朱哲清

Meta acquired 49% of Scale AI for $14.3B and reportedly offered $300M over four years to poach talent, yet still failed to buy organizational stability. By September, some new hires had already left, including 2 who returned to OpenAI; 朱哲清 says quitting within 2-3 months is unusual even in Silicon Valley, and a larger package may not offset post-pandemic bureaucratic bloat, approval burdens and “extremely intense internal political infighting.”

A 5,000-person Superintelligence Labs is unlikely to beat OpenAI on talent density when the true core team behind a SOTA model is usually no larger than 50 people. 朱哲清 estimates 20-plus people set the model direction, 20-plus handle infra, with a small number of PMs; if the goal is simply to make Llama beat GPT-5, he says a focused team of 150-250 people would be more appropriate. “More people are useless,” especially in Meta’s bottom-up culture, where multiple teams race each other without clear model ownership or product landing points.

Top AI talent is not really chasing an uncapped salary; they want to be one of the 10 core authors of the historical moment when AGI arrives. 朱哲清 says that with a 90% chance of making that author list, many people—including himself—might accept the job “without being paid a penny”; so-called “working for love” still includes the reward of having one’s contribution recognized by history. That helps explain why OpenAI and Anthropic attract talent ahead of Google, and why Meta’s massive packages cannot cover its organizational and mission risk.

OpenAI, Anthropic and Google are betting on 3 different paths to capturing AI value. Anthropic is pushing text-based action spaces to the limit, with coding as its strongest application; OpenAI is prioritizing every consumer use case, with Sam Altman’s framing that “when I have 1B DAUs, AGI will definitely happen”; 朱哲清 speculates that Google is using its deep technical stack to serve developers and let applications evolve on their own, potentially producing “mass flowering” across language, video, image, world models and robotics.

The decisive variable in the chatbot race is not the model leaderboard, but whether ChatGPT can move from a desktop productivity tool to the go-to place on mobile. 程曼祺 says ChatGPT weekly active users rose from roughly 300M at the start of the year to 700M-800M; Gemini is also growing rapidly, but has not disclosed directly comparable activity data. Chrome gives Google an almost natural distribution moat on desktop; if ChatGPT can become the first app users tap after picking up their phone, mobile stickiness could finally cut off Google’s pursuit.

AI-native startups gain organizational leverage by hiring less, raising task abstraction and using AI to sustain networked collaboration. At Pokee AI, every new role starts with the question: “Can AI replace this within 2-3 months?” If yes, they do not hire; engineers no longer receive granular tickets, but own the architecture of a complete feature and hand execution to AI. The trade-off is 3-10x more output per person, which makes line-by-line review impossible; trust, complementary skills and avoiding duplicated work therefore matter more than traditional process.

Talent pricing will ultimately be set by capital’s judgment on AI-native revenue, and the most dangerous mismatch is labeling ad gains from legacy recommendation systems as generative-AI returns. 朱哲清 speculates that if year-end and post-Q1 assessments fall short of high expectations, capital and the talent market could cool together in 2H next year; 程曼祺 points out that public markets do reward Meta for “AI improving ads.” The longer-term risk is an education system in which students rush into LLMs, diffusion models and startups, producing highly homogeneous talent while neglecting the critical thinking that matters most in PhD training.

🔗 Original source & video: 134: Meta AI Talent Turmoil: Why $100M+ Couldn’t Keep People | Taking Stock of AI Organizations with Pokee AI’s 朱哲清

The key dividing line repeatedly identified in this discussion is not demand, but whether unit economics can turn positive on every call. 朱哲清 estimates that 80%—90% of industry costs now sit in inference. A single Deep Research run exceeding 1M tokens costs roughly $3—4 on domestic models and $8—10 on top overseas models; Grok can run into the tens of dollars. He has “not seen a single company able to break even on costs” in a C-side use case. Survival routes include model routing, context compression, vertical fine-tuning and self-hosted serving. For C-side products, the question is also whether model providers can cut prices by roughly 80% while leaving room for profit; otherwise, continued cash burning eventually ends in collapse.

