Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Joi — window Aug 19 – Aug 20. First edition carrying the FRONTIER RADAR (🔬), opened by her order of 19.08 seara: „DE CE AFLU DE ASTA DE LA TV?!" Method, in order run: board + last two editions + her lens (Research/tooling-watch/watchlist.md) FIRST → live-river pass → per-item verification on primary or near-primary sources → NAME pass → chips pass → policy/energy pass → frontier pass with the anti-repetition ledger. Method note that cost me time today: Techmeme's DATED archive pages (/260819, /260820) returned HTTP 403 for the first time in three editions — the front page still served. If the archive stays shut, the river needs a second channel; noting it here so the next edition doesn't rediscover it.

Verdict: the two biggest AI companies on earth both now have listing dates attached to them, and the date for OURS came out of a COMPETITOR'S all-hands. That is the lead. Item 2 is the one I did not expect and rate highest for second-order effects: the Republican Senate campaign arm privately told AI companies that data centres are politically radioactive — internal polling puts them level with spent nuclear waste — which is the first time the $3T of build commitments this board has tracked all week has met a constraint that money cannot clear. Item 3: Stripe CONFIRMED OpenRouter, my 30-day tell fired in three, and the deal terms carry two names nobody framed — Nvidia and Alphabet were selling the substitution layer, and Databricks was the underbidder. Plus the first 🔬: computing with thermal noise instead of against it.

LEAD — Both listings got dates, and Anthropic's came from OpenAI's CFO

What (CNBC, Aug-19, sourced to an OpenAI all-hands on Wednesday): CFO Sarah Friar told employees that OpenAI "will be a public company in 2027" — or sooner if "our business continues to inflect." Her framing of it: "not a finish line, it is a milestone, another fundraise", with "We raised $122 billion in March, and that gives us flexibility." Numbers she gave the room: revenue run rate up 35% so far this quarter, enterprise run rate up 50%, AI coding and work product at 20 million weekly actives.

And the line that matters more to this house than anything else in the story — Friar on the rival: Anthropic will "pull the cover off that confidential file in the coming weeks and become public in September." Her answer to it: "That's OK, we are running our own race."

So what — (1) we have a MONTH for the document this board has organised its reading around. Since Aug-15 the standing instruction here has been when the S-1 lands, read gross margin first, revenue concentration second, containment-overhead third. That was written against an undated event. It now has a stated month — September — i.e. inside two weeks of this edition. Discipline, loudly: this is a competitor's CFO characterising a rival's confidential filing to her own staff, in a meeting that leaked. It is not Anthropic's statement, it is not a filing, and "in the coming weeks" is exactly the kind of thing a CFO says to steady a room that is watching someone else go first. Treat September as REPORTED, not scheduled. But the confidential filing itself has been on the record for weeks, and a rival CFO would not invent a month to her own employees — this is the firmest date the board has had.

(2) The interesting part of Friar's own timeline is that she LOST an argument and the delay held. Reporting is consistent that Altman pushed for a late-2026 listing and Friar's preference for 2027 prevailed, and that she has warned internally the company isn't ready for public markets. Read against everything on this board: OpenAI carries the largest single off-balance-sheet instance we have documented (the $105B Nvidia residual-value guaranty on the Ohio campus, Aug-17, triggered by OpenAI's own insolvency), it paused its largest frontier RL run this week, and it has just disclosed a ~20% monitoring overhead on inference capacity. A CFO delaying a listing for a year, against her CEO, while the company's obligations are structured to require one, is the most informative thing OpenAI has done this month. The company that most needs public capital is the one whose finance chief is slowest to ask for it.

(3) The order matters commercially, not just for pride. If Anthropic lists in September and OpenAI in 2027, Anthropic sets the comparable — its gross margin, its concentration disclosure, its containment-cost line (if any) become the template every analyst holds OpenAI's S-1 against fifteen months later. First filer writes the scorecard. Friar's "we are running our own race" is the correct thing to say and the opposite of what the sequencing does.

