August 14th 2026
Curated AI news and stories.
Z.ai's GLM-5.3 says its cyber skills outgrew its training
Beijing-based Z.ai released GLM-5.3 this morning, and the headline is not the coding. The model keeps the same 743-billion-parameter base as June's GLM-5.2 and takes every gain from scaled-up , which Z.ai says lifted its own code about 50% and pushed Terminal-Bench 3.0 from 4.6 to 28.3 and DeepSWE from 46.2 to 66.9. The number that stops you is security: Z.ai reports GLM-5.3 scored 84.5 on CyberGym, edging its own read of Anthropic's Mythos 5 at 83.8 and OpenAI's GPT-5.6 Sol at 83.6, and doubled ExploitBench to 54.4. It says the model has found 2,436 vulnerabilities across 269 open-source projects since GLM-5.2, 1,097 of them rated critical or high, with 53 already assigned CVEs and 2,383 still under embargo. Every one of those figures is Z.ai's own, unverified by a third party, and the company is holding the back roughly two weeks for safety hardening before releasing them, which is itself the tell: a Chinese lab treating its model's offensive-security capability as something to gate. It lands four days after OpenAI said it could not rule out its next model reaching the Critical cyber threshold. SourcesAB
Google shipped Gemini 3.7 Flash while its flagship stays late
Google released Gemini 3.7 Flash on Wednesday, three weeks after 3.6 Flash and still ahead of the delayed Gemini 3.5 Pro, and priced it to move. The model posts 43.6% on FrontierCode 1.1 against 34.4% for its predecessor, 65.3% on DeepSWE against 49.0%, and 30.4% on AutomationBench against 17.0%, with a WebDev Arena Elo of 1588. Google is charging $0.75 per million input and $3.75 output through the end of the year, half the launch price of 3.6 Flash, before both double on January 1st 2027. Third-party scoring on the Artificial Analysis Intelligence Index puts it at 56, a point behind GPT-5.6 Terra. The strategy is legible: while the Pro-tier model that would headline slips, Google keeps shipping the workhorse tier that developers actually wire into , and keeps cutting what it costs to do so. Most of the benchmark comparisons are Google's own; the Artificial Analysis figure is not. SourcesAB
OpenAI put GPT-5.6 Sol on Cerebras at 14 times the speed
OpenAI previewed an "Ultrafast" tier on Wednesday that runs its most capable model, GPT-5.6 Sol, at up to 14 times standard speed, generating as much as 750 output tokens per second against a standard baseline near 53, with no drop in quality. Ultrafast runs on Cerebras infrastructure, and the hardware choice is the story beyond the speed: it puts the wafer-scale company on the other side of the same OpenAI whose scale defined the Nvidia era. Cerebras spent this week getting punished by its own investors over a widening loss, and now has the highest-profile customer it could name. The tier is a waitlisted limited preview, API-first, with no pricing or general-availability date; OpenAI has been testing it with early customers in coding, commerce, support and financial research, the interactive workloads where , not intelligence, was the binding constraint. When speed stops costing you quality, the model moves into the time-sensitive parts of a business it was locked out of. SourcesAB
LG and Nvidia signed a physical-AI pact from robot to factory to car
LG Corp Chairman Kwang Mo Koo and Jensen Huang signed a broad physical-AI agreement at Nvidia's Santa Clara headquarters on Wednesday, and its breadth is the point. LG will build a next-generation bipedal on Nvidia's Isaac GR00T foundation model with Jetson Thor onboard compute and the Halos safety stack, targeting a public unveiling in the first quarter of 2027, with parts drawn from LG Electronics, LG Innotek and LG Energy Solution. Before that, LG will validate its wheeled CLOiD robots on a washing-machine line at its Tennessee plant by the end of 2026, stand up an AI-factory reference site on Nvidia DSX and Vera Rubin in the first half of 2027, and complete an 80-megawatt AI factory in Cheonan, South Korea by the first half of 2028. A vehicle platform on Nvidia DRIVE Hyperion rounds it out. One conglomerate is committing to Nvidia silicon across the humanoid, the plant that builds it and the car. Nvidia wants every industrial partnership to take that shape, and LG just signed the template. SourcesAB
California killed its AI copyright-transparency bill and kept almost everything else
