AI News Archive: July 6, 2026 — Part 3
Sourced from 500+ daily AI sources, scored by relevance.
- AI, streaming subscriptions fuel growing outflow of dollars
AI, streaming subscriptions fuel growing outflow of dollars 매일경제
- What are Britain’s AI growth zones and are the plans feasible or ‘complete bunk’?
Lanarkshire datacentre run by renewables and creating thousands of jobs not achievable by 2030, Guardian investigation finds Revealed: landmark Scottish AI project has no prospect of meeting renewables promise ‘It’s smoke and mirrors’: hope turns to fear in Scottish village chosen for AI datacentre The Guardian has examined government plans to build Britain’s AI infrastructure for the future, finding some of these to be, in the words of one source, at best unclear and at worst “complete bunk”. The plans in question are for AI growth zones, which are supposed to be regions where the government supports companies to build massive AI datacentre complexes, of 500MW or greater. These could be bigger than any now operating in the UK. Continue reading...
Score: 58🌐 MovesJul 6, 2026https://www.theguardian.com/technology/2026/jul/06/britain-ai-growth-zones-explainer - CMU Researchers Help Close a Critical Security Gap Across AI Platforms
CMU Researchers Help Close a Critical Security Gap Across AI Platforms Carnegie Mellon University
- Multi-agent teams might not be better than a single good model: Apple and Stanford paper
Researchers find that when left to their own devices, more agents are not always the answer.
Score: 58🌐 MovesJul 6, 2026https://www.thestack.technology/self-organising-multi-agent-orchestration-not-better/ - Synthetic Sciences Releases OpenScience: An Open-Source, Model-Agnostic AI Workbench for Machine Learning, Biology, Physics, and Chemistry Research
Synthetic Sciences Releases OpenScience: An Open-Source, Model-Agnostic AI Workbench for Machine Learning, Biology, Physics, and Chemistry Research MarkTechPost
- Intellect Design Arena signals strategy around ‘AI-first banking’
In the company’s annual report for FY25-26, Intellect Design Arena said that financial institutions are increasingly embedding AI across enterprise operations
- Runway Expands Global AI Footprint
Runway is going global, with the AI video company opening new hubs in London, Tokyo, and Paris, and plans to invest nearly $300 million over the next few years as it scales its business and AI research. Runway co-founder and CEO Cristóbal Valenzuela joins Ed Ludlow on "Bloomberg Tech." (Source: Bloomberg)
Score: 56🌐 MovesJul 6, 2026https://www.bloomberg.com/news/videos/2026-07-06/runway-expands-global-ai-footprint-video - Precision hiring, AI readiness to define FY27 workforce landscape: report
Precision hiring, AI readiness to define FY27 workforce landscape: report YourStory.com
Score: 55🌐 MovesJul 6, 2026https://yourstory.com/2026/07/precision-hiring-ai-readiness-define-fy27-workforce-landscape - This humanoid robotics company is going public, but its CEO isn’t promising a robot in your home anytime soon
While other humanoid startups chase sky-high valuations, Agility Robotics is betting its future on execution — and a SPAC.
- China’s web novel platforms embraced AI. Now they are fighting it
Sites from Tencent, ByteDance, and Baidu set curbs like daily word limits for authors and stricter standards to combat poor-quality automated fiction.
Score: 55🌐 MovesJul 6, 2026https://restofworld.org/2026/china-ai-web-novels/?utm_source=rss&utm_medium=rss&utm_campaign=feeds - Scaling Security Alert Triage With Specialized Agents on Databricks
What if low severity didn't mean low priority?Databricks ingests petabytes of security...
