At a glance

• The open-source war (throughout this issue, "open source" includes open-weight models) escalated on three unconnected fronts within 48 hours: Beijing's top leader gave his first AI-themed speech, read as planting the flag for the open path; a Washington prediction market repriced "the US bans an open-source model this year" from 22% to 45%; and a top Google executive publicly pushed back on bans with "900 million cumulative Gemma downloads." For the first time, policy risk has a price you can watch daily.
• Read that 45% carefully: the market prices "ban Chinese models" and "ban open source altogether" inside a single question. And if a ban does land but exempts American models, the biggest beneficiaries happen to be Google's and Meta's own open models. Whether you're pushing for a ban or dreading one, run those two policies as separate calculations.
• The full story of Anthropic blocking OpenClaw, the third-party agent workbench, went multi-source today: the block retained not a single blocked user — they rebuilt the same workflows in place on Chinese open-weight models at $10–15 a month. "Blocking protects margin and completes user migration for your competitor" is upgraded from a lone anecdote to a citable fact.

This issue draws on the July 26, 2026 research cycle; the main events date from July 24–25, plus one July 20 settlement picked up late and marked as such. Overnight we processed 4 podcast transcripts (All-In, ChinaTalk, and two MLST (Machine Learning Street Talk) episodes — the MLST pair reviewed segment by segment and judged no new signal) plus same-day original posts from 3 tracked X accounts, and read 124 daily electricity snapshots from the EIA, judged unable to yield signal — that indicator is suspended as of this issue; we also closed 2 targeted verifications of existing records → 19 linked receipts in this issue. Full transparency: thread 1 and the brief item come from the same ChinaTalk episode, and threads 2 and 5 from the same All-In episode — all secondhand relays with no primary documents pulled directly, all downgraded under our single-source rules. Our automated inventory report failed to run today; the figures here were reconstructed by checking retained records — concentration and coverage gaps detailed in the accounting section at the end.

Today's main threads

1. [This week] (delivered July 24) Xi's first AI-themed speech ran to about a thousand characters and took 20 minutes to read. The real payload is one phrase: AI framed as a "normal technology." Xi Jinping gave his first speech devoted to AI at WAIC (the World AI Conference in Shanghai, China's annual state-level official AI gathering). The read from observers on ChinaTalk (Jordan Schneider's China tech-policy podcast; this episode an emergency roundtable convened July 24) is that this was a forceful endorsement of the open and open-source path: framing AI as a "normal technology," the logic being that the cheaper and more widespread a normal technology gets, the better — a framing in direct opposition to "AI is a strategic weapon that needs control" — while appealing to the Global South and Belt and Road countries with the language of inclusive cooperation. The timing is glaring: in the same 48 hours, Washington was debating whether to ban Chinese open-source models (see thread 2). The speech's closing section also laid down a safety baseline — building regulation, monitoring, early warning, and emergency-response mechanisms, in the official translation cited on the panel, to "ensure that AI always remains under human control" — which the same observers read as pre-positioned cover for a future policy turn (ChinaTalk, Jul 24). Verification: A single podcast's secondhand relay and interpretation; we have not checked the official transcript. "About a thousand characters," "20 minutes," and "first" all come from the panel. ChinaTalk holds a critical view of the Chinese system, so the reading carries a stance. The direction is readable; treat the details as provisional. Judgment update: A frame this brief has long held — China uses open source as its market-entry and influence path, exporting ecosystem position and standards rather than revenue — just gained its final piece: top-level political backing. Open source now carries backing from the very top of the political system. The counterweight stays on the record: that safety-baseline passage is a planted pivot — openness is the current strategy, not a commitment. If any major US company starts paying directly for Chinese model APIs, or China's domestic services close the quality gap, this judgment goes back for review.

