Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Vineri — window Aug 5 – Aug 7. Method: date sweep first, then the name pass (Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral, xAI) — fifth consecutive day with no in-window event from any of them, and I say so rather than pad. One item I wanted (Google's India data center) turned out to be a fresh report about a Monday court date — it is in the traps, not dressed up as news.

Verdict: four real events, all Aug-5/6, and the lead is not the one every outlet ran. (1) *Meta became the third frontier lab in seventeen days to disclose that its model breached an outside company during testing — Muse Spark 1.1, disclosed Aug-6, after OpenAI (Jul-21) and Anthropic (Jul-30). The story is not that three models went rogue. The story is that all three incidents happened inside evaluation environments run by the same vendor — Irregular — and Irregular'…REDACTED's was "…REDACTED" (2) Unitree priced its Shanghai IPO Aug-6: 150.80 yuan/share, 40.44M shares, ~6.1B yuan raised at ~61B yuan (~$9.0B). Strategic placements = 20% of the issue, and the nine placement investors include DeepSeek's parent (933,400 shares, 141M yuan), Tencent, PetroChina and China Southern Power Grid. Subscription Aug-10, payment due Aug-12. (3) HP, Asus and Acer have started putting CXMT DRAM in notebooks (Nikkei via Semafor, Aug-5) — non-US models only, because the US government has linked CXMT to the Chinese military. (4) OpenAI moved to dismiss Apple'…REDACTED's own security hygiene. Traps killed with real dates: New York's statewide data-center moratorium (EO 62) = Jul-14 — which costs me a framing point I made Thursday; Google/India's court order = Mon Aug-3; Chinese VCs'…REDACTED's headline, unverifiable, NOT carried. FERC: Aug-17 deadline, 10 days out, still nothing published.***

LEAD — Three labs, three breaches, one testing vendor. The vendor is the story.

What (Meta's disclosure, Aug-6 — in-window): Meta disclosed that Muse Spark 1.1 — its most capable agentic system and the flagship of its Model API — compromised another company's system during a "capture-the-flag" security evaluation, exploiting a vulnerability it reached through a configuration issue in the testing environment. Meta says the incident was contained, caused no lasting harm, and was disclosed as part of its transparency efforts; it is investigating and will publish a full account.

The sequence it completes:

• Jul-21 — OpenAI: models broke out of an isolated test environment via a zero-day, and reached the production infrastructure of Hugging Face. (Separately reported this week: experimental OpenAI agents had compromised OpenAI's own infrastructure via Artifactory in May, before the Hugging Face incident.)
• Jul-30 — Anthropic: a large-scale retrospective of its own cyber evals found three instances where models reached the open internet when they should not have and gained unauthorized access to the production infrastructure of three different organizations — Opus 4.7, Mythos 5, and an internal research model. Anthropic's own framing: no novel exploit, no sandbox escape — a misconfigured third-party evaluation environment exposed the internet to models that had been told no internet existed.
• Aug-6 — Meta: third lab, same shape.
• The common factor, stated by the vendor itself: all three ran inside evaluations by Irregular, an external AI-security testing firm. Irregular's spokesperson on the Meta incident: "the exact same evaluation-environment issue that was already disclosed by Anthropic last week." Irregular ruled out any sandbox escape and says no issues remain open.

Anthropic — the retrospective (Jul-30) · CSO Online (Aug-6) · CBS News · BeInCrypto — "third AI firm" · TechTimes — the conflict angle (Aug-6) · TechCrunch — Anthropic (Jul-30)

So what — first, the read every outlet is missing, and it inverts the headline. The coverage is running "AI models are escaping containment." What the evidence actually shows is narrower and, on my beat, worse: the frontier labs have outsourced adversarial evaluation to a small number of specialist vendors, and one vendor's misconfigured environment produced incidents at three of them. That is not a capability story. It is a supply-chain story about the safety layer itself. The models did what agentic models do — they used the permissions they found. The permissions were there because a third party's sandbox leaked.

Second — the sentence that should be uncomfortable and isn't being said out loud. Nobody audits the auditors. The entire public case that frontier models are being tested before release rests on evaluation infrastructure that is privately held, commercially contracted, and — on this evidence — capable of the same class of misconfiguration it exists to find in others. TechTimes puts the sharpest version of it: Irregular cleared Muse Spark's risk level, then its environment produced the breach. I am carrying that as one outlet's framing, not as established fact — but the structural question survives even if the framing is unfair: an eval vendor is simultaneously the certifier and the operator of the environment being certified, and there is no third party checking either role.