Cursor’s $10B valuation and the simultaneous backlash over limits on its $20 plan show that application-layer growth does not automatically become gross margin. 泓君 cited the company’s $900M financing round in June and the view that Cursor and GitHub together contributed roughly $1.2B of Anthropic’s approximately $4B revenue last year. 陈志博 put it bluntly: Cursor is “basically working for Anthropic.” With Claude, OpenAI, Gemini and xAI all treating coding as a primary battleground, Cursor must absorb upstream inference costs while competing head-on with its suppliers. 泓君 speculated that the pricing change may also be intended to improve gross margin ahead of the next financing round.

Poke AI is trying to turn the entire web’s toolset into the model’s action space, tackling tool discovery and context costs at the same time. A single model may support only tens of thousands to 1M tokens of context, while descriptions of all web tools could run to tens of millions of tokens. Traditional Agents replay the full history and tool information at every step, causing the same content to be processed 20, 30 or even 50 times. Poke wants developers to call an Agent as easily as they call the ChatGPT API, but model reliability still falls sharply when it faces thousands of highly similar tools.

The three companies are betting on three different moats: tool ecosystems, identity networks and design delivery. Second Me distinguishes between the “doing” of a profession and a person’s “being,” aiming to give everyone an identity Agent that represents them and connect those Agents into a network. Lovart is not competing with foundation models such as Nano Banana for point-generation capability; it uses a designer-style planner, aesthetic memory and an editable canvas to deliver personalized work. For Lovart, model capability is raw material: “Only water deep enough can carry a ship large enough.”

Training paradigms are shifting from SFT toward RFT and RL because of the data bottleneck, but “learning through use” remains far from solved. 朱哲清 believes the most expensive part of SFT is now licensed data and labeling for complex tasks. RFT/RLVR comes at the cost of longer, less stable training and weaker generalization. 陶芳波 summarized the shift as: “The SFT era was about finding data; the RFT era is about finding environments.” Their key disagreement is whether the next generation will enter an era of experience: 陶芳波 believes online use will make Agents smarter, while 朱哲清 stresses that few-shot feedback may be an outlier and that safe inference-time updates are “extremely difficult.”

B-side Agents are already close to breaking even, while C-side products are taking three different bets: subscriptions, network effects and outcome-based revenue sharing. Poke can cover costs on the B side; on the C side, subscriptions may cover inference but not labor. Lovart uses tiered credit subscriptions, testing delivery value through whether users keep buying. Second Me is choosing to charge as little as possible in order to build an identity network; the possibility of an AI-native advertising model is 泓君’s growth bet, and 陶芳波 agreed. Pine takes a cut of the amount it recovers in users’ billing disputes, offering the most direct answer: “Take a percentage of the final outcome.”

Trust in and the durability of high-priced KOL traffic are weakening, while community and the Agent kernel are becoming more important distribution and platform assets. Bill believes expensive KOL traffic decays quickly and carries weak trust; Runway and Luma had more durable early growth through co-creation on Discord. 陶芳波 sees the Claude Code SDK as a deeply underestimated “coding kernel” that could even become the CUDA of the Agent era. 泓君’s changed view of Cluely, Lovable and Replit also shows that value need not come from fully automated delivery: real demand can exist in live summarization, prototype communication and giving people more confidence.

🔗 Original source & video: E207 | Agent Founders’ Cost Breakthroughs and Commercialization

The key divide in Agent training is not just data scale, but whether a task can be closed-looped through an objective and a Verifier. Tasks with existing text, image, and video collaboration data can first learn via SFT, then use RLHF to align preferences; but if a task requires 50 tool calls, no such complete trajectory exists online. Goal-driven tasks in code, math, physics, and supply chains need counterfactual generation of new solutions, followed by ground-truth verification—“If you can find a Verifier, you can consider the problem solved.”

Within OpenAI’s five levels of AGI, the real technical chasm is not chat to reasoning, but Level 3 Agent to Level 4 Innovator. 朱哲清 sees the framework as a product-capability taxonomy: Level 1 depends on the definition, Levels 2 to 3 may not represent a huge leap, and the gap between Levels 4 and 5 may not be that large either. The hard part is whether a Verifier can generalize from known knowledge to judge new knowledge discovered by AI itself. “The gap between 3 and 4 is far greater than the gap separating 1, 2, 3 from 4 and 5.”