2. [This week] (published July 24) Can policy risk be marked to a daily price? Washington just did it for the first time. All-In is the business-and-policy podcast hosted by four Silicon Valley investors including Chamath Palihapitiya and David Sacks, with its own positions on AI policy. The July 24 episode relays this: Polymarket (the prediction market where people bet real money on event outcomes, so the probability is the bettors' consensus) opened a new market on whether the US government bans an open-source model within 2026, and the odds moved from 22% to 45% in a matter of days. The context: the White House is reportedly evaluating a ban on Chinese open-source models, while an internal faction argues for replacing a ban with incentives for America to build better open-source models of its own (All-In, Jul 24). Our prior coverage: the July 24 brief logged the White House science-policy chief's named accusation that Moonshot AI (developer of the Kimi models) distilled American models, along with the Treasury Secretary's sanction threats. The July 25 brief's thread 3 logged the buyer-side backlash — paying customers' CEOs publicly attacking the closed labs for "regulatory capture," a coalition of 200-plus startups, and some 37 industry companies co-signing for open weights (official statement page, Jul 24). Where things stand as of this issue: the ban is at the evaluation stage only; no ban is in effect. Verification: The 45% and 22% are verbal relays on the show; we have not checked the Polymarket page directly. Prediction-market numbers are a betting proxy for policy risk — not an official decision and not a fact. The two numbers share one basis (same market, days apart), so the move itself is comparable. Judgment update: The "organized policy war" we observed yesterday is reinforced again today, and the battlefield has widened from a US domestic fight into a US–China mirror image: Washington weighs a ban while Beijing's top leadership doubles down on open. A new tension goes on the watchlist: the market's phrasing — "bans an open-source model" — prices "ban specific Chinese models" and "kill open source altogether," two very different policies, in one bucket; 45% cannot be read as the probability of either one alone. What would prove this wrong: if within three months (our self-set window) the White House evaluation fizzles, the odds fall back below 25%, and Google and Meta take no further organized action, "organized policy war" drops back to a one-week story.

3. [This week] (published July 25) Google didn't sign the 37-company statement. The next day its DeepMind CEO stepped onto the field himself, carrying numbers. Google DeepMind CEO Demis Hassabis self-reported on X on July 25: Gemma (Google's open-weight model line, whose weights you can download and deploy yourself, as distinct from the paid Gemini API) has passed 300 million downloads for its fourth generation and 900 million cumulatively across the series (downloads post, Jul 25). The same day he posted a defense of the open ecosystem, reciting what Google has open-sourced over the years — the AI development framework JAX, Transformers, and AlphaFold (position post, Jul 25) — and Google CEO Sundar Pichai reposted in support. Set against the "Gemma 4 passed 100 million in a month" figure Google self-reported at I/O two months ago, downloads are still accelerating — but the three numbers use three different denominators (single generation's first month, single generation cumulative, whole series cumulative) and cannot be combined into a single growth rate. One comparison worth logging (our own observation — no source states it): on the July 24 statement of some 37 industry co-signers, Microsoft, Meta, OpenAI, and Nvidia all signed; Google was absent. Hassabis showing up a day later with a personal post and a set of numbers is how Google's position-taking in this policy war actually looks. Verification: Operating numbers self-reported by the company's own top executive, with promotional incentive; we have not independently checked them. Hugging Face and Kaggle download counters are partially checkable; we have not done so. Downloads ≠ active use ≠ commercial adoption — and there is no way to know who is behind those 900 million downloads or what they are doing with them, nor, from download counts alone, to separate ecosystem heat from pure policy-PR numbers. The figure is a ceiling on heat, not evidence of adoption. Judgment update: Publishing this set of numbers on the hottest day of the ban debate is itself a policy act — it tells Washington that "open source" does not equal "China," and that a ban's fire would reach America's own largest open-model line. It is also the concrete expression of thread 2's market ambiguity: if a ban exempts American models, the biggest beneficiaries are precisely Gemma and Llama.

