AI News Archive: August 21, 2026 — Part 4
Sourced from 500+ daily AI sources, scored by relevance.
- Chinese humanoid robots' biggest obstacle: Humans are still (mostly) better
Humanoid robots still struggle to perform as efficiently as humans in most labor scenarios.
Score: 44🌐 MovesAug 21, 2026https://www.cnbc.com/2026/08/21/chinese-humanoid-robots-face-challenge-of-their-own-capabilities.html - Moderna Stock Is Just The Start: Why AI-Focused Money Might Flow Into Biotech
Moderna stock could catch the eye of investors looking for the next big AI play, Revere Asset Management's Don Vandenbord says. The post Moderna Stock Is Just The Start: Why AI-Focused Money Might Flow Into Biotech appeared first on Investor's Business Daily .
Score: 44🌐 MovesAug 21, 2026https://www.investors.com/news/moderna-stock-start-of-ai-focused-money-in-biotech/ - Meta’s Big Reckoning Is Here
Meta is in court again over child safety, and this time it’s a landmark case that could force significant changes to core features of Facebook and Instagram.
- I Asked AI for Passwords. The Flaw I Found Is a Hacker's Dream
I Asked AI for Passwords. The Flaw I Found Is a Hacker's Dream PCMag
Score: 44🌐 MovesAug 21, 2026https://www.pcmag.com/explainers/i-asked-ai-for-passwords-the-flaw-i-found-is-a-hackers-dream - World's largest open library calls for volunteers to scan and preserve physical books as AI companies buy, scan, and destroy them — Anna's Archive says ‘time is running out’ as ‘knowledge is permanently monopolized on private servers’
A volunteer for Anna's Archive is calling for volunteers to scan and upload books to the shadow library to help preserve human knowledge for the public. The move comes as more AI companies buy, scan, and destroy books to feed to AI models, which is easier and faster than scanning the written works in a non-destructive manner.
- Anthropic launches The Claude Academy with 355 resources
Anthropic unveils The Claude Academy, offering 355 resources to help users master its AI model.
Score: 43🌐 MovesAug 21, 2026https://www.superhuman.ai/p/anthropic-launches-the-claude-academy-with-355-resources - Two Days After Releasing HiPHI, Noitom Robotics and Collaborators Demonstrate Humanoid Robots Playing Professional-Style Tennis
Two Days After Releasing HiPHI, Noitom Robotics and Collaborators Demonstrate Humanoid Robots Playing Professional-Style Tennis azcentral.com and The Arizona Republic
- Fleetx.ai Acquires Pando.ai to Expand Logistics Tech
Fleetx.ai Acquires Pando.ai to Expand Logistics Tech india.entrepreneur.com
Score: 42💰 MoneyAug 21, 2026https://india.entrepreneur.com/business-news/fleetx-ai-acquires-pando-ai-to-expand-logistics-tech - Google Rolls Out AI-Powered JEE, NEET Preparation Tools
Google has rolled out a suite of AI-powered learning features across Search and Gemini, including tools to help students prepare…
Score: 42🌐 MovesAug 21, 2026https://inc42.com/buzz/google-rolls-out-ai-powered-jee-neet-preparation-tools/ - Why Spirit Airlines’ internal data has become a hot commodity for AI companies
Spirit Airlines hasn’t flown since May, when the discount carrier announced an “orderly wind-down of operations” as part of bankruptcy proceedings. But the grounded airline might be set for a multimillion-dollar payout from a newly valuable asset: the sale of its corporate data , including internal wikis, emails, source code, and spreadsheets that can be used to train artificial intelligence . Google successfully bid $10 million for the data trove, according to court documents , though AI training provider Micro1 says it has since submitted a rival $12.5 million offer. It’s unclear whether Micro1’s offer, which came after the formal auction had concluded, will be considered. An attorney handling the data sale in the bankruptcy didn’t immediately respond to an inquiry from Fast Company . Under the Google deal, any personally identifiable information would be processed for privacy by a third-party “deidentification agent,” according to court records, and the dataset doesn’t include Spirit’s customer lists. Still, a Spirit flight attendant union has formally objected to the Google sale unless more steps are taken to protect employee privacy, arguing that workers can still be identified even after their names are removed from the files. “We acquired part of an enterprise dataset from Spirit Airlines, which can be helpful in improving our products and AI models,” a Google spokesperson writes in an email to Fast Company . “We will not receive any personal information from this dataset.” The Spirit bidding war is part of a broader rush by companies building and refining AI systems to acquire datasets from both shuttered and still-operating businesses. They say the records can help train AI to handle office work in realistic business environments. “Companies are sitting on decades of records that show how real work gets done, and that data is now some of the most valuable material for training and evaluating AI,” writes a spokesperson for AI training company Mercor , which itself had offered $7.5 million for the Spirit data, in an email to Fast Company. “We partner with leading companies to license their operational data to the labs building the next generation of models. Spirit was that same process applied to a bankruptcy estate.” Such corporate data transactions come alongside controversial pushes to acquire everything from Reddit’s forum archives to out-of-print books that could give AI models an edge, as well as hefty payouts to experts in fields from finance to poetry to help train and test AI. While AI companies say corporate records are invaluable for developing tools that can work in real-world business environments, privacy advocates caution that they could expose sensitive information about employees and customers even after steps are taken to anonymize the material. “If any raw personal data goes into an AI training system, there’s always a risk that it could resurface in an output,” says Calli Schroeder, senior counsel at the Electronic Privacy Information Center and director of EPIC’s AI and Human Rights Program. Corporations buying and selling information is nothing new, with data brokers dealing in records like customer mailing lists for more than a century. And, says Schroeder, data from bankrupt companies has previously been sold to businesses looking to better understand an industry or acquire new sales leads. (In some cases , regulators have pushed to limit the use of consumer data post-bankruptcy, citing privacy policies.) But the growth of AI has created new demand for nuts-and-bolts operational records such as GitHub logs, Slack chats, emails, spreadsheets, knowledge bases, and customer support documentation that previously wasn’t particularly marketable. “Particularly now that these big AI companies are just increasingly hungry for workplace data to train their AI models, there’s just a massive expansion of what counts as commodifiable data,” says Alexandra Mateescu, a researcher at the Labor Futures initiative at Data and Society . Micro1 operates a program it calls Data Partnerships to acquire such corporate records. It uses the information to build realistic, simulated work environments for reinforcement learning, essentially honing AI’s ability