Daily Brief SecondSource Morning Brief · August 16, 2026 · Aug 16, 2026
This issue rests on the internal research digest compiled on August 15; no newer one arrived this morning. The events fall between July 15 and August 13. Last night's batch was a backfill sweep of mid-to-late July material rather than new events from the past 24 hours, and each item carries its event date beside the tag. Overnight, 76 pieces arrived in the unread queue → 38 clickable receipts here; the full accounting sits at the end. This is the email edition; the full edition of this issue is the website archive of record.
Why this claim is worth checking. On July 31, 2025, the venture capitalist Bill Gurley (a retired Benchmark partner) described model pricing on the BG2 podcast in a sentence with no exceptions in it: model companies were all buying share at a loss, price was the only valve left that could throttle demand, and nobody was willing to turn it. He named Anthropic: "Anthropic doesn't need to throttle. They just raise price, but they're not willing to do that because they're afraid to lose share" (BG2 Pod). Across the 529 named voices we track, this line came back 19 times — more than any other claim in the same window. The receipts are unusually hard: every model vendor's official price page, which you can open yourself. And the sentence that could be overturned is equally clear — that nobody dares raise prices.
Verdict: it does not hold. Inside a year, prices went up, and not only on the sticker. The steepest increase came from the company he named. The company repriced the flagship line outright, opened a more expensive tier above the flagship, and made its least-noticed change to the billing unit itself: Anthropic's own documentation states that its new-generation tokenizer produces about 30% more tokens from the same text (Anthropic's pricing documentation). A tokenizer is the ruler a model uses to cut text into billable units; shrink the ruler and the same document bills 30% higher while the sticker price never moves. No price-list comparison shows this. The hardest receipt we had staked out was a promotional rate due to expire and revert to standard pricing — but the official page now says that increase will not happen, and the promotional rate has become the standard rate, so that receipt is void.
The half of his claim that survives is the other one: down at the mature-capability tier, the price war is still raging, and harder than he described it. Hold capability fixed and the price of doing the same thing falls between 9x and 900x a year, median 50x (Epoch AI). What is rising is the entry price for this generation's best capability. What is falling is the price of doing the same thing.
How to use it. Do not budget model costs on a single line that says tokens keep getting cheaper. Budget two. Workloads you can move down to a good-enough tier get cheaper every year; the portion competition locks to the flagship tier gets more expensive, and not from the sticker alone. What would prove this wrong: vendors rolling flagship sticker prices back, or a new-generation tokenizer returning to the old ruler. Verdict date: September 15 — what to look at that day is flagship sticker pricing and tokenizer definitions on each vendor's official page.
Two things pointing in opposite directions happened on the same July 16 earnings call. On the results side: quarterly revenue of US$40.2B, landing at the very top of the company's own guidance, an operating margin of 60.3% that came in 1.8 points above the upper end of guidance, and net income up 77.4% year on year (TSMC's press release). This is a company that has sold out every wafer it can make. On the Q&A side: asked about raising prices, Chairman and CEO C.C. Wei said he would not suddenly take prices to the 4x or 5x he might want, because no customer survives an increase on that scale. Asked whether TSMC would follow peers into investing in its customers' customers, he said that so far it does not do these kinds of financial arrangements. Asked how large the shortfall was, he declined to give any number and said only that the gap is "very big" — the 30% to 50% figure circulating outside came from the analyst's question, not from him. He also volunteered a line almost nobody has quoted: he is genuinely envious of memory companies' 86% gross margin.
Verification: what we held before was a single reporter's live transcription from the room. Today it is the official call transcript posted on TSMC's investor relations site, where all four passages line up, with Motley Fool's independent transcript as a cross-check. The reconciliation took down half of our own claim: TSMC is raising prices 5% to 10% across more than 75% of its wafer revenue base (Culpium exclusive), and even mature nodes go up by as much as 10% from January 2027 (Nikkei exclusive). "It collects no rent" fails. What survives is the narrowed version: what it refuses is memory-style rent extraction. The comparison numbers are all public. Micron posted an 84.9% gross margin last quarter and guides to roughly 86% for the next (Micron's results). SK hynix ran a 76% operating margin in the second quarter — that measure comes after operating expenses and normally sits below gross margin — on revenue up 257% year on year (SK hynix's press release). TSMC's gross margin over the same period was 67.7%, on revenue up 33.7%.
Why we dug now: our own record held two claims that fight each other — "TSMC gave up collecting rent" and "TSMC is raising prices." Until the accounts are reconciled, both get quoted, and they cannot both be true as stated.