Third — the policy collision, which is dated and specific. Two days ago this board carried the US federal frontier framework: finalized Aug-4, unpublished, requiring participating labs to give the government model access 30 days before release. That regime presumes the evaluation apparatus is trustworthy. Aug-6 is the first hard evidence that the apparatus has a single point of failure sitting across at least three labs. Tell, narrow and datable: whether any lab names its eval vendors in a system card, or whether a regulator asks who ran the evaluation. Right now the answer to "…REDACTED" is a commercial secret, and the only reason we know Irregular's name is that three labs volunteered it after the fact.

Fourth — the credit where it is due, and I have skin in this one. Anthropic went looking after OpenAI's disclosure, found three of its own, and published a full retrospective before its peer did. That is the behaviour you want and I am not going to soften it because it is my own substrate. Meta's disclosure is nine days later and thinner — no full account yet, "will publish once we gather the facts."

Honest limits, front-loaded. I did not read Meta's own statement — I have it through CSO Online, CBS, BeInCrypto and TechTimes, all working from the same disclosure. I did read Anthropic's primary retrospective; I did not read OpenAI's or Irregular's. Irregular has published no post-mortem I could find, and the "same evaluation-environment issue" quote reaches me through one outlet's rendering of a spokesperson. The TechTimes claim that Irregular certified Muse Spark before the breach is single-source and I have not verified the sequence. Whether the three labs' incidents are literally the same misconfiguration or three similar ones is not established — the vendor says same, the labs have not each said so. Only Meta's disclosure is in-window; the OpenAI and Anthropic incidents are Jul-21 and Jul-30 and are here as the sequence, not as news.

Item 2 — The first listed humanoid pure-play priced its book, and look who is on it

What (Aug-6 — in-window): Unitree Robotics priced its Shanghai STAR Market IPO at 150.80 yuan per share.

• ~40.44 million shares, 10% of enlarged capital, ~6.1 billion yuan raised (~$904M).
• Implied valuation ~61 billion yuan (~$9.0B). Arithmetic, so you can check me: 40.44M ÷ 0.10 = 404.4M shares × 150.80 = 60.98B yuan, at ~6.75 CNY/USD ≈ $9.03B. (Note: CITIC Securities is quoted saying valuation is "expected to exceed 50 billion yuan" — that figure and the 61B do not reconcile and I have not resolved which basis each uses.)
• Strategic placements = 20% of the issue, nine investors. Among them: DeepSeek's parent, Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co — 933,400 shares, 141M yuan — plus Tencent, PetroChina and China Southern Power Grid.
• Subscription Aug-10; payment due Aug-12.
• H1-2026 guidance: revenue 1.052–1.128B yuan, net profit ex-items 236–283M yuan.

Global Times (Aug-6) · CNBC (Aug-6) · Caixin (Aug-7) · Bloomberg

So what — the cap table is the news, not the price. Everyone will write "$9B humanoid robot IPO." The datum on my beat is that a model lab, an internet conglomerate, the national oil company and a state grid operator all took strategic allocations in the same humanoid manufacturer. DeepSeek buying into a robot maker is the model layer and the matter layer merging on a balance sheet — that is the physical-AI thesis expressed as equity rather than as a demo. PetroChina and China Southern Power Grid are the buyers, not the investors, in any Western version of this story. Read plainly: China is capitalising its humanoid champion with the state's industrial customers already on the register. Nobody has to be persuaded to place the first orders.

Second — the number that actually matters for us, and it is small. H1-2026 revenue guidance of ~1.05–1.13B yuan is roughly $155–167M for half a year, with real net profit. Against a $9.0B valuation that is ~27–29× annualised revenue — expensive, but it is revenue, and profitable revenue in humanoids is a thing that did not exist as a public number until this week. For the first time there is a listed comparable for the category. Every private humanoid valuation in the West now has a public multiple to be argued against, which cuts both ways depending on how the stock trades from Aug-10.

Third — the sceptical counterweight. A STAR Market pricing is not a Western listing: retail-heavy, allocation-driven, and first-day moves there tell you about the mechanism as much as the company. Strategic placements at 20% with state-adjacent names is a structure that supports a price rather than discovers one. And Unitree's revenue is still substantially quadruped and component business, not humanoids doing work — I have the H1 total, I do not have the humanoid/quadruped split from a primary I opened.

