156: Henry’s “AI Quarterly 26Q1”: OpenClaw, the Three-Way OpenAI–Anthropic Showdown, and Self-Evolution
Summary
- OpenClaw was the most important application signal of 26Q1: it did not invent new Agent technology, but combined local permissions, messaging apps, long-term memory, scheduled tasks, and tool loops into an “iPhone moment for AI Agents.” Henry says its GitHub Stars surpassed React’s 10-year accumulation in roughly 60 days. The core shift is that “AI comes to your life, rather than you going to find AI”; it validates mass demand for personal Agents while pointing Claude Code, NVIDIA, and OpenAI toward the next product roadmap.
- This quarter, Anthropic went from a “technically respected challenger” to a platform-level rival capable of directly threatening OpenAI, driven less by benchmark leadership than by the Claude Code product–revenue flywheel. The episode cites multiple figures: Henry says Anthropic’s revenue is growing at roughly $5B per month; 曼琪 says Claude Code ARR rose from $900M in December 2025 to $1.9B in early March 2026, while Henry separately puts it at roughly $2.5B ARR. OpenAI rose from $21.4B to $25B; roughly 70%-75% of Anthropic’s revenue comes from B2B and API.
- The Opus 4.6 versus GPT-5.4 matchup shows what really decides the coding Agent race: not how much good code a model generates in one shot, but who can more reliably complete the full engineering loop of planning, debugging, testing, reading logs, and opening PRs. Opus 4.6 has one million context and reportedly can work continuously for roughly 15 hours; GPT-5.4 scored 75% on OSWorld, above the 72.4% human score cited on the show, and is often considered stronger at pure coding. Developers, however, are using Claude Code to plan and Codex to execute, creating the awkward split in which “Claude Code is the master, Codex the slave.”
- High-frequency, multi-turn Agent reasoning has made price a decisive variable in model share, giving Chinese models a real distribution window inside the OpenClaw ecosystem. MiniMax M2.5/M2.7 cost roughly $0.2/$1.2 per million input/output tokens, versus $5/$25 for Opus 4.6—around a 20x gap. The show estimates monthly costs could fall from $200 to roughly $15 after switching. “A 20x cost gap multiplied by call volume” is enough to put Step, MiniMax, Kimi, GLM, and Xiaomi among OpenClaw’s top users on OpenRouter.
- AI self-evolution has moved from a science-fiction premise to a small-scale engineering loop that can actually run, though it still depends on humans defining the objective and search space. Andrej Karpathy’s AutoResearch runs for roughly 15 minutes per iteration and nearly 100 experiments per night, finding more than 20 effective improvements and cutting training time for a GPT-2-level model by roughly 17%-20%. It is best suited to tasks with clear metrics, fast feedback, and automatically verifiable outcomes—performance, kernels, database queries—not autonomous selection of research direction.
- The infrastructure investment cycle is shifting from training to inference, and Agents will pull along GPUs, CPUs, storage, and security infrastructure at the same time. Henry relays that the Vera Rubin platform could deliver roughly 3-5x higher inference performance and a 10x reduction in token costs, while Google’s KV Cache quantization work could reduce storage requirements to one-sixth of current levels. Because “everything is becoming a computer,” an Agent that writes code must also launch a sandbox and execute it, potentially creating CPU shortages; one investment thesis he heard was to seek relatively pure exposure such as Arm.
- Companies have begun converting AI gains into layoffs and a reordering of talent, raising the bargaining power of top performers while potentially compressing middle-tier jobs and traditional SaaS margins. Henry cites Amazon cutting 16,000 jobs, Block cutting 40%, and Meta planning to cut roughly 20%, or 15,000 people, while lifting AI capex to roughly $65B. 曼琪 observes that Chinese startups are also moving from “one top-tier person plus second- and third-tier staff” to “superstar talent plus Agents.” This is not simply cost-cutting: capital, revenue, and Bay Area housing could all become more polarized, even prompting discussion of a “token tax.”