Open ?
Judgment update: the three big items on an AI hardware bill of materials now carry three completely different risk shapes. The wafer line is measurable: public reporting puts it at 5% to 10%, and Wei himself ruled out anything on a multiples scale, so this segment can be treated as a semi-constant. To be precise about what that is: it is a behavioural band we measured, not a ceiling TSMC declared. Memory has no such upper bound. Its cumulative increases run into multiples, its suppliers guide 85%-class gross margins higher still, and it prices differently to customers with weaker leverage. Advanced packaging pays a third kind of price — a time tax in lead times, allocation and queue position, rather than a sticker one. Procurement hedging and long-term contract firepower belong on the memory line, all of it. One more thing worth carrying. This cycle's AI supply chain has developed a habit: suppliers take customer equity or warrants in exchange for supply, downstream players hold each other's shares, and customers spend that money back on compute. The result is a chain that props itself up, and it breaks at the foundry. TSMC takes cash for its wafers and holds no customer equity, which makes its order book one of the few demand thermometers this cycle that insiders have not contaminated. And when you build the cost model for a hardware-bound product, remember that the memory line cannot inherit the wafer line's band.
Investor note: the market prices "AI hardware costs are rising" as a single bundle and calls it chips getting more expensive. This evidence splits the bundle: the wafer end has an upper bound its own supplier ruled out, and the end with no such bound is memory — which weakens the working assumption that foundry takes the largest share of the increase.
What would prove this wrong: any single list-price increase of 20% or more on an advanced node; or disclosure of an equity or warrant arrangement between TSMC and an AI customer, or that customer's end customer. Verdict date: TSMC's third-quarter call in mid-October 2026 — watch whether "pricing" appears for the first time in the attribution bridge for gross margin.
When people follow the money through an AI rack, they follow GPUs, memory, packaging, optics and power. But every server in that rack has to boot, has to be manageable remotely, and has to prove its own firmware has not been tampered with. Who supplies that layer? Companies like AMI, which builds platform firmware and manageability software. Their work is what runs after you press the power button and before the operating system starts. On July 27, the programmable logic maker Lattice Semiconductor closed its acquisition of AMI, with consideration filed as about US$1B in cash plus roughly 5.2 million shares (the full 8-K exhibit).
Verification: this is the only item this issue backed by a statutory filing, so the fact of closing and the deal structure carry legal liability behind them. ⚠️ But "accretive" and "a US$1B annual revenue run rate" in that same document are the company's forward guidance, not realised numbers. On AMI's share, its hyperscaler customer list, and the value captured per rack, we hold not one figure. The press release offers adjectives.
Judgment update: the number is not the thing to track. The neutrality pledge in the release is: AMI's firmware and manageability solutions will continue to be "developed and delivered without preference for any silicon vendor, including Lattice." Writing "including Lattice" into a neutrality clause is the company naming this deal's largest integration risk itself — once a neutral layer belongs to a company with silicon interests, will hyperscalers still treat it as neutral? Something you can act on today: when you buy AI servers or negotiate specifications with a contract manufacturer, put supplier neutrality for firmware and remote management — the trade calls it out-of-band management — on the due diligence list. Readers get their first set of numbers to reconcile against only when Lattice's next results break out AMI's revenue scale and customer retention. ⚠️ And the unproven part deserves saying plainly: "every machine has to carry one" is not the same as "this layer earns money." Mandatory attachment and value capture are two different things, and we hold no figures on the second. Read with main line item 3: that item is the same argument one layer up, where it is already possible to measure who collects the rent and who merely gets installed.
Investor note: the argument that value lives in the mandatory non-accelerator layers is being amplified in the market right now. The Lattice deal hands that argument an instance, and flags its weak point in the same document: mandatory attachment guarantees volume, not price, and the two cannot be blended in a valuation model.
On July 26, at the "AI Meets EDA" preshow session of DAC — the chip-design tool industry's annual conference — NVIDIA laid out its whole stack for chip-design agents in one go. At the same event, the chip-design tool oligopoly — Synopsys, Cadence and Siemens EDA — each showed their own AI agents. The positioning was explicit. Synopsys's thermal simulation agent was logged on the night as "underneath, accelerated by CUDA-X". Siemens's stack "uses Nemotron", NVIDIA's own model family. (The live relay throughout comes from semiconductor analyst Ian Cutress, independent analyst at MoreThanMoore and former editor-in-chief of AnandTech.)
