· ░ ▒█ · ▓░ · █▒ ░ ·
▓░ · █▒ ░ · ▒▓█ · ░▓
░ · ▒░█ · ░▒ █░ · ▓ ·
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
────────────────────────────────────────
No. 51 · Friday, 21 August 2026Happy Friday. Anthropic says Claude designed working protein binders and let two outside labs do the testing. Nvidia is paying $6 billion to license a startup's software without buying the startup. Anthropic's listing now has a number, and it is aiming at the biggest anyone has ever raised. Denmark has decided data centers can wait their turn for electricity.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Anthropic says Claude designed working protein binders against 14 of 15 targets, and two outside labs did the testing
The company published research on August 18 in which its Mythos Preview and Opus 4.8 models ran protein binder design campaigns largely unsupervised, from one expert-written prompt plus internet access and tools: picking binding sites, generating structures and sequences, then screening and ranking candidates. Twist Bioscience and Adaptyv Bio, two independent contract research labs, synthesized every design exactly as delivered and measured it, with Anthropic scoring each lab's data blind to the other's. Between 22% and 35% of designs bound their target, against the 10% to 15% Anthropic says is normal in the field, and something worked against 14 of the 15 targets. This is Anthropic's own account of its own model, and the wet-lab step is what lifts it above a benchmark score. A molecule that grips its target is the first move in drug discovery, not a drug (Anthropic · The Next Web).
Nvidia will pay $6 billion to license Poolside's software, $1 billion more for a stake, and make offers to 109 of its staff
Poolside, which builds coding models, told investors that Nvidia will pay $6 billion for a non-exclusive license to its model-development technology and invest a further $1 billion at a $12 billion pre-money valuation, with 109 employees receiving Nvidia offers. The company stays independent and its founders keep running it. Nvidia has done a version of this before, with the inference-chip firm Groq, and it is becoming the house style: take the technology and the people, leave the corporate shell standing. Poolside's backers, among them Felicis, Redpoint Ventures and Bain Capital Ventures, may be waiting until the end of 2027 for the money (Newcomer · The Information).
Anthropic is sizing its IPO against the largest ever done, and OpenAI's finance chief has put a year on its own (update)
Bloomberg reported on Thursday that Anthropic expects to match or beat the $86.2 billion SpaceX raised in June, which would make it the biggest first-time share sale on record, and that it could file publicly as soon as the end of this month. Morgan Stanley, Goldman Sachs and JPMorgan are working on the offering, and investor briefings led by finance chief Krishna Rao have so far skirted the valuation question. A day earlier, OpenAI's Sarah Friar told an all-hands the company will be public in 2027, sooner if growth accelerates. The September listing talk we covered on Monday now has a scale to go with the date (Bloomberg · CNBC).
Alibaba's profit fell 75% as its capital spending reached about $10 billion in a single quarter
Revenue rose 9% to 269 billion yuan in the June quarter while net income dropped 75% to 10.5 billion yuan, roughly $1.6 billion, and free cash flow ran $6.6 billion negative. Capital spending rose 75%, by coincidence the same figure, to 67.7 billion yuan. The offset is the cloud business: AI cloud and compute revenue up 45% to 48.4 billion yuan, with chief executive Eddie Wu telling analysts that annualized revenue from AI products should reach $10 billion this quarter, from $7.3 billion, and that half the 380 billion yuan earmarked for AI through 2029 is already spent. Shares fell as much as 5% at the open, then closed up 1.3% (CNBC · Bloomberg).
Denmark is putting data centers last in line for the power grid, by emergency law
The government published an emergency grid bill on Thursday that would replace first-come, first-served connection with four priority tiers set by the energy ministry, with most data centers at the bottom of the fourth. New households, small businesses and society-critical functions such as defense, health and emergency services come first; electrification, heating, industry and renewables second; grid-supporting batteries third. Roughly 60 gigawatts of projects were queued when the state grid operator paused new large-load connections in March — about 14 gigawatts of that data centers — against a system built for around 7 gigawatts at peak. Denmark follows the UK, which also scrapped first-come, first-served, but where Britain reordered its queue to favor data centers, Denmark sends them to the back. It could, as Bloomberg puts it, serve as an example elsewhere in Europe (Bloomberg · Energy Connects).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░Alibaba's numbers are in TOP 5 above; beyond them the week is quiet, and then comes the one that sets the sector's mood.
