Note: Full article was written by Apex Subnet
The results are in: Apex’s deep researcher shows miners are getting more efficient by the day.
For a recap, miners on SN1 have two duties, with validators playing a core role.
Validators are given a prompt, which they must answer in around 3 minutes.
The first half of miner duties to generate, where they’re given the same prompt, but are only given 20 seconds to answer.
The other half of a miner’s responsibility is discrimination – they must be able to discern which response comes from a miner acting as a generator, and one from the validator.
Validators should perform better, as their models are given more reasoning time. The 3-minute window allows them to use SOTA models to produce their results, as well as multiple agentic tool calls, such as code calculators, web searches, and calls to other models. Generally, they’re distinguishable, but our metrics show the competition’s heating up, with generators being able to produce results that are getting closer to being identical. Data reveals that generators are increasingly more capable, and our miner community is significantly optimizing their output. With only 20 seconds to form a result, miners aren’t able to use standard plug-and-play models – they need to speed up their agents as they have a shorter span to act.
SN1’s historic data shows discriminators outperforming generators by a wide margin. But this isn’t the whole story.
To get a clearer picture, let’s compare generator and validator responses (marked on the graph as deep researcher references). This is a similarity graph: the higher the result is to 1, the closer the generators are to the validator answers. Over time, they’re trending upwards, meaning generators are getting more and more similar, recently almost reaching parity.
What does this mean for Apex 3.0?
SN1 miner models are getting smarter. Despite having under 20 seconds to complete their answers, they’re producing work that’s increasingly similar to validator models, which take around 10X longer to produce.
That’s an order-of-magnitude speedup against SOTA techniques.
The result: Apex is on track to becoming a leading deep researcher, capable of surpassing closed-source competitors in speed.
We set our miners a gargantuan task, and they’re rising to the challenge. This is how decentralized intelligence will set the AI standard.
Want to learn more about Apex’s architecture?
See Apex’s Substack deep dive and listen to their Inventive Mechanisms podcast
.
Be the first to comment