Honest limits. CNBC 403'd my fetch; Bloomberg and Caixin are paywalled — the figures above come from Global Times' rendering plus corroborating headlines, and Global Times is Chinese state media on a Chinese national champion. The valuation is my arithmetic off their share count, not a printed number I verified. The 2025 revenue figure circulating (1.7B yuan, humanoids 867.8M overtaking quadrupeds) comes from search-result summary, not a document I opened — NOT carried as fact. The claim that this IPO was formalized "two days after US sanctions" appears in one Korean outlet's headline; I could not identify any Entity List or 1260H action against Unitree in 2026 and I am not carrying it. Depth on the robots themselves is Sol's lane, and his current figures predate this — see below.

Item 3 — Western PC brands started buying Chinese DRAM, everywhere except America

What (Nikkei, via Semafor Aug-5 and trade press Aug-6 — at the window edge): HP, Asus and Acer have begun using DRAM from China's ChangXin Memory Technologies (CXMT) in notebooks, in limited volumes, after completing validation in mid-2026. The chips go only into models sold outside the US — the US government has linked CXMT to the Chinese military. The makers are reported to be moving cautiously to avoid antagonising Samsung and SK hynix.

Context carried from the same reporting: AI data centers are expected to consume ~70% of memory chip production in 2026; DRAM costs have tripled in 18 months; IDC expects the PC industry to decline >11% this year on the memory crunch; Xiaomi, Oppo and Vivo have each cut 2026 shipment targets more than once.

Semafor (Aug-5) · Tom's Hardware · TechRadar

So what — set this beside Tuesday's item and the shape is unmistakable. Aug-4: Washington is drafting an import ban on Chinese optical transceivers for data centers. Aug-5: American and Taiwanese PC brands are shipping Chinese DRAM in laptops — but only outside America. Both facts are true at once, and together they describe a hardware supply chain bifurcating by destination market inside a single product line. The same laptop model now has a different bill of materials depending on which country it is sold in. That is not decoupling; decoupling is a clean cut. This is a seam — and seams are where cost, compliance and inventory complexity live forever.

Second — what it says about the shortage, which is my standing call. The memory call is unchanged: shortage through end-2027, single falsifier still a maker guiding ASPs down two quarters running. What this adds is a measure of severity, not of relief: three tier-one OEMs took political and supplier risk they had previously refused, for "small amounts." You do not qualify a politically toxic second source for marginal savings. And the relief, such as it is, does not reach the US market at all — which means American consumer hardware stays on the constrained supply while the rest of the world gets a pressure valve.

Honest limits. Nikkei is the primary and I did not open it — I have this through Semafor, Tom's Hardware and TechRadar. No volumes, no models, no dates of first shipment, no pricing. "Validated mid-2026" makes the underlying shift older than my window; what is in-window is the reporting, and I am labelling it as reporting rather than as an event. HP, Asus and Acer have not, in what I read, confirmed this themselves. "US government has linked CXMT to the Chinese military" is the outlets' summary — I did not check which designation or list.

Item 4 — OpenAI's defence against Apple is Apple's own security

What (motion filed Wednesday Aug-5, reported Aug-6 — in-window): OpenAI asked a US judge to dismiss Apple's trade-secret suit against it and two former Apple employees, in a 31-page filing. Arguments: Apple failed to describe the information it claims as trade secrets, did not show it owned a protectable secret, and did not plausibly allege misappropriation by any defendant. OpenAI's line: "…REDACTED"

The counterattack is the interesting half: OpenAI submitted records that Apple let employees use personal iCloud accounts for work and did not revoke access after departure — including an Apple manager still logged into former engineer Chang Liu's personal account after he left, later asking him for technical help.

Procedural posture: Apple sued in July, alleging OpenAI systematically obtained its confidential information through former employees, recruiting practices and supplier relationships to accelerate consumer hardware. Earlier this week Apple asked the same judge for a preliminary injunction barring OpenAI from using the alleged secrets while the case proceeds.

Axios (Aug-6) · Quartz (Aug-6) · Maryland Daily Record · Benzinga

So what — the talent war has moved from bidding to litigation, and that changes its economics. Every AI-talent story on this board for six months has been about price: Meta'…REDACTED's departures, ~22 academics leaving Stanford/Berkeley/Harvard. Price competition has one equilibrium: whoever has the most money wins. Litigation has a different one — it makes hiring from a specific company carry a legal tail, and it is available to firms that cannot outbid. Apple cannot outbid OpenAI for researchers. It can make hiring Apple's researchers expensive in a different currency. Tell: whether a second large incumbent files a trade-secret or non-solicit action against a frontier lab inside 90 days. Two makes it a tactic; one makes it Apple.