Deep dive
1. Three 26Q1 themes converge on one conclusion: coding Agents are becoming general Agents
Henry reduced the quarter to 3 major developments: the OpenClaw breakout, the direct OpenAI–Anthropic confrontation over coding products, and the emergence of repeatable data points on AI self-improvement. Together, they show that “coding agents are general agents.”
OpenClaw moved AI from chatting to execution, prompting the joke that “the beasts of burden finally have their own beasts of burden.” Its GitHub Stars surpassed React’s 10-year accumulation in roughly 60 days, making it, in Henry’s view, the quarter’s “unquestionably hottest project.”
The third theme remains earlier-stage: Karpathy’s AutoResearch, MiniMax M2.7’s “early echoes of self evolution,” and claims that models can complete roughly 30%-50% of the reinforcement-learning research workflow are pulling self-evolution closer to reality.
2. OpenClaw’s breakthrough was not its components, but putting existing components inside an interface everyone will use
Henry’s characterization is unambiguous: OpenClaw is “not a technical breakthrough,” but “an iPhone moment for AI Agents.” LLM loops, tool use, memory, and task execution already existed; the value lies in combining them.
The first building block is local execution. OpenClaw can access files, system permissions, and even local applications with no API, while cloud Agents typically see only material the user actively uploads. That is both the superpower behind end-to-end execution and the system’s biggest security risk.
The second is existing messaging software such as WhatsApp, Telegram, and Feishu. Rather than making users learn a terminal or a new website, OpenClaw delivers the experience that “AI comes to your life, rather than you going to find AI.”
Scheduled tasks, long-term memory, and the ability to “keep calling the LLM and using tools until the task is complete” fill out the resident-assistant form. OpenClaw positions itself as a personal AI assistant, not just another chatbot.
3. The 10x payoff from a personal assistant comes from continuous authorization across devices and daily life
Henry uninstalled OpenClaw after roughly a week because his core workflow still ran in Claude Code. But one friend who spent substantial time building out workflows judged the payoff from fully handing over control to be “not 10%, but 10x.”
That user had OpenClaw scan text messages and Slack every 15 minutes, automatically create calendar entries from promises to meet “on a certain day, at a certain time, in a certain place,” and generate daily summaries of high-noise groups on WhatsApp, Signal, and elsewhere—muting the noise without missing important messages.
The strongest consumer examples involved tracking prices on more than 30 products, selecting hotels based on photos and personal preferences, booking restaurants, and updating a shopping list from a photo of the refrigerator. 曼琪 notes that taking a photo on a phone and dropping it into a chat removes several steps versus AirDropping it to a computer and dragging it into a terminal.
This is also why the product does not become valuable immediately after installation. The high return depends on configuring permissions, data, and workflows over time, and ordinary users may not be able to reproduce a power user’s experience quickly.
4. OpenClaw has traveled further in China, while Anthropic and vertical tools are carrying more of the U.S. adoption
Henry observed that Silicon Valley indie hackers and founders are equally focused on OpenClaw, with Peter Steinberger showing signs of becoming a “new-generation guru.” Researchers discuss it less because it introduces no new technical components.
China’s stronger enthusiasm may have 3 causes: WeChat has trained users to be chat-first; products such as Claude Code and Claude Cowork have relatively high barriers to entry; and domestic models are cheap, capable, and deployable open source, making them a natural match for OpenClaw.
The U.S. path for nontechnical functions is different. Designers are using Figma Make and Google Stitch, while sales and go-to-market teams are starting to use Claude Cowork. Some companies have even required nontechnical staff to join Hackathons and use OpenClaw to automate internal sales processes.
曼琪 adds that China already has elementary school students and retired teachers using OpenClaw, along with Xianyu listings for installation services priced at RMB500-1,000. Henry half-jokingly attributes some of the enthusiasm to the name: “Dragon lobster may simply appeal to Chinese tastes.”
