People can simultaneously be getting exploited while that exploitation helps AI systems to get more capable.
As AIs get more capable, misuse risks start increasing like hacking and bioweapons, and if they continue improving the risks become much bigger.
Shouldn't we prepare for this possibility?
The world is big enough for people to work on addressing different problems without needing to dunk on the others.
AIs objectively act with agency at this point, you can ask them for an app and they can work for hours, pose and answer hypotheses, etc...while their capabilities continue to increase?
I'm not and never have been a rationalist or EA. In fact, I think these groups are largely to blame for the AI situation.
I grew concerned about AI while studying physics because the trajectory of AI is concerning, and ASI would obviously go horribly wrong.
Ignoring this is not good strategy.
Back from holidays and I can't stress enough how much basically everyone I've met outside of the tech sphere hates "AI" as a monolith (even if they use it for work).
Grok undressing everyone is turning it up to 11.
SF better be ready to wake up in 2026.
ControlAI has just launched our new campaign pushing for a ban on superintelligence! We've gathered some big-name endorsements and more are on the way. Check it out at campaign.controlai.com
We also put together a release video for the launch! www.youtube.com/watch?v=oAJU...
Honored to have been interviewed for
@washingtonpost.com!
While they missed my overall point about how GPT-5 is still very much "on trend" towards superintelligence *despite* no step-change, it's great to see mainstream media casually discussing the dangers of superintelligent AI.
wapo.st/3JleHgE
The most vile "we must risk ASI" argument is we're "doomed" without.
When you've given up on humanity and have no vision for the future beyond "summon a demon to see what happens" it's no wonder you end up pro-extinction.
Traitors to humanity should be kept far from the wheel.
This is a critical issue we cannot afford to ignore, despite the chaos around the world right now. Contacting your representatives is quick and easy, send them an email today!
In his article he explicitly calls for the US to take over the world with self-improving superintelligent AI systems (that he has no idea how to control)
The negative game theoretic incentives here are strong. The only hope for humanity to survive is via international coordination to immediately stop our suicidal race to build uncontrollable superintelligent aliens in the hope that we can wing a solution to chain them to our will.
@anthropic.com is not the "AI Safety Lab"
CEO Dario Amodei is now explicitly calling for a global US takeover via superhuman AI systems Recursively Self-Improving far beyond human capabilities.
This is unbelievably dangerous. We cannot control such systems. Humans do not "win" this arms race.
These models don't abide by our intuitions about intelligence. Just because they can't count the r's in "strawberry" doesn't mean they can't be really good in other areas.
Even if they aren't "true" intelligence or whatever, they still can posses worrisome capabilities.
My take on this is that because Math+CS are easier to verify, they are easier to implement in the current RL paradigm.
My concern is that we're on the verge of "spiky" AGI that is superhuman in some but not all ways. It might be that Math+CS are all we need for RSI to take over and spiral away.
Currently using TS + Node.js, but have been looking into backend services like Vercel, Appwrite, and Supabase. Do you have experience or thoughts on these versus custom backends?
That makes sense, I started out doing data science with python as well. Despite mostly working with typescript now, I still miss the more readable python (and no semi-colons haha)
Now that you're more established have you found any drawbacks to this setup, any major annoyances?
This is a really cool idea. Do you know of any major sites that do this? I recently learned that 1password works offline which was convenient in a pinch.
Can something like this be implemented using backend services like Vercel, Appwrite, or Supabase?
That's awesome! What drew you to Vercel vs other options like Supabase? Which features are you implementing in your backend?
Nice!
Yeah everything has worked so far, it's taken a bit longer to set up but is probably cheaper than using a service.
Other than auth, what do you need your backend to do? Have you considered which service you would look into?
Big move! What made you pick NextJS on Vercel vs other backend services like Supabase?
How's your experience been with Vercel so far? I'm looking around at different backend solutions and it seems highly recommended
Wow, that's a lot for one role! How do you manage it all? Do you build custom backends or use backend services like Vercel or Supabase?
Nice, were you able to implement the entire backend in Supabase, or did you need to do anything outside of it?
Sounds great! What have you built so far with it?
Slick, are you able to do your entire backend with Supabase or do you have anything outside of it? I've talked to some people who end up with a Next.js backend linking to the Supabase backend
Oops yeah, thanks!
Also curious why you go for python + firebase?
Nice! What are you using? Have you built functions to do backend things like DBs, users, and pyaments that you use between projects?
This is really cool! What's your stack for the project?
We've built the backend from scratch on our project, it's definitely been a bit of a pain to get everything integrated tho. Why did you choose the custom route vs. backend services?
How has your experience been with Supabase? Have you tried any other offerings or coding your own backends? Any tradeoffs that you wish you didn't have to make with it?
Was this the primary issue you ran into? Could there be an SaaP that you would use if it simplified the dev process without imposing limiting restrictions?