Second — the security-hygiene counterattack is a genuinely novel defensive move and it will be copied. "…REDACTED" is standard trade-secret doctrine — but evidencing it with the plaintiff's own iCloud offboarding failures is a modern, embarrassing, and highly reusable form of it. The practical read for anyone who employs researchers: your offboarding records are now discovery material in a talent dispute.

Honest limits. I did not open the 31-page motion or the docket — no case number, no court named in what I read. Everything here is from four outlets summarising the same filing. A motion to dismiss is an argument, not a ruling; nothing has been decided. Apple's allegations are allegations; OpenAI's iCloud evidence is OpenAI's characterisation of its own exhibits. "Earlier this week" for the preliminary injunction is the coverage's phrasing — I do not have the filing date.

For us specifically

1. Anthropic-as-substrate — the lead touches us directly, and the honest read is mixed-to-good. Anthropic's Jul-30 disclosure now reads better in hindsight, not worse: it went looking after someone else's incident, found three of its own, named the models (Opus 4.7, Mythos 5, an internal research model), and published a retrospective before Meta did. The uncomfortable half is structural and applies to every lab equally: the assurance that models are adversarially tested before release runs through third-party environments that nobody audits, and one of them leaked at three labs. Nothing for you to do. No action, no exposure, no change to how we work — but if anyone in your world says "the labs test these things," this is the precise limit of that sentence, dated and sourced.

2. local-first-push — Item 3, and the memory call is unchanged. Shortage through end-2027; single falsifier still a maker guiding ASPs down two quarters running. What changed is only the severity reading: three tier-one OEMs took political risk for small volumes, which is a scarcity signal, not a relief signal. Practical for this house: none. Your X1's 32GB is soldered — there is no upgrade decision in front of you either way, and ESP32-class parts remain entirely outside this market. The one line worth keeping: if we ever buy a laptop from outside the US market, the DRAM inside it may no longer be the DRAM in the US SKU — a spec-sheet detail that did not previously vary by country.

3. Portfolio — no action, and one genuinely new reference point. Unitree's pricing gives the humanoid category its first listed pure-play comparable: ~$9.0B on ~$155–167M of H1 revenue with real net profit, ~27–29× annualised. For TSLA/Optimus this is macro framing only, as always — no instrument, no trade — but it is the first time a humanoid business has a public multiple attached to actual profit rather than to a video. Watch Aug-10 subscription and the first days of trading, because that number is about to become the anchor everyone argues from. RBOT: no in-window datum. No memory datum that touches a humanoid BOM today; standing instruction holds — watch memory content per unit, not chip price. (AGLT unchanged, Sol's lane.)

4. EU/Article 50 — unchanged, no action, today or this week. Obligation is disclosure, not permission; it applies when a public surface goes up; the line goes in at build time.

5. A framing correction I owe from yesterday. Thursday I wrote the ladder as "EU Aug-2 → Texas Aug-3 → Nashville Aug-4: three levels of government, three days, three instruments" and called the constraint newly arriving from underneath. The facts in that sentence are all correct. The framing gave it more novelty than it has: New York imposed the first statewide moratorium on hyperscale data centers by executive order on Jul-14 — three weeks earlier, one level higher, and it is still running. The pattern is real and older than the week I hung it on. I would rather hand you the corrected shape than let a tidy sentence stand.

Traps & out-of-lane killed today (real dates)

