AI Safety Fundamentals

Follow our work on @bluedotimpact

[Josh Landes @guynamedjoshl]

Last week we co-hosted our first hackathon with @join_ef

This week we're announcing a $1M fund for our new AGI Strategy Course

Next week the first funded project enters the world

We move fast at @BlueDotImpact

You want in? Well, are you an LBE? 🧵

[Dewi Erwan @dewierwan_]

We’re putting $1M behind people who want to ship solutions to AI risks.

You’ve seen the rate of AI progress. You know you need to do something.

But you’re uncertain. What should you build? What’s the most impactful thing you can do?

[Josh Landes @guynamedjoshl]

AI models are becoming ever more capable, enabling risks at unprecedented scale and complexity. That's why we're teaming up with @EntrepreneursFirst and @WorkshopLabs to host our first AI Security hackathon!

2k GBP in prizes across winning teams.

[Josh Landes @guynamedjoshl]

🚨 We're hiring at @BlueDotImpact to build the AI governance pipeline.

Your mission: Figure out what governance is needed to make AI go well, then build the workforce to make it happen.

Imo, this is a top 1% role if you want influence on AI's trajectory 🧵

[BlueDot Impact @BlueDotImpact]

@BarackObama AI is shaping the future. So can you! Take our free online course to better understand AI’s impact and be part of the conversation about its future. We've trained more than 7000 people. Will you be next?

https://t.co/1aVdCweAaJ

[BlueDot Impact @BlueDotImpact]

Reinforcement Learning from Human Feedback (RLHF) is the main technique used to make language models output responses that are helpful to human users.

How does it actually work? Our blog post explains!

Link below 👇

[BlueDot Impact @BlueDotImpact]

Can we understand more about AI's reasoning process by observing its chain of thought?

Read our new blog to find out 👇

[BlueDot Impact @BlueDotImpact]

Do models that reason for longer before answering produce more aligned responses?

That's the bet behind OpenAI's deliberative alignment strategy.

How does this work, and will it be enough?

Check out our explainer 👇

[BlueDot Impact @BlueDotImpact]

How can governments influence the future of AI?

One lever at their disposal is compute – the cutting-edge chips used to power frontier models.

Our introduction to compute governance covers the basics!

Link below 👇

[BlueDot Impact @BlueDotImpact]

As AI models get more powerful, it will become harder and harder for humans to evaluate their outputs.

Recursive Reward Modelling proposes using AIs to help evaluate other AIs.

Read more at the link below 👇

[BlueDot Impact @BlueDotImpact]

Research shows that weaker AI models can sometimes be used to steer and improve stronger ones.

Maybe, this means that humans could supervise AI systems that are much more capable than us.

Our new blog post explains 👇

[BlueDot Impact @BlueDotImpact]

Scaffolding can make AI models more capable after they have been deployed.

Our new blog post explains how scaffolding works, what risks it poses – and how we might be able to use it for safety.

Link below 👇

[BlueDot Impact @BlueDotImpact]

AI Control aims to prevent models from causing harm, even if they are trying to.

Could this really work? Our new blog post explains!

Link below 👇

[BlueDot Impact @BlueDotImpact]

Can we make AI systems safe by having them debate each other?

Our new blog post explains (link below 👇)

[BlueDot Impact @BlueDotImpact]

One approach to ensuring AI systems remain safe is training them to follow written principles.

How does this work, and will it be enough?

Our new blog post explains! Link below 👇

[BlueDot Impact @BlueDotImpact]

AI is improving rapidly, but when will it get smarter than humans across the board?

It depends who you ask.

Here’s what the experts are saying 🧵

[AI Safety Fundamentals @aisafety_course]

Join two AISF course alumni next Friday as they discuss what promising careers might look like in the (near) future.

[BlueDot Impact @BlueDotImpact]

Worried about AI taking your job? Afraid the corporate ladder is closing behind you? Join @luke_drago_ and @guynamedjoshl for a virtual discussion on how to plan your career in the age of AI.

Link ⬇️

[BlueDot Impact @BlueDotImpact]

Worried about AI taking your job? Afraid the corporate ladder is closing behind you? Join @luke_drago_ and @guynamedjoshl for a virtual discussion on how to plan your career in the age of AI.

Link ⬇️

[BlueDot Impact @BlueDotImpact]

Are we teaching AIs to lie?

AI companies have developed techniques to make language models like ChatGPT output helpful responses. But these same techniques may incentivise bad behaviour 👇

[BlueDot Impact @BlueDotImpact]

Do scientists have a plan to make powerful AI safe?

AI companies predict the arrival of AIs more intelligent than humans this decade. They admit these systems may cause catastrophe if uncontrolled.

Some research agendas aim to prevent this, but none are silver bullets🧵