The Hard Problem of Controlling Powerful AI Systems - Computerphile

Comments · 333

• DA

  @davidswinstead · 9 months ago

  You ask an agent to find a machine's IP address, go for coffee, and when you come back the agent has SSHed in and upgraded the operating system..... my guy you are WAY TOO CHILL about this haha

  202

• RI

  @ricoreyes6044 · 9 months ago

  It's like the Simpsons episode where Homer puts the drinking bird at the computer - "Why did I leave you in charge?!"

  81

• BO

  @BobbyKingOfTheSeas · 9 months ago

  Be it morning, afternoon or evening, whenever I get notification I just can't resist watching it. I like the way how simple these guys explain no fancy editing, animations, sponsors, AI voice, just a dude with pen and paper sharing his knowledge.

  31

• JO

  @joshua_tobler · 9 months ago (edited)

  Imagine if we took an LLM that was specifically trained on strategic game theory, and could fine tune itself by playing a series of iterated strategic games. And then we hooked it up to military systems like, for example, our missile defense system and let it make autonomous agentic decisions.<br><br>What could possibly go wrong?

  60

• JO

  @jokmenen_ · 9 months ago

  Great video, great explanation and a great concept. This really inspired me, thanks!

• EZ

  @ezgarrth4555 · 9 months ago

  Look, I have experience using Claude as a coding agent, I use it every day for my job (they insist upon it, even). It is because I have that experience that I would never let an AI agent act unsupervised like you propose I should intrinsically want

  174

• LO

  @Locut0s · 9 months ago

  I’m increasingly more worried about the unintended side effects of AI that have nothing at all really to do with constraints or alignment. There is nothing “wrong” with the attention algorithms that are behind much of the net right now before AI. All they are designed to do is keep people on a site. Unintended side effect, radicalization, social fabric fraying at the edges AI is likely, and already is, accelerating this and that’s when they are working perfectly.

  16

• HO

  @HowieStephens · 4 days ago

  These somewhat older videos are actually super interesting to watch a year later 🫠

• AN

  @AnthonySennett · 9 months ago

  I used the GH agent recently and it created and included a binary in the PR. When I pointed it out it silently removed it, but I have no idea where it got that binary from. The code change was pure business logic and associated tests.

  37

• B3

  @b3night3d · 9 months ago

  That&apos;s on you for giving it sudo access.

  3

• AN

  @anthonyhart7878 · 9 months ago

  It can run long commands but its reading logs like crazy which fill up the context and eat up your tokens...

• MR

  @MrTightFilms · 9 months ago

  Teams notification at <a href="https://www.youtube.com/watch?v=JAcwtV_bFp4&amp;t=502">8:22</a> got me