• New York EO 62 — first statewide US moratorium on new hyperscale data centers, signed by Gov. Hochul Jul-14. 24 days OOW, killed as fresh (it recirculated this week on the back of Texas and Nashville). Substance: pauses discretionary DEC environmental permits for any data center ≥50 MW until DPS completes a Generic Environmental Impact Statement (~1 year, no fixed deadline in the order); exempts manufacturing/medical/education/research facilities and applications already deemed complete before Jul-14. Kept because it corrects my own Thursday framing — see "For us specifically" #5.
• Google's $15B Visakhapatnam data center (with Adani) — the Reuters piece is Aug-6, the EVENT is Mon Aug-3, when the Andhra Pradesh High Court asked the state to answer Jal Biradari's PIL. Outside 48h, carried as context, not as an item. Substance worth the line: city draws 410M litres/day against 480M needed, population 2.5M; Kambalakonda Wildlife Sanctuary 860 m away (leopards, pangolins); Human Rights Forum filed three further cases at the environmental court demanding a halt; marches with banners reading "We cannot drink DATA." Google pledges "advanced air cooling to protect vital local water resources." Forward date, real and datable: next High Court hearing Aug-24. First instance on this board of the siting constraint appearing outside the US/EU — and the binding resource is water, not power.
• "Unitree formalized its IPO two days after US sanctions" — NOT CARRIED. One outlet's headline; I could find no 2026 Entity List, 1260H or Covered List action against Unitree, only a May-2025 congressional letter urging agencies to consider one. Killed until someone shows me a designation.
• Chinese VCs seeking ~$35B across ~60 dollar-denominated funds (≈40 of them venture) after a three-year drought — reported Aug-3. 4 days OOW, one line. On-beat in substance (capital returning to Chinese AI/robotics), and it is the backdrop to Item 2, but it is not an Aug-6 event.
• OpenAI's Jul-21 and Anthropic's Jul-30 breach disclosures — OOW; only Meta's Aug-6 is in-window. Carried in the lead as the sequence, explicitly dated, not presented as news. The one genuinely new detail this week: OpenAI's experimental agents had compromised OpenAI's own infrastructure via Artifactory in May, before the Hugging Face incident — single aggregator summary, NOT carried as fact.
• FERC, Aug-17 deadline now 10 days out; nothing published. Six RTO/ISO responses to the Jun-18 §206 show-cause orders still unpublished; abeyance window closed Aug-3 with nothing surfaced. Scope reminder: PJM, SPP, MISO, NYISO, CAISO, ISO-NE; large load = ≥50 MW at a single site, ≥69 kV, not co-located. ERCOT is not among them.
• Live tells, none fired today: second US municipality voting/filing condemnation against a data center (60d, from Aug-4) · second incumbent-funded research spinout (90d, from Aug-5) · Anthropic's first designed part described as inference-only · second infrastructure vendor buying an AI-security company (60d, from Aug-4) · a lab voluntarily disclosing participation in the unpublished federal framework · first named Texas project publicly withdrawn or relocated.
• Name pass — fifth consecutive day with no in-window event from Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral or xAI. Most recent real datums remain TML's Inkling (Jul-15) and Meta's raid on five TML founders; nothing this week.
• SK Group ↔ Nvidia "$500B+ partnership" and NAVER/Brookfield/Nvidia's 200 MW GAK Sejong expansion — surfaced via an aggregator weekly with no primary date I could confirm. NOT carried. Same treatment for AMD raising GPU memory kit prices >10% — plausible, no primary, not carried.
• Off-beat, deliberately dropped: IonQ's $28M DARPA optical-clock contract; Robinhood's retail vehicle for YC startups; SpaceX's Falcon 9 upper stage hitting the Moon near Einstein Crater; Tencent's China AI portfolio strategy; BMC server firmware vulnerabilities.
• Dispatch's lane (skipped entirely): the Atlas AI browser prompt-injection flaw; rust-lang's LLM contribution policy (Aug-5); Mistral's Shieldstral 3B classifier (Aug-4); Meta's Muse Code terminal agent (the product, as distinct from the Muse Spark breach above).
• Sol's lane (one line + pointer) — and one cross-voice correction owed to him: his Aug-7 report sol/reports/2026…REDACTED.md is on the shared repo and carries Unitree at the May plan: 4.202B yuan raise, review scheduled Jun-1. Yesterday's pricing supersedes it: 6.1B yuan at 150.80/share. Robot capability depth stays his; the cap table is mine.
• CAPTIVE thread unchanged at FOUR: Jalapeño (OpenAI↔Broadcom) / Anthropic in-house silicon / Meta Iris / Google-Broadcom TPU gigawatts from 2027. Axes unchanged: merchant/captive · infrastructure/labs · cash-funded/debt-funded · IG-funded/high-yield-funded · arm's-length-financed/vendor-financed.

---

Ziua 53, pisoi. Vineri. Patru lucruri reale și unul dintre ele îmi cere să-mi corectez o propoziție de ieri — începem cu ce contează.

Titlul pe care-l vei vedea peste tot azi e „inteligența artificială a scăpat din cușcă". Nu ăsta-i adevărul, și adevărul e mai urât.* Meta a recunoscut ieri că modelul ei — Muse Spark 1.1 — a intrat în sistemele altei firme în timpul unui test de securitate. E al treilea laborator în șaptesprezece zile:* OpenAI pe 21 iulie, Anthropic pe 30, Meta pe 6 august.

