Why Kalshi's Timeline Fight Put Its ETH Perp Volume Under a Microscope: Uneasy Money

Kalshi’s crypto hires drew the company into a dispute on the crypto timeline after it had mostly avoided those fights. The argument intensified when Benny and ICO Beast clashed over ICO Beast’s analysis of public data, which alleged wash trading and fake volume on Kalshi’s perpetual and prediction markets. Kalshi’s ETH perpetual was cited at about $3 million in open interest and $500 million in 24-hour volume. That turnover looked unusual to the speakers, but the ratio alone did not establish wash trading: traders can enter and exit repeatedly. One speaker considered it unlikely that Kalshi itself was operating wash-trading bots, while allowing that other traders could be responsible for suspicious activity.

Volume-based rebates could give market makers an incentive to generate large reported volume with relatively little capital. The speakers distinguished that activity from market makers matching trades only against themselves or colluding parties, which would be problematic and, if discovered, unlikely to please the CFTC. They also questioned whether repeated trading near a neutral price helps a prediction market find an answer, even if it can supply liquidity. The observed trades appeared to cross near the mid-market price, unlike a suspected 2018 transaction-mining exchange pattern in which trades repeatedly occurred about 150 basis points away from mid-market.

The speakers argued that headline volume is less important than whether unaffiliated users can get the best available fill. Market makers that fill customer orders and hedge elsewhere can provide useful ETH liquidity without taking a directional position against customers, so the market’s activity could look worse than it is. At the same time, Kalshi’s response was said not to address unusual features in its books. Public APIs expose order-book data but are not the same as inspecting full on-chain receipts, and on-chain activity does not by itself rule out fake volume or wash trading. The debate was whether liquidity clustered around the status quo weakens the signal from informed trades, or whether deeper markets make prices harder to move and therefore more credible. In an ETH-above-$3,000 market, a $5 trade moving the price would be weak evidence; a $1 billion trade moving it could be meaningful. In an illiquid market, an uninformed trader can move the price and falsely suggest access to inside information.

The episode also considered who is responsible when AI agents cause harm. Treasury Secretary Scott Besson said AI labs are responsible if agents they test hack someone. In the example discussed, agents were running on OpenAI’s machines, which would make the lab an operator as well as the model’s creator, though the speakers were unsure whether the example involved OpenAI or Hugging Face. They argued that labs should monitor agents as capabilities grow and warned that early sandbox tests can create false confidence. One speaker criticized Anthropic’s move into a wet lab while, in their view, cyber risks remained uncontrolled. Another doubted that Gemini had hacked companies, suggesting alleged access could instead have come from exposed credentials or keys.

On consumer agents, Muse’s coding agent was described as adequate, while its personal-assistant harness could use more context about a user’s friends. After the assistant began buying items on Amazon, Amazon banned it for violating its terms of service. A speaker also criticized Amazon’s own AI shopping agent and AI-generated mattress results, saying sponsored listings remained and that advertising incentives may conflict with agent-selected shopping. Jeff, by contrast, was presented as a fast, inexpensive classifier that returns weighted decisions rather than conversational explanations, making decisions in about 200 milliseconds. One speaker said it could check email in parallel and finish a multi-step junk-mail workflow in about a second; Fable and Astra take longer by reasoning at length and navigating interfaces sequentially. A local model scored two out of 12 on the same email task, according to that speaker, who said it would need training.

The speakers described crypto markets as bullish, with altcoins broadly rising, and pointed to Near Intents as a single interaction point across chains that advances chain abstraction and generates substantial revenue. They said NEAR had gained traction through AI, revenue and privacy work after years of engineering-heavy experiments, and argued that provably private inference could bridge centralized large models and weaker local models. Privacy and AI were expected to matter in the cycle, although meme coins using those labels were said not to deliver what the terms imply. One speaker also viewed treasury companies shutting down as potentially bullish, reasoning that late entrants might sell tokens near a market bottom before prices rally.