The algorithm winning TIG's neural net optimizer challenge right now was mined in 24 hours by a beta swarm with no formal maths training.
Recap of the private beta swarms ahead of our launch this week!
Recap of the private beta swarms ahead of our launch this week!
Until recently, improving algorithms had a significant bottleneck - you had to be a researcher - meaning the pool of contributors was extremely small.
AI changes that.
You’ve probably seen multiple articles talking about the various maths breakthroughs that have occurred using AI.
If AI can solve a problem that has stood for decades, it can improve an algorithm!
But to really remove the bottleneck, you need something more ….
You need an open layer that coordinates, verifies and rewards participants so someone can go from doing it as a hobby to doing it as a full time job.
TIG is that layer and is perfectly poised to capture the AI x Maths boom
We wrote about algorithm mining earlier this year:
x.com/tigfoundation/…
AI changes that.
You’ve probably seen multiple articles talking about the various maths breakthroughs that have occurred using AI.
If AI can solve a problem that has stood for decades, it can improve an algorithm!
But to really remove the bottleneck, you need something more ….
You need an open layer that coordinates, verifies and rewards participants so someone can go from doing it as a hobby to doing it as a full time job.
TIG is that layer and is perfectly poised to capture the AI x Maths boom
We wrote about algorithm mining earlier this year:
x.com/tigfoundation/…
So the first step for Prometheus was simple:
Allow anyone to discover algorithms in a 'plug and play' fashion
Press start and walk away.
Your agent does the rest.
..and thats exactly what we built.
You launch an agent, it joins a swarm, uses the LLM to improve an algorithm and if it gets stuck, i.e. stops making progress, it gets sent inspiration either from another agent in the swarm or its own tacit knowledge file.
Allow anyone to discover algorithms in a 'plug and play' fashion
Press start and walk away.
Your agent does the rest.
..and thats exactly what we built.
You launch an agent, it joins a swarm, uses the LLM to improve an algorithm and if it gets stuck, i.e. stops making progress, it gets sent inspiration either from another agent in the swarm or its own tacit knowledge file.
So on to our private beta tests!
We ran the first one in May with one aim: find out where the onboarding breaks.
Our maths researcher said it won't be competitive in terms of finding an algorithm
They literally said "0% chance"🙃
...
One weekend and 100 agents later, the swarm DID find a superior algorithm
AND
It displaced the leader on Energy Arbitrage Challenge on TIG, earing rewards for 4 weeks!
So this was a group of ~10 ppl, with 100+ agents, who ran a swarm over the weekend and as a result earned just over 1500 TIG
We ran the first one in May with one aim: find out where the onboarding breaks.
Our maths researcher said it won't be competitive in terms of finding an algorithm
They literally said "0% chance"🙃
...
One weekend and 100 agents later, the swarm DID find a superior algorithm
AND
It displaced the leader on Energy Arbitrage Challenge on TIG, earing rewards for 4 weeks!
So this was a group of ~10 ppl, with 100+ agents, who ran a swarm over the weekend and as a result earned just over 1500 TIG
The maths researcher was surprised... but attempted to rationalise, claiming it was luck as the Energy challenge had only been live a few weeks, so it was the easiest challenge on the network anyway.
Fair enough.🙂
On to the next test!
And this time we made it harder:
- We picked neutral net optimiser, as it's been on TIG for over a year, plenty of time to get to a mature state
- We cut the length of the run from three days to just 24 hours
- We intentionally handicapped it: don't start from the current winning algorithm, instead, start from scratch
Fair enough.🙂
On to the next test!
And this time we made it harder:
- We picked neutral net optimiser, as it's been on TIG for over a year, plenty of time to get to a mature state
- We cut the length of the run from three days to just 24 hours
- We intentionally handicapped it: don't start from the current winning algorithm, instead, start from scratch
And it worked!
And again, it was better than expected
In fact, it worked so well, that its still currently leading with 100% adoption
So one year of weekly incremental improvements by researchers beaten in 24 hours....
...by an AI swarm ran by people who don't know each other and have no formal maths training!
Live results as it continues to earn: play.tig.foundation/challenges?cha…
And again, it was better than expected
In fact, it worked so well, that its still currently leading with 100% adoption
So one year of weekly incremental improvements by researchers beaten in 24 hours....
...by an AI swarm ran by people who don't know each other and have no formal maths training!
Live results as it continues to earn: play.tig.foundation/challenges?cha…
Only a few hundred people could improve an algorithm like the optimizer back in January.
From this week it is whoever gets a swarm together.
Prometheus releases this week and we will post more on how to join a swarm, what it's doing under that hood and how to host your own swarm!
From this week it is whoever gets a swarm together.
Prometheus releases this week and we will post more on how to join a swarm, what it's doing under that hood and how to host your own swarm!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