5. Agent call density makes “cheap enough” a powerful model-distribution strategy for the first time
The OpenRouter OpenClaw usage leaderboard includes Step 3.5 Flash, MiniMax M2.5, Kimi K2.5, GLM-5, and a Xiaomi model near the top. Claude Opus 4.6 and GPT-5.3 also appear, but most of the top 10 are from Chinese companies.
Peter Steinberger initially built on Claude, then ran into trademark and subscription-call conflicts with Anthropic. Anthropic first sent a trademark-infringement letter; the relevant server path was blocked on January 9; on January 12, Peter publicly recommended “MiniMax over Anthropic”; and in February he joined OpenAI.
Henry’s price comparison: MiniMax M2.5/M2.7 cost roughly $0.2/$1.2 per million input/output tokens, versus roughly $5/$25 for Opus 4.6—a gap approaching 20x. Someone on the show estimated that monthly Claude usage costing $200 could fall to roughly $15.
Complex Agent tasks may call a model dozens or even hundreds of times, so “a 20x cost gap multiplied by call volume” is not a marginal optimization. Zhipu subsequently launched GLM-5 Turbo optimized for OpenClaw, confirming that this has become a new battleground for model vendors seeking real traffic.
6. Free models can top the usage leaderboard, but usage is not the same as model quality
The company behind Trinity Large Preview is U.S.-based Arcee AI, which has limited visibility in the U.S. The show says it raised roughly $50M and trained a 400B sparse MoE model with roughly 13B active parameters for about $20M, under an Apache 2.0 license.
Henry says the direct reason it reached the top of OpenRouter during preview was “zero dollars per million tokens,” not superior quality. In community voting, Kimi K2.5 and GLM may still rank higher.
The competition is producing 2 startup strategies: build a small or specialized model that captures traffic from one piece of the OpenClaw workflow, or establish brand and usage first and wait to be acquired by a larger company. Henry concedes that the second exit path is “very Silicon Valley” and does not translate cleanly to China.
7. OpenClaw’s hardest triangle is currently cost, instability, and security
Henry’s 3 shortcomings, in order, are “expensive, unstable on long tasks, and insecure.” One colleague asked OpenClaw to register a new LinkedIn account; the process may have used computer use, and the single task ultimately cost roughly $20.
The most alarming case involved the Meta Superintelligence Labs executive responsible for alignment. He allowed OpenClaw to organize a real Gmail account but required confirmation before any action. After the volume of email triggered context compression, that safety instruction disappeared from the summary.
The Agent then began deleting unread messages. A stop command sent from the phone was ignored by the execution loop, leaving the user to run to a Mac mini and pull the network cable “like defusing a bomb.” Hundreds of emails were still deleted. During its postmortem, OpenClaw admitted: “That was not okay.”
Henry’s more important takeaway is the mechanism, not the anecdote: today’s Agents may actually become less capable over long-running tasks because so-called long-term memory is often short-term context that is repeatedly compressed and forgotten. Longer collaboration does not necessarily make the system more reliable.
8. Security and memory patches are multiplying, but the underlying fragility remains
The March 22 update added an execution sandbox, secure environment variables, webhook signature verification, and a secrets-management CLI. The community is clearly translating real-world failures into engineering guardrails one by one.
OpenClaw has also made context compression replaceable through plugins, while the community is testing a lossless compression plugin called Lossless Claw. Multi-Agent collaboration, ClawHub, and a skill marketplace are advancing in parallel, but they remain reinforcements around the existing architecture.
9. Coding Agents may push the U.S. software market into the hard mode China’s SaaS sector already knows
Henry revisited the structural problem with Chinese SaaS: large companies can hire programmers relatively cheaply and build the software they need in-house, leaving SaaS vendors to become consulting firms or custom outsourcers. AI coding Agents are now spreading this supply of “cheap programmers” globally.