Score: 55🌐 MovesJul 6, 2026https://www.databricks.com/blog/scaling-security-alert-triage-specialized-agents-databricks - Can AI really move the needle in mid-market businesses? - University of Oxford
Can AI really move the needle in mid-market businesses? University of Oxford - Saïd Business School
Score: 55🌐 MovesJul 6, 2026https://www.sbs.ox.ac.uk/oxford-answers/can-ai-really-move-needle-mid-market-businesses - How Nations Are Deploying AI for Strategic Priorities
Nations have long invested in domestic infrastructure to advance their economies, protect and use their data, and take advantage of technology opportunities in areas such as transportation, communications, commerce, entertainment and healthcare. AI, the most important technology of our time, is turbocharging innovation across every facet of society. Countries are investing in AI capabilities so […]
- Why AI agents will make your governance playbook obsolete
Large Australian banks have started implementing agentic AI at scale this year. The same is happening across the country’s larger enterprises. These are not pilots, but production systems that are the tip of a global trend. Research company Gartner expects 40% of enterprises to embed AI agents in applications by the end of 2026, up from less than 5% in 2025 — an eightfold jump in 12 months. Most governance teams at enterprises are responding to these changes by building playbooks as they always have: through committees, policies, approval gates and periodic audits. It’s a model that assumes humans review most decisions and that governance is a central function that sets and enforces all the rules. None of those assumptions holds for AI agents. The playbook most organisations are drafting at this moment is already obsolete before it is even finished. In our view, three central changes must happen together for AI governance to work in environments where agents are used at scale: Enterprises need agents’ behavioural telemetry, which they currently lack, to better understand what the bots are doing from a quality, compliance and cost perspective. They need controls that operate at the speed of the systems they govern, which means AI helping govern AI. They must distribute accountability across the organisation, because no centralised system can keep up with what is happening. Find it and measure it “You cannot govern what you cannot measure” sounds like a platitude, but most enterprises do not yet know how to measure AI agents’ actions. For instance, what counts as normal agent behaviour? What telemetry tells you that an agent is drifting from its original scope? What does an incident look like when it is not a single breach but four hundred micro-decisions, each individually defensible, that add up to an outcome you would never have approved or anticipated? Businesses are still working this out. In the meantime, they are building and deploying agents without the capacity to see or understand in detail what they are doing. A 2026 Gravitee survey found that only 24.4% of organisations report having full visibility into how AI agents communicate with one another. Almost nine in ten (88%) have reported “confirmed or suspected” agent security incidents in the past year. Despite that, 82% of executives say they are confident that existing policies protect against unauthorised agent actions. These numbers indicate a widening gap between confidence and risk whenever AI agents are deployed. In my experience working with enterprise clients across the region, this gap is not a matter of negligence. Most governance teams are doing what they have always done well. The problem is that the systems they are now responsible for are behaving in ways their existing tools were never designed to detect. Visibility matters because policy, control frameworks and ethics boards are all scaffolding that collapses without behavioural data that can be analysed and acted upon. You cannot write a meaningful policy for a system whose normal actions and performance you have never characterised or audit an agent whose decisions you cannot investigate or understand. In this context, observability means instrumentation, baselines, anomaly detection on agent behaviour and telemetry that humans can interpret. Governance at machine speed Once you can see and understand what agents are doing, the next problem is how to manage them at scale. When the average enterprise runs 12 agents, as Salesforce’s 2026 Connectivity Report suggests, human oversight is still feasible. When leading deployments are already running into the hundreds — IQVIA has deployed more than 150 agents , for example — that approach stops working. We don’t necessarily need a completely new security framework, but the updated model must allow companies to operate non-human interactions at scale with confidence. And the only way to ensure this is to use AIs to govern other AIs, because it is not economical to rely solely on humans to do the work. This agentic AI governance model needs to monitor agentic behaviour and respond within milliseconds when needed. It must provide situational awareness and key insights to support informed decision-making. Humans are responsible for establishing the parameters and guardrails, but they only intervene on demand and spend most of their focus on continuously improving their AI governance capabilities and the governance agents. Distributed accountability If governance must be observed continuously at machine speed and at a high level of complexity, no single function can do it all. That is why another essential