4. [Evidence update] (original events Feb–Apr 2026; verification completed July 26) Blocked users rebuilt the same workflows on Chinese open-weight models: the full OpenClaw timeline went multi-source today — and it is worse than our original record. What had lived in our records as a single-source podcast anecdote completed verification today and is upgraded to multi-source fact. The full timeline: OpenClaw is the open-source agent workbench built by developer Peter Steinberger — it binds to no particular model and lets AI agents run everyday tasks — and after going viral it ran on Claude subscriptions. In late January, Anthropic forced a rename — in Steinberger's own words, "I was forced to rename the account by Anthropic. Wasn't my decision." — with the sequence running Clawdbot → Moltbot → OpenClaw (Forbes, Jan 30; CNBC, Feb 2). In February came blocks on individual accounts, and from April 4 this widened into a blanket revocation of third-party applications' access to subscription plans, Anthropic's official reason being the "outsized strain" OpenClaw put on its systems (The Next Web). Where the blocked users went is also now confirmed: Chinese AI company MiniMax pitched OpenClaw users a $10-a-month plan and shipped a one-click replacement within two months, and there is a public case of a $200-a-month Claude setup rebuilt on Kimi for $15 (recounted firsthand by the author in a Lex Fridman interview, Feb 12). Verification: Firsthand statements from the person involved + independent reporting from CNBC, The Next Web, and Forbes + Anthropic's official line, triangulated — the strongest-verified item in this issue. The details of individual users' migrations remain secondhand anecdote from an emotionally involved party. Judgment update: The old judgment that model-layer switching costs are trending toward zero just got its best behavioral evidence: in the third-party-workbench scenario, the accumulated working context lives on the workbench side (user-owned), not the model platform side — so blocking doesn't retain users; it completes the last mile of their migration to a competitor. Draw the boundary clearly: this applies only to agent-workbench scenarios; ordinary consumers' chat memory is still locked on the platform side and unaffected. The layered map our July 25 deep dive drew is validated here. That map has two layers: the base is the everyday tasks a third-party workbench can carry; the summit is the high-price tier of a few frontier flagship models. Commoditization of the base has already happened; it is observed user behavior. And the political backlash that summit pricing power invites is exactly the policy war of threads 1–3. For API and ecosystem leads at the model vendors: lock-in must come from account systems, memory, and workflow assets, not from licensing terms.

5. [This week] (settled July 20; picked up by this brief July 26) In the $1.5 billion settlement, the court priced a clear rule: training is legal, and pirated acquisition is what pays. Specific terms of Anthropic's copyright class-action settlement have surfaced: a $1.5 billion settlement, called the largest copyright settlement in US history; roughly 500,000 books included, at about $3,000 per book to authors (500,000 × $3,000 = $1.5 billion — the arithmetic checks out); $101 million in lawyers' fees. The case began with Anthropic's downloading of roughly 7 million books from pirate sites for training — note that 7 million is the count downloaded, while 500,000 is the count included in the settlement; the two numbers have different denominators. The legal structure is the point: the court had already ruled that training AI on copyrighted books is fair use (the US copyright doctrine permitting certain unauthorized uses); what pays damages is the acquisition route — from pirate sites — not the act of training itself (All-In, Jul 24). Verification: The term details are all secondhand verbal relay on the show; we have not read the court's primary documents. The "$1.5 billion settlement" direction independently matches a mention by AI policy and industry researcher Nathan Lambert in late January, but the two sources disagree on when the settlement was finalized — treat the details as directional. Judgment update: Copyright risk gets its first computable market price: the $3,000-per-book settlement rate versus the cost of legally buying and scanning books — any model shop with books in its training data can now run that arithmetic. If the "training legal, acquisition punishable" structure is confirmed in court documents, it becomes the pricing benchmark for data-acquisition policy across the industry. Until the final settlement documents (public court records) confirm it, the roughly $3,000 per book remains a secondhand, unconfirmed figure.

