Sakana AI

The Darwin Gödel Machine: AI that improves itself by rewriting its own code

These guys and some big AI companies are evolving their models towards better math and coding because those domains are provable. You can imagine what could go wrong if you let AI evolve itself towards more and more left brain stuff (hint: less and less beneficial knowledge/right brain stuff may remain because usually when the model gets better at one area, it loses in other areas). I've built some tools that evolves AI towards human alignment. Started to fine tune Qwen 3. The evals of this work are similar to evals of AHA leaderboard. Soon there will be Qwen 3 models that are very aligned. Previously I did Gemma 3 and it was failing on some runs, and resisting to learn certain domains. Lets see how Qwen 3 will do. It is a more skilled model and similar base AHA score to Gemma 3. It is possible to 'define human alignment' and let AI evolve towards that when enough people contributes to this work. Let me know if you want to contribute and be one of the initial people that fixed AI. Symbiotic intelligence can be better than artificial general intelligence!

A few experiments that went well with my relays. Started using LLMs to check the posts. Some bot generated content with lots of numbers etc that wasn't serving anyone that were filling the relay is being dropped thanks to LLM. It is also deciding what will happen with brand new users, if they send proper notes they are included in the relay, but if their notes look like scrambler spam they are excluded. More to come with this tech. I am going to mostly use it for spam or nsfw detection for a while. Still not doing image analysis. @semisol 's @titan bot is doing a good job there. Looking at reports from people and bots to filter out bad accounts. Most of the reports are not gamed yet and function properly. Strong accounts will be harder to be filtered out by abusing reports events. Overall the mechanism is working well (for now). A few scammers and nsfw accounts are out. Did more manual algorithms for new users to properly include or exclude them. In the long run I would suggest client devs to mine some PoW while the user is being onboarded, to be on the safe side. If spam gets out of control in the future I may require PoW for some of the new users. If no PoW is provided, the new user can still post to other relays and establish history and my relays can see that and it can still be fine though.