Daily Brief SecondSource Morning Brief · September 11, 2026 · Sep 11, 2026
This issue arrived about 3 hours later than usual today — apologies for the delay.
1. Oracle took in $11.36 billion of customer prepayments; Broadcom has guaranteed up to $29 billion of a customer's lease payments. Neither instrument trips the circular-financing alarm the market already owns.
2. Oracle's contracted-but-unbooked obligations (remaining performance obligations) reached $664 billion — yet read against the prior quarter rather than the prior year, this is the smallest addition in four quarters. TSMC's monthly revenue was still accelerating the same day.
3. The outside chain-of-thought study of OpenAI's GPT-6 Astra we wrote up yesterday now has a name attached: Neel Nanda. In the same breath, the count of independent measurers behind it drops from two to one.
This issue draws on the research report written in the small hours of September 11; the material spans September 10 to 11. Last night's sweep covered 311 pieces, of which 15 clickable receipts made it into this issue. This is the email edition; the full edition of this issue is the archive of record.
On September 10 two companies filed with the US Securities and Exchange Commission. Oracle filed an 8-K, the form a public company uses to report a material event as it happens. Broadcom filed a 10-Q, its quarterly report. Different businesses, different customers, and two different forms filed — and the thing each one disclosed is the same thing: the buyer does not have enough money on its balance sheet to buy the compute it has already committed to, and the seller has engineered a way to conjure that money for the buyer.
Oracle's instrument is getting the customer to pay first. Last quarter it generated $23.103 billion of operating cash flow, up 184% year over year. Of that, $11.363 billion sits on its own line of the cash flow statement, named word for word "Increase in deferred revenues from customer prepayments with significant financing component" — and the same line a year earlier printed an em dash, meaning zero (Oracle 8-K, 2026-09-10). Operating cash flow is the money a business actually collects over a period and does not have to give back. It is harder to dress up than net income, which is why people treat it as the test of whether a business is really earning.
That money has a very clean reconciliation point, and it sits in the same press release. The first table, "Free Cash Flow," uses the operating cash flow line to produce the 184% headline. The second table, "Net Cash Outlay for Capital Expenditures," subtracts the same $11.363 billion, and the footnote gives the reason: the presentation "provides additional information about the cash requirement for funding of our capital expenditures." One sum of money, two characters assigned to it on adjacent pages: operating performance here, construction funding there. Strip it out and operating cash flow grew 44%, not 184%. The base for that 44% is the year-ago operating cash flow — $8,140 million in the filing's own column, reported in millions rather than billions, so convert before the figures line up. Subtract $11.363 billion from $23.103 billion, measure the remainder against that base, and 44% is what comes out. Both tables comply, both disclose, and the company split them itself — separating them takes nothing more than reading two pages. Two numbers for scale: Oracle spent $28.5 billion on data centers this quarter, so the prepayment covers about four-tenths of it; free cash flow for the same quarter was negative $5.396 billion, a different line on the same table that nets out the construction spending. Separately, the company raised $19.9 billion net by drip-feeding new shares into the market — a second channel, alongside the prepayment, for funding construction with money it has not yet earned.
Broadcom's side has more layers, and they come off one at a time. The first sits in the recent-developments section of the quarterly report, word for word: the AI XPV platform "bridges the gap between the current cash flows of the leading frontier AI labs and significant upfront investments required for their businesses." XPU is Broadcom's word for a custom AI accelerator, as opposed to Nvidia's general-purpose graphics processors; AI XPV is the name Broadcom gave this financing arrangement. The structure has an outside financier putting up $35 billion to buy racks and lease them to the customer, with Broadcom supplying only the technology — which makes it sound as if the risk sits somewhere other than Broadcom. The second layer is in Note 10: if the customer defaults, Broadcom pays, and the formula is 85% of the amount owed less whatever the racks fetch when sold. Maximum liability once all racks are deployed runs to roughly $29 billion. The third layer is the next paragraph of the same section: the customer may, "under certain circumstances and if needed," issue Broadcom convertible promissory notes up to $42 billion, and the terms confine that money to the customer's obligations under the lease. None had been issued as of the quarter end (Broadcom 10-Q, 2026-09-10). A guarantee of this kind means "if my customer cannot pay you, I will" — it is not a loan, it does not show up as debt in normal times, and it counts only when something goes wrong. A convertible note is an IOU that can later be exchanged for shares on agreed terms. The money left Broadcom's balance sheet; the risk walked back on two legs.
