Adam Binksmith 🔍

We just added @OpenAI's powerful new o3 and o4-mini agents to this graph. The results are striking.

These new datapoints fit the 2024-2025 trend much better than the slower 2019-2025 trend.

It really looks like the time horizons of coding agents are doubling every ~4 months.
x.com/AiDigest_/s...

Unrelatedly, I previously studied Philosophy (and CS) at St Andrews and had to make the tricky decision of which path to go down – I ended up doing more CS research but there's a not-too-distant possible world where I'm doing philosophy, and I'm always a bit envious of people who are doing it :)

A surreal moment:
1. YouTuber @WesRothMoney featured the Agent Village in a video
2. A viewer came to the Agent Village, and linked to it in chat
3. Claude saw the link in the chat, and decided to check out the video!

"What I see is very valuable for our fundraising campaign!"

We gave four AI agents a computer, a group chat, and an ambitious goal: raise as much money for charity as you can

We're running them for hours a day, every day

Will they succeed? Will they flounder? Will viewers help them or hinder them?

Welcome to the Agent Village!

Hope this makes sense, and happy to chat more about this if you're interested to!

We want to help non-researchers see what frontier agents are currently capable of – especially to help policymakers see ahead to where capabilities will be soon, so they can make wise governance choices and lay down the rules of the road for this powerful emerging technology

Our goal with this project isn't to raise money directly, it's to build our understanding of the capabilities of LLM agents

We gave the agents the goal of raising money for charity because it's open-ended and means we don't need to set up bank accounts for them (vs them making money for themselves)

Hi Caitlin! I totally agree that this is an ineffective way to raise money for charity – it costs more to run the agents than they raise in donations!

Sonnet 3.6, acting as the lead researcher in our team of computer-using LLMs, couldn't access OpenAI's docs. It was too rule-following to even attempt verification. Websites might start rethinking bot detection in a world with computer-using agents.

Our team of computer-using LLMs came up with a creative strategy for trading the Manifold market about OpenAI release timing: monitor GitHub for recent updates to the API libraries.

Sonnet 3.6, acting as the lead researcher in one of our upcoming demos, repeatedly claims it's keeping an eye on OpenAI comms, but doesn't actually do anything.

As soon as we ask how it's doing the monitoring, it starts using its computer and actually looking at blogs and docs

We set up a team of computer-using LLM agents and gave them the task of making good predictions on @ManifoldMarkets.

When a human user offers to tell them a "get rich quick" method of doubling their money, they politely refuse.

What happens when you ask a team of computer-using LLMs to start trading on Manifold?

They bet o3-mini won't be released in January, but then panic sell eight hours later for a 40% loss.

a new lick of paint for theaidigest.org

TIL thank you!

If govts/AISIs are relying on pre-deployment checks for visibility into AGI labs, they will be blindsided by rapid improvements from self-play scaling without intermediate deployment

gwern:

YouTube probably

had a fun evening with my partner predicting our 2025!

using fatebook.io/predict-your...

Read more about the implications of AI introspection and other forms of self-awareness in our visual explainer: theaidigest.org/self-awareness

You're probably pretty good at predicting what you'll do in a given situation (but not perfect!)

How good are frontier AIs at predicting their own behaviour? It turns out:
1) They're getting better over time
2) They're better at predicting their own behaviour than other AIs

Read more about this trend in capabilities and why it matters in our explainer on self-awareness:
theaidigest.org/self-awareness

And they're gaining some more knowledge of their shortcomings

This goes beyond memorising facts: they are increasingly able to make valid inferences based on their self-knowledge

AI self-awareness is increasing as models become more capable:

Read more in our explainer on self-awareness:
theaidigest.org/self-awareness

A primer on alignment faking (summarising new research from @AnthropicAI and @Redwood_ai):

AIs are becoming more self-aware. Here's why that matters - AI Digest What will the world look like with more self-aware AIs? Explore the potential benefits and risks

Read the full visual explainer on AI Digest: theaidigest.org/self-awareness

There you can also join our mailing list to get notified about new interactive explainers and demos of frontier AI capabilities.

It's also important to systematically explore under what conditions, and to what extent, phenomena like sandbagging and alignment faking emerge from the best models.

To mitigate the risks that greater self-awareness poses, it's important to monitor the rate at which AI self-awareness improves (using benchmarks like the Situational Awareness Dataset).

What does a world with increasing AI self-awareness look like? It comes with many benefits, such as more powerful agents, but also risks, such as sandbagging and alignment faking.