One line from that session is worth more than the layering. NVIDIA's speaker explained on stage why agents in engineering domains are unusually hard to build, and the answer was data rather than models: "Data doesn't exist to train these models. Can't just download the Internet" (the original relay).
Verification: ⚠️ all of this comes from one person, one event, one thread of live notes; nobody turned the slides into a transcript, and most of the posts are headlines with no body. That supports "the layering looked like this at this event." It does not support an assertion about how an industry is layered. ⚠️ Cadence, the third of the three, showed only ChipStack, with nothing on the record about NVIDIA underneath, so we do not count it in this layering. But that "data doesn't exist" line has a rare evidentiary structure: it testifies for a competitor. Three months earlier, Cadence's CEO listed "large language model companies do not have enough data in this domain" as the third pillar of his own moat. That was a self-serving claim, and it should have been discounted. Now NVIDIA has confirmed the same thing from the stage. Two companies start from the same scarcity and arrive at opposite ways of monetising it: one says the data is mine and so is the rent, the other says you need synthetic data tools and I make the tools.
Judgment update: a vertical AI moat has to be split into two independent questions — who can collect the rent (who owns the tools, who monopolises the data) and what the margin on that rent is (the bargaining power of the supplier one layer down). The chip-design tool industry gives us a sample: the answer to the first is the big three, and the answer to the second is a single supplier who could climb up and become their competitor at any point. That produces an expectation running against the market's default, and one that future results will either confirm or embarrass: as agent revenue rises as a share of these three companies' business, gross margins may be diluted rather than lifted. When you assess any vertical AI position, ask two questions instead of one: who holds the agent orchestration layer on top, and who holds the models and compute beneath it, with how many substitutes. ⚠️ This is a question worth watching, not an established judgment. The evidence is live notes, and we have not been through the three companies' own financial disclosures. Verdict date: results from the big three across the second half of 2026 and into 2027 are the first window in which the margin direction of agent-related revenue can be read.
Investor note: the market's default for "software companies go agentic" is that gross margins rise. What was logged at DAC points the other way, and it is testable: when the layer underneath an agent is bought in from a single supplier, the more agent revenue grows, the more likely margins get diluted.
GPUs mostly talk to each other over copper today. Past a certain node count, copper's cost and power draw blow up, and optical interconnect is the answer — except the economics have never worked. The whole optics supply chain — optical engines, glass-core substrates, external laser sources, packaging capacity — buys ahead on a cycle longer than a chip generation. That makes "when does this actually get used" a live financial question for everyone on the chain. In January this year, AMD CTO Mark Papermaster's answer was a rolling window: photonics would cross into economic viability "within the next three years", starting with the largest clusters (MoreThanMoore interview). ⚠️ Said in January, that meant 2029. Said in July, it still meant 2029. It never comes due, so it can never be proved wrong. At AMD's annual event on July 23, the Helios system AMD unveiled was indeed still 72 GPUs on copper — but the roadmap wrote optics into a named generation: "MI500 with opticals", with the next line reading "MI600 in 2028" (live relay). The same session announced a rack-level annual cadence: not just a new GPU every year, but a whole new rack every year (live relay).
Verification: ⚠️ both ends were relayed by the same person — the interview and the live posts both come through Ian Cutress's channel — so they do not constitute two independent sources. And "MI500 with opticals" is a handful of words that do not distinguish co-packaged optics from in-rack optical interconnect from optical switching. That distinction is the crux: the three hit completely different supply chains. Which year MI500 lands in was not given on the night either.
Judgment update: what happened here is that a schedule became arguable for the first time, not that optical interconnect is arriving. It now has an anchor: a supply chain building meaningfully ahead of the MI500 generation is front-running, and one building meaningfully behind it is carrying shortage risk. ⚠️ One alternative explanation, recorded honestly: treating the CTO's January remark about an industry-wide economic crossover and his July remark about his own product schedule as one timeline is our splice, and AMD has never put those two sentences together itself. Until both the optical form factor and the MI500 year are confirmed, this must not be used as the basis for a judgment about any specific supply-chain name. Verdict date: the next public occasion on which AMD states MI500's year or its optical form factor — the annual event or an earnings call both count.
Investor note: the optics narrative has swung for a year between "optics in, copper out" and "optics deferred, copper in", with neither side able to produce a date anyone can reconcile against. This evidence settles nothing between them, but it collapses the demand-side timing from a rolling window into a named generation, which means the argument can now be had against a specific product generation. Anyone planning clusters and racks can use the anchor this way: line the procurement and rack-interface decision calendar up against the MI500 generation, where far too early is front-running and far too late is shortage risk. Before the form factor is settled — co-packaged optics, in-rack optical interconnect or optical switching — this is still not enough to pick which part of the supply chain to back.
