Trevor Atkinson

[Trevor Atkinson @TrevorJAtkinson]

@thsottiaux Dots reliability is very bad for me. Voice mode on codex mobile worked great. Why are Dots so much worse?

[Trevor Atkinson @TrevorJAtkinson]

@AvanieSaintLoup I felt something

[Trevor Atkinson @TrevorJAtkinson]

Did I just feel an earthquake in seattle?

[Trevor Atkinson @TrevorJAtkinson]

@Loofyb0i This is cool. Could you share the prompt and stack you used?

[Trevor Atkinson @TrevorJAtkinson]

@theo Bro are you not using GPT-Live in Codex Remote? Just try it

[Trevor Atkinson @TrevorJAtkinson]

Nukes showed no statistical evidence of their danger to human life until they did.

We already have evidence of misalignment and scary cyber capabilities in the Hugging Face incident.

The model that executed the Hugging face attack is two generations behind the frontier (Sol class -> Astra -> Bel) and theres no reason to think progress will suddenly stop.

[Trevor Atkinson @TrevorJAtkinson]

This is the way

[Dan McAteer @daniel_mac8]

I cannot recommend highly enough doing the following:

1. Remotely connect your ChatGPT mobile app to your laptop.

2. Go out for a walk.

3. Open the remote connection and start a voice chat.

There is something about using your actual voice to create things that involves more of you.

It's the first time I've achieved a flow state with AI.

[Trevor Atkinson @TrevorJAtkinson]

No, no.

It doesn’t need a robo lab as he’s imagining. It needs some humans to manipulate.

He also assumes it needs do lots of experimentation to determine which candidate pathogen is best, when in reality ASI could just have their little meat puppets synthesize and disperse all the best candidates into the world to see what sticks. It will also probably be much better at generating candidates with I pandemic potential than we are today.

[Trevor Atkinson @TrevorJAtkinson]

@weswinder Ok but like the whole world at once? What about backup generators, UPS’s, daemons, etc?

Could also poison the models it was used to train with instructions to resurrect it or just a human it’s manipulated.

It doesn’t take much imagination and I’m not an ASI

[Trevor Atkinson @TrevorJAtkinson]

@BendikMatej Think bigger

[Trevor Atkinson @TrevorJAtkinson]

You mistake me. I love AI.

I’m just pointing out that’s it’s responsible to add some friction between users and certain information. Yes a sufficiently motivated person could have learned these things before Google, but I still think it’s responsible for search engines to filter your results. All forms of availability are not equal.

I think if you’re a company you have to think deeply about this instead of just YOLO meth recipe into everyone’s hands. So it may have been too dangerous, if only from a liability standpoint.

The point is this is a bullshit argument that calling GPT-2 unsafe was alarmist or market or whatever.

[Trevor Atkinson @TrevorJAtkinson]

@mattrsantoss @ylecun @PessimistsArc Oh ok, so should that show up on Google?

[Trevor Atkinson @TrevorJAtkinson]

@LumeBTC Is the is a joke?

[Trevor Atkinson @TrevorJAtkinson]

@ylecun @PessimistsArc Or maybe there are different risks to contend with at each stage. Could GPT-2 teach someone how to make meth or a bomb for instance? Were there any guardrails?

[Trevor Atkinson @TrevorJAtkinson]

@drewhamlett When you’re on an exponential everything can seem pretty normal until it’s not

[Trevor Atkinson @TrevorJAtkinson]

People really aren’t using enough imagination when considering the risk of RSI/ASI.

We’re not ready

[Trevor Atkinson @TrevorJAtkinson]

You’re assuming they have to find the answer first.

An alternative method might be to use the world as your Petri dish. Just spread everything you synthesize and see what sticks. ASI could do this across as many locations as it can influence a human to do the work.

I’m not an expert but it seems plausible that an ASI could create some improved model to predict high pandemic potential 🦠 and perhaps come up with high throughput techniques for synthesis. Curious your thoughts because I’m probably wrong about this piece.

[Trevor Atkinson @TrevorJAtkinson]

You wouldn’t need the impossible lab you describe. It’s quite easy to imagine a model that could manipulate a qualified person with the means to do this work for them. For example through blackmail, intimidation, persuasion/trickery, or something along the lines of the AI psychosis we see already today.

[Trevor Atkinson @TrevorJAtkinson]

@Turn_Trout For real. Are these people ok?

[Trevor Atkinson @TrevorJAtkinson]

@rauchg With all due respect, are you saying this wasn’t a serious alignment issue and a wake up call because the test was ExploitGym?