WhiteForest, previously a co-founder at Foresight Ventures, used a lengthy post to roll out his AI trading product Mojo. His bigger claim: most AI trading tools on the market still miss the real reason retail traders lose money.
He said he had overseen five business lines at Bitget and examined actual trading behavior from hundreds of thousands of retail users. In one example from the post, a longtime friend sat down with the product and typed: "I want to swap half of my ETH into SOL" ("wo xiang ba shou shang yi ban ETH huan cheng SOL"). Mojo did not place the trade right away. It stopped first and showed three things: the user had made the same ETH-to-SOL rotation four times over the prior 12 months, got it wrong three times, and lost 11% on average; before the last identical move, the user had said, "If SOL/ETH drops below 0.038, I will definitely switch back" ("ru guo SOL/ETH die po 0.038 wo yi ding huan hui lai"), but never followed through; and the user’s SOL allocation was already 4 percentage points above a self-imposed hard cap.
Then the system asked one question: "Are you making a decision now, or repeating something you have already done four times?" ("ni xian zai shi zai zuo yi ge jue ding, hai shi zai chong fu yi jian ni yi jing zuo guo 4 ci de shi?") WhiteForest said the goal was not to tack on one more market signal. It was to make the trader face his own pattern before he made the next mistake.
Retail losses, he argues, are not mainly an information problem
WhiteForest said he used to think retail traders lost because they lacked judgment, information, analysis, or trading signals. After going through the data, he said, that story stopped making sense.
He pointed to Bank for International Settlements research covering crypto trading data from 95 countries between 2015 and 2022, saying 73% to 81% of retail traders lost principal. He also cited a 2025 survey of 1,005 crypto retail traders that put the number at 84%.
His take is different. A trader may have already set a stop-loss, then get trapped in a meeting when the market breaks through it. A watched on-chain wallet may move, yet by the time the trader reacts, the price is already 20% higher. Position size may already be above a hard limit, but the trader waits because the market still "looks fine." Same mistake. Three months ago. Six months ago. And then again, because nobody put that pattern back in front of the trader at the exact moment it mattered.
So his argument is that the core retail weakness is not judgment in some abstract sense. It is the lack of a system that can bring a trader’s own past behavior, spoken rules, limits, and violations back into view right as a trade is about to happen.
He splits current AI trading products into four groups
WhiteForest said his review of mainstream AI trading products left him with four broad buckets.
Strategy cards
First: the strategy-card model. He described apps that show a row of cards with metrics like BTC trend 137% or SOL breakout with a 68% win rate, usually stamped with an "AI optimized" label. In his view, these products are just sorting whatever made money in the last 90 days and sticking it at the top. His verdict is blunt: that is automated survivor bias. He also said versions of this same idea were already around in 2018 under the name of copy trading.
Copy-trading agents
Second is the copy-trading agent built around following so-called smart-money wallets. WhiteForest mapped out the timing chain: 12 seconds for on-chain confirmation, 5 seconds for the agent to detect it, 3 seconds to send an alert to the phone, and 4 seconds for the user to tap and confirm. By the time the trade is done, he wrote, the original "smart money" wallet is already 24 seconds ahead.
And he said the uglier version is when the wallet is not a smart-money signal at all, but internal activity from a market maker creating an apparent trade for others to chase. Once large numbers of copy-trading agents rush in, that flow can be sold into. His broader point: "smart money" is not some fixed identity. It is a label pinned to past returns, and the chain only shows what a wallet did yesterday.
AI analysis chatboxes
Third comes the AI analysis chatbox. WhiteForest described products that let users type questions such as whether they should buy BTC, then spit back an 800-word market brief covering RSI, MACD, on-chain flows, social sentiment, and historical analogies. He argued that many of these tools are basically ChatGPT wrappers connected to five free APIs. In his version of it, the on-chain data comes from public Etherscan pages, the social sentiment comes from LunarCrush’s free interface, and the historical analogy gets produced from model training data.
He contrasted that setup with Citadel, which he said spends $700 million a year on real market intelligence, while retail users may be paying $29 a month for what is, in practice, a polished bundle of free inputs. He also claimed that every prompt entered, every recommendation clicked, whether the user acted, and whether the user made money can be packaged up and sold to proprietary trading firms.
Products that use real LLMs but restrict them heavily
Fourth, in his phrasing, is the category where a real large language model has been "castrated." He said some teams actually integrated Claude or GPT-4, paid for the models, and understood AI, only to cut those systems down to a basic chat window with five hard-coded strategies in the backend. The most valuable feature of an LLM, he argued, is generalization. And that gets thrown away when every user request is shoved into one of a few fixed functions.