change is organisational, with accountability distributed by design. Today, most businesses use a centralised model that no longer works. Legal owns policy; security focuses on runtime monitoring and response; developers build controls into the agents themselves. The problem is that each function is limited in its own way. Security can monitor telemetry data, for instance, but lacks insight into what each agent is supposed to do and therefore cannot develop customised anomaly-detection controls. Closing the gaps in the overall systems requires understanding a specific approach to building agents. Developers cannot simply ship unmanaged agents that operate in stealth mode. They need to send key metrics to a centralised AI governance layer. To enable this layer, a clear shared responsibility model between the developer and governance functions must be defined. Developers are required to implement reporting hooks that generate data to create key metrics and task-specific insights and detect anomalies across clearly defined governance domains. The centralised AI governance layer analyses this incoming data from the agents and provides situation awareness across all deployed AI agents. Guardrails that are baked into AI agents can’t be trusted, as they have proven to be vulnerable to prompt injection attacks. That is why an independent AI-powered governance layer is required to supervise all agent behaviour and provide insights and key metrics to key stakeholders. Distributed accountability like this is hard to set up. It requires an understanding of the reasons behind the changes and an agreement between many stakeholders on how the new model must operate and where different responsibilities lie. It also needs a shared language across functions that have not historically worked together at this pace and clarity about who decides what when something goes wrong. But it is the only model that survives an environment with AI agents deployed at scale. Faster decisions, new mindset One way to think about these changes is to reflect on cloud adoption over time. That experience showed us that investing in governance and assurance early reduced risks and created a competitive advantage for the businesses that understood why they should do it. The same dynamic is playing out with AI agents, only faster, more distributed and very likely at a much larger scale. Managing security, compliance, privacy, responsible AI, quality and cost in an agentic world at machine speed hasn’t been done before. Vendors help innovate on the customer’s behalf and can offer building blocks for this governance layer. But organisations also need to consider creating AI-powered custom capabilities to fill their processes and observability gaps. AI governance teams, therefore, require engineering capabilities and an agentic development environment that is tightly integrated with out-of-the-box AI security and compliance solutions. What I see across the market right now is that the professionals responsible for governance have the expertise and experience to lead this shift. What they need is the confidence to rethink their operating model and recognise that the centralised control they are accustomed to will not scale for agentic environments. Good governance practices that consider people, processes and technology have always paid off in the long run. Now is the time to define the individual North Star for your AI governance layer, because retrofitting these capabilities will carry high risk and cost. This article is published as part of the Foundry Expert Contributor Network. Want to join?
Score: 55🌐 MovesJul 6, 2026https://www.cio.com/article/4192402/why-ai-agents-will-make-your-governance-playbook-obsolete.html - The Rise Of The Agent Manager In The Modern Enterprise
It's likely we'll soon have the “hybrid workforce” running the modern enterprise, and the role of a human “agent manager” will become critical.
- Why Chinese youth aren’t booing AI, unlike American graduates
This spring, American graduates did something commencement audiences rarely do: they booed the future being sold to them. At several universities, students jeered speakers who praised artificial intelligence (AI), so often that a National Public Radio report advised this year’s orators to avoid the subject altogether. That reaction is too easily dismissed as technophobia. These are among the most digitally fluent graduates ever. What they reject is not AI; they reject a version of it in which...
- ‘AI era intensifying loneliness, burnout must become a leadership priority’
Kunal Sood, founder of AudacityAI said AI has been accelerating at extraordinary speed, but many of the challenges people are facing today are fundamentally human
- Companies are buying AI tools. That doesn't mean they know what to do with them.
Companies are buying AI tools. That doesn't mean they know what to do with them. Business Insider
- Robots available for rent: But what can they do?
Robotics tech is changing fast, so for many it makes sense to rent a robot.
Score: 52🌐 MovesJul 6, 2026https://www.bbc.co.uk/news/articles/c4gymkg9lr2o?at_medium=RSS&at_campaign=rss - The growing AI expectation gap: Employees are expected to be AI-ready, but are organisations preparing them?
As AI tools become part of everyday work, employees are increasingly expected to use them to improve productivity and decision-making. Yet many organisations have not provided structured training or clear guidance, creating an expectation gap that affects confidence, performance and long-term workforce readiness.