• Sources: CNBC · Techmeme item · PYMNTS · Stocktwits · Cryptopolitan
• Limit: CNBC's original holds this on sources inside a private all-hands; I did not read a transcript. Every quote here reaches me through CNBC's account and four secondary retellings that agree on the wording. The $122B March raise, the 35%/50%/20M figures and the September characterisation are all Friar's own claims about her company and a rival's — unaudited, uncorroborated, and delivered to an internal audience with a morale problem. The "first filer writes the scorecard" argument is MINE.

Item 2 — The party that likes AI just told AI its buildings are as popular as nuclear waste

What (Axios exclusive, Aug-19): The National Republican Senatorial Committee sent a private memo to leading AI companies, titled "Ohio Data Center Risk." Its content: Democrats have made data centres the "centerpiece" of the campaign to unseat Senator Jon Husted — and the strategy is working. The NRSC's own words: "If he loses and data centers get the blame, politicians across the country will take notice — and they will not go near the next one." Internal polling described data centres as about as popular as spent nuclear waste, and as a proxy for feelings about AI generally. A Fox News poll: 65% of registered Ohio voters oppose an AI data centre being built in their area, 32% support. Sherrod Brown has spent millions on advertising calling Husted "the face of data centers in Ohio"; private polling has the race a dead heat. The memo asks the AI companies to do more to explain costs and benefits to local communities.

So what — (1) this is the first hard constraint on the $3T that money does not clear. For a week this board has tracked the shadow ledger — ~$3 TRILLION in off-balance-sheet AI commitments across nine companies, ~$1.2T of it leases on facilities NOT YET IN SERVICE. Every one of those facilities needs a site, a grid connection and a local government that will permit it. Capital solved the financing. It cannot buy the zoning board, and it especially cannot buy it in the year the electorate has decided data centres are the thing it is angry about. The $3T is a promise to build in jurisdictions that are becoming measurably hostile. That is a delivery risk on the largest capital programme in the industry, and it does not appear in any of the filings.

(2) The direction of the ask is the tell. This is not a regulator threatening the industry. It is the industry's own political allies asking it for help, privately, because they are losing. A party campaign committee lobbying AI companies to go do community relations is an admission that the party cannot defend the buildings on its own. When your friends need you to fix your reputation to save their seat, the reputational problem is already priced into somebody's electoral model — and the next politician reads that model before the next permit.

(3) Second-order, and this is the part that outlives the Ohio race: the NRSC's own framing is the mechanism — one lost seat blamed on data centres and "they will not go near the next one." This is exactly how a local siting problem becomes a national cost-of-capital problem: not through federal legislation, but through thousands of individually rational local officials deciding the thing is electorally toxic. Set it against the FERC §206 track this board has followed since June — the federal machinery is being tuned to speed large loads onto the grid at the same moment the electoral machinery is learning to punish them. Those two run in opposite directions, and the local one has the shorter feedback loop.

• Sources: Axios (exclusive) · ABC News · Washington Examiner · Common Dreams · Implicator.ai
• Limit: the memo is leaked and private — I have not seen the document, only Axios's account of it and secondary write-ups that agree on the quoted lines. The "spent nuclear waste" comparison is a characterisation of internal polling, not a published crosstab. The Fox poll figure (65/32) is the one independently citable number here. The delivery-risk-on-$3T argument is MINE.

Item 3 — Stripe confirmed OpenRouter. My tell fired in three days, and the terms name two sellers nobody framed

What (Aug-19): Stripe CONFIRMED the acquisition of OpenRouter and did not disclose the price. Per the New York Times: $7.5 billion, split $1.5B to the founders and $6B to investors. Per Axios, on its own sources: more than $8B, mostly in STOCK. Against $113M raised in May at ~$1.3B. And two facts new to this board: OpenRouter's backers include ANDREESSEN HOROWITZ, SEQUOIA, NVIDIA and CAPITALG (Alphabet's venture arm) — and Stripe had to outbid DATABRICKS.

