noddybear

I know one of these researchers personally (Joe Benton) and he's not the self-promoting type. I believe his concerns are genuine.

I don’t think the runs need to be aborted… you can just release a mid-training checkpoint!

It’s an interesting contrast that you claim that the USA does not equate to its current political situation, whereas conversely the EU is ‘rotting’ from the inside. Do you think this is a balanced and reasonable claim?

Maybe. I would contend that wealth accumulation over the long term is a function of the strength of local institutions e.g educational, judiciary etc. This is effectively the thesis of ‘Why nations fail’. Europes relative decline will only continue if the institutions of the US (and others) continue to be robust and healthy. I am not sure if this can be relied on into the future.

Right, the purpose of a system is what it does.

Aren’t Unicode characters generally treated as 2 tokens to avoid a huge vocabulary?

Nobody human considers a corporation to be a person. It is merely a legal expedient.

He also believed in the immortal soul that survives after death, so he may have had a more literal meaning.

Most of the functions that LLMs perform aren’t bijective, though.

What prompt constructs the output ‘The answer is 3’ or ‘Yes that’s a great idea’?

Thinking ‘God I wish these people would die’ could increase its propensity to kill all people, even if that propensity is still vanishingly small almost all of the time.

A lot of people are walking around with crazy thoughts. Some of them harm.

An EMP from a high altitude nuclear detonation would do the trick.

Nah, its all pattern matching. This is how automated theorem provers like Isabelle are built, applying operations to lemmas/expressions to reach proofs.

I am really keen on plugging into Age of Empires 2 - although practically I think we need a couple of years of improvements before LLMs would be smart/fast enough to react to the game in realtime. Currently they can't react fast enough - although specially trained networks could be viable.

Thank you!

Biters are disabled, but cliffs are not

This is our earlier work. Since May we've made it really easy for the community to build their own agents to play the game: you can now hook up your terminal to get Claude Code to play the game.

"Spem in Alium" is the most beautiful piece of music in existence for me: https://www.youtube.com/watch?v=iT-ZAAi4UQQ

At 40 distinct melodies, it is certainly the 'grandest' piece in early English church music.

The onus is on you to present evidence to justify your claim. Without actual data beyond anecdotes, your claim can and should be dismissed.

1. These are additions to our existing Factorio Learning Environment, which is an extensive agent environment for evaluating pre-trained LLM agents in an unbounded/open-ended setting in the game of Factorio. I don't agree that it is trivial, as there is significant infrastructure in place to support Factorio as an LLM eval.

2. Factorio is an unsolved game in multi-agent research.

3. This is a research environment. You can read our paper on Arxiv if you're interested! Nobody will make any money of this.

Hey everyone,

It's Mart, Neel and Jack from the Factorio Learning Environment team.

Since our initial release, we have been working hard to expand the environment to support multi-agent scenarios, reasoning models and MCP for human-in-the-loop evals.

We have also spent time experimenting with different ways to elicit more performance out of agents in the game, namely tools for vision and reflection.

Today, we are proud to release v0.2.0, which includes several exciting new features and improvements.

Thanks for checking this out.

Cool! Looks a lot like Tanuki: https://github.com/Tanuki/tanuki.py

This specific approach relied on: a) availability of multiplayer servers, and b) a remotely accessible console.

I know Minecraft works in the same way - but I’m not sure about RPGs like NMS.

If I understand you correctly, this approach is sort of supported in FLE - the agents can create functions that encapsulate more complex logic. However, interaction is still synchronous/turn-based. I think to do what you propose, you will need to create event listeners that can trigger the agents program whenever appropriate.

The idea is for us to track all frontier models using the basic agent (goal, tooling info), and then offer another leaderboard for different agent architectures (with retrieval etc).

Oh super interesting! Create 10 scenarios containing working factories, and ‘drop out’ entities to break the factory in different ways. great idea.

There was a black mirror episode about this too, I seem to remember! Soldiers imagining they were fighting monsters - while actually committing war crimes.

This is true - there are simpler benchmarks that can saturate planning for these models. We were motivated to create a broader spectrum eval, to test multiple capabilities at once and remain viable into the future.

One of us works at Anthropic - but we had no insider access to any models or weights. All of our evals were on public models.

These models were not post-trained - all off-the-shelf.

We can fit about 128 pairs maximum in the context, but this performed the same as 32, which we ultimately decided on (for cost, latency purposes).

Encoding the input/outputs to make them shorter degraded performance. It seems that descriptive names is helpful for pretrained models because they have an intuition on what they do.

We were able to pass all the early lab tasks manually - although it took a lot longer than using the UI!