Introduction

It's hard to believe it's been more than 2 months since my last programming diary entry. This is not because I've been inactive, but rather because time spent describing progress takes away from actual productivity - and programming is much more fun than writing. However, describing progress is important, too. So, today, I'm making this a priority.

In the activity and reflections section of today's post, I'll be talking about the Thoth project, since that has basically been my only area of focus when I had time for programming.

Before I get in to the main body, I'd suggest that the most important contribution of the last 2 months was the completion of this June 19^th post: Here is a complete framework for a Generation-5 voting service. I invite you to bookmark and read it if you haven't already. Whether the Thoth project succeeds or fails, I believe the framework it uses is a roadmap that anyone can follow to create positive change in the Steem ecosystem.

The key assumption that (I believe) makes this framework valuable is my long-standing and often-stated opinion that:

rewards + audience is more valuable than rewards alone

This framework attempts to harness that observation to the benefit of both authors and investors.

So now, on to the diary...

Background

I guess nobody will have been surprised to learn that nearly all of my programming time has been spent working with Thoth. Since it has been so long between updates, I asked perplexity.ai to review my git log, and the AI said that I have completed the following activities during the intervening time.

Stats:

• Total commits: 44 (Jun 14 – Aug 3, 2025)
• Merges: 13+ feature branches and PRs

(There are also some more changes pending that still need to be tested, pushed, and merged.)

The AI also listed these as the top-level accomplishments:

Key Achievements:

• Stronger reliability via strict config and validation.
• Secure API key handling.
• Improved fairness in delegation/beneficiary logic.
• Better AI integration and individualized user engagement.
• Continuous, well-documented iterative development with frequent testing.

So let's dig a little deeper into those topics

Activity Descriptions

Stronger reliability via strict config and validation.

One of the frustrating things about testing Thoth has been when I let a typo sneak into the configuration file that it's using to determine the amounts of its beneficiary settings. This is bad because a single run can take 2 or 3 hours (or more), and many times the error wouldn't be discovered until Thoth tried to post.

So, I added some validation right at the front so that it will find errors and fail immediately, instead of waiting and failing after a few hours. This will help me with my testing activities, but it's also important for any operators who might eventually make use of the code to run their own Thoth instance.

Secure API key handling.

Another challenge with testing is that API keys and posting keys can be stored in a config file. This is fine if it's sitting, protected, on a private computer. But, when I'm pushing changes to the github repo I'm always afraid of accidentally including my personal keys in the repo's configuration template file. I'm pretty sure that I haven't done this yet, but I suppose it was just a matter of time 'til I slipped.

So, I changed the code to pull values from environment variables. They can still be included in a config file, but that will not be a recommended configuration. This means that I (and anyone who contributes to the code) will be less likely to accidentally leak keys that should remain private.

Improved fairness in delegation/beneficiary logic.

This one has been a real challenge. The problem with randomness is that it's random. So, you can't look at one post and tell if a random pick was fair or not. You can't even look at five or ten posts and make that determination.

In the daily posting, however, I was seeing some picks that made me question the fairness, so I took a careful look at it. I didn't find any problems with the selection mechanism, but I was still suspicious so I also asked Gemini, Microsoft Copilot, Claude, and perplexity all to work through it with me a few times. The AIs sometimes hallucinated bias problems and suggested improvements, but in the end they have all concluded that the selection mechanism is fair.

Additionally, I wrote a short driver program that runs the selection algorithm tens of thousands of times and reports its results. From those results, the algorithm also seems fair.

I'm still not 100% convinced that it's right, though (my wife says my name should have been "Thomas"), so I invite others to look through the logic. The goal is that someone should be selected as a beneficiary with a frequency that is in proportion to their delegation amount. The relevant code is here for the shuffling, here for posts, and here for replies.

Fairness to delegators and fairness to authors are critical to this project, so I would appreciate as much scrutiny as possible.

Better AI integration and individualized user engagement.

This was another substantial challenge. When I started this project, the only free AI that I knew of was Mistral-Nemo-12B-Instruct-2407 from ArliAI, so that's what I was using. Somewhere along the way I learned that gemini-2.0-flash is available for free, so I switched to that.

After a while, though, I wanted to make sure that none of the changes had broken the ArliAI model, so I switched back and ran a test. Wasn't I surprised to learn that Mistral-Nemo-12B-Instruct-2407 wasn't available any more? And worse, their free tier models are now reasoning models from the Qwen3-14B family.

Thoth did not handle reasoning well at all. It was so bad that I eventually had to make separate prompt templates for the Qwen3-14B family and the gemini family.

I also learned that gemini-2.5-pro is also available for free, so that's what Thoth is using now. It should also work with Qwen3-14B, but Gemini seems far superior with the current system and user prompts. Eventually, I'd like to get away from the practice of customizing prompts for each model, but maybe it's necessary.

The rate limit for gemini-2.5-pro is 100 calls per day. With 2 posts and 5 replies per day, Thoth is consistently under 60, so for now it is comfortably under the limit for free API calls. ArliAI has no daily cap, but it does not support any parallelism (which Thoth doesn't need, anyway).

