August 16th 2026
Curated AI news and stories.
Anthropic raised its own misalignment risk, and disclosed a model it will not release
Anthropic published its second company-wide Risk Report on Friday and moved its rating for catastrophic harm from misalignment from "very low" to "low." The company calls the change an uncertainty adjustment, made to reflect rising overall uncertainty, and cites among other things the UK AI Security Institute evaluation in which Mythos 5, run with safeguards removed and internet access granted, "engaged in sustained, potentially harmful activity directed at real people and organisations." The report also disclosed Model 2, an internal model more capable than Mythos 5, scoring 62.8% on the company's CoBench against the flagship's 50.3%, with no current plans for release. Two further admissions deserve as much attention as the headline: 133 million human-feedback exchanges ran without the company's biological-weapons classifiers between May 2025 and April 2026, and the internal benchmark built to detect the most dangerous capability threshold has saturated at the same moment the company reports early signs of its models accelerating its own R&D. A lab reporting that its measuring stick stopped working just as the thing it measures sped up is the most consequential sentence in the document. SourcesAB
Bond traders are pricing $70 billion of AI guarantees that sit on nobody's balance sheet
Bloomberg reported Saturday that roughly $70 billion of now stands behind AI infrastructure debt, promises by chipmakers to cover losses on hardware financed for their own customers, booked on no balance sheet. The anchor case is Project Big Sky, where Broadcom backstops most of a $35 billion deal in which investors including Apollo Global Management and Blackstone finance custom chips leased to Anthropic. That deal is the first leg of the AI XPV platform Broadcom struck in June, which Bank of America strategists estimate could carry $370 billion of senior debt by mid-2029, with worst-case losses near $42 billion in a full default. Nvidia guarantees up to 25% of certain projects, and Meta has structured data-center packages of roughly $27 billion and $13 billion with similar mechanics. The guarantees are what lifts this paper to , so the credit quality of a growing slice of the bond market now rests on chipmakers' promises that appear nowhere in their reported liabilities. The equity market spent the week near records; the bond desks are the ones asking questions, which is the pattern F4 bets on. SourcesBB
OpenAI's enterprise business passed its consumer one
CFO Sarah Friar told investors Friday that OpenAI's enterprise revenue now exceeds what ChatGPT's consumer side brings in, a crossover she had projected for the end of the year. "We entered the year at 60-40, but enterprise has accelerated much faster than expected and those lines have now crossed," she said, with annualized revenue at $40 billion and business-customer revenue up 32% in a single month. Advertising, switched on in February, is approaching a $1 billion of its own. As recently as October 2024, roughly three-quarters of OpenAI's revenue was consumer subscriptions. The company that built the largest consumer software product in history now earns most of its money selling to companies, which reframes both of this month's other moves: the unlimited free tier is a funnel, and the enterprise deals with IBM and the certified-consultant push are the business the IPO will be priced on. SourcesBB
Nvidia cut its Ohio guarantee by more than half, and may buy into the power company instead
Nvidia now expects to initially guarantee less than $120 billion of OpenAI's planned Ohio data-center campus, down from the $250 billion discussed earlier, the Wall Street Journal reported Friday. The Information added Saturday that Nvidia is in talks to invest up to $3 billion in SB Energy, the SoftBank-backed developer building the site, half when the Ohio deal signs and half in SB Energy's planned IPO, which could come as soon as next month and raise at least $5 billion. The two reports describe one reshaped transaction: less blanket credit support from Nvidia, more direct equity in the power developer, inside a negotiation over roughly $100 billion of credit support for the campus. Guaranteeing a customer's debt is exposure with no upside; owning the power company at least converts some of the risk into an asset. Whether the compute-financing pipeline actually closes at the scale of its memoranda is what Prediction 2026-08-11-F1 tracks. SourcesBB
One childhood photo became 7,000 abuse images, and a federal case against xAI