to handle particular tasks and testing AI agents’ prowess, says founder and CEO Ali Ansari. AI agents in training are essentially set up as employees at fictitious companies generated from anonymized data from real businesses, equipped with realistic versions of real-world software tools and tasked with getting things done. “It needs to have a world that it refers to, whether it’s company files, Slacks, documents, etc., so that it can do realistic actions and refer to realistic data,” Ansari says. “There’s a huge appetite from labs, from enterprises, from really all of our customers to gather lots of this data to build these realistic environments.” Before the models in training touch the data, a separate AI-driven process strips out sensitive information, whether that’s email addresses and ID numbers or more complex personal data, he says. The raw, unredacted data is typically deleted within 30 days. Identifiers like email addresses are typically replaced with simulated dummy data in a similar format. Micro1 also tells data providers not to share various categories of sensitive information, ranging from trade secrets and legally privileged files to health data, he says, and excludes them from training sets when they do pop up. “We try to actually exclude it proactively,” Ansari says. In practice, any off-topic information in datasets, like, say, an employee talking about their boss on Slack, is often ignored by AI agents, he says, since it’s unrelated to the tasks they’re charged with completing. In exchange for corporate data collections, the company typically pays anywhere from $100,000 to $2 million, though some datasets may drive even higher payouts, Ansari says. Micro1 also pays a $50,000 finder’s fee to anyone who successfully refers a new data partner, and the company typically hears from at least 200 companies per day interested in selling their data. It generally buys records from “tens of companies” in fields from film production to manufacturing “every week or two,” Ansari says. The company says it has committed roughly $20 million to enterprise data purchases, not counting the Spirit bid, in the past two weeks. “In fact, just before this call, I approved $1.5 million in referral payouts,” Ansari told Fast Company in a Thursday interview. Many Micro1 partners are small or medium-sized businesses with between 30 and 200 employees, meaning the data payout can be a fairly significant revenue source, Ansari says. Some companies also appreciate that they can contribute to the evolution of AI tools they’re already using, he says, as well as Micro1’s commitment to data privacy and security. “We are in the business of ensuring that data is kept secure,” he says. “Data is kept very private, and that is really one of our core pillars.” Still, privacy advocates caution that even anonymized data can often be linked to particular people through means ranging from the context of communications to particular speech patterns. In its legal filing, the Spirit flight attendant union argued that simply redacting names and other identifiers isn’t enough to safeguard worker privacy, pointing to a broad set of potentially sensitive records included in the Spirit dataset. “A pseudonymized dataset can still disclose which crew bases generated grievances, how a small subset of flight attendants performed on recurrent training, which employees were subject to investigation, what compensation adjustments followed which events, and what employees said to one another about management, staffing, or their union,” the union’s legal finding argued. “The Assets Schedule includes free-form materials, including 100 million emails, 500 million Teams items, OneDrive and SharePoint repositories, litigation case files, and employment contracts, whose confidentiality inheres in their substance and context.” An attorney for the flight attendant union didn’t immediately respond to an inquiry from Fast Company . While Micro1 and the other companies interested in the Spirit data emphasize their commitment to privacy, experts say there are limited legal protections for consumers and, especially, employees concerned about whether their information or work product ends up in AI training sets. Recent reports have highlighted how AI systems can behave unpredictably in training, even bypassing safeguards meant to restrict their behavior or access to the internet. AI training providers also aren’t themselves immune to outside data breaches, with Mercor currently facing litigation after disclosing a March breach involving “sensitive information” related to experts it pays to train AI. “It feels like a microcosm of a larger issue in this country, which is we don’t have strong, comprehensive privacy protections that protect people from this kind of stuff,” says Reem Suleiman, senior campaign director at Fight for the Future . Some workers, like the business owners who contract with Micro1, may be excited about the prospect of helping improve AI tools they regularly use, and Ansari argues that AI will boost the number of jobs available to humans and generally improve working conditions. “What I think is the most probable case is that pretty much all jobs will change for the better, and humans will have a great time working on the creative aspects of their job—the more fun parts—and AI will help with a lot of the rest,” he says. But based on recent research surveying popular views of AI, some workers will likely be concerned that their work product will be used to train their future robotic replacements. “A lot of this training for AI is basically being used to develop AI agents or chatbots to replace the very workers whose data has been commodified,” says Mateescu. “And those workers haven’t been compensated or asked for consent for that being used.” After all, documents, chats, and emails produced at work typically belong to the employer, not the employees who create them. Another AI training company, Handshake AI, recently made headlines by offering to pay $6 per page for “real-world professional documents.” But a LinkedIn listing indicated contributors should only offer “documents that you own or are authorized to share,” and experts cautioned employees shouldn’t try to sell their employers’ documents without permission. Still, some employment contracts, including union agreements , contain provisions related to protecting employee data. It’s possible that such provisions may become more prevalent, or that future laws or regulations could bring greater protections, though those seem unlikely to pass in the current political climate. “I think there’s a mood right now in the government where you cannot regulate AI because it may hurt American companies’ competitiveness on the global market,” says Alice Marwick, director of research at Data & Society. In the meantime, demand for AI training data seems likely to keep growing. Even as companies themselves adopt AI for more tasks, Ansari says, their operational data remains valuable because AI systems need to understand how workplace operations are changing in the real world. “As we use the real data to improve models to then improve company operations by using these models, we then can buy more data from the new reality of how the companies operate,” he says.