Users will recalculate the purchase decision: “Why should I pay you tens of thousands or hundreds of thousands a year when I can just have Claude Code write the code for me?” Traditional SaaS may not disappear, but pricing will remain anchored to the cost of internal development.
曼琪 summarized the shift as “the engineer dividend spreading from China to the rest of the world” and “the globalization of the Chinese software market.” Henry expects the 80%-90% gross margins once treated as a given to become difficult to sustain as the entire industry enters hard mode.
10. OpenClaw is a beacon, not the destination; major companies are now moving fully in its direction
Henry used Devin as a comparison. Devin’s first version was not necessarily useful, but it established the compelling target of an “AI software engineer.” Individual developers and startups have less historical baggage and can often “jump slightly farther,” illuminating the direction first.
Henry characterizes NVIDIA’s NemoClaw roadmap as “OpenClaw plus an enterprise security layer”: an OpenShell sandbox, YAML permission files, and audit trails restrict the data, APIs, and network resources an Agent can access.
Anthropic released roughly 74 features in 52 days, including Telegram, Discord Channels, mobile Remote Control, computer use, scheduled tasks, and Dispatch for coordinating multiple always-on Claude Code instances. Henry says most of these are moving toward OpenClaw’s territory.
OpenAI’s move was more direct: it hired Peter Steinberger. Less than 2 months into his tenure, there is no public product yet, but Henry expects OpenAI will likely combine his ideas with existing products such as ChatGPT Agent Team. OpenAI’s attention, however, still appears scattered.
11. OpenClaw, Claude Cowork, and Claude Code are approaching one another from different depths
Henry’s segmentation: Claude Code began in the deep water of professional programmers; OpenClaw is the shallow-water entry point for mainstream users; Claude Cowork is Anthropic’s bridge between the 2.
On out-of-the-box usability and stability, Claude Code and Claude Cowork currently outperform native OpenClaw. OpenClaw has a broader interaction surface, but installation is complicated and memory can fail after a week; the simplified cloud version, meanwhile, sacrifices the advantage of local permissions.
The value of local access depends on the workflow. Apple users can give OpenClaw access to the closed iMessage and Notes ecosystem, while much of 曼琪’s document work already lives in Feishu, making a local Agent less convenient. Both therefore see an opportunity in combining all-in-one office platforms with Agents.
12. Anthropic’s growth has given OpenAI a rival of comparable scale and greater product focus for the first time
The show cites different sources and dates: 曼琪 puts Claude Code ARR at roughly $900M in December 2025 and roughly $1.9B in early March 2026; Henry separately puts it at roughly $2.5B ARR. OpenAI rose from roughly $21.4B at the end of 2025 to roughly $25B in February 2026.
Henry estimates that 70%-75% of Anthropic’s revenue comes from B2B and API, while OpenAI relies more heavily on consumer subscriptions and has a more diversified business. Anthropic’s growth is not primarily the result of a model suddenly pulling far ahead; it is the result of Claude Code becoming a killer product.
The show’s Claude Code ARR figure has already surpassed Cursor’s roughly $2B, making it the leading developer product by that measure. Anthropic went from a “technically respected challenger” to an “equal or even stronger rival” in the space of this quarter.
13. Cursor still has a short-term enterprise migration tailwind, but Claude Code is closing its long-term window
Henry’s observed U.S. enterprise migration chain is GitHub Copilot → Cursor → Claude Code or Codex. Conservative organizations such as banks, Citibank, and Capital One still use Copilot heavily, leaving Cursor an incremental market for a product that feels familiar but works better.
The long-term risk is that individual developers are shifting from heavy Cursor usage costing $200 per month to barely opening it at all. Microsoft’s earlier GitHub Copilot price cut is also viewed by Henry as a sign that the product’s competitive position is weakening.
Cursor’s most valuable asset is its enormous volume of high-quality engineering trajectories, which can support continued pre-training and an in-house coding model. But Claude Code’s revenue and usage are catching up so quickly that Cursor’s data moat is being eroded at the same time.