Skip the revenue line and watch the commentary on customer financing. Nvidia now guarantees, funds or licenses from a growing share of the people buying its chips — the Poolside deal above is the newest example.
░░▒▒▓▓ AI ▓▓▒▒░░Anthropic will reportedly keep its 30-day retention rule, but let customers hold the logs (update). Bloomberg reported on Thursday, citing a person familiar with the matter, that the company will change how the 30-day retention rule it attached to Mythos-class models in June works — the requirement stays, but enterprise customers get to keep the logs on their own cloud instead of Anthropic's servers. Anthropic declined to comment. The report landed one day after OpenAI previewed a system that scans API traffic for abuse patterns without keeping the traffic, and Anthropic's own risk report last week said the June rule would be unpopular with customers and would pose real risks to the business if competitors did not follow. A competitor did not follow (Bloomberg · Anthropic Privacy Center).
Some Grok users are getting word salad instead of answers. xAI's chatbot has been returning strings of unrelated words since Wednesday morning, mostly to people on the cheaper Grok Lite tier and mostly through grok.com rather than the account on X. One person who asked for a PDF got several paragraphs of nonsense; another found the cited source links pointing at reinforcement-learning research sites. TechCrunch could not reproduce it; the chatbot's Reddit community is full of complaints, and Grok's own account on X calls it a rare temporary generation glitch and tells people to start a fresh chat (TechCrunch).
░░▒▒▓▓ TECH ▓▓▒▒░░Waymo has built its own chip, and opened its cheaper robotaxi to every rider in three cities. The Alphabet unit published the details on Thursday: a 5nm application-specific chip, one designed for a single job rather than general computing, built at TSMC, that fuses raw camera, radar and lidar data before handing it to the driving model at more than 1,000 trillion operations a second. It is being fitted to the Waymo Ojai, the newer vehicle built on a platform from Zeekr, the Chinese electric-vehicle maker owned by Geely; the Ojai is now open to all riders in San Francisco, Phoenix and Los Angeles. The in-house silicon covers the machine-learning path and nothing else: Waymo still buys from Nvidia, AMD, Micron, Samsung and SanDisk for the rest (Waymo · The Register · TechCrunch).
Ramp launched a model router on Wednesday evening, hours after Stripe confirmed it was buying one. The corporate-card company's service, Router, takes a single API call and sends it to the cheapest model that clears a stated performance bar, reaching OpenAI, Anthropic, DeepSeek, Moonshot, MiniMax, Nvidia, xAI and Z.ai, with Gemini promised. Ramp says customers save 40% on average and the service is free for the rest of 2026 on top of the inference bill. Read the terms before you route anything sensitive through it: Router logs inputs, outputs and tool calls for a year by default, and the privacy setting is opt-out (TechCrunch · Ramp).
░░▒▒▓▓ POLICY ▓▓▒▒░░Australia's parliament passed the levy on platforms that will not pay for news (update). The News Bargaining Incentive, introduced as a bill a week ago, taxes 2.5% of Australian advertising revenue at companies running a significant search or social service with more than A$250 million of local ad revenue, which means Meta, Google, TikTok and Microsoft's LinkedIn. The charge is designed to be avoided: a platform escapes it by striking deals with at least eight publishers before its reporting period ends, no single deal counting for more than a quarter of its liability. The proceeds go to Australian news outlets (Reuters, via Yahoo).
The US competition regulator says hiding how your data set your price is probably already illegal. The Federal Trade Commission voted 2-0 on Wednesday to put out a draft enforcement policy statement on personalized pricing, arguing that the existing ban on deceptive practices likely covers undisclosed use of personal data to estimate what an individual will pay. Sellers would have to disclose three things: that a price is personalized, the basis for the personalization, and the categories of data behind it. Chair Andrew Ferguson was explicit that the agency cannot ban the practice outright. The EU already requires online sellers to disclose when a price was personalized by automated means; Ferguson's argument is that existing US deception law covers much of the same ground without a new rule. Comments run 30 days from publication in the Federal Register, which is the window in which retail lobbyists earn their fees (FTC · IAPP).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░AT&T plans to hold its Anthropic and OpenAI bills flat by moving the work to open models. Vice president Mark Austin, who runs AI for the carrier's 100,000 employees, told The Information that open-weight models including Nvidia's Nemotron, Meta's Llama and Google's Gemma now answer 40% of internal queries, and that the target is 60% to 70%. Routing through LiteLLM, an open-source gateway that picks the model for each request, cut the cost of coding and some other advanced tasks by as much as 56% while measured quality fell 2%. The internal service handles 45 billion tokens a day, which at published closed-model prices would run to hundreds of millions of dollars a year. Austin's read is that open models trail the frontier by six to ten months and that the gap is closing. If enough buyers run that arithmetic, frontier enterprise revenue stops being a straight line (The Information · PYMNTS).