On August 11, Mitchell Hashimoto (co-founder of HashiCorp, author of the open-source terminal Ghostty) posted a new daily routine and set it against his own version from September 25, 2025. The old list's first item of the morning was "open computer, check email"; the new list fills the same slot with reviewing the agents that ran overnight and dispatching the next round; he dispatches another round at eight in the evening. His own account: "I regularly have at least 2 agents running constantly with many running through most of the night" (the post, with the old routine alongside).
Verification: ⚠️ this is a self-reported integer lower bound, with no execution logs, no task types, no success rate and nothing on how much of what ran all night was usable the next day. ⚠️ He is founding a company building a multi-session agent workbench, so the post runs in the same direction as his commercial narrative; and across those ten and a half months he changed jobs and had a newborn, so more than AI moved his routine. This is n=1, and one of the most extreme users in the category. But the number itself cuts against the marketing: someone who has written agents into his daily schedule, and who builds tools in this space, runs a low single-digit number of them at once, not a double-digit one.
Judgment update: the shape matters more than the number. The human role is being pushed into a twice-a-day scheduler and reviewer, and agent execution time has come unhooked from human availability. That swaps out a badly posed question — how many times faster does AI make an engineer — for a well-posed one: how many times a day does a person come in, and what do they do each time. When you estimate AI's effect on engineering capacity, do not stop at multiples per head; count the agent-hours produced per day and how many of them a human can get through. The bottleneck is moving to the second number. It also implies a concrete staffing question: who reviews the overnight output first thing in the morning? Without that role, what runs overnight simply piles up untouched. ⚠️ The number we most need and do not have is how much of a full night's output is genuinely usable the next day. Without it, "attention is the bottleneck" is only a story that sounds right.
Investor note: agent tooling narratives generally rest on the assumption that concurrency scales up quickly, and the only measured reading available differs from that by an order of magnitude. This weakens usage forecasts that extrapolate linearly from concurrency, without disturbing the separate observation that agents are running longer.
1. [This quarter] SpaceX is reported to have placed its first AI server order with Foxconn, planning to buy 13,000 units. ⚠️ Anonymous sourcing, the dollar figure is the media's estimate, and the quantity is a plan rather than a booking. What stands is the contrast: the company that talks about data centres in space is ordering standard ground racks (supply-chain reporter Dan Nystedt, 07-19).
2. [This quarter] Taiwan's Legislative Yuan reviewed a bill on July 22 requiring large electricity users with contracted capacity above 5MW to build their own supply facilities. This is not rationing; it pushes responsibility for building power plants back onto the people using the power. ⚠️ We did not track down the outcome of that review (Dan Nystedt, 07-19).
3. [This quarter] A former TSMC deputy manager has been charged with stealing trade secrets, with prosecutors seeking seven years. The material lands in failure analysis rather than process recipes, and the seniority is not high. ⚠️ A sentencing request is not a verdict, and we did not track the outcome (Dan Nystedt, 07-20).
4. [This quarter] The open-source fine-tuning toolchain Unsloth compressed the 975B-parameter Inkling model to 270GB, from roughly 1.9TB. ⚠️ We checked the size against the parameter count and it reconciles, but the capability-retention rate claimed alongside it has no definition and names no benchmark (Daniel Han, 07-15).
5. [This quarter] A circulating claim that a particular model designed a chip was publicly questioned by a semiconductor analyst working through the specification arithmetic. ⚠️ He asked, he did not rule; we cannot find the original claim's source, and this must not be read as "refuted" (Ian Cutress, 07-20).
[This quarter] (event dates 07-16 to 07-19) One week, three different actors, three different routes, all pointing at the same bottleneck: advanced packaging. Advanced packaging is the back-end step that bonds GPU dies to high-bandwidth memory, and through this AI build-out the capacity limit has sat here rather than in front-end wafer fabrication. Three things from mid-to-late July, side by side: TSMC's chairman publicly welcomed a successful packaging offering from his competitor Intel so that it could take some of his own load (see main line item 1); Broadcom and Taiwan's Powertech announced a US$400M joint venture to build a packaging substrate plant in Singapore, with Powertech having developed fan-out panel-level packaging for large dies — a separate technology route using square panels rather than wafer-shaped substrates, not an expansion of the existing mainstream process (Dan Nystedt, 07-16); and UMC, Taiwan's second-largest contract chipmaker, is planning a plant expansion on advanced packaging orders spilling over, silicon interposers among them (Dan Nystedt, 07-19). ⚠️ All three were relayed by the same reporter, so they do not corroborate each other, and most of them give no amounts, no capacity and no timelines. A counter-intuitive reading: orders spilling over to second-tier houses while an alternative technology route enters at the same time is an early signal of a bottleneck easing, not worsening. Read the other way round, it is also the reason the price and margin peak for advanced packaging may land in 2026 rather than later.