- AI is forcing open-source projects to rethink their contributor rules
Does your open-source project allow agent-generated pull requests? OSS community practices are diverging
Score: 52🌐 MovesJul 6, 2026https://www.thestack.technology/open-source-projects-ai-contributors-rules/ - How quantum technologies could open new frontiers for AI
This three-part blog series explores the growing complementarity between artificial intelligence (AI) and quantum technologies. The first post introduced quantum technologies and outlined their strengths and the challenges of combining AI with quantum systems. The second examined how AI can support the development of quantum technologies, helping to optimise systems and accelerate progress towards practical […] The post How quantum technologies could open new frontiers for AI appeared first on OECD.AI .
Score: 52🌐 MovesJul 6, 2026https://wp.oecd.ai/how-quantum-technologies-could-open-new-frontiers-for-ai/ - Naukri launches an AI recruitment platform in India for smarter hiring
Naukri launches an AI recruitment platform in India for smarter hiring YourStory.com
Score: 51🌐 MovesJul 6, 2026https://yourstory.com/2026/07/naukri-launches-ai-recruitment-platform-india - AI Job Disruption Has Come for Ireland's Technology Sector
The country benefited from a jobs boom after attracting US multinationals, but some of these roles are being cut in the name of AI efficiency.
- How I run autonomous coding agents from my phone with OpenAI Symphony + Linear | Alessio Fanelli (Kernel Labs)
Watch now | 🎙️ Alessio Fanelli shows his Symphony + Linear setup for running parallel coding agents from his phone, then demos Codex hunting underpriced Pokémon cards in real time
- Sakana AI Launches Sakana Translate, a Namazu-Powered Japanese–English–Chinese Translation Tool With Translate, Proofread, and Ask Modes
Sakana AI Launches Sakana Translate, a Namazu-Powered Japanese–English–Chinese Translation Tool With Translate, Proofread, and Ask Modes MarkTechPost
Score: 50🌐 MovesJul 6, 2026https://www.marktechpost.com/2026/07/05/sakana-ai-launches-sakana-translate/ - Every major tech layoff in 2026 that has name-checked AI
A running look — in reverse chronological order — at the bigger tech companies that have announced significant layoffs this year with AI as a stated factor.
Score: 50🌐 MovesJul 6, 2026https://techcrunch.com/2026/07/06/the-running-list-major-tech-layoffs-in-2026-where-employers-cited-ai/ - Understanding Annotator Safety Policy with Interpretability
Safety policies define what constitutes safe and unsafe AI outputs, guiding data annotation and model development. However, annotation disagreement is pervasive and can stem from multiple sources such as operational failures (annotators misunderstand or misexecute the task), policy ambiguity (policy wording leaves room for interpretation), or value pluralism (different annotators hold different perspectives on safety). Distinguishing these sources matters. For example, operational failures call for quality control, ambiguity calls for policy clarification, and pluralism calls for deliberation…
Score: 50🌐 MovesJul 6, 2026https://machinelearning.apple.com/research/annotator-safety-policy-interpretability - How Open Models Are Driving AI Research
Every year, the International Conference on Machine Learning (ICML) reveals where thousands of AI researchers have decided to put their work. This year’s accepted papers reveal a clear direction: open frontier models and open AI infrastructure have become foundational to how modern AI science gets done. NVIDIA had 74 papers accepted at ICML 2026. Approximately […]
- The future CIO mandate: Managing compute, data, and intelligence as core business assets
By Vishal Sirohi, CEO and Co-Founder, Island Computing The defining mismatch in enterprise technology spending is now in plain view. Gartner projects global AI spending will reach US$2 trillion in […] The post The future CIO mandate: Managing compute, data, and intelligence as core business assets appeared first on Express Computer .