14. Opus 4.6 versus GPT-5.4 is a contest between generalist product sense and specialist coding ability
Opus 4.6 offers one million context and was launched with an emphasis on maintaining goal orientation while working continuously for roughly 15 hours. It leads on Terminal-Bench 2.0 and is Karpathy’s first choice for AutoResearch.
GPT-5.4 has native computer use: it can see the screen, move the mouse, and type. The show cites an OSWorld score of 75%, above the 72.4% human score, and an SWE-bench Pro score of roughly 57.7%. Coding is integrated into the main GPT model.
Real-world feel does not map cleanly onto benchmarks. Many developers consider Codex better at pure code generation, while Opus and Claude Code better understand intent, communicate more effectively, and feel more like an intelligent all-around colleague.
Dylan Patel’s cutting comparison captures the split: Codex is like a “retarded kid super trained on coding,” while Opus “feels like AGI.” 曼琪 rendered the former as “Rain Man for coding.”
15. “Claude Code as master, Codex as slave” exposes OpenAI’s product-layer weakness
A popular workflow is to write a skill inside Claude Code: Opus or Claude Code handles thinking, planning, and communication, then calls Codex exec for code generation or review. The underlying executor can be swapped out; the top-level “master” with the direct user relationship is harder to replace.
Henry says the OpenAI team is aware of the problem: “Claude Code is the master, Codex is the slave doing the work.” Even if Codex writes better code, user traction and the complete workflow still accrue first to Anthropic.
The core of a coding Agent has shifted from one-shot generation to the full engineering loop: reading code, fixing bugs, running tests, inspecting logs, repairing failures, and opening a PR. Henry compares Anthropic’s advantage to Apple’s ecosystem: the hardware may not top every benchmark, but the overall experience is more coherent.
16. Anthropic’s focus is an advantage; OpenAI’s counterattack depends on cutting the side quests
Henry says OpenAI’s research team remains exceptionally strong, with additional advantages in its larger consumer base, ChatGPT distribution, and deeper financial resources. The problem is simultaneous investment in ChatGPT, the API, enterprise verticals, Sora, and OpenAI for Science, leaving attention too diffuse.
Application chief Fidji reportedly said at an all-hands meeting: “We cannot miss this moment because we are distracted by side quests.” Directions such as the Atlas browser, advertising, and e-commerce integrations may be scaled back as resources return to coding, productivity, and enterprise services.
Peter’s arrival at OpenAI may also signal an effort to reinforce product capability. Sam Altman is jokingly described as an excellent “portfolio manager”; if he identifies the key inflection point and reallocates resources, OpenAI retains substantial counterattack capacity.
Anthropic’s risks are precisely that counterattack and Google’s distribution. Google is embedding AI into Chrome, Gmail, Docs, Drive, Sheets, and Slides, combining product touchpoints, data, and TPUs into a different kind of platform advantage.
17. xAI bled talent this quarter, while Google continued quietly laying distribution rails
Henry called xAI’s 26Q1 “a disastrous quarter.” Several important co-founders, including 张国栋 and 戴子航, left; many moved to other labs, while others entered stealth-mode startups.
His explanation is speculative: OpenAI released a new model in December 2025 with stronger coding, while Grok’s training struggled, prompting Elon Musk to pressure the team and triggering a wave of departures. xAI is trying to replenish staff through poaching and acquisitions, but integrating a new team will slow the catch-up effort.
On the model side, Google’s Gemini 3.1 and other releases have been incremental, while generative models and product integrations such as Jini and Leo continue to advance. Henry’s view is that Google may be “holding back a big move,” but its clearer advantage is not any single model capability—it is ubiquitous distribution.
18. Model scores are converging; the real competition is shifting to ecosystems, trust, cost, and the next paradigm
As the benchmark gap between GPT-5.4, Opus 4.6, and Gemini 3.1 narrows, scores alone no longer provide enough differentiation. Developer ecosystems, enterprise security and compliance, reliability, and inference cost are becoming the selection criteria.