Both of California's appropriations committees ran their final suspense-file hearings Thursday, the last gate before the floor, and the marquee casualty is AB 412, the Bauer-Kahan bill that would have forced generative-AI companies to disclose the copyrighted material in their training data. It was held in committee and is dead for the session, along with a labor-impact reporting bill, a critical-infrastructure AI bill and an agentic-AI measure. Almost everything else survived. SB 813, which would create a first-in-nation AI Standards and Safety Commission, advanced as amended; so did a stack of companion-chatbot safety bills from Senator Padilla, a workplace-surveillance bill, an AI-auditor registry, and the narrowed successor to the automated-decision bill Governor Newsom vetoed last year. Anything that clears the floor must reach Newsom before the legislature adjourns August 31st, and he has until September 30th to sign or veto. The pattern is clear: California will regulate how AI touches children, workers and consumers, and it will not make the labs open their training sets. These outcomes come from the committees' own unofficial results documents, ahead of any outlet tally. SourcesAA
Unitree's Shanghai IPO drew a record retail book
Unitree's listing, priced August 6th at ¥150.80 for a roughly $9 billion , closed its retail subscription this week 5,526 times , a record retail book that Bloomberg puts above the previous STAR Market high, and the allocation lottery results landed today ahead of a trading debut expected between August 17th and 21st under ticker 688836. The company said Tuesday its cumulative humanoid production has reached about 18,000 units, with the G1 alone past 11,000, against 6,500 total robots in all of 2025. The valuation works out near 219 times earnings on first-half revenue guided to ¥1.05 to 1.13 billion, and Unitree's own prospectus concedes its robots cannot yet perform useful work at scale. Chinese media reported scalpers offering winning allottees a 170% premium before the stock has traded a single share. This is retail demand pricing a story about embodied AI, years ahead of the cash flow that would justify it. SourcesBB
The rest of China's humanoid makers are lining up to list too
Unitree is the first of a wave. X Square Robot, a Shenzhen humanoid startup, has submitted a confidential filing for a Hong Kong IPO, the South China Morning Post reported this week citing sources. AgiBot, which shipped more than 5,100 humanoids in 2025 and over 15,000 embodied robots cumulatively by June, has begun its own Hong Kong process, hiring CICC, CITIC Securities and Morgan Stanley as sponsors and targeting a valuation reported at HK$40 to 50 billion, roughly $5.1 to 6.4 billion. Galbot, which called itself China's highest-valued unlisted humanoid firm after a ¥2.5 billion March round, is reported preparing a Hong Kong filing as well, though on thinner sourcing and after previously denying listing plans. Hong Kong's first-half IPO proceeds already ran 92% ahead of last year. The embodied-intelligence buildout and a hospitable listing window have the whole Chinese sector running for public money at once. SourcesBB
Samsung and SK Hynix are set to hand back a record $144 billion
The Korean memory rally we have tracked all week crossed a threshold Friday, the KOSPI briefly topping 7,000 for the first time before closing up 2.41% at 6,977.34, capping an 11.5% weekly gain, with SK Hynix up 3.26% and Samsung Electronics up 2.43%. The new fuel under the move is capital returns: Seoul Economic Daily reports the two chipmakers will unveil their largest-ever shareholder-return plans as soon as late August, potentially topping a combined ₩200 trillion, about $144 billion, as both target roughly half of for dividends and buybacks. Singapore's Temasek added to the bid, with Bloomberg reporting it is buying into both names and calling SK Hynix "severely undervalued." The read is that the memory makers now generate enough cash from the cycle to fund record payouts and their capital spending at the same time. The return figures are analyst and press estimates, not company , and range widely. SourcesBB
HBM4 will cost roughly double HBM3e, and DRAM keeps climbing
TrendForce, in a report Wednesday, forecast conventional contract prices rising 13% to 18% quarter over quarter in the third quarter and up 10% to 15%, with no loosening in sight. The number that reprices the memory thesis is HBM4: TrendForce puts it at $31 to 32 per gigabyte for Nvidia's against $17 to 18 for HBM3e, and $35 to 36 for other GPU and custom-ASIC buyers. Samsung's HBM4 revenue is projected to more than triple sequentially and exceed 60% of its total HBM sales in the second half. This is the supply side of the same story the shareholder-return plans capture on the cash side: memory is scarce, priced accordingly, and the AI accelerator makers are paying it because they have no substitute. The figures are TrendForce's market-research estimates, not company disclosures. SourcesA
Intel's $20 billion raise came with the CEO's own $12 million