[This week] (event date 08-13) Mitchell Hashimoto: he bounces off "low-effort AI-designed web pages" on sight, and the product never gets evaluated at all. On August 13, the HashiCorp co-founder and Ghostty author posted that he closes pages with thin lines, glowy styles, inconsistent fonts and heavy monospace on sight, and so "never get[s] a chance to judge the product itself" (the post). The mechanism at work is probably not the intuitive one: the four features he lists are also the dominant house style of developer tooling from 2024 through 2026. AI did not invent that look. AI made it free. The question worth asking is what happens when a look that used to require investment becomes free: it flips from a positive signal of taste to a negative signal of effort not spent. The signal did not go to zero. It changed sign. That supports and extends the academic Ethan Mollick's judgment about text as a proxy signal — that writing as a proxy for effort and quality has been punched through by AI in one keystroke, decoupling the signal from the underlying contribution. His line is about decoupling; this one describes an inversion, and it lands earlier, before the product is ever opened for assessment. The general form: any signal that once served as a proxy for quality because it was scarce — complete documentation and a thorough README — needs its polarity rechecked once AI makes it cheap. ⚠️ The honesty this requires: n=1, a personal account rather than market data, from a speaker whose stance on AI output has been persistently negative, and resting on a premise never tested — whether he can reliably tell an AI-designed page from a human one built in the same fashionable style.
The academic crawl produced nothing new today: the same digest came back identical six days running, so we judged it a static snapshot rather than that day's product, and ran no new paper scan this issue. The two items below are older research pulled back out to reconcile this week's vendor claims, with years marked. They are not news. They are rulers.
1. [Trend] Self-improvement: able to use feedback is not the same as able to self-examine (CMU and KAIST, November 2025). NVIDIA claimed at that July 26 event that internal research lets its agents "improve themselves over time". There is already a measurement for exactly this. RefineBench runs 1,000 problems across 11 domains in two modes: one where the model revises with no prompting, and one where it gets explicit feedback first. Unprompted, even the strongest model of the day, Gemini 2.5 Pro, scored 31.3%, and GPT-5 scored 29.1%, with additional rounds adding almost nothing. Given explicit natural-language feedback, the mainstream large models fix their way to near-perfect within five rounds (arXiv 2511.22173). ⚠️ Those are readings from the November 2025 generation, and we hold no 2026 retest, so they measure how hard this is rather than where the water sits today. It is also a single benchmark, on open-ended problems, with a scoring process that carries its own subjectivity. Which changes the question to put to a vendor: not how good the self-improvement is, but what multi-round training method they used to put it in.
2. [Trend] Whether synthetic data breaks a model depends on the workflow, not the ratio (Stanford, Google, CMU and Yale, October 2024). NVIDIA's recipe at that same event was to start with synthetic data, add real data, then add more synthetic. This study compared three workflows systematically: replace real data with synthetic outright and every task collapses; accumulate real and synthetic together and the model stays stable even as the real share tends towards zero (arXiv 2410.16713). NVIDIA described the second of those, so the recipe has academic support behind it rather than nothing. ⚠️ But the gap needs stating: the study's experimental domains are statistical estimation and language model fine-tuning, not chip design or physical simulation, so extrapolation takes a discount — and it proved only that the process does not diverge, never that it learns well.