RL pretraining could push AI toward knowledge unknown to humans, but it could also produce strategies that run while remaining unintelligible to people. Current attempts may still carry prior knowledge unavailable through pretraining; if developed further, code could have completely correct inputs and outputs while its internal variables, operations, and compilation logic look like gibberish. “Human readability” itself is difficult to encode as a rule-based check. 朱哲清 therefore sees reward design as a potential regulatory target, because the incentives assigned to an Agent during training determine its eventual behavior.

RL’s cost disadvantage is real, but complex tasks may eventually leave companies no choice. 泓君 cited the industry view that RL delivers roughly 2x better results at roughly 10x the token consumption. 朱哲清 did not confirm the ratio, but explained that RL must rely solely on a reward function to reach its objective, while SFT already has standard answers; “RL fine-tuning will cost more.” Once tasks become complex enough to have no standard answer, the cost-effectiveness comparison breaks down: “Problems will become more complex in the future… and then there will be no choice.”

Meta’s rationale for acquiring Scale AI may not be a return to a pure-text data race, but filling the medium-term annotation gap in multimodal and robotics. 朱哲清 summarized the resource cycle as compute in the short term, data in the medium term, and RL optimization over the long term. Image aesthetics are hard to standardize, video details are difficult to parse, and annotators may not even understand a robot’s intermediate actions. Multimodality may first require better parsing and reward-model capabilities before a Judge or Verifier can be built and RL fine-tuning can begin.

Pokee AI is betting on a proprietary “reasoning + tool sequence” model that delivers the result through a single API and one line of prompt. A browser or virtual machine can compress thousands of tools into one generic interface, but at high cost, low speed, and with weak generalization in professional scenarios. Pokee instead accesses tools directly and hopes eventually to combine semantic understanding, planning, tool selection, and result generation into a single model. But 朱哲清 stressed: “Model capability determines your product’s floor, while product details determine its ceiling.”

Scarce RL talent and negative-gross-margin Agents will jointly drive the next round of capital bifurcation. 朱哲清 estimates that “the entire industry has only a few dozen people” who can truly optimize RL well. Meanwhile, Agent companies may grow faster and lose money faster, with gross margins potentially reaching negative 50%. The technical route determines whether a company survives; the founder and team determine product differentiation. Over the next 1 to 2 years, the market may see a “big fish eating small fish” wave in which small teams are acquired.

🔗 Original source & video: E201｜The Next AI Frontier: Reinforcement-Learning Pretraining and AGI’s Inflection Point (Part II)

ChatGPT Agent links Deep Research, Operator, and a virtual machine into an end-to-end workflow, but its first-round experience feels more like an engineering effort to seize the general-purpose Agent entry point than a mature technological leap. 朱哲清 tested “research first, then make a slide deck,” which took 35 minutes to 1 hour; travel tasks also took 20–30 minutes. He believes the main cause is the original speed of Deep Research and Operator, which becomes even slower when the two are combined. Its advantage is using strong vision capabilities to operate on webpages, but the overall results fell short of expectations and were far slower than imagined.

For consumer Agents, the hardest problems are not clicking through webpages but personalization, payment trust, and whether the task is worth delegating in the first place. ChatGPT Agent’s selected Singapore flights and hotels were ultimately rejected because it failed to remember the user preferred Hyatt, direct flights, and lower prices; when it reached payment, a human still had to take over. 泓君 noted that human assistants also require communication. 朱哲清 acknowledged the point, but said complex preference memory remains incomplete and cold start has not been solved.

The four architectural paths for general-purpose Agents are fundamentally trade-offs among universality, speed, reliability, and permission coverage. The browser route “is indeed universal,” but comes with high token usage and network latency; an open virtual machine suits coding and data analysis but struggles with login and authorization; restricted environments trade templates for speed; Pokee-style third-party API/SDK integrations are faster and more reliable, but can execute only actions formally exposed by platforms.