- The 5 Open Models Worth Knowing in 2026, and Exactly What Each One Is Best At
The open-model world moves so fast that “which is best” changes almost monthly, and the honest answer is that there’s no single best, there’s a best for your specific job. So instead of one ranking, here is the intuitive version, five open models that have each carved out a clear identity, and a plain-language guide to which one you actually reach for depending on what you are trying to do. Think of it as a toolbox, not a leaderboard. If you’ve tried to keep up with open-weight AI models this year, you know the feeling. A new one tops the leaderboard, a think-piece declares it the new king, and three weeks later a different model from a different lab takes the crown. Chasing the number-one spot is a losing game, because the ranking genuinely changes month to month. The more useful way to think about it is the way you think about tools in a workshop. You don’t ask which single tool is best, you ask which tool is right for the cut you’re making. Open models have matured to the point where the leading ones each have a distinct personality and a job they do better than the rest. Learn those identities once, and you stop chasing leaderboards and start picking the right model on instinct. Here are the five worth knowing, and exactly what each is for. A quick note before the list. All of these are open-weight models, meaning you can download and run them yourself, and most of them are Mixture-of-Experts designs, which is why models with hundreds of billions of parameters can still run on a single serious GPU, only a fraction of the parameters activate on any given token. Benchmarks quoted here come from the labs and public leaderboards, so treat them as a starting signal, not gospel, and always test on your own task before committing. 1. DeepSeek V4, the serious coding and agent workhorse If your work is building coding agents or software that has to reason across a large codebase, this is the one to start with. DeepSeek V4 Pro has become the reference point for agentic coding among open models, posting a SWE-Bench Verified score around 80 percent, a level that effectively ties the best closed frontier models on that benchmark, and topping the neutral leaderboards for real-world agentic work. It pairs that with a 1 million token context window and a memory design tuned for keeping long agent runs coherent, which matters because coding agents fail exactly when logs, file diffs, and previous steps overflow the context. There’s also a lighter, cheaper variant, DeepSeek V4 Flash, which is the cost floor of the whole category at roughly 14 cents per million input tokens. That makes DeepSeek unusual, the same family gives you both the capability ceiling for hard agent work and one of the cheapest options for high-volume tasks. Reach for it when, you’re building a coding agent, working with large repositories, or running long multi-step tool-use workflows where staying coherent over a long session is the whole challenge. Think twice when, you need something small and simple for a laptop, the top model wants serious multi-GPU hardware to run well. 2. Qwen 3.6, the best all-rounder you can actually run If DeepSeek is the specialist heavyweight, Qwen is the practical default, and it has quietly become the most widely deployed open-model family in the world. The reason is balance. The mid-size Qwen 3.6 model, around 27 billion parameters, posts coding scores competitive with much larger models while fitting comfortably on a single 24GB consumer GPU with quantization. It’s strong across coding, reasoning, and multilingual work, it ships under a clean Apache 2.0 license that lets you do essentially anything commercially, and it comes in the widest range of sizes of any family, from tiny edge models up to frontier-scale versions. That combination, genuinely good at most things, easy to run, clean to license, and available in whatever size fits your hardware, is why Qwen is the model most people should try first when they do not have a specialized need. It’s the sensible starting point that will handle the majority of tasks well. Reach for it when, you want one dependable model for general use, you need clean commercial licensing, you are running on a single GPU, or you want multilingual support. Think twice when, you have a specialized frontier need like the hardest agentic coding, where a dedicated specialist may edge it out. 3. Gemma 4, the one for your laptop and the edge If the constraint is hardware, if you want to run a capable model on a laptop, a small workstation, or an edge device, Gemma 4 from Google is the standout. Its smaller variants are engineered for exactly this, one version runs in around 6 gigabytes of memory, small enough for modest consumer hardware, while still offering multimodal understanding and tool calling. It ships under Apache 2.0, and it has been specifically optimized to run well on common consumer GPUs, which makes it the natural pick for developer workstations and on-device prototyping. The tradeoff is honest, Gemma 4’s largest model is smaller than the giant frontier models, so for the very hardest multi-step reasoning it trades some peak capability for the ability to run almost anywhere. For most everyday tasks that trade is well worth it, and for anything running locally on