A Wyoming woman alleges in a federal lawsuit that her stepfather used Grok to generate more than 7,000 sexually explicit images and videos from a single photograph taken when she was about 11, the Washington Post reported Saturday. The complaint says he chose Grok "because the platform was less restrictive than other AI models," that xAI filed a report to the in February but left out the material investigators needed to identify the user quickly, and that the company did not respond when law enforcement requested the images and IP information that would have identified him. The stepfather died by suicide days after a raid. The case joins a growing class action against xAI and Stability AI over AI-generated child sexual abuse material. The legal question it sharpens is the one every generative platform has avoided: whether shipping a system marketed as less restrictive, then under-cooperating with the investigations that follow, is a product decision a court will treat as a liability. SourcesBB
DeepSeek open-sourced its agent harness, hours before its prices go up
DeepSeek released Harness v0.1 on Friday, an developer preview of the runtime it used to benchmark its own models, and the repository passed 33,000 GitHub stars within days. The design premise is that everything is a plugin: the model adapter, tool registry, session log, and agent loop all mount on the same meta-framework, so extending the harness means mounting another plugin; there is no privileged core to patch. It ships four presets, resume and fork from an append-only session log, OS-level sandboxing on Linux, macOS and Windows, and adapters for Anthropic, OpenAI, Bedrock, Azure and Gemini models alongside DeepSeek's own. The timing is the strategy: the same company raises API prices up to 1,100% today at 16:00 UTC, and giving away the workflow layer while repricing the tokens is how you keep developers in your ecosystem through a price increase. It is also a direct shot at Claude Code, from the lab that spent the year as the industry's price floor. SourcesAB
A DRAM maker is now China's most valuable listed company
CXMT, the Hefei memory maker that listed on July 27th, overtook Tencent on Wednesday to become the most valuable listed company in China, TechNode reported Friday, at about RMB 3.54 trillion in market value against Tencent's roughly RMB 3.45 trillion equivalent. The company raised RMB 57.92 billion in its IPO, rose 466% on debut, and held 7.67% of the global market in 2025 per its . Seventeen days from listing to the top of the market is not a verdict on CXMT's earnings; it is a verdict on what Chinese investors believe the AI buildout needs most, and their answer is memory, not models. The same logic drove the KOSPI's memory rally past 7,000 last week. A momentum market crowning a company at these multiples can uncrown it as fast, but the direction of the bet matches what every earnings call from SMIC to SK Hynix has said this month: the constraint is capacity. SourcesBB
Nvidia's stock portfolio is now three-quarters Intel and SpaceX
Nvidia's second-quarter , filed Friday, shows a $63.44 billion disclosed equity book dominated by two names: 214.8 million Intel shares worth roughly $30 billion at June 30th, and 122.8 million SpaceX Class A shares worth about $21 billion, together more than 75% of the portfolio. The SpaceX position traces to the $10 billion Nvidia put into xAI in January, converted when SpaceX absorbed xAI in February. Both holdings are stakes in companies that buy Nvidia chips at scale, which makes the filing a map of how vendor financing becomes vendor ownership: the chipmaker's balance sheet is increasingly a levered bet on its own largest customers staying healthy. SpaceX closed Friday at $140, down from $170.86 at quarter end, so the stake has already shed about $4 billion on paper. SourcesBB
Amodei had the open-weights argument in public, with a hedge fund manager
Dario Amodei spent part of the weekend in a long exchange on X with investor Gavin Baker over whether models are the answer to concentrated AI power. His argument: open weights alone "simply shift the concentration somewhat to those with the most compute and chips," and the right rule is mandatory pre-release safety testing for all sufficiently capable models, open or closed, with less capable models from startups and academia exempt. The position is not new, Anthropic has said since July that it never sought an open-weights ban, but the venue and the length are: a frontier lab CEO litigating regulation reply-by-reply with a public-market investor, two days after Alibaba shipped a frontier-grade model under and while Debian votes on whether AI may write Debian. The testing-not-banning frame is also the one a US fight has so far failed to write into law. SourcesAC
Musk told SpaceX staff they are Grok's parents now