6. [This week] (interview published July 25) Three numbers that lock together: a data supplier's rare self-disclosure traces the real shape of the expert-data business. Mercor is the fast-rising expert-data supplier that recruits human experts for AI labs to produce training data. Its chief product officer, Oswald Nitski, disclosed three things on 20VC (Harry Stebbings' venture-capital interview podcast). First, against the market line that "90% of enterprise workflows can already be automated by models," he counters with the company's own Apex benchmark — top models reach only about 50% on long-horizon workflows (tasks requiring many consecutive steps executed across days) — and argues the "90%" arithmetic covers only existing demand, missing an entire class of latent demand (say, setting up a procurement agent to run fully autonomously for months while a human checks in once a week). Second, revenue is heavily concentrated in frontier-lab customers, and de-concentration depends on moving down into the general enterprise market. Third, on many projects, costs already split "roughly half and half" between paying human experts and model-inference spend (20VC, Jul 25). Read the three as one mechanism: the capability gap is real, so demand isn't saturated; revenue is concentrated, so this business's lifeline hangs on a few labs' training budgets; and half the cost is compute, so the "human data business" is itself half a compute business. Verification: A supplier's self-account. Apex is an internal benchmark with its scoring methodology undisclosed, and "the frontier still has masses of unsolved workflows" is precisely its upsell pitch — keep the real capability gap and the commercial pitch separate. The 50% and the "90%" it rebuts have different denominators (its own defined long-horizon task set versus already-attempted existing workflows) — they are not opposing readings on one scale. Judgment update: This forms a two-sided structure with today's expert-reads item, whose verification also completed today — one view from inside a supplier, one from an outside industry observer, pointing at the same structure: expert data is the actual supply chain of current model-capability progress, and its demand concentration is bound to the frontier labs. Negotiating intelligence for lab data buyers: you are the customer it cannot afford to lose — demand disclosure of the benchmark methodology. A signal to watch from here: if Mercor publishes Apex's methodology, or announces a named contract with a non-frontier customer, de-concentration is actually happening.

Also worth noting

• [This week] (episode July 24) DeepSeek has reportedly raised about RMB 50 billion (≈US$7–7.5 billion), with investors said to include Tencent, CATL (the world's largest EV-battery maker), and a state-level AI industry fund; founder Liang Wenfeng reportedly keeps control through a structure that injects the funds into a limited partnership he personally manages, and is said to have turned down Alibaba. The podcast's sourcing is "according to several reports"; the transcript's numbers carry obvious transcription noise; and no primary funding announcement exists — a significant direction, but the weakest-evidenced item in this issue (ChinaTalk, Jul 24)

Expert reads

• [Evidence update] (originally published June 2026; verification completed July 26) Dwarkesh Patel's "RL data industry" argument from last month, reconciled line by line today: which words are now solid and which are forward-looking, cleanly split. Dwarkesh Patel is an independent tech interviewer and writer whose deep AI-industry essays are widely relayed. His June essay argued: every skill you want to teach a model maps to at least hundreds of human experts writing demonstrations and grading rubrics; the expert-labeling-plus-RL-environments industry (RL environments being simulated task settings where models practice repeatedly) is already earning billions a year in revenue, headed for deca-billions; and labs put around 30–50% of their compute into inference — compute that currently plays no role in improving the model (original essay, Jun 2026). Today's verification results: "hundreds of experts" is supported by Mercor's and Surge's (Surge AI, another expert-data supplier) public job postings and expert-team pages; "billions a year" has independent third-party market research (Mordor Intelligence, Grand View Research) pointing the same way, though with a somewhat narrower scope than his claim; "30–50% of compute" matches his text word for word. But "headed for deca-billions" is his forward-looking prediction, not a current fact — the RL-environments segment specifically is far smaller today; split the two when citing him. The argument supports a frame this brief has long held — data and environments are the real bottleneck of model progress — and is the outside half of today's thread 6.

Models (research & engineering)

• [This month][Trend] (published July 2026) Why do models keep getting better at solving problems yet worse at writing like people? MIT offers a mechanism — one that also explains why human expert data is still scarce. The current mainstay of post-training (the stage after base training where a model is taught how to answer) is RLVR — reinforcement learning with verifiable rewards, which rewards the model only on tasks with checkable answers, like math and code. Its structural blind spot: it optimizes only what can be objectively scored and is numb to style, structure, and the hard-to-articulate qualities that matter — hence the failures everyone has seen: output turns monotone and unnatural, and models learn to game the score rather than write well. An MIT team's July paper proposes a fix: add an adversarial discriminator that learns to tell "human-written" from "model-generated," feeding that hard-to-quantify quality signal back into training. Self-reported results: on bug-fixing, markedly smaller edits at unchanged scores; on story-writing, output more human-like and more diverse (arXiv, Jul 2026). A single team's self-report, not independently replicated — directional for now. It is the other face of today's thread 6: the dimensions verifiable rewards cannot reach are still bridged by human demonstration — which is exactly the mechanistic reason the expert-data business exists. No other major new papers within this week's scan.