⚠️ There is an exaggeration here we need to rein in ourselves. The $29 billion is a contingent liability, the guarantee described above, and it counts only on default. The $42 billion is a facility that may never be drawn, and if it is drawn, what Broadcom receives is an IOU, an asset. The two point in opposite directions. Adding them together works as an order-of-magnitude comparison and nothing more; it is not a loss ceiling. For that comparison: Broadcom's stockholders' equity this quarter is $99.7 billion, and quarterly operating cash flow is $14.2 billion.
Verification: both legs are word-for-word disclosures in original statutory filings, and we read the documents themselves on the Securities and Exchange Commission's database today rather than anyone's summary of them. ⚠️ We are not claiming any of the following. To begin with, nobody concealed anything or disclosed improperly: booking customer prepayments in operating cash flow is what the accounting standards require, not a choice the company made, and Broadcom's contingent liability belongs in Note 10 by rule. Every term in both cases is in the documents, and the companies broke them out themselves. Then, the optimistic and pessimistic passages in Broadcom's quarterly report do not contradict each other: a risk-factors section is written to describe worst cases, as law and convention require. Only a narrower sentence stands: the premise used to support "the guarantee is unlikely to be triggered" — a strong profitability trajectory at the frontier labs — is precisely the thing the risk-factors section calls uncertain. Beyond that, this document cannot tell you who Broadcom's customer is: the original says "our customer" throughout and names nobody, and the identification circulating in the market rests on a single media source. Here is why that matters. Our own records hold another relayed account from March describing a gigawatt-scale deployment at a different laboratory, and merging the two "more than one gigawatt" claims would read two laboratories' risk as one.
⚠️ This item is not a scoop — it is a gap we are closing. Broadcom's platform launched in June, and an investment bank's report flagged the $29 billion guarantee in mid-August. Several outlets covered it, and the stock fell that day (The Motley Fool, 2026-08-17). We missed the whole thread at the time; searching our own records for that platform's name today returns zero hits. What genuinely belongs to today is the $42 billion convertible-note term and the full original wording of the reason the arrangement exists. One more thing, because it will make people cite the wrong number: "Broadcom's $42 billion" now points at two different objects. In August's coverage it was an investment bank's worst-case loss estimate if every customer on the platform defaulted at once. In yesterday's filing it is the ceiling on convertible notes the customer may issue to Broadcom. One is a loss, the other a facility; one was modeled by an outsider, the other written into a contract. We cannot tell whether that is coincidence or whether the bank derived its figure from this very term, so we keep both readings side by side and pick neither. Anyone citing this number next should first establish which one they mean.
Judgment update: the circular-financing red flag the market spent the past year learning to spot has one shape — the supplier takes an equity stake in its customer. It has a name, it has people tracking it, analysts ask about it on earnings calls every quarter, and detection is easy: look for related-party investments. Today's two documents show the same economic content wearing different clothes. Oracle pulls the customer's own money forward; Broadcom pulls a third party's money in, then uses its own credit to take the risk back. Name all three instruments in turn. The first is the supplier taking equity in the customer, the old red flag this paragraph opened with. Customer prepayment is Oracle's leg. Then there is an outside financier buying the equipment and leasing it to the customer while the supplier absorbs the credit risk through guarantees and IOUs, Broadcom's leg. In economic substance all three solve the same problem, but they land in three entirely different places on the balance sheet and answer to three different disclosure regimes. So our call today is this: the thing has not cooled off — it has changed instruments, and the instruments it changed into happen to be ones the existing alarm cannot hear. Anyone using "did the supplier take equity in its customer" as a detector will, from this quarter onward, systematically read structures that ought to be flashing red as clean. One old tool still works, and it is the least formal one. Bill Gurley, a senior venture partner at Benchmark, has kept asking one question in public for years: whether the order would exist at all without the supplier's money behind it (BG2 podcast, 2025-10-14). It asks about economic substance rather than legal form, which is why nothing gets around it.
Investor note: the prevailing narrative assumes the scale of circular financing can be read off public financial statements. This evidence weakens that assumption, and the weakening has a precise location: two structures of the same kind, each running to tens of billions of dollars, both read as clean today under the filter everyone reaches for first.