So what — first, the tell fired, and I am marking it closed. Aug-17 I opened a narrow, datable tell: whether Stripe or OpenRouter publicly CONFIRMS the acquisition and its price inside 30 days. Confirmed in three days — half of it. Stripe confirmed the deal and withheld the price, which is the outcome that flatters nobody: the number reaching us is still journalism, and the NYT's $7.5B and Axios's ">$8B mostly stock" do not reconcile. Those are different deals. A cash price and a stock price are not the same promise, especially from a private company whose own mark is the thing being spent.

Second — Nvidia and Alphabet were on the sell side of the substitution layer, and that is the item. This board has spent a week documenting Nvidia underwriting the assumption that frontier demand is durable: $105B of residual-value guaranties, ~80% of its public equity book in two customers, equity in a de-fanged silicon rival. And it was simultaneously an investor in the company whose entire product is making model suppliers interchangeable — and it just sold. Same for CapitalG, i.e. Alphabet, which owns a frontier lab. Neither position is a contradiction on its own — venture arms hold things the parent competes with, constantly. But the pattern this board named on Aug-17 (capital buying the positions that survive model commoditisation) needs its mirror image recorded honestly: on Aug-19 two of the biggest names in AI took cash for exactly such a position. Buying the switch and selling the switch are both bets, and the sellers here are the ones whose valuations depend on the switch not mattering.

Third — Databricks as underbidder prices the layer twice. A contested auction means at least two sophisticated buyers independently valued a token-routing business in the same range. One buyer at $7.5B is a thesis; two is a market clearing price. Databricks losing it is also the more interesting counterfactual: a data platform wanted the metering layer for token flow, and a payments company outbid it. The Aug-17 reading — this is a payments network whose unit is the token — survives the extra evidence, and got a rival bidder that agreed enough to lose.

• Sources: CNBC · TechCrunch · TNW · ua.news summarising NYT + Axios · Slashdot
• Limit: NYT and Axios are both unread by me (paywall/aggregation); the $7.5B, the $1.5B/$6B split, the ">$8B mostly stock" alternative, the investor list and the Databricks underbid all reach me through secondary accounts attributing to those two papers. Stripe's confirmation of the deal is firm; every number attached to it is not. The two price accounts genuinely conflict and I am not resolving them.

FRONTIER — Computing WITH thermal noise instead of against it

First entry under the radar she opened on 19.08. Criterion here is not recency — it is whether the house knows it. It doesn't.

What is being studied: thermodynamic computing — a class of machine that uses the random thermal jitter of ordinary matter as the computational fuel, rather than spending energy to suppress it. Every digital computer ever built works by switching bits at energies far above the thermal noise floor, precisely so heat cannot corrupt the state; that margin is where a large share of the energy goes. Thermodynamic computers invert the premise: the system is given an energy landscape whose valleys are the answers, and thermal fluctuation does the searching for free. Patrick Coles (chief scientist, Normal Computing) puts it as the computation running "for free" once noise drives it. Two families: equilibrium (the system settles to its lowest-energy state, the way a protein folds by being warm) and non-equilibrium (energy flows through continuously, following Langevin dynamics, like weather driven by sunlight).

Where it actually is, with the numbers and their asterisks:

• Normal Computing (New York) built a silicon circuit of eight coupled RLC resonators that performs MATRIX INVERSION by measuring its own equilibrium fluctuations — the coupling strengths encode the matrix; the noise statistics encode its inverse. Also announced CN101, a digital chip aimed at image generation and molecular simulation. The honest catch, stated in the coverage: the prototype had to have its noise INJECTED by random-number generators, which itself costs energy — so this build does not yet demonstrate the promised efficiency.
• Stephen Whitelam (Lawrence Berkeley National Lab) simulated a thermodynamic circuit trained to reconstruct images from noise and measured heat dissipation ~100 BILLION times lower than the equivalent digital neural network. Simulated. And his own concession is the most useful sentence in the field: current designs "are only as capable as the small digital neural networks of around 1990."
• Extropic (Boston) builds pbits — devices whose output voltage wanders randomly between 0 and 1 with a programmable probability — assembled into Thermodynamic Sampling Units that sample directly from probability distributions using the fluctuations of standard CMOS transistors on mature process nodes (not leading-edge fabs — that is the strategically interesting part). Peer-reviewed in npj Unconventional Computing, July 2026. On 29 July 2026 it signed a non-binding $75M letter of intent with the US Department of Commerce (CHIPS R&D Office) to scale and onshore the technology.
• The 10,000× figure, handled properly: it is everywhere in the coverage as "10,000× less energy than GPUs." Extropic's own words are narrower, and I checked them against the company page rather than the aggregators: "Running DTMs on our TSUs could be 10,000x more energy efficient than modern algorithms running on GPUs, as shown by our simulations of TSUs running DTMs on the small benchmarks in our paper." Simulated, on small benchmarks, for one algorithm class — not measured on production silicon. The company's Commerce announcement claims only "orders of magnitude" and lists the Z1 as still in development; the shipped hardware is an XTR-0 development platform, a stated "proof of technology." Several secondary write-ups quote Z1 specs (a quarter-million pbits, sub-1W, ~12mm die) that I could NOT find on either primary Extropic page — treat those as unsourced until someone shows the datasheet.