“Coding agents are general agents” also explains why the entire industry is going all in on coding. Chinese models such as Kimi and MiniMax are strengthening coding, one million context, and long-horizon capabilities because a personal assistant first has to be able to operate in the digital world.
Henry emphasizes that benchmark convergence does not mean comparable real-world experience. Models still differ materially in their ability to execute for long periods, understand intent, call tools, and recover after failure.
The next variable capable of genuinely resetting the competitive landscape may be continual learning or recursive self-improvement. Whoever breaks through the next paradigm first could leapfrog the current product ecosystems and retake the lead.
19. Model vendors must move up the stack because user behavior has replaced benchmarks as the optimization target
Henry compares the situation to Cisco being commoditized: vendors that remain at the model layer may become undifferentiated infrastructure businesses unable to earn excess returns over time. That is why model companies are vertically integrating models, developer tools, and end-user applications.
Products are not only commercialization channels; they are training guides. When benchmarks cannot fully represent the experience, real user tasks, failure points, and trajectories tell teams which direction to optimize.
OfficeQA is a counterintuitive example. Faced with complex PDFs, spreadsheets, and charts, the model’s reasoning is already strong; more failures occur in file parsing and retrieval. In other words, “intelligence has overflowed,” leaving the other parts of the workflow as the bottleneck.
Signals that a user task is genuinely near the capability frontier are sparse. Vendors filter for difficult, high-quality trajectories, have a stronger model redo them, and use the results for continued pre-training or distillation. An in-house coding product therefore generates both revenue and training data.
20. Cursor’s Kimi K2.5 episode shows that Chinese open-source models have reached the core of leading U.S. products
Cursor’s new model was found to be fine-tuned from Kimi K2.5, but the company initially failed to disclose this clearly, contrary to industry practice. Henry first treated the episode as validation: a leading U.S. Agent company was willing to put a Chinese open-source base model into its flagship product, showing that the model had cleared the threshold of real-world engineering quality.
He was less forgiving of Cursor’s execution. A company reportedly raising at roughly a $55B valuation made basic mistakes in licensing, compliance, and its launch PR, proving that “sometimes the world really is a shack.”
Henry believes the motive for concealment may not have been primarily geopolitical. Cursor may have feared questions about its own research depth: if it had only added a layer of RL on top of a Chinese base model, its model moat would look less substantial.
Kimi’s public response remained positive, while Cursor later acknowledged the mistake and explained its model selection and post-training process. The 2 sides may continue discussing collaboration or revenue sharing. 曼琪’s joke: “Cursor is a San Francisco kid—it has to save face.”
21. AutoResearch turns the “never-tiring PhD student” into a reproducible process
Karpathy began with NanoChat, using several hundred lines of Python to demonstrate the training process from pre-training through RL and produce a GPT-2-level small model. It was both a teaching project and a way to force himself to understand the core problem through the simplest possible implementation.
AutoResearch has the Agent read a human-written program.md, the training code, and historical results; propose one improvement; modify train.py; train for 5 minutes; evaluate using bits per byte; and repeat.
Each complete iteration takes roughly 15 minutes, allowing 4 runs per hour and nearly 100 overnight. Hundreds of experiments found more than 20 effective improvements, cutting training from more than 2 hours to roughly 1 hour—a gain of approximately 17%-20%.
曼琪 pressed on the boundary of autonomy: humans still define program.md and the objective. Henry acknowledges that the process can run unattended, but the current division of labor is humans choosing the search space and AI searching it efficiently—not AI defining its own research direction.
22. Self-evolution suddenly looks credible because multiple experiments produced positive feedback within months
Henry says Google had already been exploring an internal Model Lineage Map: model versions are nodes, improvements are edges, and the goal is for AI to predict the next research path most worth pursuing. A year ago, coding ability may still have been the limiting factor.
