Conference news · Published September 30, 2026
Exploit Conference Catch Up · Day 1
Missed Montreal? Here’s the team-by-team catch-up from September 28, Exploit Summit’s first day: what they discussed, what actually changed, and why it caught our attention.
The thread running through it all: make the work useful. Better models are part of that. So are paying customers, honest evaluations, private infrastructure and products people can use.
Continue to the Day 2 catch-up →Const · Gamma
ProposalCan subnets become each other’s customers?
Const’s opening pitch was about connecting the economy inside Bittensor. Gamma would let teams purchase other subnets’ services with part of their emissions. His follow-up describes burning alpha for dollar-denominated usage credits. The target he outlined was 2027; these are proposed mechanics, not a live payment network.
The Shizzy takeThis is the part to watch: a subnet buying useful work from another subnet gives us something more concrete to measure than token attention. The test will be actual service demand.
SN9 IOTA · Macrocosmos
Announced · early accessTraining across scattered hardware
Will Squires and Steffen Cruz introduced the iota SDK and Liquid Compute. The team describes software that coordinates geographically distributed machines as a training system. Its own announcement says commercial rollout follows in the coming weeks and points developers to early access.
The Shizzy takeThe hard part is making scattered GPUs useful together. A developer-friendly SDK could matter as much as a headline model size if it removes that coordination headache.
SN68 NOVA · Metanova Labs
Research updateTaking molecule competitions into the lab
Micaela Bazo discussed candidates reaching laboratory synthesis after additional processing, according to the organizer. Metanova’s current work uses open competitions for drug discovery; its own recent update expands the Blueprint search space. An ADME competition was also outlined, covering how compounds behave in the body.
The Shizzy takeA promising score needs to survive contact with a real experiment. Lab work is a meaningful next step, but it does not establish clinical success or approval of a new drug.
SN4 Targon · Manifold
Presenter-reported figuresA cloud business with reported customer activity
Rob Myers put numbers behind the confidential-compute story: more than 27,000 workloads this year and a $12 million run rate, as reported in the organizer’s recap. Targon subsequently announced Bare Metal and Sandboxes, extending its product range beyond its confidential virtual machines.
The Shizzy takeCustomers and repeat usage are the interesting part here. A run rate annualizes a current pace; it is not the same as audited revenue already earned over a year.
SN64 Chutes
DemonstrationParallax puts an 8B model on a phone
Jon Durbin demonstrated local inference. Chutes reports 59.6 tokens per second for its trained 8B model on a phone CPU. Its larger 40.75B speed demo used random weights, so that result should not be read as evidence of useful answers from a trained model.
The Shizzy takeAI that can work without a cloud connection opens up privacy and offline use cases. Getting it onto ordinary devices is exciting; model quality and production readiness still matter alongside speed.
SN114 SOMA · Dendrite
Builder lessonsDesign for the miner who finds the shortcut
Piotr Barbachowski’s session addressed incentive failures. The organizer’s highlights emphasize transparent scoring, competition and ground truths that miners cannot manipulate. Dendrite’s own profile identifies it as SN114 SOMA’s owner and a builder and miner across several other subnets.
The Shizzy takeIf the easiest way to win is a loophole, that is where effort will go. Better incentive design means paying for the outcome you actually want, even when someone tests every edge case.
SN90 KubeTEE
Technical presentationProof before trust
Pierre-André Manseau focused on confidential AI and verifiable infrastructure. The team describes hardware-enforced protection for workloads, while the conference discussion emphasizes evidence customers can check before sending private data or model weights to somebody else’s GPU.
The Shizzy takeThe customer question is simple: what can the machine owner see? Attestation is part of answering it with evidence. It should not be mistaken for a guarantee that every application running inside the environment is secure.
Taostats · Marcus Graichen
Founder perspectiveThe human cost of closing subnets
Marcus discussed closing three unsuccessful subnets, according to the organizer. The session brought the focus back to the people behind projects: the time, work and judgment involved when continuing no longer makes sense.
The Shizzy takeWe talk a lot about winners. An honest ecosystem also needs room to explain what failed and why. Closing a project can be a hard business decision rather than something a price chart explains on its own.
SN25 URnetwork
Infrastructure discussionAn open internet should work across borders
Brien Colwell argued for open, global connectivity. URnetwork’s own focus is private internet access; the conference conversation put that work in the context of access increasingly depending on where somebody lives.
The Shizzy takeAn AI tool is not useful to someone who cannot reach it. Connectivity is an important part of the decentralized-AI story, even when it gets less attention than the model running at the other end.
Manifold + TAO.com
DebateHow fast should Bittensor change?
Rob Myers and Yoav Cohen debated protocol speed. Myers pushed for faster building in the organizer’s published highlight; the broader session weighed that urgency against the stable foundations product teams need.
The Shizzy takeA protocol can improve while still creating expensive work for its builders. The useful question is how to ship necessary changes with enough predictability for teams to keep serving customers.
TAO Templar + Learn Bittensor
EducationFollow the scoring rules to understand the subnet
TAO Templar and Maciej Kula walked through incentives, from evaluations to miner rewards and the shortcuts competitors discover. Their session supplied the foundation behind many of the summit’s more technical discussions.
The Shizzy takeBefore asking which subnet has the best story, ask what its miners are paid to do. Understanding the score is often the fastest route to understanding the business—and its weaknesses.