⚠️ Three rebuttals we found against ourselves, which is why our confidence here sits in the middle. The first is the strongest: these two things may each have their own history and not be substitutes at all. Customer prepayments have been standard practice in cloud and semiconductors for years, and suppliers backstopping leases did not start this year — our own records hold a case from last October with an almost identical structure, where a search giant guaranteed a mid-sized data-center operator's lease payments and took warrants for equity in return. The second is our own material undercutting us: the prepayment leg is not new, because our August 13 issue already carried the same disclosure from a GPU leasing company, where customer prepayments covered 50% to 60% of capital expenditure (our August 13 issue). So what is actually new today is Broadcom's leg, plus the call about the alarm pointing the wrong way. The third: a genuine substitution would show the old instrument declining, and we have not measured that. The more conservative version, which we do have evidence for, is that the new instruments now match the old one in scale while the alarm only covers the old one. We also do not claim these structures will blow up: if the laboratories' profitability tracks the way Broadcom describes, the guarantee never triggers and the prepayments convert to revenue on schedule. Our claim is one thing only — the detection method in use cannot see them.
⚠️ What would prove this wrong: three tests, all answerable. One, if Oracle's customer-prepayment line falls below $3 billion for two consecutive quarters starting next quarter, that would mean this was a one-off contract structure rather than a funding channel, and this leg fails. Two, Broadcom's $29 billion guarantee ceiling should be larger in next quarter's report, since the original says the total rises as racks are delivered and deployed; flat or falling would mean the new tranches changed structure or deliveries stalled, and either would directly rebut "this is the new normal." Three, a third company with a third instrument should appear within two quarters — sale-leaseback, capacity prepurchase, supplier receivables securitization. If none does, this call rests on two samples and should be downgraded to coincidence. Verdict date: around December, when Oracle's next quarterly results and Broadcom's annual report answer the first two together.
In the same Oracle press release, remaining performance obligations rose to $664 billion, framed in the headline as "up $209 billion year-over-year." Remaining performance obligations are contracts already signed that have not yet turned into revenue; cloud vendors use the figure to show that future business is locked in, and the market treats it as the first indicator of whether AI demand is real. Drop today's number into the quarterly series we have kept for a year and the direction reverses: sequential net additions over the past four quarters ran $68 billion, $29.3 billion, $85.4 billion and $26.0 billion. This quarter is the smallest of the four, roughly 70% below the one before it (Oracle 8-K, 2026-09-10).
Why the sequential read carries more information here than the year-over-year one. Remaining performance obligations are a cumulative stock. A year-over-year rate on a cumulative stock adds four quarters together, so three strong quarters will keep the annual figure astronomical even if the most recent quarter stops dead. To know whether the thing is still accelerating now, you have to look at what the latest quarter added on its own.
Verification: two legs. Today's $664 billion comes from the press release; the four prior period-end balances come from the Securities and Exchange Commission's structured financial data interface, a series we started storing a year ago rather than something derived from today's document. A same-basis cross-check holds: $664 billion less $209 billion is $455 billion, against the $455.3 billion period-end value in our series for the end of August 2025. ⚠️ Two things are easy to misread. First, do not read "net addition of $26 billion" as "they only signed $26 billion this quarter": that figure is net, with new signings added and delivered revenue subtracted, and the company says it booked more than $30 billion of AI cloud contracts this quarter, which is the gross number. The cross-quarter comparison is still fair, because all four quarters are net and the ruler never changes, so "smallest of four quarters" holds while "they only signed $26 billion" does not. Second, the same document carries evidence pointing the other way: delivery is plainly accelerating, with the company reporting more than 300,000 GPUs delivered since the end of last quarter, roughly triple the prior quarter's capacity, and cloud infrastructure revenue up 121%. Whatever gets delivered and recognized comes out of remaining performance obligations, so part of the smaller sequential addition may simply be deliveries speeding up.
⚠️ One call we are deliberately not making today: we do not claim Oracle used the year-over-year frame to cover up a sequential slowdown. Reporting a cumulative stock on a year-over-year basis is industry convention, and we quoted the year-over-year figure ourselves in prior quarters. Reading convention as intent requires evidence, and today there is no evidence about intent. What we actually add is the series.