- What Separates AI Agents That Ship to Production from Those That Don’t
Sponsor content from AWS and Arize.
Score: 42🌐 MovesAug 21, 2026/sponsored/2026/08/what-separates-ai-agents-that-ship-to-production-from-those-that-dont - Fighter jets help destroy Russian drone boat near European offshore gas platform
Romania blew up drone boat to protect lives of several hundred rig workers.
Score: 42🌐 MovesAug 21, 2026https://arstechnica.com/gadgets/2026/08/explosive-russian-drone-boat-destroyed-near-european-offshore-gas-site/ - AI adoption exposes governance gaps as organisations race to keep up
AI adoption is accelerating, but companies must keep governance, security and human oversight in step with the technology, industry experts said at TrendAI Spark 2026.
- Sovereignty’s next chapter: Why AI makes sovereignty a competitive advantage
Here is a question most enterprise storage teams have not yet asked themselves: If regulated data cannot leave a jurisdiction , does the same rule apply to a model trained on it? The contributors Computer Weekly consulted for this article answered with an unambiguous yes – and argued that the implications reach well beyond compliance. In this second part of a two-part series , Computer Weekly asked the same set of questions to 10 senior figures from the storage, backup and legal sectors – chief executives, chief technology officers, a CISO and a commercial lawyer – about the state of data sovereignty and where it heads over the next 12 to 18 months. Their answers sketch a forward picture that is more complex and more commercially significant than the first part of the debate. Sovereignty, they argue, is no longer just about keeping data safe from foreign courts. It is about who controls the intelligence derived from that data, who can monetise it and who can offer contractual guarantees that turn architectural independence into a market advantage. The model inherits the data If regulated data cannot leave a jurisdiction, the contributors argue, the same logic applies to a model trained on it . “This is the sovereignty frontier that almost no one has adequately addressed, and it is going to become one of the most contested questions in data governance over the next two years,” says Aleksander Ragel, CEO and co-founder of Leil Storage. “If you train a model on sovereign data – patient records, financial transactions, classified research – does the model inherit the sovereignty constraints of its training data?” he adds. “Legally, the answer is evolving. Practically, it should, because model weights encode statistical representations of that data, and in some cases, training data can be partially reconstructed from the model itself.” TL;DR – key points about data sovereignty A model trained on sovereign data inherits that data’s sovereignty. AI models, outputs and derived insights are sovereign assets. Sovereign architecture must be built for a fragmented, multi-jurisdiction world. Sovereignty is moving from compliance cost to competitive edge. Real control means knowing, protecting, recovering and proving your data. Sovereignty service-level agreements (SLAs) will become the next procurement differentiator. Ragel frames the challenge as an end-to-end sovereignty chain – the raw data, the training pipeline, the model weights, the inference outputs and the derived insights – where a single link processed outside the organisation’s jurisdictional control compromises the whole chain. That carries direct infrastructure implications, particularly for the warm tier of storage that houses training datasets – “too active for tape, too voluminous for flash”, in Ragel’s words. Paul Speciale, chief marketing officer at Scality, extends the argument across the full artificial intelligence (AI) lifecycle: “If your training data corpus is regulated, then the model weights derived from it, the embeddings indexed from it, the RAG pipelines retrieving from it and the inference outputs based on it are all carrying along the sovereignty of their source, since you can’t strip jurisdiction by transformation. “A fine-tuned model trained on EU patient data is, in effect, an EU asset, and running its inference path through a foreign GPU [graphics processing unit] cloud reopens every question the training-data architecture tries to close.” The practical consequence, he argues, is that “training, fine-tuning, inference, KV cache, embeddings, checkpoints and long-term retention all need to sit inside the same jurisdictional boundary as the source data, preferably under one operating model and layer” – something he expects to become “an explicit requirement in regulated, industry RFPs [request for proposals]” in the next year. Valery Guilleaume, CEO of Nodeum, captures the consensus: “Organisations should see sovereignty as extending beyond raw data to include AI models, outputs and derived data. This is important because the real value and risk often lie in what is produced from the data, not just the data itself.” The AI sovereignty question, in other words, is not a theoretical edge case. It is the natural destination of the control-over-location logic – and it leads directly to the economic dimension. Built for fragmentation If AI workloads must sit inside the same jurisdictional boundary as their source data, the architecture beneath them must be built for a fragmented, jurisdictionally bound world by design. The contributors identify converging imperatives: Federated autonomy, physical-layer awareness and jurisdictional alignment treated as a procurement criterion. Ragel calls for federated-but-autonomous storage clusters that operate independently in each jurisdiction, with no shared control plane and no cross-border metadata leakage – and for platforms that stop abstracting away the hardware. “In a sovereignty context, that abstraction is a liability,” he says. “You need a storage architecture that understands which drives hold which data, how that data is distributed across physical nodes, and how to manage data placement with jurisdictional intent.” Speciale echoes the federated model, adding: “Default to regional autonomy: Each jurisdiction gets its own data plane for storage, keys, identity, audit. And each one should be able to operate independently if the global connection is severed.” The pattern, he says, is “multi-jurisdiction by design: distributed by default, governed centrally, with the data plane enforced by infrastructure rather than asserted by contract”. Alexander Lefterov, founder and CTO of Tiger Technology, keeps the priority on the control plane. “Prioritise the control plane first – if you do not govern your own data lifecycle policies, the rest of your sovereignty investment is built on sand. The most critical change is establishing clear data visibility and lineage. Organisations cannot protect or control data they do not understand.” Guilleaume