Khala Research · Jack Ai Leung
Research perspectiveOpen-source thinking meets robotics
Jack argued for open models and broader access to robotic hardware. Khala’s published research also follows the gap between demand for robotics and the systems being built to meet it. This session was an argument about access and development, not proof of a new robot launch.
The Shizzy takeRobotics needs more than a model download: data, hardware and places to test matter too. Opening those doors could give more builders a chance to contribute useful work.
Latent Holdings · Cameron Fairchild
Security educationReduce dependence on one private key
Cameron explained multisig approvals. Latent’s product ecosystem already discusses multisigs and other wallet controls; the conference session focused on spreading authorization across signers instead of putting every consequential action behind a single key.
The Shizzy takeOperational security gets more important as a team grows. Multiple approvals can help, but teams still need a workable signing process and a plan for unavailable signers.
SN15 ORO
Product and evaluation discussionFresh shopping tasks for AI agents
ORO showed how its reported 11-million-product catalog supports agent tasks. The team had already announced ORO Bench on September 22: a generator of new synthetic shopping environments powering SN15. The conference explained the work behind the competition rather than marking that benchmark’s first release.
The Shizzy takeA shopping agent needs to succeed when the task changes. Fresh evaluations are more useful than teaching it to memorize one fixed test.
SN38 ChronoLLM
Live product · planned upgradeA model that cannot peek into the future
Jean Hérelle discussed keeping future information out of historical prediction tasks. SN38 ChronoLLM separately announced its time-aware model was live. The move from a 3B to a 16B baseline was described as upcoming, so the two milestones should stay separate.
The Shizzy takeA model that already knows the result can look brilliant in a backtest. Time boundaries make forecasting claims more meaningful; a larger model alone does not prove better predictions.
Seby Rubino · SN46 Instant
Business perspectiveStart with the customer’s result
Seby challenged teams to start with the customer outcome instead of a generic speed-and-price pitch. He is associated with Zipcode and SN46 Instant; his current feed identifies Instant’s inference work and its September 29 mining announcement.
The Shizzy takeA customer does not buy a subnet number. They buy a job getting done. That is a useful filter for every demo in this catch-up: who needs it, and what becomes easier for them?
SN93 Bitcast
Applications announcedInfrastructure other creator teams can build on
Will and Tom presented Subnet Summer and TAPE as applications using SN93 Bitcast. The team subsequently confirmed Subnet Summer as the second external-team application on its network. That builds on its creator-marketing infrastructure rather than requiring every application to become a separate subnet.
The Shizzy takeThis is a useful kind of adoption: other teams choosing to build on what already exists. Watch whether those apps attract sustained creator and customer activity.
SN111 Claims
Production milestoneTurn papers into claims with traceable evidence
Philipp Koellinger and the Claims team presented a system for extracting structured claims and evidence from scientific literature. The organizer dates V1 to September 27. The team had announced mainnet in August and discussed production quality before the event, so V1 should not be confused with its first-ever mainnet activity.
The Shizzy takeA useful scientific AI needs to show what supports its answer. Building a structured evidence graph tackles that problem more directly than simply feeding a model another pile of papers.
SN124 Swarm
PreviewSentinel and autonomous solar-site patrols
Swarm introduced Sentinel in its physical-AI presentation. The organizer describes solar-site security patrols, and the team calls the session a preview. Its earlier announcement names a German Solar Power surveillance project in Greece; that background does not establish Sentinel’s general availability.
The Shizzy takeA defined job at a real type of site is easier to assess than a vague robotics promise. Field reliability and completed deployments are the milestones to follow next.
SN71 Leadpoet
Product discussionSales agents improve through open competition
Gavin Zaentz explained a competition where winning sales-agent work improves a shared baseline. Leadpoet’s own material describes its Lab Arena, and its September 24 announcement identifies a pilot supporting enterprise outreach for SN80 OpenRoboto Shift.
The Shizzy takeThe business test is whether a better agent helps find relevant customers. An improving benchmark is useful, but successful outreach is where the product has to deliver.
SN63 Enigma
Mechanism and research updateLet rewards build around an unsolved problem
Bob Wold explained Enigma’s challenge-based rewards. The team also reported at Exploit that two miners independently solved its Hardening Quantum Proof Milestone 3 after months without a solution. Winning work is intended to become available for others to build on.
The Shizzy takeA clear problem and a checkable solution make progress easier to judge. The quantum challenge result is research progress, not proof that a practical quantum-advantage product has arrived.
SN46 Instant / Zipcode + SN54 Yanez
Token and equity debateWhat belongs in the company—and in the subnet?
Seby Rubino and Jose Caldera discussed the division of work between companies, miners and validators. The recap frames it as a token-and-equity discussion. It did not announce that holding a subnet token grants shareholders’ rights in an operating company.
The Shizzy takeInvestors and builders should be clear about what each side owns and does. Decentralizing a useful task and running a customer-facing business are connected decisions, but they are not interchangeable.
Philippe Beaudoin
Research perspectiveWhat keeps humans and AI cooperating?
Philippe explored cooperation using records of human and AI interactions. His own research profile focuses on long-running human–AI relationships and stable multi-agent systems. This was a research discussion, not a subnet product announcement.
The Shizzy takeMore capable agents still need ways to work together. For a network built around competing participants, understanding when cooperation lasts is a useful question alongside raw performance.