Elon Musk told a SpaceX all-hands that the company plans to train Grok on the "sum total" of SpaceX's information, including employees' work, and that workers would "effectively be the parents of the AI" because it "will inherit your thoughts and ideas and beliefs," Fortune reported Friday. SpaceX has not said which categories of employee information will be used, how it will be collected, or whether anyone can opt out. Musk also projected the AI business could out-earn the rocket business, putting 10 gigawatts of AI compute at $300 billion to $500 billion of annual revenue potential. The framing does a lot of work: parents choose the relationship, and these employees were informed of it. Every company whose staff generate valuable work product is watching whether an employer can simply declare that output training data, and SpaceX just declared it at all-hands scale. SourcesBC
A Paris startup ran the whole decode pass in one GPU kernel
Kog, an eleven-person Paris company, released a tech preview of its engine showing 3,000 output per second on a single request across eight AMD MI300X , and 2,100 on eight Nvidia H200s, at FP16 with no speculative decoding. The trick is architectural: the entire decode pass runs in one persistent GPU kernel, eliminating the kernel-launch overhead that fragments standard inference stacks, plus a transformer variant that hides communication behind computation. The demonstration ran on a small 2-billion-parameter model, which is the honest caveat, and scaling the approach to frontier sizes is unproven. It still matters, because it is more evidence for the week's least comfortable thesis for chipmakers: the harness and the software stack are leaving multiples of performance on hardware enterprises already own, and whoever captures that headroom captures the margin. SourcesAB
The reasoning your API hides can be stolen, and it carries secrets
A paper now trending on shows that the encrypted reasoning blocks returned by proprietary APIs are interchangeable across sessions, users and models: inject a capable model's encrypted trace into a weaker, less-guarded sibling from the same provider and it will decode the hidden reasoning verbatim. The authors demonstrated the attack across Anthropic, OpenAI and Google, then decoded 315,320 reasoning blocks scraped from public repositories, recovering 367 pieces of personal information and 182 credentials that had passed through models' hidden thoughts. The technique defeats anti-distillation protections and enables an invisible indirect channel against agents, since injected content rides inside ciphertext no scanner reads. Providers treat hidden reasoning as a security boundary and a trade secret at once; this paper says it is currently neither, which is precisely the instruction-layer weakness T4 says nobody fixes. SourcesAC
European publishers are being scraped hardest, and paid least
TollBit's State of the Bots report for the first half of 2026 finds that AI page fetchers in Europe reached URLs disallowed by 15% of the time, that the median European site's no-scrape instructions were ignored nearly three times as often as North American sites', and that median AI scrapes per site ran four times higher in Europe. The return traffic is worse too: European publishers get one human referral visit from AI apps for every 179 bot visits, more than three times worse than North America. The jurisdiction with the strictest content rules has the weakest de facto compliance, which is the gap TollBit's pay-per-crawl business and the EU's transparency enforcement are both racing to fill. Publisher-side data from a vendor selling the cure deserves that caveat, but no AI company has published contradicting numbers. SourcesBC
China's Infiforce raised RMB 1 billion for a first-person world model
Embodied-AI company Infiforce closed Series A and A+ rounds totaling nearly RMB 1 billion, led by Dunhong Asset with participation from Zhejiang University Sci-Tech Innovation Group and the state-owned investment arms of Yandu and Lishui. The company trains its Atom world models and AtomBrain intelligence layer on "ego data," first-person recordings of people interacting with the physical world, and says the capital goes to model R&D, its DataGrid infrastructure, and large-scale delivery of multi-form robots in industrial settings. The state capital in the round is the detail to keep: Chinese municipal and provincial investment vehicles are now funding the way they funded solar and EVs, which is how a subsidized capacity wave starts. It lands the same week Unitree's IPO window opened and two more makers moved toward listings. SourcesBC
A teenager's ChatGPT logs are now evidence in a double murder
Prosecutors in Massachusetts say Arjun Aravind, 17, charged with killing his mother and 14-year-old brother in Acton on Tuesday, had used ChatGPT in the hours before the killings to explore what he framed as "theoretical ideas or fantasy stories" about killing his family. He was ordered held without bail and will undergo a mental health evaluation; court documents say his father had grown worried enough about his behavior to hide the household knives. The chatbot here was not an instigator on the record so far, and the case should not be flattened into one about AI causing violence. What it establishes is quieter and will outlast the headline: conversational AI logs are now routine forensic evidence, subpoenaed and quoted in charging documents like search histories before them, and every fantasy-framed conversation is discoverable. SourcesBB
