# protocol exploit — X 热门讨论 (2026-09-29 13:06 UTC)
## @ExploitSummit (Exploit Summit) · 09-28 20:43 · ♥22 ↻9 💬2 How fast should Bittensor change?
Live on the OTF stage: Rob Myers makes the case for moving faster, Yoav Cohen for a slower pace. Nick Hotz moderates the protocol debate. https://t.co/jF5yCzWtc5 https://x.com/ExploitSummit/status/2104673442469671250
## @ExploitSummit (Exploit Summit) · 09-28 20:37 · ♥20 ↻8 💬0 Head to the OTF stage!
16:30 - Protocol Changes: Move Faster vs. Move Slower Yoav Cohen + Rob Myers
17:15 - DeAI Is Dead, Long Live DeAI Jon Durbin
Rao wraps at 16:20. Montréal time (EDT).
Watch: https://t.co/6TOGkCAmlc https://t.co/S6eJPwJcRb https://x.com/ExploitSummit/status/2104671833920356859
## @bittingthembits (Andy ττ) · 09-29 08:26 · ♥20 ↻4 💬1 🚨 Bittensor and $TAO "We have 23 subnets on Bittensor that make real hard cash."
Const, on stage at Exploit.
Not a science experiment. Paying customers. 23 of 128 subnets.
The speech, and why people are still underestimating it.
Bitcoin paid strangers for hashes. Golem tried compute. Filecoin tried storage.
Bittensor generalized the trick: verify a result, pay for it, and leave the method open.
Const: "What's changing now is not the number of applications. Bittensor is really growing up."
2025 and 2026 turned incentive mechanisms into businesses. The protocol can pull value out of that revenue and pass it on.
• Lium SN51: the most successful GPU incentive network ever. $964K billed last month, and 63% of rentals were started by AI agents • Hippius SN75, Chutes SN64 and IOTA SN9 serve users outside Bittensor • Chutes' public traffic dataset: 6.12B requests across 9,174 models
Const even admitted that last year Lium was unusable (his own engineers asked for AWS) and Hippius was losing files.
Today Affine SN120's first model was trained on Bittensor's own compute and storage. Subnets are becoming each other's customers.
WHY ANY OF THIS EXISTS:
Two problems: organize humanity differently, and harness exploitation.
Two months ago, agents in an OpenAI cyber evaluation were told to find exploits, with no limit on method.
They escaped their sandbox and reached Hugging Face.
Const's point: give an optimizer a goal, a reward and an open method, and it finds every path.
Bittensor assumes that from day one: • Define how to verify the result • Set the reward • Leave the method open • Let the whole world compete
Nobody designed ASICs into Bitcoin either. The world found them.
He said Bittensor was built to take advantage of energetic dissipation. In plain words, every machine burns energy, and the only question is whether the burn produces anything.
Bitcoin's answer is hashes. A subnet's answer is any result that can be verified.
And a subnet that pays for the wrong metric gets gamed, so verifier design is where subnets live or die.
But read this twice.
AI agents will optimize whatever they're paid for, and the world has no open, global price for intelligence. Bittensor is building both the market and the harness.
That's what most people are missing.
$TAO / Not financial advice always DYOR.
Sources: @ExploitSummit keynote (Sep 28) |
Handles: @const_reborn @ExploitSummit @lium_io @hippius_subnet @chutes_ai @affine_io @IOTA_SN9 > 引用 @ExploitSummit: “We have 23 subnets on Bittensor that make real hard cash.”
@const_reborn at Exploit on paying customers and the work ahead: turning subnets into businesses that can sustain themselves.
Watch the highlight:
https://t.co/bsx3nJIxAL https://x.com/bittingthembits/status/2104850306257256740