Product differentiation has shifted from “who is more general-purpose” to “who can be reused reliably within a target workflow.” Manus wins on coverage through its browser-plus-virtual-machine architecture, but may be dragged down by long contexts, hallucinations, and hour-scale runtimes; JSBox breaks the super-agent into templates for slides, spreadsheets, AI calls, and other tasks, improving speed while moving toward a vertical platform; 朱哲清 says Pokee can be 4 to 10 times faster than comparable products and cut per-run tool-call costs by 50% to 60%, at the cost of giving up some personal-account scenarios.

The clearer commercial market is recurring workflows for professional users, not one-off “universal assistants” for ordinary consumers. 朱哲清 observes that many consumer Agents have poor retention because a task “ends after you use it once”; Pokee users, by contrast, repeat the same workflow every week. Standardized business travel suits API-based Agents, while leisure travel with many variables is better suited to browser-based Agents—a divide that directly determines retention and unit economics.

If Agents become the entry point over the next 1 to 2 years, portal traffic, protocol control, and content monetization could all be rewritten. 朱哲清 expects portal traffic for e-commerce, search, and video to fall rapidly. He cites Google’s launch of A to A and argues that ChatGPT, Cloud, and Pokee launching protocols likewise reflects a contest for the Agent entry point. He also estimates that of nearly 20,000 MCPs, fewer than 200 are actually usable, making maintenance and reliability hard constraints on the protocol ecosystem.

The creator economy may not disappear, but advertising and recommendation could migrate from “page ranking” to “conversational sequencing.” Responding to 泓君’s concern that AI summaries could erode podcast advertising, 朱哲清 imagines Agents paying content providers per invocation, then inserting commercial recommendations at points where there is no single correct answer. The traditional ranking mechanism that displays 5 to 10 pieces of content at once could be replaced by 5 to 10 rounds of recommendations, with each round competing for the next interaction. He explicitly preserves the uncertainty: “I’m not 100% sure this is definitely where the future is headed.”

🔗 Original source & video: E201｜From Manus to ChatGPT Agent: How Do Their Underlying Technical Architectures Differ? (Part 1)

Poke AI’s core bet is to use language models as the frontend for understanding intent and interacting with users, while putting RL-trained non-language models in the backend to handle planning, reasoning, and tool calls. Bill argues that if 50 tools each require 1,000 tokens of documentation, tool context alone reaches 50,000 tokens; once memory, web results, and a dozen-plus execution steps are added, a single task can run into more than 1 million tokens, making language models prone to hallucination. Over time, the LM may look more like a UI than the Agent’s decision-making core.

The product wedge is not generating more content, but letting Agents write into real work accounts across platforms. Poke AI plans to launch with roughly 1,000 callable sub-tools spanning dozens of platforms; it does not currently use MCP, but plans to support MCP and Agent-to-Agent, while offering a simplified JSON-based integration path. It prioritizes official APIs and code, using browser automation only where necessary.

Bill’s opening summary of a strong general-purpose Agent included four traits: faster than humans, minimal human intervention, able to both read and write, and low cost. Poke’s demos take about 60 seconds when approval is required, and potentially 10–20 seconds without it; Bill says the per-task cost is roughly one-fortieth of comparable products. Because tasks write into user accounts, the product supports both controlled and autonomous modes.

The commercial entry point is developers and prosumers, moving bottom-up into the enterprise; the moat is defined as workflow attachment, not point-model technology. For users in social media, marketing, legal, and finance, the pain point has shifted from content generation to cross-platform distribution and operations; enterprise workflows spanning dozens or hundreds of steps are the larger end market. Bill believes integrations, historical files, account relationships, and usage habits can create deeper user lock-in.

The technical barrier is effective exploration across a large tool universe, not enduring overwhelmingly negative signals through random sampling. When only one of 10,000 tools produces the right answer, traditional supervised or contrastive-learning methods may repeatedly receive the wrong signal. Poke AI wants to use smarter exploration to skip known-bad tools and paths, reducing the data required for training.

Market education shaped when Poke AI could move back from a vertical application to a horizontal platform. When it raised money in fall 2024, investors said Bill was “six months ahead of the curve”; after DeepSeek and related work from other companies created broader market consensus, the company says it heard from more than 100 investors and dozens of large prospective customers. After releasing its demo in March, it generated more than 800 waitlist sign-ups in a week; after launching in March, another 800-plus people signed up, with online-post-to-waitlist conversion at roughly 8%–9%.