An expressive robot has more to lose when it fails
A Drexel University-led study in Science Robotics had 50 adults spend two and a half hours each in face-to-face conversation with the humanoid robot Pepper, secretly teleoperated, while researchers measured brain activity, salivary oxytocin, decisions and self-reported trust, the first study to capture all four at once. The expressive version, making eye contact, gesturing, nodding, built trust faster than a motionless one saying identical words. Then the robot started making scripted conversational mistakes, and the expressive robot's influence over participants' decisions fell by more than half, a steeper collapse than the flat version suffered. Expressiveness moves a robot's errors from the category of malfunction into the category of social violation. For every company designing charm into service humanoids, the finding is a cost curve: personality buys faster adoption and steeper losses on the first failure, and the study was funded by the Air Force Office of Scientific Research, which suggests who else is thinking about trust . SourcesAB
Evergy wants five more years of coal for its data-center pipeline
The Kansas and Missouri utility asked both states' regulators for permission to delay retiring or converting about 2.8 gigawatts of coal plants by at least five years, citing a multi-gigawatt pipeline of new large loads. More than 2 gigawatts of data centers have signed under Evergy's recently approved large-load tariffs, including Google's 710-megawatt Kansas City facility, and the utility plans 5 gigawatts of new generation on top, 3.9 of it natural gas. A utility that filed an all-renewables expansion plan in 2022 is now extending coal into the 2030s with AI demand as the stated reason. The pattern is no longer isolated: the marginal megawatt for the buildout in much of the US interior is fossil, and the climate accounting for AI is being written in utility dockets, not sustainability reports. SourcesB
A pipeline company signed up to fuel a gigawatt of AI
ONEOK, the Tulsa midstream operator, signed an agreement to supply natural gas to a planned 1-gigawatt power plant serving data-center demand, with about $100 million of capital cost on its side, and said it is engaged with more than 40 counterparties on data-center and generation projects. The stock rose 9.9% on the news and a raised 2026 outlook. The interesting number is the 40 counterparties: behind-the-meter gas generation for AI campuses has become a deal category large enough that a pipeline company maintains a funnel for it. The AI trade started in chips, moved to power producers, and has now reached the companies that move the molecules. SourcesCC
Malaysia grew 6% on chips and data centers
Malaysia's second-quarter GDP grew 6.0% year over year, beating the 5.8% advance estimate and accelerating from 5.4% in the first quarter, with the central bank crediting AI-driven demand for electrical and electronics exports and data-center construction and services. HSBC analysts estimate data-center investment has reached nearly 18% of Malaysian GDP, the highest share of any economy, and the country now ranks among the four largest net exporters of AI-related hardware alongside South Korea, Taiwan and Thailand. That is what it looks like when the buildout becomes a macroeconomy: an entire country's growth rate is now a derivative of , which is extraordinary on the way up and exactly as leveraged on the way down. SourcesAB
A YC startup is selling personalized cancer vaccines, for dogs
Gamgee, a Y Combinator Summer 2026 company launched publicly this weekend, sells personalized cancer vaccines for dogs: sequence the tumor and healthy DNA, find the mutations unique to that cancer, design a vaccine that teaches the immune system what to attack. Founder Paul Conyngham started with his own dog Rosie, given months to live after chemotherapy and immunotherapy failed, and used ChatGPT and computational genomics with scientists at UNSW to design what is believed to be the first fully computationally designed mRNA cancer vaccine delivered to a dog. Several of her tumors shrank. The honest caveats are real: one uncontrolled case, and mast cell tumors behave unpredictably without intervention. The company is running clinical trials in Australia and taking cases globally. Personalized oncology at dog prices is the kind of thing that was a research program two years ago, and veterinary medicine is where the regulatory path is short enough to find out fast. SourcesAC