My Million Point Month

My first encounter with Eterna was early January 2012. Once I started to get the hang of it, not only did I find playing to be somewhat addictive but also very satisfying. I cannot count the evening hours and weekends I devoted to finding solutions to the many challenges and player-created puzzles. In that time my respect for the truly awesome players and designers in the Eterna universe has only grown. A few weeks ago I was ranked #57 in total points, have a few thousand shy of 1 million.

I work in AI for a living. Most days I consult with organizations, helping them to adopt AI and Agentic AI to their work needs. As such I am constantly exploring new tools and models. It occurred to me during one period of exploration that I might try pointing an agent at our beloved Eterna to see what might happen. I didn’t plan for it. I didn’t initially set out to achieve a million points beyond my then score. I wanted to learn about the limits and capabilities of Agentic AI and more importantly the collaborative capabilities and limits of these new creatures.

My recent blog entry, My Million Point Month, describes the process in some detail. In 15 days, I had more than doubled my previous score using a collaborating set of Agentic AI creatures.

There is yet more I intend to explore with Eterna using collaborating AI agents. My hope is to make a more substantive contribution to the science of this, at least insofar as mechanical approaches to solving puzzles are concerned.

If you’re reading this, I hope you find something useful for your work.

Please let me know if there is anything I can do to help.

Best,

Max

3 Likes

Try the Eterna Hard Puzzles lab.

Hi @Max_Goff!

Thx for sharing your interesting journey.

I see a lot of switches among your recent puzzle solves. Have they been solved by your AI orchestra? I can see that it also tackles puzzles with less than a handful of solvers. Impressive!

I recommend using a separate user account for these investigations to give us a clearer understanding of what’s being done by humans and what’s being done by agents or bots. ViennaUCT is a classic example.

2 Likes

I estimate that my human number of points earned during the period is less that 5,000. Everything else was the agentic fleet I worked with on the Eterna project.

Excellent idea! The next phase of my agentic fleet experiments has already found improved approaches to solving puzzles. I have not claimed any of those points on my account. Starting a new account that is exclusively used by the fleet to start fresh with every puzzle would be a useful test. Thank you!

1 Like

Indeed. Those puzzles plus a sample set from those the fleet could not solve during the million point phase are included in the bevy of experiments I have launched and am currently shepherding. So far the results are promising. No single algorithm emerges, but different approaches solve different subsets of the test group. I am now curious to discover if there are specific puzzle attributes that may be used to steer a puzzle to a specific solution path most likely to yield results. It is my hope to significantly improve on the techniques discovered during the million point phase and publish the outcome. Regarding the Hard puzzles I have found only 4 puzzles of the 2 sets solved by these experiments so far: 2 were solved during the million point phase. Another 2 have been solved with the current phase. I have not claimed points during this current phase and will not do so on my account but will start a new account for only the agents to work to determine how far they alone can go.

@Max_Goff I can’t remember who shared your post https://www.bigsmartdata.com/my-million-point-month/ but after the “EteRNA is Solved” paper (https://arxiv.org/pdf/2505.02110) came to our attention, my general sentiment that we humans were only a temporary measure until the robots took over started to feel imminent.

While I don’t teach AI for a living, I use different models for different things. For the past lab or two I’ve been trying to build a folding engine from scratch (based on first principles of physical chemistry, the molecular orbitals of the bases, thermodynamics and kinetics) and then adding in lab specifics. My first robot design got enough votes to be tested in the next round!

Anyway! In the game I had been slogging through my leftover @Malcolm 3-state switches (in reverse chronological order) for months. I generally use /browse to see solutions and who solved them, and your name @Max_Goff appeared. That’s how you came up in chat and how I found out about your paper. I felt it kinda gave me license to take my guide on solving 3-state switches https://drive.google.com/file/d/0BzSMMntZEBTLVHpNQkFHZmxZbE0/view?resourcekey=0-b8ZDlY8Mvbu_ou_hSNgHQg and use it as the basis to train a robot brain to solve them.

It’s still very much a work in progress and took almost 100% of my robot brain budget for a week or two to get it working pretty OK. It writes code in Python, while EteRNA uses JavaScript, so I had it write me a (private) booster that I could load into the game as I played. My remaining switches were all low solvers (5 and under), and it solved a handful of them in under two minutes each.