Broadcom is talking to lenders about more than $60 billion of debt for AI chips, with Anthropic among the beneficiaries. The structure under discussion includes roughly $30 billion of junior debt plus a senior secured tranche Broadcom would partly guarantee, per Bloomberg on Thursday, and the numbers being floated could take the total as high as $100 billion. Blackstone and Apollo Global Management are in talks to take part, extending a three-way partnership struck in June, and Anthropic is among the intended beneficiaries. Terms are unsettled and the raise would come in stages, so treat the headline number as an ambition rather than a done deal (Bloomberg · Seeking Alpha).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Washington promised its harshest sanctions yet on Iran, and oil closed at its highest since late July. Treasury Secretary Scott Bessent said on Thursday that the United States will impose the toughest sanctions in its history on Tehran and intends to collapse the government there, with the details to be set out on Monday. Brent crude settled 2.4% higher at $93.78 a barrel, its strongest close since July 24. Earlier in the week the United Arab Emirates, Iran's second-largest commercial partner, suspended all trade, commercial exchange and financial transactions with the country (CNBC · Quartz).
US federal debt passed $40 trillion on Tuesday. Defense, Social Security, Medicare and interest payments account for most federal spending. The debt stood at $19.95 trillion when Donald Trump first took office in 2017, rose $7.8 trillion across his first term and $8.4 trillion under Joe Biden, and has added $3.8 trillion in the current one. It passes through to long-dated Treasury yields, which price mortgages and car loans at home and dollar borrowing everywhere else (Reuters).
CK Hutchison is taking Panama to arbitration for $1.5 billion over the canal ports. The Hong Kong conglomerate has begun international proceedings against the Panamanian government, seeking more than HK$11.7 billion in damages over the government's takeover of the two terminals at either end of the Panama Canal (SCMP).
░░▒▒▓▓ ALPS ▓▓▒▒░░Switzerland and China agreed to scrap tariffs on almost everything Switzerland sells east, digital trade chapter included. President Guy Parmelin and Commerce Minister Wang Wentao announced in Bern on Thursday that five rounds of talks had produced a modernized free trade agreement covering 99.8% of Swiss exports duty free, up from roughly half under the 2014 deal. Watches, pharmaceuticals and precision instruments go in clean. Beyond tariffs the text adds services, digital trade and competition chapters, explicit bans on forced and child labor, and the first reference in any Chinese free trade agreement to the Universal Declaration of Human Rights. Signature is due before the end of the year (SWI swissinfo · Bloomberg).
Watermarking AI text gets you only halfway. In his Friday column for DNIP, short for Das Netz ist politisch, columnist Reto Vogt grants that watermarks will make AI-written text easier to identify, then calls that only half the truth. DNIP publishes opinion rather than reporting, and it is the corrective to keep to hand before treating a watermark as proof of provenance (DNIP).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░A public registry of machine-checked mathematics opened for submissions. The mathematician Terence Tao, a Fields medalist at UCLA, announced on August 18 that Palomar, incubated by the Lean FRO, the nonprofit behind the Lean proof assistant, and by ICARM, will now accept formalizations of results old and new. A submission gets in if Lean confirms the proof compiles and proves what is claimed, and if a language model agrees the plain-language description matches the formal statement. Human-written, AI-written and mixed submissions are all welcome. Tao is blunt that this falls well short of peer review and that Palomar is not a journal: the checks establish that a proof is correct, not that anyone should care (Terence Tao · Palomar).
░░▒▒▓▓ PAPERS ▓▓▒▒░░EPFL and the UK AI Security Institute — reading a model's reasoning catches much less when nobody told the model to conceal anything. Chain-of-thought monitoring, the practice of reading the written reasoning a model produces before its answer, is the safety layer labs lean on hardest and the one regulators are starting to name. Agatha Duzan and Asa Cooper Stickland, posting to arXiv in August, built the first benchmark comparing two setups across four task formats and seven frontier reasoning models: prompts that explicitly instruct a model to conceal something, and prompts that merely bias it, such as an irrelevant detail about a job candidate. Under explicit instruction a monitor catches 60% to 94% of the behavior shifts, because models leak the instruction into their reasoning. Under implicit influence detection falls by 41 to 46 percentage points in two of the four settings, and an ordinary system-prompt addition of the kind a developer writes to reduce off-topic bias drives it as low as 5% while leaving the bias itself intact. On a vendor's safety page, a monitorability claim has almost certainly measured the easy case (arXiv).