Continuous, well-documented iterative development with frequent testing.

Nothing to say about this. Thanks perplexity!😉

One more thing: Posting frequency

The AI didn't pick up on this, but there's one more change that I want to highlight.

When I first imagined the Thoth project, the idea was to discover a bunch of posts and write a single post that contained summaries of all of them. In June, it occurred to me that this would require Thoth to post 10 top-level posts per day in order to maximize stake usage for its delegators.

Obviously, however, ten posts per day in someone's feed would be far too many. So I adopted a different strategy, which I think is superior. Now, Thoth finds all of the posts and writes a summary blog post as it's top-level post. It follows that post with an individualized reply for each of the curated articles. So now, Thoth can maximize voting for delegators while limiting its feed appearances to two per day. (If we had more beneficiary slots, it could be done in just one daily feed appearance.)

The AI doesn't understand this context, but I think this was actually a very important change.

Reflections

Three topics for reflections today: appreciation for delegators, voters, and positive feedback; the arrival of downvotes; and unexpected side-effects.

Thank you to the community

Community feedback

Everyone who has taken the time to provide textual feedback to the project has expressed a positive viewpoint. Authors have confirmed that Thoth's summaries have been accurate, and readers have expressed appreciation for articles that they would never have seen without Thoth.

Thank you, especially to @massola, and @solperez for ongoing and consistent feedback.

If Thoth is going to be a change agent for improving the Steem ecosystem, this sort of feedback is necessary.

Delegators

The @thoth.test account is up to about 22,000 SP, and more than half of that is not from my accounts. So thank you very much to @massola, @steempromoution, @cub1, @chriddi, and @the-gorilla. Thoth can only prove itself with backing from the community, so your delegations are vital to proving the project.

Voters

The key for a generation-5 voting service to be able outcompete a generation-4 service rests in its abilities to attract organic votes. It will have a hard time winning unless real people are voting for the authors that it finds.

To that end, Thoth is trying to find content that real people will actually read and vote for.

In a sense, I guess we have a three-legged stool, where authors, delegators, and voters make the service possible.

So, I am very appreciative towards the people who are supporting @thoth.test and the authors it finds with your votes.

The arrival of downvotes

Unfortunately, the @thoth.test account seems to have picked up an anti-fan club, too. These voters haven't told me the reasons for their downvotes, but I suspect it's more about retaliation for participation in an anti-spam downvote trail than it is about the things that @thoth.test is posting.

At any rate, it's disappointing because the collateral damage is that authors are blocked from participating in the #lifetime-rewards experiment. The @thoth.test account is only keeping 2.5% of rewards. Everything else goes to authors, delegators, and @null. Thus, many of the people being harmed by this downvote campaign are not the people being targeted.

I'll do what I can to try to counter the downvotes with my own accounts, and I have noticed that other voters are already doing the same, so thank you to those voters. At present downvote levels, I don't see this as a threat to the project's continuance. Hopefully, that will remain true as time passes.

Unexpected side-effects

As I've been developing Thoth, I have noticed two benefits from the account that I totally did not expect when I started. First is that in testing the program I'm finding all sorts of interesting content that I would never have seen without it. For example, this was a fascinating post with an interesting video from 2016 and here's a nice rendition of "O mio Babbino caro" from 2020. Without Thoth, I don't know that those posts exist.

The second benefit is related, but it has to do with languages. If I can't tell what a post in my feed is about, my personal tendency is to scroll past foreign language posts that I can't read without translation. However, Thoth gives me English language summaries. I can't tell you how many times I've read a summary by Thoth and decided that I'd like to read it, and only then discovered that the post was written in Spanish or German. At that point, I'm far more likely to click through and go through the effort of translating it to English.

It may be that one of Thoth's most compelling use-cases is to provide single-language summaries of multi-language articles. This is something that wasn't on my radar at all when I started the project.

Conclusion

I think the past two months have ushered in good deal of useful progress, including improvements to Thoth in aspects of reliability, security, and fairness. This progress has involved improvements in handling config file errors, more secure handling of API keys and wallet passwords, updates to prompt handling to work with Gemini and Qwen3-14B families of models, scrutinizing the beneficiary selections for fairness, and adopting a new approach to optimize voting while minimizing feed appearances.

Overall, positive feedback has been dominant from the community, with the only exception being a downvote trail which has attached itself to the test account.

I hope the community will check out the @thoth.test account and see what it's doing. Whether it's done through Thoth or some other vehicle, I think that the Generation 5 voting service can be a positive change agent for the Steem ecosystem in the future.

I also renew my invitation for other developers to contribute to the effort. I honestly think this has tremendous untapped potential for the ecosystem, and it deserves to move faster than what I can do alone. The repo is here.

Thank you for your attention!

Postcript:

I just noticed, a moment ago, that ArliAI has now replaced Qwen3-14B with the Gemma3-27B family in the free tier, so... I have no idea whether Thoth will work with that API now^😉.

---

You can check out my projects here