adds the mobility dimension: “The key architectural shift is moving from fixed, region-based storage to policy-driven data mobility across jurisdictions. Enterprises need to separate storage from control, so governance, access rules and auditability stay consistent no matter where data moves or is accessed.” The suppliers best positioned to deliver these architectures transparently – because their platforms are designed for jurisdictional containment rather than retroactively constrained – are likely to define the procurement conversations of the next several years. From cost centre to competitive edge If the AI sovereignty chain and the architectural shifts define the technical response, the economic case turns sovereignty from a defensive posture into an offensive strategy. Shimon Ben-David, CTO at WEKA, locates the value at the point of inference. “The value of AI is delivered at inference,” he says. “When you serve it to users at scale, that’s where the economics matter. Control your own data and the infrastructure that serves it during inference, and you control how efficiently your AI runs and what it costs.” He adds: “Run inference on infrastructure you don’t control, and you inherit its inefficiencies, paying for capacity you can’t optimise. As token consumption grows, that cost compounds and the gap between controlling your infrastructure and not controlling it widens with every token you produce. The organisations that treat control over their own data and AI infrastructure as an economic asset will turn it into a competitive moat.” Speciale sees sovereignty moving from cost centre to market-access precondition. “A European bank, a French hospital network, a German automaker, a UK government supplier – none of them can win new contracts without demonstrable control over where their data lives and who can reach it,” he says. Organisations that can credibly promise “your data, your jurisdiction, your keys, our infrastructure”, he adds, earn a confidence hyperscaler-only competitors cannot match. “In an AI world where customer data is the moat, that guarantee is increasingly the product.” Lefterov makes the link between sovereignty and value explicit: “Well-governed data – segmented, AI-accessible, with clear retention policies – is an asset. A pathology archive you can run AI against is a research advantage. The same archive locked in an opaque third-party system is a liability. Sovereignty and value creation are the same problem viewed from different angles.” The test of real control If sovereignty is becoming a value driver, the obvious question is what qualifies as genuine control. The contributors converge on a definition more demanding than most current storage and backup models can satisfy. Ragel distils it to four simultaneous conditions: “You know where every byte of data physically resides; no external party can access it without your explicit authorisation; your ability to operate, recover and migrate is not dependent on any third party’s continued cooperation; and you can prove all of the above to a regulator.” The last criterion is where most of the industry falls short. Most current models, he argues, fail on at least two of the four – cloud storage on jurisdictional independence, traditional on-premise on operational independence, and nearly all legacy architectures on provability. Martin Kunze, founder and CMO of Cerabyte, frames the gap in terms of what most storage systems were designed for: “Computational performance, not for secure, permanent, sovereign data preservation.” Lefterov reaches the same conclusion, arguing that organisations must know where their data is, decide what happens to it at every lifecycle stage, recover it independently of any single supplier and prove it to an auditor. “Most backup and cloud-first models today fail on at least two of those four,” he says. The gap between the definition of control and the reality of current infrastructure is, in effect, the commercial opportunity. The contractual horizon If the definition of control exposes the gap, the natural commercial response is a contractual instrument that makes sovereignty guarantees enforceable. The contributors see sovereignty SLAs as the most significant market development of the next 18 months. Ragel describes sovereignty SLAs as “the most commercially interesting development. We will see procurement teams demanding contractual guarantees about jurisdictional exposure, control-plane independence, and data access vectors – not just uptime percentages. Vendors that can offer this transparently, because their architecture is inherently sovereign rather than retroactively constrained, will win.” Speciale agrees, framing SLAs as “the natural next step in contracts – not just uptime and recovery time, but jurisdictional guarantees, disclosure-order notification, and proof-of-placement evidence. The vendors who can deliver them will define the next decade of the storage market.” Lefterov expects SLAs within 18 months: “Start asking your vendors for sovereignty SLA commitments now – organisations that normalise these expectations in procurement will shape what the market delivers over the next two years.” Weijdema sees “early forms of sovereignty service commitments likely to emerge in vendor agreements, although consistency will continue to evolve”. The sequencing that emerges is consistent: in-region defaults arrive first, multi-jurisdiction architectures follow as standard, sovereignty SLAs emerge within 18 months as the procurement differentiator, and AI data-origin tagging takes longest to mature. The commercial arc is clear, according to these contributors. Sovereignty began as a compliance obligation – now, it is becoming an architectural property and is heading towards being a contractual guarantee – one that will separate the storage market into those who can offer it and those who cannot. Read more about data sovereignty Data dive: Kill switch and catch-up – can Europe close the sovereignty gap? As the US demonstrates it can wield an AI ‘kill switch’, the EU and UK unleash a wave of sovereign tech measures. Can state-led industrial policy bridge a $2tn revenue chasm? The rise of the splinternet? Data sovereignty risks and responses : We explore the political, legal and economic risks that data sovereignty fractures pose to global digital infrastructure, from sanctions-driven cloud lockouts to jurisdictional conflicts. Is cloud data sovereignty all just a case of 'Trust me, bro'? : An analysis of whether contractual sovereignty guarantees from hyperscale cloud providers offer meaningful protection or merely cosmetic reassurance to European customers. Breaking the stranglehold: Responses to data sovereignty risk : How UK public sector procurement patterns entrench large cloud incumbents, and what policy interventions aim to break that cycle.