[Trend watch] (event dates 07-20 and 07-23) AMD's attack on the CUDA moat has moved forward, from persuading data centres to switch to making sure developers start on its side. Evidence on AMD's software ecosystem used to cluster around large inference frameworks; the past two weeks produced two independent sources pointing somewhere else. From the edge segment of AMD's July 23 annual event: the next-generation Halo development platform will offer 192GB of memory in a very small box and has secured direct support from the open model hub Hugging Face, and every Halo machine ships with a year of Hugging Face Pro (Ian Cutress's live relay). On July 20, the open-source fine-tuning toolchain Unsloth announced an official collaboration with AMD, bringing training and inference to more than 500 models and claiming training twice as fast on 70% less memory (Unsloth co-founder Daniel Han). Those last two numbers are vendor-reported, with no baseline attached to either. Read together, they say this: what decides an ecosystem is often not the large deployments but the long-tail entry point — the first notebook a graduate student opens when they want to fine-tune a model. What to watch to test this is whether AMD hardware's coverage list on open-source fine-tuning toolchains and model hubs keeps getting longer, rather than which new data centre customers AMD announces. ⚠️ Neither number has a comparison baseline (twice as fast as what, 70% less memory relative to which configuration), and all of this sits in inference and fine-tuning, not large-scale pre-training. On the moat around core training workloads, these two data points say nothing at all.
Can the scaffolding on the outside be pulled apart from the model? (Our deep dive of July 16.) People building AI applications broadly believe the model itself commoditises and the real craft lives in the layer you wrap around it: how you feed in context, break the steps up and validate the output. The trade calls that scaffolding, and a good number of startup valuations rest on the assumption. That report's conclusion was that the dividing line moves: models absorb the parts as components, climbing from easy tasks upward by difficulty, while the integration premium at the hard end thickens at the same time. Why we are raising it today: Anthropic's chief product officer wrote on July 24 that on the same class of task he had gone from a bespoke harness and verification system to roughly four sentences plus heavy use of the product's built-in workflows (the post). ⚠️ The correct reading is that the instruction burden moved rather than disappeared. A built-in workflow is itself a layer of scaffolding, relocated from something the user builds to something the product ships — and those two facts carry opposite implications for which layer the moat sits in. When you assess a company that says its value lives in the scaffolding layer, the question is not whether models get better. It is how much of your layer the model vendors will build in. This data point suggests that happens on a timescale measured in half-years.
The past 24 hours. Overnight we went through 76 candidate pieces and used 38, with every receipt this issue drew on sitting in the brackets above. By category: 15 podcasts (Dwarkesh 12, Colossus 2, Gooaye (Taiwanese podcast) [a Taiwanese markets podcast] 1); 85 newsletters, the largest being Latent Space 16, Zvi Mowshowitz 13, Gary Marcus 10, Stratechery 8 and SemiAnalysis 4 (the last two are paid subscriptions, from which we take direction only and quote nothing); 4,984 academic papers into the index; 28 companies' filings under watch, of which this issue used the Lattice 8-K in main line item 2. X posts are the bulk of this issue's material, and every one of them is linked in the item that uses it.
One-time backfill. No new backfill batch this issue. The items tagged with July event dates above come from an overnight sweep of mid-to-late July material. They are not things that happened in the past 24 hours, and each carries its event date.
Source-concentration warning. Two things to know today. First, this issue leans on two analysts on X: among this issue's candidate material, semiconductor analyst Ian Cutress accounts for 37% and supply-chain reporter Dan Nystedt for 33%, seven in ten between them — and nine of Cutress's are concentrated in a single set of live notes from the same two events. Multiple records from one person at one event are not multiple pieces of evidence, and we have marked that item by item in the body. Second, you are getting nothing today from official company channels: the 19 AI company blogs we track published 165 posts yesterday, and last night's batch read none of them, so no material in this issue comes from a company's own publication. Put together, the consequence is this: not one piece of material today clears the multi-source verification threshold, and the only thing in this issue backed by a statutory filing is the Lattice 8-K in main line item 2. We would rather say that outright than paper over it with a count of how many sources we reviewed.
The sources we track. 529 named voices across platforms. 302 on X (Mark Zuckerberg, Arvind Narayanan, Sergey Levine, Graham Neubig and others); 46 in newsletters (Zvi Mowshowitz, Benedict Evans, Ian Cutress, Dean Ball and others); 48 in papers (Yann LeCun, Noam Shazeer, Ion Stoica, John Jumper and others); 26 in results and earnings calls (C.C. Wei, Jensen Huang, Lisa Su and others); 90 podcasts, 23 keynotes, 48 blogs, 51 media outlets; plus 76 institutional and company accounts (NVIDIA's technical blog, Google Research, Hugging Face and others).
Correction: the August 14 issue said "32 sources / 32 clickable receipts" in both its opening and its footer, while the public edition actually carried 27 clickable receipts, 2 of which were our own back issues. That 32 was counted before the layout trim and never recounted after it. The count is now produced by machine after the cut, with trimmed links subtracted, and is no longer filled in by hand.
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 · 38 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.