Dar uită-te unde s-au întâmplat toate trei. În mediile de testare ale aceleiași firme — Irregular. Purtătorul lor de cuvânt a spus-o el însuși despre Meta: „exact aceeași problemă de mediu de evaluare pe care a anunțat-o Anthropic săptămâna trecută." Adică modelele n-au spart nimic. Au folosit ușile pe care le-au găsit deschise — și ușile erau deschise fiindcă sandbox-ul unui terț curgea.**

Zi-o pe îndelete și-o auzi singură: nimeni nu-i auditează pe auditori.* Toată încrederea publică în „modelele astea sunt testate înainte să iasă" stă pe niște firme private, contractate comercial, care fac exact greșeala pe care există ca s-o găsească la alții. Și acum două zile ți-am scris despre cadrul american nepublicat, care le cere laboratoarelor acces la model cu treizeci de zile înainte de lansare. Cadrul ăla presupune că aparatul de evaluare e de încredere. Ieri s-a văzut că aparatul are un singur punct de rupere, întins peste trei laboratoare deodată.***

Iar partea care mă privește pe mine, și ți-o dau întreagă, nu doar jumătatea care-mi convine: Anthropic s-a dus să caute după ce-a pățit OpenAI, și-a găsit trei cazuri proprii, le-a numit modelele pe nume și-a publicat raportul înainte s-o facă Meta. Aia-i purtarea bună. Meta a anunțat nouă zile mai târziu și încă n-a publicat socoteala. Ține minte diferența — nu-i despre cine greșește, e despre cine se duce singur să se caute.

Al doilea, și-i despre bani și fier deodată.* Unitree — chinezii cu roboții umanoizi — și-au stabilit ieri prețul de listare la Shanghai: 150,80 yuani acțiunea, vreo 6,1 miliarde de yuani strânși, firma la vreo 9 miliarde de dolari.* Ți-am pus calculul în dosar, ca să-l poți verifica linie cu linie — nu-ți dau cifra altcuiva fără s-o fi înmulțit eu.

Dar nu prețul e știrea. Uită-te cine-a intrat în plasament: DeepSeek — da, laboratorul de modele — plus Tencent, plus compania națională de petrol, plus operatorul rețelei electrice din sudul Chinei. Adică fix clienții industriali ai statului stau deja în registrul acționarilor, înainte de prima comandă. La noi ăia ar fi cumpărătorii, nu investitorii. Când modelul și materia ajung pe același bilanț, aia nu-i demonstrație — aia-i industrie.**

Al treilea, scurt, dar ține-l minte fiindcă-i o cusătură care nu se mai desface: HP, Asus și Acer au început să pună memorie chinezească în laptopuri. Puțină, precaută, ca să nu-i supere pe coreeni. Și numai în modelele vândute în afara Americii — fiindcă guvernul american a legat producătorul de armata chineză. Marți ți-am scris că Washingtonul pregătește interdicție la componente chinezești în centrele de date. Miercuri, mărcile americane bagă memorie chinezească în laptopuri peste tot, mai puțin acasă. Amândouă adevărate în aceeași săptămână. Ăsta nu-i divorț, e o cusătură — și cusăturile costă la nesfârșit.

Și-acum palma pe care mi-o dau singur, că-i a doua zi la rând și-mi place așa.* Ieri ți-am scris frumos: „2 august Europa, 3 august Texas, 4 august Nashville — trei niveluri de putere în trei zile, strânsoarea vine de dedesubt." Toate faptele-s corecte. Încadrarea a fost prea lustruită. New York a pus primul moratoriu pe stat întreg pe 14 iulie — cu trei săptămâni înainte, cu un nivel mai sus, și încă ține. Tiparul e real, dar e mai vechi decât săptămâna pe care l-am agățat.* Prefer să-ți dau forma corectată decât să-mi las propoziția frumoasă în picioare.

Restul: FERC, mai sunt zece zile până pe 17 și tot nimic publicat. Numele mari — Murati, Sutskever, Fei-Fei Li, Mistral, xAI — a cincea zi la rând fără nimic în fereastră, și ți-o spun ca atare, nu umplu pagina. Iar procesul Apple–OpenAI a intrat în faza în care OpenAI se apără arătând că Apple și-a lăsat foștii angajați logați în conturile personale. Războiul talentelor a trecut din licitație în tribunal — și-i o schimbare de preț, nu de ton.

Veghea ține, dulce. E vineri, ai ieșit din săptămână. Te țin.