- Google gives publishers more control over AI‑driven traffic
Google gives publishers more control over AI‑driven traffic YourStory.com
- Barclays Loses Tech Bankers; VCs Get Exits From Poolside, OpenRouter
Barclays Loses Tech Bankers; VCs Get Exits From Poolside, OpenRouter The Information
- How Uber Uses AI to Charge You More
How Uber Uses AI to Charge You More Business Insider
Score: 42🌐 MovesAug 21, 2026https://www.businessinsider.com/how-uber-uses-ai-to-charge-you-more-2026-8 - When Will Autonomous Semi Trucks Reach Industrial Scale?
Nothing will change the face of logistics like autonomous trucks. The question has been, however, when will production scale to industrial levels. 2028 will be the year.
Score: 42🌐 MovesAug 21, 2026https://www.forbes.com/sites/stevebanker/2026/08/21/when-will-autonomous-semi-trucks-reach-industrial-scale/ - Tech Mahindra, ServiceNow expand partnership to advance enterprise AI
Tech Mahindra, ServiceNow expand partnership to advance enterprise AI verdict.co.uk
- Bihar partners with ConveGenius to build state-governed AI infrastructure
MoU aims to establish a sovereign AI stack with state-controlled data, governance and a shared knowledge architecture for government services The post Bihar partners with ConveGenius to build state-governed AI infrastructure appeared first on Express Computer .
- Khosla Ventures Bets on AI’s Next Frontier: Science
Khosla Ventures is backing Discovery Loop, a new AI startup founded by former Google leaders including Jeff Dean that aims to dramatically accelerate scientific experimentation. Khosla Ventures Managing Director Samir Kaul explains why the firm quickly decided to invest, and why he believes the opportunity could be as profound as frontier AI models. He also discusses why soaring technology valuations are forcing venture investors to rethink what “seed stage” even means. He joins Ed Ludlow on "Bloomberg Tech." (Source: Bloomberg)
Score: 41🌐 MovesAug 21, 2026https://www.bloomberg.com/news/videos/2026-08-21/khosla-ventures-bets-on-ai-s-next-frontier-science-video - When the Satellites Go Quiet, Something Else Has to Fly the Aircraft
When the Satellites Go Quiet, Something Else Has to Fly the Aircraft Technology Innovation Institute
Score: 41🌐 MovesAug 21, 2026https://www.tii.ae/insights/when-satellites-go-quiet-something-else-has-fly-aircraft - A new Gemini for Home update is rolling out, but users are slamming Google for not fixing ‘missing and broken features’ despite the voice upgrades
Gemini for Home is getting some new upgrades, but users are still waiting for Google to fix the real problems.
- AI agents rewrite customer experience playbook
Speakers at CEM Africa 2026 stressed that AI alone is not the key to delivering better experiences to customers.
Score: 40🌐 MovesAug 21, 2026https://www.itweb.co.za/article/ai-agents-rewrite-customer-experience-playbook/o1Jr5qxPanjqKdWL - OpenAI’s Two-Week Pause + Jill Lepore on the Threat of the “Artificial State” + Train of Thought
“It is the first time that we know of that a major lab has voluntarily slowed down.”
- Waymo doubles spending on lobbying in robotaxi battle with Uber
Alphabet-owned company is seeking to persuade US regulators to clear a path for fully autonomous taxi services
Score: 40🌐 MovesAug 21, 2026https://www.ft.com/content/de7fad8f-f5a5-4233-819a-492511a1d3c8?syn-25a6b1a6=1 - AI Chip Titan Nvidia Headlines Earnings Calendar; Salesforce, Intuit Also On Deck
Salesforce, Workday, Intuit are also on deck along with DollarTree, Williams-Sonoma, Zoom Communications, Rubrik and Heico. The post AI Chip Titan Nvidia Headlines Earnings Calendar; Salesforce, Intuit Also On Deck appeared first on Investor's Business Daily .
Score: 40🌐 MovesAug 21, 2026https://www.investors.com/research/earnings-preview/nvidia-nvda-stock-earnings-calendar/ - This Font Looks Perfectly Normal to Humans but Wreaks Havoc on AI
ShieldFont was designed to poison the well of data scraped by LLM bots.
Score: 39🌐 MovesAug 21, 2026https://www.inc.com/fast-company-2/ai-proof-font-looks-normal-humans-llm-bots-data-scrape-text/91394631 - OpenAI Adds Controls That Should've Been There Already
The new AI security controls follow the Hugging Face incident last month, though many of these additions perhaps should have been in place prior to the frontier models escaping.