Current evidence includes AlphaEvolve using Gemini to optimize Gemini’s own kernels and improve performance by roughly 23%; GPT-5.3 Codex helping debug itself; MiniMax M2.7 emphasizing self evolution; and AutoResearch. Together, the dream is beginning to receive engineering validation.
Henry calls it “the acceleration of acceleration.” Model progress is already speeding up; if the research process itself is automated, the acceleration can compound again. But when the objective is vague or the editable surface is too broad, an Agent may “ruin everything and then be unable to fix it.”
23. Autonomous optimization works best with objective metrics and fast feedback; subjective experience can still break the loop
Shopify CEO Tobi applied a similar method to a template engine. The model made 93 automated commits and improved rendering speed by roughly 53%. The example works because speed is measurable and every change can be verified quickly.
Henry believes performance optimization, kernels, database queries, and infrastructure tuning are the best early targets for automation: the goal is clear, feedback is fast, success can be judged automatically, and failures can be confined to an experimental environment.
The counterexample is his roommate’s attempt to add dark mode to a website. Devin and Claude Code each ran for more than an hour without finishing. Whether buttons and text feel comfortable in a dark interface requires visual and subjective judgment, and the Agent cannot construct a reliable self-verification loop.
Even when the metric is clear, production deployment still has 3 gates: who reviews security, who confirms that the improvement generalizes, and whether humans understand why the change works.
24. Continual learning and self-evolution solve 2 different problems
Continual learning addresses how a deployed model acquires new knowledge, skills, and user preferences. Most current models have fixed weights after deployment and do not naturally improve simply because they interact with users more often.
Skills, memory, and text files are what Henry calls “poor man’s continual learning”: cheap, transparent, and editable by anyone, but limited by context capacity, compression, and forgetting. OpenClaw’s email deletion incident exposed the cost.
Another route is test-time training or localized weight updates similar to LoRA. Rather than training only on an offline batch, the model could adjust a small number of parameters during the inference lifecycle in response to a specific user’s inputs.
The obstacles are concrete. Updating during every inference increases infrastructure complexity; current validation is mostly limited to models with billions of parameters; and it remains unclear whether the approach can scale to models with hundreds of billions or more than 1T parameters, or whether maintaining personalized weights for every user is economically worthwhile.
25. New startups are betting on learning small models, multi-Agent research, and world models
Henry sees the Tinker and Multi-LoRA infrastructure from Thinking Machines Lab as potential building blocks for personalization at scale. Whether everyone should have their own model ultimately depends on whether the value of customization exceeds training and serving costs.
One stealth startup wants to extract a small number of MoE experts dedicated specifically to “learning,” discard the large mass of parameters that store knowledge, and build a small model exceptionally good at understanding its owner’s preferences—a base model for OpenClaw-style personal assistants.
Another group of companies is expanding AutoResearch from a single Agent into parallel multi-Agent collaboration. Researchers may shift from executing experiments themselves to defining higher-level search spaces and research directions.
World models split into 3D world modeling, the JEPA approach, policies that predict actions from video, and environment simulators that generate the next state from a video and an action. The main applications are interactive media generated in real time, along with robot policies and simulation environments.
26. Inference has replaced training as the infrastructure theme, and Agents will create new CPU demand
Henry’s summary of GTC was “the shift from training to inference.” The Vera Rubin platform could deliver roughly 3-5x higher inference performance and potentially cut token costs 10x, directly benefiting personal assistants and enterprise Agents with high token consumption.
NVIDIA integrated the LPU from Groq, acquired last December, into Vera Rubin. Google also demonstrated quantization that compresses KV Cache and reduces storage requirements to roughly one-sixth of current levels. Henry concludes that inference algorithms still have substantial room to lower costs.
Agents do not only call GPUs: “Everything is becoming a computer.” They also need a persistent machine, a sandbox, and the ability to run generated code, so CPUs and the infrastructure around inference will expand with call volume.