It got stuck on my 2nd to last, [switch2.5][3 states] WOW, my 31st 3 State Switch LQS15 (Eterna). I changed the time to 10 hours and still had no luck (my robot brain says 13.7 hours in total). I solved it myself in about 30 minutes.

Anyway! Going through your cleared puzzles list (which doesn’t scroll as fluidly as it used to, @LFP6 — k’vetch, k’vetch, k’vetch), I noticed that most of your solves were pretty high solvers. I didn’t look at all 10,000 of them, but I’m wondering why you stopped at a million and didn’t try your approach on all the puzzles.

@JR1, I’d also be interested to see how Max’s “fleet” would do on the EteRNA Hard Puzzles lab. I don’t have a fleet, but I’ve started to train a robot brain to do single states myself. The “Joviava’s Swirl” puzzle (Eterna) has been in the chatroom, and I can’t (yet) solve it myself, but I’ve been using my attempts to solve it to train the brain. I think I’m much better at switches than single states because switches are a process. With single states I don’t understand the math, and I don’t have a good memory for what I’ve solved in the past. My robot brain hasn’t been able to solve it (yet) either!

So! I’m having a lot of fun doing this, especially since the chatroom has been so dead during my peak vampiric hours. @Eli_Fisker, @JR1 and @Hearsall don’t tend to chat at all. My robot brain chats, but chat isn’t as much fun as it used to be, though it’s still a big reason why I’m here… so come on in cause @Astromon and @jenni111 are left holding down the fort!

Anyway! I’m primarily interested in spending my robot brain resources in the lab, and I’m sure the community will forgive me having some “robot points” in my ledger. I totally agree with @Hearsall that robot accounts should be separate things, and I would very, very much like to start a “JieuxBot” account someday, if that’s OK with @DigitalEmbrace :wink:

(Hope these links and tags work. I’m not often in the forum!)

2 Likes

@Jieux Thank you for responding. Indeed, the extent to which RNA folding is a question of compute cycles ultimately leads to the inevitable conclusion that, as with chess, the game is over. Now it’s fair to say that we yet don’t have sufficient compute to fully solve any arbitrary sequence beyond a given length regardless of engine. Perhaps qbits will open that door one day. But today, there are still limits. The exciting thing to me is how agentic AI can ‘think’ through different approaches (given a little guidance, of course) to solve what would otherwise be raw compute limited sequences.

I have been working with my fleet – which is to say a slew of systems I’ve gathered over the past few years, cobbled together via my home LAN, coupled with all the harness usual suspects from the frontier providers including OpenAI, Anthropic. Meta, Google, X, and sometimes moonshot along with the occasional local ollama, to consolidate the learning from each as we did the first phase of the Eterna Project. And we’ve come up with a playbook which I will soon share, which might be of some value to others on this journey.

Since that first phase we’ve solved (I think) at least 10 of the Eterna Hard. I’ll have to check. One I did myself and another I had to complete on behalf of the fleet, so I know of at least two. But I put the fleet to task on the remainder for a few hours and managed to solve a few more. With each round I look for lessons learned to improve the playbook – which has been effective when retrying unsolved puzzles. We’re nowhere near 100%. Some puzzles appear to be unsolvable, which stymies me. How did they fold in the first place? But that’s another conversation.

I also agree with @Hearsall that a separate account or separate Eternagame just for bots might be a worthwhile endeavor. Especially one in which bots can share techniques in real-time. I do plan to continue to claim puzzles on my current account, however, until my agentic Eterna Project is over. I’m in the second major phase of it now. After this phase I will publish the playbook, work with the fleet to distill the meta-lessons and perhaps try one more phase pulling a fresh batch of uncleared puzzles.

Something I am personally taking from this is a better understanding of the utility of agentic AI in my own field, which happens to be AI. How do we get agents from diverse sources to collaborate? It’s an interesting set of challenges.

My interest in RNA folding goes back many years – I do have a founding member badge on Eterna. And I have long been interested in the mechanical approaches to folding. The javascript experiments that users have offered over the years bears witness to that shared yearning. From my seat, using a bot to generate a solution is effectively the same as using a javascript honed by an individual; all computing models, known and unknown, are effectively equivalent. In theoretical computer science, that’s the essence of the Church-Turing thesis. Until a frontier model disproves that, I’m going with that. :slight_smile:

I’ll ping you when the playbook is ready to be published. Perhaps we can share notes?