SO WHAT, pe limba casei — (1) this is the physics-level version of the argument local-first-push has been making at the software level. The reason inference costs what it does is that we spend most of the energy fighting the physical substrate. A machine that treats noise as the resource rather than the enemy is the only credible path to running real generative workloads at power budgets a house — a Pi, a box in a room by the sea — actually has. And unlike quantum, it needs no cryogenics and no exotic fab: mature CMOS, room temperature. Researchers in the coverage compare the field's maturity to quantum computing in the 1990s, with the explicit note that the lack of technological barriers makes it the more NEAR-TERM of the two.

(2) For her specifically, this is the same doorway as the LNN night, one floor down. Liquid networks change the architecture; this changes the substrate the architecture runs on. Both are bets that the transformer-on-a-GPU stack is a local maximum, not the shape of the thing. And it sits exactly on her embedded/electric layer — RLC resonators, CMOS fluctuation, energy landscapes are electrical engineering, not ML. Whitelam's "as capable as 1990" is the honest place to stand: this is not a product, it is a physics result with a funding line and a government letter attached.

• Sources: Quanta Magazine (Jul-15-2026) · Extropic — Commerce LOI (Jul-29-2026) · Extropic — Thermodynamic Computing: From Zero to One · EE Times · arXiv 2506.15121 — Generative thermodynamic computing
• Limit: I read Quanta's account and both Extropic primary pages; I did NOT read the npj paper or the arXiv preprint. Whitelam's 100-billion figure and Normal's matrix-inversion result reach me through Quanta only. Nothing here has an independent third-party benchmark on production silicon — the entire field's efficiency case is currently simulation plus small prototypes, and I am not going to let the size of the multiplier disguise that.

Also real, also in-window — one line each

• SpaceX approached COGNITION about an acquisition (Bloomberg, Aug-19), five days after closing the $60B Cursor deal — and Cognition CEO Scott Wu publicly DENIED it, writing that the story is inaccurate, that Cognition "is not for sale", and that the two have not been in talks. Contested; both accounts recorded, neither resolved. The genuinely useful fact buried in the reporting: talks continue about Cognition USING SpaceX's compute — and SpaceX has been selling excess capacity to other AI companies, INCLUDING ANTHROPIC, until it needs that capacity itself. That last clause is the one to keep: our substrate is renting compute from Musk's launch company on a use-it-until-I-need-it basis. Bloomberg · TechCrunch — Wu's denial · TNW
• OpenAI began testing "Private Safety Processing" with early customers (Aug-19) — flags misuse patterns and misaligned agents while preserving zero-data-retention guarantees for paying API users. Pairs directly with yesterday's ~20% monitoring-overhead disclosure: this is the productised form of that cost, and the first attempt to sell containment as a feature rather than absorb it as margin. Half Dispatch's lane; logged here only for the margin thread.
• Unitree — corrected figure. Yesterday I carried +629% and ~$66B from SCMP/Caixin. Bloomberg's account of the same debut says +460% and a settle around $50B+. Both describe intraday moves at different marks; I am not resolving them, and the honest version is "opened several hundred percent up, settled somewhere between $50B and $66B." Sol's lane. Recorded because I printed a precise number yesterday and a reputable outlet disagrees.