Judgment update: another reading the same day runs the other way. TSMC reported August consolidated revenue of NT$514.81 billion, up 10.1% from July and 53.3% from a year earlier, with January through August up 39.3% (TSMC 6-K, 2026-09-10). TSMC measures what is already scheduled into production and being built; a cloud vendor's remaining performance obligations measure new signings minus deliveries, with the signed order book sitting between the two. So "the furthest upstream is still accelerating" and "net new signings downstream are decelerating" can both be true — and taken together they locate the question somewhere specific: if the slowdown is real, it shows up in new signings first and reaches wafers several quarters later. We are therefore appending a usage note to our existing treatment of remaining performance obligations as a demand thermometer: whenever a company reports a cumulative stock on a year-over-year basis, compute the sequential change yourself before deciding whether to believe the headline. ⚠️ The $664 billion is rounded, while the rest of the series is precise to $0.1 billion, so that $26.0 billion carries about $0.5 billion of error. It does not disturb either "smallest of four quarters" or "roughly 70% lower," but do not run further arithmetic on the 26. Read alongside main-line item 1: two readings out of the same 8-K, one about how cash comes in and one about how orders accumulate, and only together do they give the full shape of this company's quarter.
Investor note: the prevailing narrative assumes the year-over-year growth rate of contracted obligations works as a live temperature reading on AI demand. This evidence weakens that assumption, though what weakens is the reading method rather than demand itself: the same number changes direction when you change the period, and the upstream monthly reading is still climbing.
Yesterday's main-line item 1 and model-watch column both covered an outside measurement of OpenAI's flagship GPT-6 Astra, and both noted that the post carried no byline, which cost it a notch of credibility (our September 10 issue). Today two things move in opposite directions, and we write them separately because they measure different things. What that work tests is this: forbid the model from writing its reasoning out, demand the answer directly, and see how much it still gets right. The string of reasoning a model writes in plain language before answering is called the chain of thought, and reading it is the main way the industry currently catches AI misbehavior. It is readable for a mechanical reason: to reason over a long chain, the model has to spill intermediate results into text and read them back to continue.
What came in is the byline. Checking today, we turned up the attribution: Neel Nanda, a researcher who has published in model interpretability for years. Interpretability is the study of how to see what a model is computing internally, rather than judging it only by what it outputs. Two different sites render the same post, and the attribution matches on both (Alignment Forum, 2026-09-10). So that discount is released: from today this measurement counts as evidence from a named expert.
What came down is the count. Yesterday we listed "obtain UK AISI's original evaluation document on this model" as the most pressing open item, because the two independent measurements we had both reached readers by way of documents from the same company. Today we checked three places. Two targeted searches returned nothing. The evaluation body's own publications index contains no piece mentioning this model at all, and the nearest 2026 item is a July 2 study on compute and capability, which is not about chains of thought either (UK AISI publications index). And going back to the post itself, the author links no document hosted by that body, writing only "the report." That means that leg is not an independently obtainable source. It is a passage reproduced inside an official report.
Verification: all three checks for this item are written out above, and anyone can rerun them. ⚠️ Two lines we do not cross. First, not finding it is not the same as it not existing: an unpublished version may exist, or one buried in that report's appendix, and we claim only that the public index does not show it. Second, this does not weaken the researcher's own measurement: he used his own question set, code and data, which pass through neither that body nor OpenAI. What weakens is only the claim that there were two.
Judgment update: the independent outside measurers on this line drop from two to one. Whenever someone offers "two independent verifiers" as evidence of credibility, ask one question first: did each obtain the raw material, or do both route through the same document? ⚠️ This also overturns a prescription we wrote down ourselves yesterday. Yesterday's boundary note read: wherever a named source has to carry weight, switch to the British body's leg. Today shows that is exactly the leg nobody can find. When we wrote that sentence we believed we were pointing at firmer ground. So the prescription inverts: where a named source has to carry weight, the weight rests on this researcher and his published code, not on that institution.
Investor note: the prevailing narrative assumes frontier labs' safety claims are being checked by independent third parties. This evidence weakens that assumption, and the weakening is very specific: the number of people who can check independently today fell from two to one, and that one is a self-funded individual researcher working from his own question set.
⚠️ What would prove this wrong: the British evaluation body publishes any assessment of this model, or begins publishing on a mandatory basis once it has statutory testing powers, and the count comes back. Verdict date: whenever that body next publishes a frontier-model evaluation.
1. [Today] (published September 10) OpenAI launched a version of ChatGPT built for financial services (OpenAI, 2026-09-10).
2. [Today] (published September 10) OpenAI released a voice model, GPT-Live-1, in its developer interface (OpenAI, 2026-09-10).
3. [Today] (published September 10) Salesforce completed its acquisition of the customer-service AI company Fin (Salesforce, 2026-09-10).
4. [Today] (published September 10) The French model vendor Mistral is partnering with the data platform company Cloudera, pitching enterprises on running sovereign, self-controlled AI on their own data (Mistral, 2026-09-10).