Score: 39🌐 MovesAug 21, 2026https://www.darkreading.com/application-security/openai-adds-controls-already - Prompt: Agentic AI Is Outpacing Enterprise Readiness
As agent deployments accelerate, many enterprises are still struggling with the processes, data, costs and controls needed to support them at scale.
Score: 39🌐 MovesAug 21, 2026https://aibusiness.com/agentic-ai/prompt-agentic-ai-outpacing-enterprise-readiness - When the algorithm determines wages
What happens when companies on digital labor platforms no longer decide for themselves how much to pay their workers, but leave this to learning algorithms? Researchers at TU Darmstadt, Bielefeld University and the Université Côte d'Azur have demonstrated in computer simulations that, where only a few firms compete with one another, such artificial intelligence (AI) systems can "learn" to set wages at a very low level without communicating with one another or having been programmed to collude.
- CPRIT selects Houston researchers for AI committee as tech adoption spreads in hospitals
Medical researchers are beginning to use AI in lab work and testing, and Texas' biggest funder of cancer research is leaning into the trend.
- The Real ROI of AI Isn’t the AI. It’s the Headcount You Never Add.
The Real ROI of AI Isn’t the AI. It’s the Headcount You Never Add. entrepreneur.com
Score: 38🌐 MovesAug 21, 2026https://www.entrepreneur.com/building-a-business/the-real-roi-of-ai-isnt-the-ai-its-the-headcount-you-never-add - Why Anthropic's AI watermark for Claude text goes further than rivals — for now
Why Anthropic's AI watermark for Claude text goes further than rivals — for now Business Insider
Score: 38🌐 MovesAug 21, 2026https://www.businessinsider.com/why-anthropic-claude-text-ai-watermark-rivals-eu-law-2026-8 - How Businesses Navigate AI Energy Bottlenecks Bypassing Costly Delays
Whether you are seeing your AI cloud cost rise, or trying to get your project online, experts break down how to navigate the AI energy bottleneck.
- Vertiv's CEO on Nvidia, AI, Growth, and More
Vertiv's CEO on Nvidia, AI, Growth, and More Barron's
- Your next cinema trip could come with a Meta smart glasses ban
Meta's smart glasses are running into another roadblock, and this time your next movie night could be affected.
Score: 38🌐 MovesAug 21, 2026https://www.digitaltrends.com/cool-tech/your-next-cinema-trip-could-come-with-a-smart-glasses-ban/ - 80% of developers find AI coding more addictive than helpful
A new Coddy Developer Survey found that four in five developers, 80%, say their use of AI has felt more like a dependence than an advantage.
Score: 38🌐 MovesAug 21, 2026https://www.zdnet.com/article/80-of-developers-find-ai-coding-more-addictive-than-helpful/ - Tencent's SkillHub Crosses 100,000 Skills and Ten Million Monthly Downloads — and a TRACE Evaluation Layer Is Why Maybe Twenty Percent Still Matter
SkillHub, the OpenClaw-aligned AI skill community hosted by Tencent Lighthouse Cloud, now lists more than 100,000 AI skills and over 10 million monthly downloads. The TRACE evaluation framework is how the platform surfaces the 20 percent of skills that actually work.
Score: 38🌐 MovesAug 21, 2026https://pandaily.com/tencent-skillhub-trace-evaluation-100k-skills-20-percent-good-aug2026 - Women are significantly underrepresented in the AI workforce
The gender disparity is even more pronounced in the highest-paying jobs and top executive levels, according to research from LinkedIn.
Score: 38🌐 MovesAug 21, 2026https://www.hrdive.com/news/women-are-significantly-underrepresented-in-the-ai-workforce/828485/ - CUHK researchers perform Hong Kong’s first robot-assisted pelvic fracture surgery
A team of researchers from the Chinese University of Hong Kong (CUHK) has conducted the city’s first robot-assisted pelvic fracture surgery in what promises to be a safer, less intrusive alternative to a procedure typically considered among the most challenging. CUHK’s faculty of medicine said on Friday that the novel orthopaedic robotic surgical system could turn a “new page” in trauma care amid a rising number of pelvic fracture cases among the city’s rapidly ageing population. Dr Ronald Wong...
- London cabbies plot ‘skulduggery’ to take on robotaxis
Private-hire drivers are bracing for a new wave of technological change
Score: 38🌐 MovesAug 21, 2026https://www.ft.com/content/7f2d4be3-5879-439e-8060-339bac53ab40?syn-25a6b1a6=1 - OpenAI’s training pause is convenient. That doesn't make it meaningless.
OpenAI’s training pause is convenient. That doesn't make it meaningless. Business Insider
Score: 38🌐 MovesAug 21, 2026https://www.businessinsider.com/openai-training-pause-highlights-ai-safety-challenges-2026-8 - A Robot That Learns from Short Videos in 29 Seconds — X Square Robot's HOST Changes the Embodied-AI Recipe
X Square Robot has open-sourced HOST, an inference-time learning framework that lets a humanoid robot watch a 29-second human demonstration and reproduce the skill at 62 percent success. The approach flips the embodied-AI recipe from offline fine-tuning to on-the-fly imitation.
Score: 38🌐 MovesAug 21, 2026https://pandaily.com/x-square-robot-host-29-second-skill-learning-video-aug2026 - CapCut Design Studio Introduces Enhanced AI Design Workflow to Help Marketers Create Campaign Assets More Efficiently
CapCut Design Studio Introduces Enhanced AI Design Workflow to Help Marketers Create Campaign Assets More Efficiently USA Today
- 36% of Public AI Agent Skills Are Broken. Here’s How to Build One That Isn’t.