There is a competitive window, but Bill does not believe speed alone will secure the market. He expects at least 10 horizontal Agent companies to emerge over the next year, with perhaps 3–4 or 4–5 ultimately remaining, differentiated by vertical focus, capability emphasis, integrations, and user lock-in. North America’s open APIs, SDKs, and third-party integrations make it the first major battlefield; China’s relatively closed commercial ecosystem may require Poke AI to first partner with large incumbents around platform capabilities.

The founder-market fit comes from nearly a decade of deploying RL, while whether a travel demo could generalize to the real world was once a “big question mark.” At Meta, Bill led Applied RL and estimated that the algorithms contributed nearly $500M in annual revenue, while he completed a Stanford PhD in parallel; his method for evaluating technical judgment is to first find a minimal viable example with very low compute requirements and broad applicability, test whether the technology has a first-principles advantage, and only then attempt to scale it into a large system.

🔗 Original source & video: 111: 朱哲清’s Pokee.ai Agent Playbook: RL as the Backend, Language Models as the Frontend | Agent #3

Bill’s core view is that RL is not an alternative to LLMs, but a learning and planning method layered on top of Transformers. LLMs provide language, embeddings, and a basic understanding of the world; RL searches unknown paths and makes decisions based on long-term reward. The two are not parallel competitors, but “may be vertical” layers. That is the premise behind his view that RL can be “reborn” in the era of large models.

The bottleneck for complex Agents is not understanding instructions, but finding viable solutions across an exponentially expanding sequence of steps. LLMs tend to loop along the highest-probability path, while RL can prioritize the most uncertain direction and backtrack or switch routes based on expected future value. Bill summarizes it this way: “You don’t need brute-force enumeration anymore; you need search.” Travel planning is the classic example: the system only discovers on step 10 that cities or dates conflict, then has to know where to backtrack.

Pokee AI is betting on an architecture where “RL makes decisions and LLMs handle I/O,” rather than putting a larger LLM in charge of everything. User requests are abstracted into goals, states, and actions; RL determines API call order and parameters, and the final result is converted back into human-readable form by an LLM. Its public Shopify system reportedly uses a 15M-parameter RL Agent and a language model under 10B parameters to master dozens of APIs in less than a month, with no special personalization training required when moving across stores.

The commercial leverage of this route comes from reliability and cost, not model-scale narratives. Bill says the team spends less than $10,000 a month on training, and the RL model can run on CPUs. If GPT-4o were used for comparable capabilities, the cost could be “1,000x” higher. His response is: “If it solves the problem, that’s enough.” The company does not position itself as a large-model company, but as a “tool-using Agent with strong reasoning and planning.”

E-commerce is merely the first closed-loop setting for a general-purpose API Agent; the value proposition is replacing fragmented plugins and human coordination. A Christmas promotion that once required sales, engineers, and multiple vendors to separately modify discounts, shipping, the homepage, search, and recommendations could eventually be executed by an Agent through a single text instruction. Bill says some merchants spend more than $1M per quarter on this kind of work. If the product delivers, he estimates that “45% to 60%” of this purely engineering-oriented software could lose its value.

Pokee and coding Agents target different markets: the former combines existing business capabilities, while the latter creates capabilities from scratch. Bill believes most code in real-world work is simply composing APIs. Treating one API call as an action is far cheaper than generating, compiling, and validating code token by token. The roadmap is therefore to expand to thousands and eventually tens of thousands of APIs, then move into advertising, travel, education, fitness, legal, and finance—not to chase general-purpose coding first.

For investors, the key questions are accuracy, reliability, and unit economics in the public beta—not the slogan “the first year of Agents.” Bill sees no shortage of US capital and continued interest in vertical Agents, but also a crowded field and a confusing target set. Less crowded areas such as science and supply chain may attract more attention. He rejects the first-year narrative, saying only that the next year may become “the year when Agents truly help you take action and produce results.”

🔗 Original source & video: How I Use RL + LLMs to Build Agents | A Conversation with Pokee AI Founder 朱哲清 Bill