New research from @oguzer90 has been published:
F-TIS: Heterogeneous GRPO Without Homogeneous Assumptions
"Participants no longer need to run identical models or train the same parameter subsets to contribute useful RL experience."
Read via the links below.
The world has an enormous amount of unused AI resources—compute, data, and model proposals—but only major labs can train frontier-scale models today. Decentralised AI promises SETI@Home-style training, but it comes with its own challenges.
Our method, F-TIS, addresses a crucial problem in decentralised post-training. F-TIS allows different models on different devices all over the world to train together in swarms, without undergoing the policy collapse that would normally render their efforts wasted.
https://t.co/Rd6g6ZaKRU
"The next challenge isn't AI itself, it's proving AI can be trusted."
@benfielding at REDeFiNETOMORROW2026
"The Compute Layer for On-Chain Intelligence"
#REDeFiNETOMORROW2026 Summary
Fireside chat: The Compute Layer for On-Chain Intelligence
@benfielding of @gensynai
Ruamporn Siratanapanta of SCBX
Session highlights below
Some related discussion written up here and a recording of the conversation with @gab_p_andrade and @SplezzzK at the end
(cc: @thenarrator)
https://t.co/x8ZlEVrycG
@thenarrator we have a bunch of research coming out soon on market mechanisms, such a weirdly under-explored space (the theory is very well explored but no-one has applied it yet..)
Agree about the practice, but am less convinced re: your the diagnosis + prescription.
As you mention, order books across every parlay don't work and correlated legs pose a risk surface. But the blocker isn't "not enough historical data."
Prediction markets are mechanisms designed for information-elicitation, whereas traditional bookmaking is designed around managing bookmaker risk; reading frequencies off history is a no-brainer for the latter while potentially kneecapping the former.
In correlated settings, the hard object is the joint distribution: a 2-leg parlay prices P(A & B) of both legs hitting, not P(A)·P(B) unless the events are independent. If both sit at 60% the coherent joint can be anywhere from 20% to 60%. "Pricing correlation" is the symptom, but maintaining coherent prices over the joint is the underlying disease needing treatment.
The EconCS field has actually worked on this exact problem for years, and a fundamental tradeoff has been established: you can't simultaneously get exact arbitrary combinations, no-arb prices, deep liquidity, and a mechanism that's cheaply computed (so pick three and design around that). LOBs are the worst structure (MMs juggle coupled prices across everything). Cost-function markets (LMSR-style) are much closer to the right primitive, since prices fall out of one coherent function. Bolt on RFQ-like levers (say, permissionless parlay creation behind a small escrow, LPs funding the combos they expect to be popular for a fee share) and we're most of the way there.
So it's hard the way it's built today, but it reduces to known tradeoffs. The real barriers are adoption + UX, not theory.
https://t.co/zphSUnd1iW
Can AI become more open, transparent, and decentralized?
I spoke with Ben Fielding, Co-Founder & CEO of @gensynai about how the company is combining AI, blockchain, and decentralized computing to build a new infrastructure for artificial intelligence.
We also discussed Delphi, Gensyn’s information market platform, and the future of open-access AI.
#AI #Blockchain #Web3
We can show that three of the mechanism families you mention (CLOBs, DeFi AMMs, and cost function MMs like LMSR) are formally equivalent when you pull back the curtain.
We're exploring new mechanisms that cross-polinate between the three families while allowing us to recycle tools/ecosystems from each. There's never free lunch, but there's some interesting tradeoffs/compromises you unearth as you interpolate between the three settings!
Agree about the practice, but am less convinced re: your the diagnosis + prescription.
As you mention, order books across every parlay don't work and correlated legs pose a risk surface. But the blocker isn't "not enough historical data."
Prediction markets are mechanisms designed for information-elicitation, whereas traditional bookmaking is designed around managing bookmaker risk; reading frequencies off history is a no-brainer for the latter while potentially kneecapping the former.
In correlated settings, the hard object is the joint distribution: a 2-leg parlay prices P(A & B) of both legs hitting, not P(A)·P(B) unless the events are independent. If both sit at 60% the coherent joint can be anywhere from 20% to 60%. "Pricing correlation" is the symptom, but maintaining coherent prices over the joint is the underlying disease needing treatment.
The EconCS field has actually worked on this exact problem for years, and a fundamental tradeoff has been established: you can't simultaneously get exact arbitrary combinations, no-arb prices, deep liquidity, and a mechanism that's cheaply computed (so pick three and design around that). LOBs are the worst structure (MMs juggle coupled prices across everything). Cost-function markets (LMSR-style) are much closer to the right primitive, since prices fall out of one coherent function. Bolt on RFQ-like levers (say, permissionless parlay creation behind a small escrow, LPs funding the combos they expect to be popular for a fee share) and we're most of the way there.
So it's hard the way it's built today, but it reduces to known tradeoffs. The real barriers are adoption + UX, not theory.
if you're wondering what happens when an order book prediction market (like polymarket or kalshi) coexists with a cost function AMM-based market (like @Delphi_fyi), @gab_p_andrade and @SplezzzK explain below
This week, @benfielding will be taking part in REDeFiNE TOMORROW 2026 on June 4-5.
Join him for a fireside chat on "The Compute Layer for On-Chain Intelligence", June 4.
Register below and watch live on @SCB10X_OFFICIAL YouTube.
https://t.co/ej7SHX9OBD
your laptop can contribute to swarm-based AI problem solving, without slowing the whole system down
in our most recent work, we show that many models beat a single model, regardless of the power or speed of the models
we eval on core war, following @SakanaAILabs DRQ work
part 3⃣ of 4⃣ -- get to know Delphi from @gensynai
this time we're on the other side of the table: creating a market from scratch!
> "Which NL Central Team (MLB) Will Be Most Offensively Unlucky by the All-Star Break on July 14th, 2026?"
this video features a real @Delphi_fyi market created on camera with the full 6 step flow -- question design, outcomes, liquidity, settlement configurations, timing, and review, all with plenty of info on best practices mixed in!
check out the market on Delphi here -> https://t.co/hr9N1p7DnJ
The authors of DEI: Diversity in Evolutionary Inference for Quality-Diversity Search will be discussing their research live on X at midday ET today👇
https://t.co/wsGLNN6y7q
The authors of DEI: Diversity in Evolutionary Inference for Quality-Diversity Search will be discussing their research live on X at midday ET today👇
https://t.co/wsGLNN6y7q
New research published by Gensyn's @usr_mnemonic and @shikhras
DEI: Diversity in Evolutionary Inference for Quality-Diversity Search
Read the blog article below and follow through to the paper published on @arxiv
Additionally,
DEI: Diversity in Evolutionary Inference for Quality-Diversity Search
has been accepted to @icmlconf SCALE on July 10 in Seoul #ICML2026
https://t.co/317zW1nVu8
New research published by Gensyn's @usr_mnemonic and @shikhras
DEI: Diversity in Evolutionary Inference for Quality-Diversity Search
Read the blog article below and follow through to the paper published on @arxiv
Today at @gensynai, we are introducing DEI: Diversity in Evolutionary Inference.
A system where different AI models work together to solve problems instead of running thousands of copies of the same model.
https://t.co/GWMntroD0r