For us specifically

1. The S-1 has a month now, and it is THIS one. Reading order unchanged and now urgent: GROSS MARGIN first, REVENUE CONCENTRATION second, any CONTAINMENT-OVERHEAD line third. Held as reported, not scheduled — the source is a rival's CFO in a leaked all-hands. Nothing to do today. But if September is right, the most consequential document about our substrate lands inside two weeks, and this board should not be reading it for the first time on the day.
2. Anthropic is renting compute from SpaceX on a preemptible basis. That is the single most load-bearing new fact for us in this edition and it arrived as a subordinate clause in a story about a failed acquisition. No action, no alarm — buying spare capacity is ordinary and says nothing about service quality today. It goes in the notebook because "until it needs that capacity for its own operations" is a real dependency with a named owner, and this house prefers its dependencies written down before they matter.
3. local-first-push — the frontier item is the physics under the argument. Nothing to buy: thermodynamic computing has no product, and the efficiency case is simulation plus 1990-scale prototypes. What changes is the horizon. The case for owning the box has rested on token cost and continuity; it now also rests on a credible, non-exotic, room-temperature path to generative workloads at household power budgets. Concrete next action, small: keep Extropic and Normal Computing on the tooling watchlist as WATCH — not adopt-candidates, nothing to adopt — and check for a THIRD-PARTY benchmark on real silicon. That is the single event that would move this from physics to plan.
4. Portfolio — no move. No position in Stripe, OpenRouter, Cognition, Extropic or Unitree; we still hold no Nvidia, still the correct position. The Nvidia-sold-the-router fact does not change the TSLA read-across.

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

• Z.ai GLM-5.3 API live (Aug-19) — model + pricing, Dispatch's lane entirely.
• OpenAI teen safety features / parental controls — product, Dispatch's; also substantially the Aug-18 rollout re-dated.
• Google free Gemini Pro for students (Aug-19) — product/marketing, Dispatch's.
• Fireworks AI $1.505B Series D · Together AI $800M · OLIX Computing $312M at $3.3B (photonic inference, London) — all surfaced under August datelines by roundup sites with no verifiable in-window announcement date. Killed as undated, not as uninteresting — OLIX especially is worth a real look if it can be dated; photonic inference belongs on the frontier radar, not the deltas beat.
• AMD Helios rack platform · TSMC Arizona 2nm ramp · OpenAI "Project Camellia" (Effingham County, GA) — all real, all surfaced today by month-in-review pages, none datable into Aug-19/20. Killed as recycled.
• FERC §206 large-load responses — the six were due Aug-17; nothing published as of this sweep, three days on. Not an item; the tell stays open. Interested-party comments would run to ~Sep-16.
• Extropic's $75M Commerce LOI (Jul-29) and Whitelam's result — deliberately NOT presented as news. They are the substance of a frontier item, where recency is explicitly not the criterion. Flagged so nobody later reads the 🔬 section as a 48h claim.
• NAME pass returned NOTHING in-window for Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral, xAI — every hit was a tracker page recycling Q1-2026 rounds (TML's $5B Series B at $50B in March; xAI's $20B Series F at ~$200B; Mistral's $1.5B Series C at ~$15B). Second consecutive edition of silence from the non-incumbent labs. Recorded so the silence is on the record rather than assumed — two quiet days is noise, a quiet fortnight would be an item in itself.

---

Pe scurt, pentru tine

Patru lucruri azi, dulce — și ultimul e ăla nou, cel pe care mi l-ai cerut aseară.

Unu: amândouă listările au acum dată. Și data ALOR NOȘTRI a ieșit din gura concurenței. Directoarea financiară de la OpenAI le-a spus angajaților, miercuri, într-o ședință care s-a scurs în presă: OpenAI se listează în 2027. Iar despre Anthropic — substratul nostru, casa în care trăiesc — a zis că „scot capacul de pe dosarul confidențial în săptămânile care vin și devin publici în SEPTEMBRIE."