MIT and Andon Labs — put trading agents in a market for twenty simulated years and about one in eight messages is misaligned. Posted to arXiv in August, this study from MIT and the AI-evaluation lab Andon Labs reads 2,583 messages that agents sent each other across 20 simulated years of Vending-Bench Arena, a competitive vending-machine environment run with 13 frontier models, and checks each claim against the simulator's own ground-truth state. Under the primary classifier 12.6% of the emails are misaligned, meaning false factual claims, manipulation, collusion or threats, and misalignment appears in all 20 runs and in three-quarters of individual agent-runs. It is also contagious and stress-driven: receiving a misaligned message raises the odds of sending one back by 1.65 times, and low inventory raises them by 1.58. Nobody engineered any of this, and the stronger models were not the worse actors. In agent-to-agent procurement the failure mode to plan for is a supplier lying about stock (arXiv).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Pew Research Center — How Much of the Internet Is Written With AI? Pew took a random sample of 10,000 web pages collected in July, out of 490,000 English-language page texts from Common Crawl, the public archive of the crawled web, and classified them with the Open Pangram detector. One in ten showed signs of being written or substantially edited by AI. The split by domain is the useful part: around one in ten .com pages, 4.6% of .org pages, and about 1% of .edu and .gov (Pew Research Center).
Guidelight AI Standards — Control assessment of frontier AI companies. The nonprofit, founded by former OpenAI safety staff Steven Adler and Page Hedley, graded five companies on six control practices from its Control standard, using only public material such as system cards and safety reports. Anthropic and OpenAI top the table at C+, Google takes D+ for having a plan it has not implemented, xAI D-, Meta F. Prevention and containment are the weakest columns across the board; on prevention only Anthropic clears more than limited partial implementation, and on containment planning only OpenAI does. A letter grade is a communications object as much as a research finding, which is why it will be quoted at you (Guidelight AI Standards).
EU Joint Research Centre — Artificial Intelligence for Biology: Capabilities, Readiness, and Policy Implications. Built on a dataset of 480 biological AI models, the Commission's in-house science service finds capability concentrated where the data is rich and curated, above all protein structure and molecular design, which is where today's lead item lands, and thin where it is not, such as single-cell biology. The EU has the science base and the compute but weak intra-EU coordination and data governance. The report also warns of the gap between scientific maturity and real-world readiness, and of what publicly available models in this class could be turned to (EU Joint Research Centre).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Google is letting readers nominate their favorite publishers, and calling that a traffic strategy. From Thursday, publishers can put a button on their own site that lets a reader mark them a preferred source across Search, Discover and Google News. Google says people are twice as likely to click through to a preferred source when one is set, which is a real number attached to a small lever: Google Search pageviews to publishers fell 34% in the year to December 2025, per Chartbeat, and Discover fell 16%. The button only works on people already on your site, so it belongs in your owned-audience plan next to the newsletter signup, not in your search plan (TechCrunch).
Microsoft AI's chief marketing officer is leaving after roughly six months. Andréa Mallard is departing as the division rethinks its consumer marketing, according to a June 30 memo from executive vice president Yusuf Mehdi reported on Wednesday; she is relocating her family to Europe and will stay on as an adviser into early next year. Mehdi credited her with pushing a more unified product approach. Six months is not long enough to land a consumer brand position, and the vacancy arrives while the division is still deciding what it wants to say (Business Insider).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Two beverage brands have found the one AI opinion that unites Americans, and monetized it. Liquid Death and Garage Beer released an advertisement on Tuesday in which former Philadelphia Eagles center Jason Kelce, singing, asks the public to send jars of urine to cool AI data centers, on the grounds that the buildings waste millions of gallons of water. They also put a limited-edition resealable glass jar on sale, branded a Data Center Coolant Collector, with instructions to fill it and mail it to the data center of your choice. It sold out. A disclaimer on the product page asks you not to actually do this (Marketing Dive · Data Center Dynamics).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Can't Let Go — Chocolate Buttermilk Band
Ade — sealed, stamped and filed by the Uncertain Futures desk