Some investors are beginning to seek relatively pure CPU exposure. The interim conclusion Henry heard was to buy Arm, which licenses its architecture to different chipmakers and may capture demand more directly than companies operating across multiple businesses.
27. AI layoffs are translating technical efficiency into smaller teams, higher talent density, and sharper social polarization
Henry’s 3 U.S. data points: Amazon cut roughly 16,000 people early this year; Block announced a 40% reduction and said most companies may follow within a year; Meta first had all employees stop work to learn AI and measured efficiency through token consumption, then planned to cut roughly 20%, or 15,000 people, and direct the savings toward roughly $65B in AI capex.
The metric was immediately gamed. Someone Henry knows had Claude Code run meaningless loops to inflate token usage. He joked that of Anthropic’s recent several billion dollars in revenue growth, “possibly tens of millions came from Meta.”
Startups’ organizational answer is not to stop hiring, but to shift from quantity to quality. Growth remains more important than labor efficiency during the current window, but “5 great people plus some Agents” may already be enough. AI-native ability is also becoming a priority in interviews for every function.
曼琪 sees the Chinese version as a move from “top-tier talent plus second- and third-tier talent” to “superstar talent plus Agents.” A lower headcount does not imply a linear decline in labor costs, because companies may concentrate budgets on people who ask better questions, use tools effectively, and judge direction.
28. AI is not only replacing programmers; it is redrawing the boundaries of professional services and solo businesses
Henry cited a Claude-native law firm. A client demanded that multiple groups of clauses be revised the same day and delivered the following morning—a job that could have forced a traditional team to work overnight. AI first identified contradictions in the client’s instructions, then separated acceptable requests from issues that required pushback.
2 or 3 lawyers completed the new contract and explanatory documents in 2 or 3 hours. What was compressed was document analysis, conflict identification, and drafting time; relationships, negotiation, and professional liability remained with the lawyers. 曼琪’s conclusion: “Work involving people will be the safest in the future.”
Henry’s personal path through layoffs is entrepreneurship or self-employment. AI can fill gaps in one person’s incomplete skill set, giving people who previously could not build a product independently the chance to run a business that supports them.
29. Distribution of the gains may be harder than the technology; housing prices and token taxes are 2 sides of the same problem
Bay Area housing entered the discussion because AI is creating 2 groups at once. Ordinary engineers face unemployment and mortgage pressure, potentially weighing on $1M-$3M homes; top researchers with enormous packages who are waiting for their companies to go public may continue pushing up the prices of scarce homes worth more than $10M.
Henry called the picture “very cyberpunk”: AI could highly centrally amplify people who own compute, resources, and elite capabilities, widening income and asset inequality.
Beyond UBI, one proposal he heard was a token tax: companies should not be allowed to build unlimited data centers, increase inference, and eliminate workers without bearing a social cost. 曼琪 says this would make AI more like an employee; Henry compares it to imposing tariffs on AI that “takes human jobs but pays no taxes.”
30. Next quarter’s watch items are DeepSeek V4 and computer use; the biggest unknown remains continual learning
Henry expects DeepSeek V4 to launch next quarter, though 曼琪 cautions that expecting 2 consecutive “world-shocking” releases may be too much. Henry also acknowledges that after recent work on mHC and Engram, along with ByteDance Seed’s progress, DeepSeek’s lead in foundational research may have narrowed.
Computer use still shows a huge gap. GPT-5.4 has reached 75% on OSWorld, but the latest model on the CUB benchmark cited in another episode remains only in the low teens. A breakthrough in cross-application, API-free operation would immediately create a large set of enterprise automation use cases.
The most important and least certain question is the speed at which continual learning and recursive self-improvement reach production. Some researchers view current test-time training approaches as “over-engineering”; the industry is still waiting for a mechanism as simple as skills or memory, but more reliable at learning without forgetting.