Intel's equity raise, first reported last week at $15 billion, priced Tuesday at $95 a share and was upsized to $20 billion after demand ran more than five times supply, for net proceeds near $19.7 billion earmarked for expansion and 14A-node tooling. The detail that moved the read is insider conviction: CEO Lip-Bu Tan and a family member bought $12 million of stock in the offering at the $95 price, and Intel shares rose 4.4% Thursday to close at $100.95, above the deal price, as Bank of America kept a Buy on what analysts called "increasing foundry success conviction." A company that raised equity rather than debt, whose foundry lost about $2.1 billion last quarter on $5.8 billion of revenue, just got its CEO to put personal money in at the clearing price. That is the cheapest credibility Intel can buy right now, and it worked for a day. SourcesBB
OpenAI hired its second revenue chief in under a year
OpenAI named Dali Rajic chief revenue officer Thursday, replacing Denise Dresser after roughly nine months and making Rajic the company's second CRO inside a year. The hire is a statement about which problem OpenAI thinks it now has. Rajic was president and COO of Wiz, the security firm Google bought for $32 billion in its largest-ever acquisition, and before that held revenue and operating roles at Zscaler and AppDynamics, an enterprise-sales pedigree rather than a consumer one. OpenAI framed the job around monetizing a base it puts at more than a billion weekly users and over two million business customers. The revolving door at the top of the revenue org, in a year the company is reportedly steering toward a public listing, is the kind of instability investors read closely; the counter-read is that a consumer phenomenon is deliberately retooling to sell to enterprises, and hiring the person who did exactly that at Wiz. SourcesAB
Cognition is already raising again, at $40 billion or more
Cognition, the maker of the Devin coding agent, is in talks to raise at a valuation of $40 billion or more, TechCrunch reported this week, up from the $26 billion it set in a $1 billion round in May. The ask rests on a reported annualized revenue near $1 billion, and no lead investor or round size is named yet. The interval is the story: a company repricing 1.5 times higher inside three months is running on a clock the annual venture cycle was never built for, and it is not alone, with a Swedish legal-AI firm and Europe's most visible company both doubling marks this week. When a category's valuations move faster than its financials can be audited, the number is a bet on velocity, not a measure of worth. The talks are early and single-sourced to unnamed investors. SourcesB
Half the companies covered by California's AI-transparency law are not complying
The California AI Transparency Act, SB 942, took effect August 2nd and requires large generative-AI providers to offer tools that detect AI-made content. Two weeks in, a test by the investigative site Indicator with the nonprofit WITNESS found that of 13 covered companies, only 7 had detection tools available, leaving nearly half out of compliance. Across 243 tests on 85 files, the researchers report that on edited files only Google's and OpenAI's detectors correctly flagged the tampering. The statutory penalty is $5,000 per violation per day, and the law pairs with EU transparency rules that became enforceable the same day, carrying fines up to 3% of global revenue. The first enforcement test of content-provenance law is not the labs failing to build detectors; it is that half of them did not bother to turn them on. The study is single-sourced through an advocacy group, and no regulator has acted yet. SourcesC
AI designed 16 working viruses that had never existed in nature
A Stanford and Arc Institute team reported in Science this month that genome language models Evo 1 and Evo 2 generated roughly 700,000 candidate genomes, from which researchers synthesized 285 and got 16 functioning viruses, none found in nature. A cocktail of the 16 killed two E. coli strains already resistant to a natural phage. The teams deliberately excluded human, animal and plant viruses from training to limit misuse, drawing only on bacteria-infecting phages, and a companion Science editorial from Johns Hopkins biosecurity experts warns that governance for this capability "does not yet exist." This is the physical-world edge of the same frontier-risk debate driving OpenAI's caution on its next model and Z.ai's decision to gate its cyber : the models are now designing functional biology, and the rules for who may run them have not been written. SourcesAB
The most-read new paper is a red-teaming arena for agents