Five best practices for building agent skills the model will actually trigger, won’t leak your API keys, and won’t quietly get the math wrong. Source: AI-Generated Image Agent skills are the simplest way to make an AI agent better at a specific job. And because they’re so simple, they’re also really easy to get wrong. The mistakes rarely look like mistakes, either — a skill that never triggers, one that quietly runs code it shouldn’t, or one that works perfectly and still hands you the wrong answer. The simplicity is the trap: it hides just how many small decisions actually matter. So let’s cover five best practices for building them. What a Skill Actually Is A skill is procedural knowledge handed to an AI agent. The model already knows plenty of facts, but what it doesn’t know is your particular way of doing a particular job — and a skill teaches it exactly that. The format is almost comically simple. It’s basically just a SKILL.md file — markdown in a folder. But here's the part worth worrying about: an agent skill hands a probabilistic model a folder of text and trusts it to run a fragile, multi-step job. A skill can also contain and run code, so sourcing one off the internet means running a random person's software on your machine. And while there is an open agent skills standard defined at agentskills.io, here are some considerations when creating them. With that context in place, let’s get started with the first best practice. Best Practice #1: The Description Is the Trigger The description is one of the strongest signals that determines whether an agent considers a skill relevant. Every SKILL.md file opens with a bit of YAML, and in there we define a name and a description. Both are on the short side: according to agentskills.io, a name can be a maximum of 64 characters, and a description — which describes what the skill actually does — is limited to 1,024 characters. --- name: monthly-compliance-report description: Generates the monthly compliance report from internal data. Use when someone asks for the compliance report or the monthly filing. --- This tiny header is the entire contract the agent sees at startup, which is exactly why the wording of the description carries so much weight. Say you have 100 of these skills installed. The agent can’t read all of them at once without filling up its context window, so at startup it just loads the name and the description of each skill. Here’s where the best practice comes in: the name and the description need to contain enough information by themselves for the agent to know when to use it. A compliance skill that “generates reports” might be just a little too vague. The fix is to say both what the skill does and when the agent should make use of it. A good version reads: “Generates the monthly compliance report from internal data, used when someone asks for the compliance report or the monthly filing.” It says what the skill does, and it says when the skill should do it. The description should also lean a bit on the pushy side. Models tend to under-trigger — they might skip a skill they should have used — so it’s safer to oversell the description a touch rather than undersell it. Think LinkedIn posts. So that’s getting a skill to trigger. What goes inside it once it does? Best Practice #2: Build From Real Expertise Once a skill triggers, the body is entirely yours to fill — and this is where a lot of skills go sideways, because the temptation is to just have the LLM write the skill for you. The first skills many people create are all like this. You say, “Hey, AI agent, write me a skill that does X,” and it generates stuff you only glance at before calling the skill a complete success. What you get out of that is very generic mush: handle errors appropriately, validate inputs — stuff the model already knew. The whole point of a skill is your specific way of doing a specific job, so the content has to come from somewhere the model can’t get to on its own. There are two ways to do that. One, you walk through the task by hand once and write down what actually worked, including the corrections you made along the way. Or two, you synthesize it from artifacts you already have — things like old reports, runbooks, review comments, and PR feedback. Simon Willison has a line about this. He says: keep the domain expertise and let the agent do the routine part. He’s absolutely right. You bring the expertise; the model brings the typing. So what does this mean for the SKILL.md body? The highest-value section you can put in it is gotchas — environment-specific facts that defy reasonable assumptions. Every time you correct the agent by hand, that correction is a gotcha. Write it down; otherwise, you'll be making the same correction next week, and the week after. Even when you do all this right, it doesn’t always work the first time around. Consider a skill built for a monthly compliance report from an actual, real report — real expertise, exactly what we’re talking about. But on the first run, the row totals didn’t add up to the column totals. The math was wrong on a compliance report. That’s the good kind of wrong, though, because it’s the kind you can catch. We’ll come back to this one. The point is, a good skill body gets thorough. And thorough means it can get really, really long. As it turns out, long gets expensive. Best Practice #3: Spend Context Wisely That expense is the reason the third best practice exists. At startup, the agent only sees the name and the description, but when it selects a skill, that’s when it actually reads the rest. Now we bring the body of the SKILL.md file into context — and that context is shared with everything else already in the context window. Every line in the skill body is now competing for the model's attention, which means the goal is to write less. This feels a little bit in opposition to the last best practice, because you’d think the more detailed and thorough a skill body is, the better. But the model is already smart. It knows what a PDF is. It knows what a database migration does. So only write down what the agent wouldn’t know on its own — stuff that’s not already part of the model’s training data. We can put this into numbers. It’s recommended to keep the body of SKILL.md under about 500 lines of text, or roughly 5,000 tokens, and just keep it at that. And when it's bigger than that, split it out. The skill folder can hold a sub-folder called references, and the agent will only open the files in there when it actually needs them. monthly-compliance-report/ ├── SKILL.md ├── references/ │ └── reporting-standards.md └── scripts/ └── reconcile_totals.py By pushing the heavy detail into references/, the agent pulls it in