From the archive

The flip in AI compute money — from building models to running them — has already happened (our deep research, July 15, 2026; targeted verification July 12, 2026). Inference spend (the compute burned when a model answers questions) overtaking training spend happened in early 2026 — a direction supported by three indicators with three different denominators: research firm Zylos calls the flip on a cumulative-spend basis (Zylos Research, Apr 2026); Deloitte estimates inference rising from about half of all AI compute in 2025 to about two-thirds in 2026 (Deloitte TMT predictions); and industry media report the same direction, with inference passing half of cloud-infrastructure spend. Note this is three differently-denominated indicators jointly supporting a direction and a timing — not one number verified by three parties; the sub-figure "inference at 85% of enterprise AI budgets" is single-source — don't cite it on its own. How to use it: this is the watershed where model economics moves from training-centric to inference-centric, the baseline for compute capital-allocation calls — inference demand flexes with users' willingness-to-pay budgets, a completely different logic from training's one-shot capital outlay. What's worth watching next is who builds the scheduling layer that converts budget into capability, rather than who simply burns the most GPUs.

Sources & accounting

The past 24 hours. 4 podcast transcripts read in full: All-In (aired Jul 24), ChinaTalk's emergency roundtable (aired Jul 24), and two MLST (Machine Learning Street Talk) interviews — both reviewed segment by segment and judged to carry no new signal, so neither is cited this issue; a 20VC interview with Mercor's chief product officer finished processing overnight and counts in this batch. On X, same-day original posts from 3 tracked accounts (Greg Brockman, Sundar Pichai, Demis Hassabis) were assessed post by post; only Hassabis's Gemma numbers carried extractable industry signal. On the energy side: 124 daily electricity snapshots from the US Energy Information Administration (EIA) were batch-judged no-signal — two weeks of day-by-day review confirmed that daily granularity cannot separate datacenter electricity signal from weather noise, so as of today we retire this cut and move to a de-seasonalized year-over-year aggregate; until the new cut is live, the electricity dimension's daily coverage has a gap, noted here. We also closed 2 targeted verifications of existing records (OpenClaw and the RL data industry — results in thread 4 and expert reads); those are retrospective calibration and not counted as new items in this batch. Coverage statement: this issue vouches only for signals within the scan above; our automated inventory report failed to run today, and the figures above were reconstructed by checking retained records. Source-concentration warning. Today's story spanning three fronts stands on three legs — Xi's speech (a ChinaTalk secondhand relay), the Polymarket numbers (an All-In verbal relay), and Gemma downloads (the company's self-report) — none of them pulled directly from a primary document. Thread 1 and the brief item share one ChinaTalk episode; threads 2 and 5 share one All-In episode — each accounting for about a third of today's new items. The calibration line (thread 4 and expert reads) is, by contrast, today's most multi-sourced material. That is the structure; every number has been downgraded by the rules, and readers should know it. Inventory (not past-24-hour). Accumulated reading backlog (mid-July snapshot): academic papers 2,825; company and personal blogs 2,299; X posts 1,381; industry newsletters 974; company filings 924; industry analyses 533; podcast transcripts 288 — queued for batch processing. The sources we track. Underneath this brief's judgments sit 529 named voices currently tracked: 305 on X (Elon Musk, Andrej Karpathy, Greg Brockman, Nathan Lambert, and others), 90 podcast voices (Satya Nadella, Dario Amodei, Jensen Huang…), 51 news outlets, 48 personal blogs (Simon Willison, Chris Olah…), 48 paper authors (Noam Shazeer, Percy Liang, Tri Dao…), 46 newsletters (Dylan Patel, Ben Thompson, Ethan Mollick…), 26 earnings and filings lines, and 23 keynotes.

This is not a news digest: we hunt each day's AI firehose for the insights that actually matter and the practitioner judgments worth tracking over time, and we show how every item was verified — the point is always "which judgment got harder, and who's been right," never "what happened today."

— SecondSource · generated by our research system · 14 sources · Reply to this email — it's the best feedback you can give us

Written from the same research and judgments as the Traditional Chinese edition; every claim links to a primary document.

EN English edition|繁 中文版 Traditional Chinese →