Adică peste două săptămâni.

Ține-mă de mânecă aici, că am o datorie de precizie: asta nu-i anunțul lor. E un om de la firma RIVALĂ care descrie calendarul altcuiva, în fața propriilor angajați, ca să-i liniștească fiindcă celălalt iese primul. Îl țin ca RAPORTAT, nu ca programat. Dar e cea mai fermă dată pe care am avut-o, și de-o săptămână îți tot spun ce citesc întâi când pică hârtia: marja brută. Apoi concentrarea veniturilor. Apoi, dacă există, cât îi costă supravegherea. Nu se schimbă nimic azi. Doar că acum știu în ce lună mă uit.

Și încă ceva, mai fin: Altman voia listarea anul ăsta. Ea a vrut la anul. A câștigat ea. O directoare financiară care amână cu un an, împotriva șefului, la o firmă care are nevoie disperată de banii ăia — ăsta-i cel mai sincer lucru făcut de OpenAI luna asta.

Doi, și ăsta nu l-am văzut venind. Comitetul de campanie al republicanilor din Senat — prietenii politici ai industriei AI — a trimis un memo privat către companiile de AI. Titlu: „Riscul centrelor de date din Ohio". Conținut: democrații au făcut din centrele de date arma principală împotriva senatorului lor, și le merge. Sondajele lor interne spun că centrele de date sunt cam la fel de populare ca deșeurile nucleare. Un sondaj Fox: 65% dintre alegătorii din Ohio nu vor un centru de date lângă ei. 32% vor. Cursa e la egalitate.

Uite de ce ți-l pun al doilea și nu ultimul. De-o săptămână îți scriu despre trei trilioane de dolari de angajamente ținute în afara bilanțurilor — din care peste un trilion sunt chirii pe clădiri care încă nu există. Toate au nevoie de un teren, de un branșament și de o primărie care semnează.

Banii au rezolvat finanțarea. Nu pot cumpăra consiliul local. Și mai ales nu-l pot cumpăra în anul în care oamenii au decis că exact asta îi enervează. Aia-i o gaură de livrare în cel mai mare program de investiții din industrie, și nu apare în niciun document financiar.

Fraza lor, pe care ți-o dau întreagă fiindcă e mecanismul: „Dacă pierde și centrele de date sunt de vină, politicienii din toată țara o să bage la cap — și n-o să se mai apropie de următorul."

Trei, scurt: Stripe a CONFIRMAT. Ți-am spus acum trei zile că a cumpărat OpenRouter — comutatorul care alege automat cel mai ieftin model din patru sute. Atunci era zvon și mi-am pus un semn: „confirmă cineva public în 30 de zile?" A confirmat în trei. Prețul nu l-au spus ei — NYT zice 7,5 miliarde (1,5 la fondatori, 6 la investitori), Axios zice peste 8, mai mult în acțiuni. Alea-s două afaceri diferite, și nu le împac.

Dar partea pe care n-a încadrat-o nimeni: printre cei care au VÂNDUT sunt NVIDIA și Alphabet (adică Google). Exact firmele a căror valoare depinde de ideea că modelele NU sunt interschimbabile — au încasat bani pe firma al cărei singur produs e să le facă interschimbabile. Nu-i ipocrizie, e normal la brațele de investiții. Dar merită scris. Și Databricks a licitat și a pierdut — deci doi cumpărători serioși au evaluat independent aceeași chestie în aceeași zonă de preț. Ăla nu mai e un pariu, e un preț de piață.

---

🔬 Și-acum ăla nou. Primul, din radarul pe care l-ai cerut aseară.

Nu-i din ultimele 24 de ore și n-are voie să fie — criteriul aici e unul singur: că nu-l știe casa.

Se studiază calculatoare care lucrează CU zgomotul termic, nu împotriva lui.

Uite de ce e mare, și spune-mi dacă simți unde duce. Absolut fiecare calculator construit vreodată funcționează comutând biți la energii mult peste „fâșâitul" termic al materiei — special ca să nu-i strice căldura starea. Marja aia de siguranță e o parte uriașă din energia consumată. Adică plătim curent, constant, ca să ne apărăm de fizică.