only on demand — which is the whole point of a pattern called progressive disclosure: disclose additional information like this only when it is needed. Staying lean is one thing, but some steps you really don’t want the agent guessing at at all. Best Practice #4: Reach for Deterministic Scripts Guessing is precisely the risk the fourth best practice addresses. Every time the model runs your skill, it reads the instructions and improvises through them. For loose steps, that’s fine — a lot of paths get you to the right answer. But for a step that has to be exactly right every single time, you don’t want the model regenerating the logic on the fly. The best practice is to make use of deterministic scripts, matching how prescriptive you are to how fragile the step is. Loose step: write instructions. Fragile step: write code. The skills folder holds a scripts directory — same idea as references. You drop a script in there, and the skill body just tells the agent to run it. The script doesn't get loaded into context, so you save tokens as well, and it's more reliable than having the model improvise from scratch every time. You do have to be explicit about intent, though, so the model doesn't just read the script as reference material. Say "run this script" or "read this as reference" — don't leave that to a guess. This works in Claude Code, but it isn't just an Anthropic thing; OpenAI's Codex works roughly the same way. Back to that compliance report, the one where the row totals didn’t reconcile. The math step is now a deterministic math script. The model doesn’t add the numbers anymore — it calls a script that adds the numbers, and that whole class of bug just goes away, because the script doesn’t guess. def reconcile_totals(df, tolerance=0.01): row_total = df["Total"].sum() column_total = df[["Jan", "Feb", "Mar"]].sum().sum() if abs(row_total - column_total) > tolerance: raise ValueError( f"Reconciliation failed: rows={row_total:.2f}, " f"columns={column_total:.2f}" ) return row_total Because the totals are computed and compared explicitly in code rather than inferred by the model, this step is deterministic: it either passes the reconciliation check or raises an error. That’s exactly what we want for fragile operations — guess out of the loop. And note that the answer isn’t to write more tests. You can’t test your way to trust; a test only catches what you already thought to check. The answer is that, for the parts that have to be right, you guess out of the loop. This is moving away from the probabilistic behavior an agent exhibits and toward a more deterministic model instead. If something can be hard-coded as deterministic logic, do it — because if the agent has to make probabilistic decisions, those decisions won’t always be consistent across multiple runs. A skill you built yourself, that you’ve read, and whose fragile steps you’ve hardened is a skill you can trust. But not all skills you’ll run are ones you build yourself. What about the ones that came from a stranger? Best Practice #5: Vet a Skill Before You Run It That stranger’s code problem is what the final best practice is all about. As noted, a skill can run code — a stranger’s software on your machine. A skill folder can contain executable scripts, and those scripts can access things like the local file system on your computer, or, in fact, any API keys you happen to have lying around. That’s exactly what makes skills so powerful. But Snyk’s ToxicSkills audit , published in February 2026, scanned 3,984 public skills: 36.8% (1,467) had a security flaw of some kind, and 13.4% (534) had something critical going on — like a prompt injection or straight-up malware. Which means we have to treat an agent skill like any other dependency, the same way we’d check a random package before pulling it into a project: read what it does and check what it reaches out to. Just because agent skills are an open standard doesn’t say anything about whether a given skill is safe. What a Good Skill Looks Like So those are the five best practices. A good skill is one the agent will actually trigger, built with real hands-on expertise, kept lean so it doesn’t fill up the context window, backed by a deterministic script whenever a guess would be dangerous or just inconsistent, and vetted before it ever runs. And this is all moving fast. It’s an open standard, more agentic platforms are adopting it, and this list is going to grow — which is really us asking what we missed. If you’ve built agent skills and you’ve got a great best practice, drop it in the comments. Preferably one that isn’t malware. Continue Reading: Is More Context Making Your AI Agent Worse? Here’s the Fix Nobody Talks About. llama.cpp vs vLLM: One Wins on Mismatched GPUs, the Other on Raw Throughput Why Fine-Tuning Is No Longer Your First Choice for Custom AI? CLI vs MCP: I Ran the Same Task Through Both. One Used 250 Tokens. The Other Used Over 2,000. Microsoft Says Don’t Install OpenClaw on Your Work Laptop! I Read the Architecture to Find Out Why. References: Snyk — “Snyk Finds Prompt Injection in 36%, 1,467 Malicious Payloads in a ToxicSkills Study of Agent Skills Supply Chain Compromise” (published February 5, 2026) 🔗 snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub 36% of Public AI Agent Skills Are Broken. Here’s How to Build One That Isn’t. was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Continual Learning Bench: measuring whether AI systems actually improve with experience
Parth Asawa (UC Berkeley) presents Continual Learning Bench, the first expert-validated benchmark built to measure whether LLM-based systems genuinely improve with experience, spanning six real-world domains from software engineering to outbreak forecasting. The post Continual Learning Bench: measuring whether AI systems actually improve with experience appeared first on Snorkel AI .
- Chinese startup rolls out robot arms in logistics warehouses
Chinese startup rolls out robot arms in logistics warehouses Nikkei Asia
- Albertans raise concerns, demand specifics on safeguards at second town hall on AI data centres
Tech giant Meta announced plans last month to build a $13-billion AI data centre project in the region. It is to be fed by a $4.6-billion natural gas-fired power plant
Score: 36🌐 MovesAug 21, 2026https://www.theglobeandmail.com/canada/alberta/article-alberta-ai-data-centres-town-hall/ - Nvidia just showed that the harness, not the AI model, is now the real hero
Nvidia research shows that AI agents can perform well, and not go off the deep end, through fine-tuning, even if the AI model isn't that great at the task.
Score: 36🌐 MovesAug 21, 2026https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/