Vishal Menon @vmcrypta

Conviction-Hybrid Oracles: why settlement infra (not another frontend) is the real bet as prediction markets go institutional. Just published a note tying @wmt_ventures recent two-sided LP move to the oracle bottleneck in subjective/long tail markets using my public GitHub proposal as one possible fix. Would love to hear feedback/thoughts if any :) itsvishalmenon.medium.com/conviction-hyb…

The panels made it crystal clear that the interesting work is expanding inputs while protecting that last mile of accountability. What remains open is whether the validation predicate can be made compositional across onchain and offchain sources while keeping execution layers robust enough for agentic loops. Curious for thoughts! And many thanks once again to the panelists and the team at 
@Mantle_Official @smu_blockchain for making this such a great experience!

Personally, some of the key takeaways that hit harder after the sessions were the following: 1) Onchain data is just the subset where V(D_on) = true. Everything else stays rich context that still needs human judgment. 2) Intent matching is the irreducible probabilistic step. No agent parses real user ambiguity without it. 3) User sovereignty stays a true non negotiable. Even the strongest agentic trading loop, API fed and CLI exposed, must route final confirmation back to the person who owns the outcome. 4) The gold function U_gold(s,a) = E[r] − λ·Var(R) only works when Codex style flows and plugin extensibility keep the human in the seat, not replace them entirely.

Attended Mantle Builders’ Night last evening at SMU and found the panels and conversations I’ve had on agents and onchain systems to be really engaging and intriguing to say the least. Came away convinced the real constraint was never the data itself. Probabilistic or deterministic, onchain or offchain the information remains the same. 
Only the validation predicate shifts and that single distinction is now how I think about agent design.

keyboard_arrow_left Previous keyboard_arrow_right Next

It’s a decent early proof point but rather inconsequential as far as I’m concerned. The change in architecture is where I’m interested in most. Quantum adapters give us a completely different representational substrate that slots in without retraining the whole model. I’d say that this is the start of hybrid systems that scale in ways pure classical can’t. What do you think the next milestone needs to be before this moves from interesting demo to production grade ready?

Recently just read the new Multiverse Computing paper (arXiv:2605.05914, published May 7) and what I’ve learned is that hybrid quantum classical LLMs have moved from simulation to measurable execution on real superconducting hardware. Researchers froze the 8 billion parameter Llama 3.1 8B backbone and inserted small Cayley parameterised unitary adapters into its projection layers. End to end inference on IBM’s 156 qubit Heron processor produced a 1.4% perplexity improvement on WikiText, shifting from 8.877 to 8.752, while adding roughly 6000 parameters. The authors of the paper have described this as the “first demonstration of end to end quantum enhancement of a production scale LLM on real hardware for autoregressive generation”. That would track with the existing evidence and here’s why:

By freezing the classical backbone and grafting on small quantum adapters executed on real hardware, the work shows that quantum circuits can slot directly into production inference pipelines without requiring full model retraining or replacement. This approach matters because classical LLMs already face hard physical constraints on memory and energy as parameter counts grow. Toolchains, compilers and deployment frameworks will likely need to evolve to support mixed classical quantum inference. That evolution tends to accelerate investment, standardisation and iterative engineering across both quantum hardware providers and AI labs. My 2 cents, I feel this would be the first time quantum circuits have been treated as a drop in modular upgrade rather than a full replacement architecture and that modular mindset is probably where real progress will come from. Would be curious to hear more perspectives on this from you guys. Full paper: arxiv.org/abs/2605.05914

4/ The architectural implication of all this is the most durable takeaway. Classical LLMs face hard physical limits because every additional parameter requires classical memory. Quantum systems supply a qualitatively different representational substrate. The adapter pattern demonstrates one workable path as it retains the scale and training investment of the classical model, and attaches quantum components only where they add expressivity or efficiency, while running those components on the QPU. This will not displace GPU clusters in the near term but it does establish the fact that the two paradigms can interoperate on today’s hardware. Incremental engineering results of this kind tend to compound fast.

@Kunallegendd Yep fair enough, makes sense

@evelynjunn Don’t hate the player (s), hate the game.

Such an excellent read this…the pivot from chat first to ambient analysis is honestly the real architectural winner here. In 2B scale models, persona management and instruction following consume way too much of the parameter budget. So when u constrain the task to feature extraction and trend analysis, you’re effectively reclaiming those weights for higher precision domain logic. It’s the actual shift from instruction tuning to functional mapping that makes the entire edge intelligence viable. This is why I’d always say that niche and domain locked specialists will beat shallow generalists every time!

The problem is that you can't "patch" a probability distribution. When an agent drifts into an unsafe state space, it’s not a code bug but a valid inference path. We need to move toward a hybrid architecture where the agentic reasoning layer is actually wrapped in a deterministic and formally verified monitor. In most high stakes environments, emergent intelligence is a liability unless it's constrained by a provable logic trace. Would be glad to hear some of your thoughts on this.

Diving into formal verification while mapping out autonomous decision paths. There’s a widening verifiability gap in how we deploy agentic AI. We’re currently making a categorical error in giving probabilistic reasoning engines deterministic execution authority over critical infrastructure. If you can’t mathematically bound the agent’s decision space, you aren't actually securing it. As a matter of fact, you're just gambling on the alignment of the next token.

Spot on. The convenience argument is actually a risk adjusted opportunity cost argument. SG’s real edge isn't as a market but as a laboratory. The outliers here treat SG as a high trust sandbox for R&D but build for global and homogeneous markets to bypass the "SEA localization tax" entirely. In the US or China capital will buy pure growth. In SEA that same capital is diverted into localization friction. The strategic escape isn't a better regional strategy rather, it’s a global first mandate that prioritizes market uniformity over geographical proximity.

@0xRak @oviohq Just replied via DMs!

@Jason I’m in