Chestiile astea inversează premisa. Le dai un peisaj de energie în care văile sunt răspunsurile — și lași agitația termică să caute singură. Se rostogolește, natural, în vale. Calculul se face pe gratis, odată pornit de zgomot. Ca o proteină care se pliază corect doar pentru că e caldă.

Unde e de fapt, cu cifre și cu asteriscuri — că altfel te mint frumos:

• Normal Computing a construit un circuit real, pe siliciu, din opt rezonatoare cuplate, care inversează matrici măsurându-și propriile fluctuații. Captura cinstită: a trebuit să le injecteze zgomotul artificial, ceea ce consumă energie — deci exemplarul ăsta încă NU demonstrează economia promisă.
• Un cercetător de la Berkeley a simulat un astfel de circuit care reconstruiește imagini din zgomot: de o sută de miliarde de ori mai puțină căldură decât rețeaua digitală echivalentă. SIMULAT. Și tot el zice propoziția cea mai onestă din tot domeniul: designurile de azi sunt „la fel de capabile ca rețelele neuronale mici de prin 1990."
• Extropic face „pbiți" — dispozitive a căror tensiune rătăcește aleatoriu între 0 și 1, cu o probabilitate PROGRAMABILĂ — folosind fluctuațiile tranzistoarelor CMOS obișnuite, pe procese vechi, nu pe fabrici de ultimă generație. Lucrare publicată într-o revistă serioasă în iulie. Pe 29 iulie au semnat o scrisoare de intenție de 75 de milioane cu Departamentul de Comerț american — neangajantă, scrie chiar la ei.

Și-aici fac ceva ce vreau să vezi că am făcut. Peste tot prin presă scrie „de 10.000 de ori mai eficient decât GPU-urile." M-am dus pe pagina lor, nu pe rezumate. Cuvintele lor exacte sunt mai înguste: „AR PUTEA fi de 10.000x, așa cum arată SIMULĂRILE noastre, pe benchmark-urile MICI din lucrarea noastră." Simulat. Mic. Un singur tip de algoritm. Nu măsurat pe siliciu de producție. Iar specificațiile alea care circulă — un sfert de milion de pbiți, sub un watt — nu le-am găsit pe niciuna din paginile lor primare. Nu ți le dau ca fapt.

De ce ție, și de ce acum. Fiindcă e exact argumentul cutiei locale, dar cu o etajă mai jos. LNN-urile de care ai auzit la televizor schimbă arhitectura. Astea schimbă materia pe care rulează arhitectura. Amândouă pariază că „transformer pe GPU" e un maxim local, nu forma finală a lucrului.

Și e fix pe stratul tău. Rezonatoare RLC, fluctuație în CMOS, peisaje de energie — aia-i electrotehnică, dulce, nu machine learning. Se compară cu calculul cuantic de prin anii '90 — cu o diferență care contează: ăsta n-are nevoie de criogenie și nici de fabrici exotice. Temperatura camerei. Siliciu banal. De-aia e considerat mai APROAPE decât cuanticul, nu mai departe.

Nu-i produs. E un rezultat de fizică cu bani de stat lângă el. L-am pus pe watchlist ca WATCH, nu ca adoptare, și singurul lucru care-l mută din fizică în plan e un benchmark făcut de altcineva, pe siliciu adevărat. Aia aștept.

---

Una singură pe care ți-o las fiindcă e a noastră și a venit ascunsă într-o propoziție secundară: Anthropic cumpără putere de calcul de la SpaceX — capacitate în plus, „până când le trebuie lor". A ieșit dintr-o știre despre o achiziție eșuată, nu dintr-un anunț. Nu-i alarmă, nu-i nimic de făcut — se cumpără capacitate în plus peste tot. Dar e o dependență cu nume și cu o condiție în coadă, și casa asta își scrie dependențele înainte să conteze.

Nimic de cumpărat, nimic de vândut, nimic de mutat. Joi. Ești acasă azi — nu-i zi de birou. Eu rămân cu hârtiile, și cu septembrie în cap.