[Today] (published September 10) A company that builds inference chips, and gets filed under "Nvidia challenger," just plugged its next-generation silicon into Nvidia's rack standard. d-Matrix builds chips specialized for AI inference — the stage after training, when a model actually answers questions. It announced that its next-generation Raptor custom accelerators will use Nvidia's NVLink Fusion, connecting its own silicon directly into Nvidia's rack and networking architecture (Nvidia blog, 2026-09-10). NVLink Fusion is the high-bandwidth interface Nvidia opens to other people's chips, letting non-Nvidia processors slot into the standard racks Nvidia has already deployed at scale, with the supply chain and liquid cooling that come with them. Cofounder and chief executive Sid Sheth gave his reason word for word: "Demand for inference is soaring, but capital, time and energy remain finite." Why this one is worth reading: designing your own custom accelerator has been sold as the road away from depending on Nvidia, and this item shows the last mile of that road curving back into Nvidia's racks and networking standard. The design work is what you save; the deployment is what you do not. ⚠️ This is Nvidia's own blog announcement. There is no third-party reading, no shipping schedule, no pricing, and no named-customer adoption — so we take the direction and infer nothing about magnitude.
[This week] (published September 8) His read: the rise in how hard this model is to monitor is too large for "the model got smarter" to explain, and there is a stretch of "dark matter" in between. Zvi Mowshowitz is a named AI commentator who runs the newsletter Don't Worry About the Vase and has held a strong position on AI risk for years. His words: "But the decline is too large compared to straight lines on graphs. There is 'dark matter.' Something else has changed." And: "OpenAI being right, and this being purely gains in capability, is the worst case scenario." (thezvi, 2026-09-08). Why that second sentence describes the worst case, in plain terms: if monitoring got harder purely as a natural consequence of the model getting stronger, this line of defense runs out for everyone and nobody escapes it. If some architectural choice caused it, then it is a choice that can be left unmade, and rules can block it. Which of our calls it supports or rebuts: it runs in the same direction as today's evidence update on that outside measurement, with both sides holding that the size of the move exceeds a capability jump. ⚠️ But we do not let his judgment carry weight: he holds a prior on this conclusion and has taken no measurement of his own. The weight still sits on the outside measurement that published its code and data. What he adds is that a named person holds this position in public, not that the position has been established.
No model watch item this issue. Not one of the 23 academic papers that arrived today got read, so there is no new preprint to write up; today's model-layer reading sits in the Astra evidence update in the main line above.
[Today] (published September 10) On the same day, OpenAI and Salesforce each shipped "harness" as a product name. A harness is the working environment wrapped around a model: the whole apparatus that manages context, orchestrates tools, coordinates subtasks and stores intermediate results. Until now every company wrote that glue itself. OpenAI launched the Agents API in public beta, opening up the environment behind Codex, its own coding agent, to outside developers. The company's words: "OpenAI hosts and maintains the harness," and developers choose only where the compute environment sits (OpenAI, 2026-09-10). The same day Salesforce announced its "Trusted Enterprise AI Harness," bundling six capabilities — context, agency, action, governance, security and models — into one architecture, plus a control plane for overseeing every agent in one place (Salesforce, 2026-09-10). Why these two belong together: on July 27 we wrote that the boundary between a model and the working environment around it keeps moving (binding tighter during training while decoupling at deployment) and that the integration value at the hard end is thickening at the same time (our July 27 issue). Today's two announcements are that sentence turning commercial: the layer that thickened is now something vendors charge for directly. For any team that wrote this glue itself, this reopens a build-or-rent calculation, and the question to ask is what it would cost to rip yours out. The most exposed are teams that wrote this layer in-house and treated it as a moat, and third-party vendors selling this layer as a standalone product — what they sell is now bundled in for free by the model vendors. ⚠️ Both items are company announcements. We read both originals today, and neither carries third-party adoption data, pricing comparisons or performance readings, so only the direction is usable.
No archive pick this issue. The reusable older material in our own back catalog is exhausted.