Topping 's trending papers is OpenART, a framework for stress-testing AI agents by evolving the environments they run in rather than scripting fixed tests. It ships more than 10,000 validated stateful scenarios across 50 domains, drawn from over 500,000 tools and skills, with tasks that require a median of 97 tool calls, and evaluates 75 agent-model configurations. Its method, an evolutionary attack that grows harder scenarios as agents solve easier ones, targets the failure mode short static benchmarks miss: cumulative risk over long, multi-step agent runs. The timing is not accidental. In a week when a Chinese lab says its model found thousands of real vulnerabilities and a Taiwan intrusion showed agents running four-day campaigns, the field has no good way to measure what agents do over a hundred consecutive actions, and this is an attempt at one. It is a preprint, not yet independently reproduced. SourcesAB
Waymo's robotaxi factory is staging a thousand cars at once
An observer's count at Waymo's Chandler, Arizona plant this week put roughly 950 of its Zeekr-based "Ojai" robotaxis on site, 684 fully retrofitted and 269 awaiting conversion, alongside the first 12 finished Hyundai Ioniq 5 robotaxis ever seen there. The fleet context is the point: Waymo now runs close to 4,000 driverless vehicles, up from about 700 early last year, across 15 cities, delivering roughly 500,000 paid trips a week and targeting a million by year-end. Ojai test cars were also spotted in Seattle in the past week, ahead of any announced launch there. The company that spent a decade proving one car could drive itself is now visibly in the different business of manufacturing them by the thousand. The unit counts come from an informal factory tally, not a Waymo disclosure, and should be read as estimates. SourcesB
One in four breaches is now AI-enabled
A CNBC piece Friday resurfaced the finding at the center of IBM and Ponemon's 2026 Cost of a Data Breach report: across 602 organizations studied from March 2025 to February 2026, 25% of malicious breaches were AI-enabled, up 56% from the prior year. Those breaches cost $6 million on average against a $4.99 million global mean, driven mainly by deepfake impersonation and AI-written malware, while defenders using AI and automation cut their own breach costs by about $2 million. The report predates this week, but it is the quantified backdrop to the week's security news, from the Taiwan agent intrusion to Z.ai's vulnerability haul. The tools that let a small team find exploits with a handful of prompts are the same ones showing up on the offense side of the breach statistics, and the cost curve is bending the wrong way. SourcesAB
Taiwan detailed its AI-agent intrusion timeline and promised new rules
Taiwan's Ministry of Digital Affairs gave its fullest account yet of the AI-agent intrusion into government systems, the Taipei Times reported Friday, saying the National Institute for Cyber Security began issuing alerts on July 20th, that the source, methods and scope have been determined, and that affected agencies "have completed their responses." The ministry confirmed the campaign combined conventional hacking with agent frameworks including OpenClaw, said it showed "clear signs of originating overseas," and said Taiwan is drafting new AI-specific cybersecurity guidelines. The underlying numbers are unchanged from Dream Security's disclosure: at least 85 government accounts compromised, more than 2,500 personnel records taken, the nuclear safety regulator and seven energy companies targeted. No formal state attribution has been made; the only pointer remains Simplified-Chinese operator communications. SourcesBB
DeepSeek's price increase goes live tomorrow
The DeepSeek API price schedule published last week takes effect at 16:00 UTC on Saturday August 16th, moving the service fully onto peak and off-peak billing. V4-Pro peak uncached input rises from ¥3 to ¥9 per million tokens and output from ¥6 to ¥27, with increases across models and token types running from 50% to as much as 1,100%; off-peak runs at half the peak rate. It is the lab that set the industry's price floor tripling its own headline rate with a week's notice, citing demand that has outrun capacity. Whether it holds is the cleanest open-weight pricing experiment in the market, and it is what Prediction 2026-08-13-B1 is testing. SourcesA
Nvidia's quarter, the month's last big read, lands August 26th
The AI tape has spent the week pricing everything upstream of Nvidia, from Foxconn and Supermicro to the memory makers and now LG, and the settlement arrives August 26th when Nvidia reports. Consensus sits near $91 billion of revenue for the quarter, with analyst price targets spanning a remarkable $180 to $743 and a near $314. The signals feeding in are uniformly hot: HBM4 priced at double HBM3e, assemblers guiding well above the Street, a fresh conglomerate committing to Vera Rubin racks. A Counterpoint note circulating Friday adds that Nvidia silicon powers about 92% of some 170 sovereign-LLM deployments across 55 countries. The setup is a quarter where the whole has already told you demand is there, which raises the bar for the print itself, since a beat is now the base case. Prediction F3 is the bet that it clears the consensus anyway. SourcesB