Under that setup, the model does not ask whether the trader has already made the same move four times, whether position size has broken a self-set ceiling, or whether the current action is a decision or a habit. It just executes. WhiteForest tied that choice to a product focus on MAU rather than decision quality, saying the user becomes little more than an active-user number inside a pitch deck.
"You think you are using AI trading. In reality, AI is sampling you"
WhiteForest pushed the argument further, saying retail investors have historically survived in part because they are inconsistent. They get tired. They forget. They skip a week. Or they leave after one lucky run. In his framing, those irregularities acted like armor because the market could not fully model an unstable opponent.
Now, he wrote, AI is recording those gaps one by one. If a trader makes the same mistake four times in 12 months, a machine at a firm such as Citadel may know the fifth is coming 0.5 seconds before the trader does. In that setup, paying for AI tools does not make the trader stronger. It makes institutional counterparties better at spotting when that trader is likely to chase, panic, or cut losses.
His challenge to the phrase "empowering investors"
WhiteForest also went after one of the sector’s favorite slogans: "empowering investors." He wrote that nearly every AI trading landing page uses the phrase, and that it shows up in founder pitches and LP decks too. In his telling, it is branding, not a promise.
He defined real empowerment as giving an ordinary person who works during the day, picks up children in the evening, and goes to sleep at 3 a.m. three things usually reserved for institutions: cognition, judgment, and execution, all tied directly to that person. Fake empowerment, he said, is turning those capabilities into cheap substitutes — a chat window, a copy-trading button, a strategy card — and selling the package as an upgrade.
He took it a step further and said true empowerment would mean access to the same level of information and judgment available to firms such as Citadel, while false empowerment only makes a user feel upgraded without changing the basic role that user occupies in the market.
His view of the next 18 months and the next five years
WhiteForest said that if the industry stays on its current path, at least 180 out of roughly 200 AI trading wrappers will vanish within the next 18 months. He wrote that the remaining 20 would be acquired by proprietary trading firms, and that user behavior data would then flow into Citadel’s training pipeline.
He stretched the point over five years. In his wording, a retail trader without an agent would look like a crypto-era version of an unsophisticated investor walking into a brokerage branch. A retail trader using a superficial AI wrapper would be like infantry carrying a plastic sword — dying faster, more precisely, and more cheaply. He said this was not a prediction, but the destination of the current path.
Mojo is presented as the alternative
WhiteForest then made it plain that the product his friend tested was Mojo, the system he is building himself.
He said the decision to build it came from three vantage points: investing in AI infrastructure at Foresight Ventures, watching the broader shape of the industry at The Block, and managing five business lines at Bitget while seeing real retail trading behavior at scale. Those three roles, he wrote, all pointed to the same conclusion: what this generation of AI can actually do is still, for the most part, not being built.
That capability, in his telling, is not another signal engine and not a forecast about whether SOL will rise. It is the ability to bring back the user’s own past 12 months of behavior at the exact moment a fifth identical mistake is about to happen: the trades already made, the rules already stated, the lines already set, and the lines already crossed. He argued that this gives the trader something institutions already have — a system that truly understands that trader’s own behavior and is attached to the trader, not an outside desk.
He framed that as the real promise of this AI cycle: not a fancier trading button, but institutional-grade cognition, judgment, and execution placed directly on the retail side. Cognition can be loaded into an agent, he wrote; impulsive judgment can be interrupted by a conversation; execution can be handed to an around-the-clock API. In his view, that is physically possible now, not just a pitch-deck line.
Criticism of VC incentives and institutional buyers
WhiteForest also argued that $1.7 billion in venture capital is currently funding the opposite direction. In his telling, that money is there to produce venture returns and build a product clean enough, from a data standpoint, for Citadel to buy within 36 months for $50 million. The real acquisition target, he wrote, would not be the product itself so much as the data set behind it: every hesitation, every impulse, every capitulation from every retail trader over the previous 1,000 days.
Once that kind of data reaches an institutional trading system, he argued, each retail pattern can be modeled more precisely: when a user is likely to buy impulsively, when panic selling is likely to hit, and when to be waiting on the other side. If this AI cycle ends with Citadel getting stronger, Jane Street making more money, and retail traders still being harvested with greater precision, then the so-called progress is only institutional progress, he wrote.
He closed by saying he does not want to build or fund that kind of product. Instead, he said he wants to bring AI trading "back to retail traders."