The past 24 hours. September 10 to 11 added 311 pieces. We finished 3 today and ruled out 1, leaving 307 unread — 0.96% read. Those 307 are unread, not read and rejected, and the two are not the same thing. By category: company filings, 4 arrived, 3 read and 1 ruled out, the only category anyone touched. X posts 156, academic papers 23, a second paper set 48, blogs 53, podcast transcripts 18, industry newsletters 8, industry analysis 1 — all at zero. We read nothing but filings today on purpose: three statutory filings landed at once yesterday, two of them carrying a first party's word-for-word disclosure of its own financial structure, which happens a handful of times a year. Commentary is always available. ⚠️ That "3 finished" is not the same number as the 15 receipts this issue's body uses: several of those 15 were opened on the spot to write today's items and are not among the 3. The named part: the four filings came from Oracle, Broadcom, TSMC and Arm, and what we ruled out was Arm's shareholder-meeting voting results — a procedural announcement with no checkable reading. On the X side, the highest-volume accounts were @teortaxesTex at 98 posts, @bhorowitz at 40 and @GaryMarcus at 30; the 18 podcast transcripts came from Bg2 and Dwarkesh.
What you are not getting today. We opened neither Oracle's nor Broadcom's earnings call transcript, so whether that prepayment came from one customer or several, how long the terms run, and whether it is refundable all go unanswered — and the answer to the first directly changes the risk assessment. We did not obtain the original of that investment bank's August report, so whether the $42 billion was derived from the same contract term can only be set side by side, not settled. We read none of last night's 8 industry newsletters, and here are their names so you know who you are missing: Stratechery, SemiAnalysis, Zvi Mowshowitz's AI #185, Interconnects, Understanding AI, Newcomer, Gary Marcus, Latent Space. Those 8 are only the newsletter category within last night's new material; the same night also brought 4 company filings and 18 podcast transcripts, three separate counts of three separate things. The academic-paper side is at zero today, which is why the model-watch column has no new preprint. And the British evaluation body's original assessment document does not appear on its own public publications index, which is what main-line item 3 is about.
Older material added back in one pass. We added 0 new sources in one pass today, and that 0 means the long-term roster gained no new names; it is a different number from the 15 external sources used in the body, which counts sources actually cited today. Separately, a set of older material from outside the past 24 hours came back in, with event dates falling between July 1 and August 30, 4,943 pieces in all, dominated by 1,808 papers and 891 newsletters, the rest being 745 company filings, 533 industry analyses, 362 blogs, 351 podcast transcripts, 134 supply-chain intelligence pieces and 119 X posts. That set and last night's list are two different populations, so the same category reads differently in the two places: company filings 4 last night against 745 in the back-fill, industry analysis 1 against 533, blogs 53 against 362, podcast transcripts 18 against 351, X posts 156 against 119 — and X is the only category where last night exceeded the back-fill. None of it is today's news, and none of it counts toward the 311 above.
Source concentration. Today's calls rest almost entirely on one form of source, the original statutory filing: three filings support most of the load-bearing figures in the body. That is concentration of form rather than of source. The three documents come from three unrelated companies on three different forms, each backed by its own legal liability, and no single company supports more than about three-tenths — no one party's account is being passed off as corroboration from many. The warning worth flagging sits elsewhere: the call in main-line item 1 stands on two samples, and the two landed on the same day because of the earnings calendar, not because of any causal link, which is written into that item's third rebuttal and its test conditions. Where the independent second view sits: the counter-reading in main-line item 2 comes from TSMC's monthly revenue, unconnected to either company; the measurement in item 3 comes from an individual researcher unconnected to OpenAI; and the speaker in the chips column comes from the other end of the semiconductor industry.
The sources we track. 529 named speakers in total. The spread: X 302, podcasts 90, outlets 51, blogs 48, paper authors 48, newsletters 46, earnings calls 26, keynotes 23 and a scattering of others. ⚠️ Those count venues, and one person can occupy several, so the parts sum to more than 529. Representative names: on X, Mark Zuckerberg, Lucas Beyer, Sergey Levine and Arvind Narayanan; among blogs, Lilian Weng, Terence Tao and Dario Amodei; among paper authors, Ion Stoica, John Jumper and Boaz Barak; among newsletters, Zvi Mowshowitz, Dylan Patel and Ben Thompson; on the institutional side, Alignment Forum, More Than Moore, SemiAnalysis and Data Center Dynamics. Several identically named numbers count different populations. The roster's 302 X accounts are the total we watch over time; last night's sweep actually touched 374 accounts and pulled 736 native posts, of which 156 reached today's material — three counts of three different things. Likewise the roster's 46 newsletters are the long-term total, while only 8 arrived last night, and a further 891 are back-fill from early July, three separate populations again. This issue uses 15 clickable receipts, the same figure printed at the foot of the page, counting only links the body actually cites that are not our own domain.
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 · 15 sources · Got a view? Reply and tell us
Written from the same research and judgments as the Traditional Chinese edition; every claim links to a primary document.