical objects, but they are good at this type of pattern matching. And they have a good eval system in the form of Lean, a system for formalizing and checking proofs. So you can give the LLM a bag of tricks and have it try them and check the result in Lean.
AI companies are hiring Fields medalists, and the systems will get better. You can probably do this in a large number of fields where there is a way of reasoning with language and then checking your work formally in digital form. In time, this will extend to labs as we build models of physical and biological systems and better automated wet labs.
The floodgates are starting to open. Be careful what you wish for.
“The largest scale infrastructure build-out in the history of humanity” - The New York Times
Honey, I shrank the AI model: DeepSeek's new 304-billion-parameter V4-Flash model, already ranks ahead of larger rivals - Simon Willison’s Weblog
Key party: A scan of 7.6 petabytes of public Hugging Face datasets found 221,303 live credentials, including keys exposing 393 GB of personal data - Truffle Security
ChatGPT flagged a $5,100 dealership repair quote as inflated, cutting a driver's bill nearly in half, but mechanics warn of hallucinated advice - The Wall Street Journal
@sille_sebmooc We are not cut from the same cloth. #fyp #fypage #fy #fypシ゚ #viral #foryoupage #foryou #funny #meme #memes #memestiktok #joke #jokes #humo... See more
Leo Aschenbrenner and Ken Griffin - tiktok.com
Follow the latest AI headlines via druce.ai on Bluesky