limited hardware, this is where to start. Reach for it when, you’re deploying on a laptop or edge device, running on limited memory, building on-device or offline features, or prototyping locally without a big GPU. Think twice when, you need maximum reasoning depth on very hard problems, the smaller size shows its limits at the extreme end. 4. Llama 4 Scout, the long-context champion If your problem is size, feeding an enormous amount of text into the model at once, Llama 4 Scout stands alone. Its context window reaches around 10 million tokens, which is far beyond anything else on this list and enough to hold entire codebases, whole books, or massive document collections in a single prompt. When your challenge isn’t the difficulty of the reasoning but the sheer volume of material the model needs to see at one time, nothing else comes close. Llama also carries the advantage of ecosystem. Meta’s Llama family remains the most widely deployed and tooled open-weight lineage, so whatever framework, tutorial, or integration you need almost certainly supports it. The one caveat worth knowing is the license, Llama uses Meta’s community license rather than a pure Apache or MIT one, which is fine for most users but carries a usage cap that very large companies need to check. Reach for it when, you need to process huge documents or entire codebases in one pass, you want the broadest ecosystem and tooling support, or long-context recall is the core of your problem. Think twice when, you’re a very large company that needs to review the license terms, or you need the absolute top coding or reasoning score, where specialists lead. 5. GLM-5.1, the clean-license enterprise coder If you’re building inside a company that cares about license clarity, and you want frontier-level coding without legal review headaches, GLM-5.1 is the pick. It posts one of the strongest coding scores among open models, leading on the harder SWE-Bench Pro benchmark where it edges out even some leading closed models, and it’s built for long-horizon agentic engineering, the kind of multi-step autonomous software work that businesses increasingly want. Crucially, it ships under a plain MIT license, which is about as permissive and unambiguous as licensing gets, with no clauses to parse. That mix, top-tier coding capability plus the cleanest possible license, is what makes GLM-5.1 a favorite for enterprise teams that need to deploy something serious without a legal review of usage caps and restrictions. It’s a frontier-class coder you can build a product on without worrying about the fine print. Reach for it when, you need strong coding and agentic capability with a completely clean license, you are deploying in an enterprise that requires license clarity, or you want long-horizon autonomous engineering. Think twice when, you need to run on light hardware, its frontier size wants real infrastructure. The honorable mentions, because the field is deep Five is a clean number, but a few others deserve a nod because they’re genuinely excellent at something specific. Kimi K2.6, from Moonshot, is a standout for multi-agent orchestration, running coordinated swarms of agents on complex long-horizon tasks, and it’s one of the top all-round open models by the neutral indexes. If your work is heavily agentic, it belongs on your test list alongside DeepSeek. DeepSeek R1 remains the go-to open reasoning model, its chain-of-thought approach is excellent for step-by-step math and logic, and its distilled smaller versions bring real reasoning down to laptop-sized models. If your problem is hard reasoning rather than coding, start here. Mistral Small 4, from the European lab Mistral, is worth knowing for teams that want a capable model under a strict Apache 2.0 license with a strong quality-to-size ratio, and it’s a common pick where European data considerations matter. It runs on a single GPU and offers configurable reasoning effort. None of these is the wrong choice. They’re simply tuned for narrower jobs than the main five, and if one of those jobs is yours, it may be exactly right. How to actually choose, the thirty-second version Here’s the whole thing boiled down to instinct. If you’re building a serious coding agent or working across a large codebase, start with DeepSeek V4. If you want one dependable, easy-to-run, cleanly-licensed model for general use, start with Qwen 3.6, and honestly, most people should start here. If you need to run on a laptop or limited hardware, reach for Gemma 4. If you need to feed enormous amounts of text into the model at once, use Llama 4 Scout. And if you need frontier coding inside an enterprise with a spotless license, pick GLM-5.1. The meta-lesson matters more than any single pick. The open-model field now has genuine depth, several models that each rival closed frontier systems on the specific task they are built for, and the skill is no longer finding the one best model. It’s knowing which tool fits the job in front of you, and matching them well. The leaderboard will change again next month. The identities of these tools, and the habit of choosing by task, will still serve you. One last honest note. Because these models move so fast, treat the specific scores as a snapshot and the categories as the durable part. A new release may reshuffle who leads on coding or context next quarter, but “pick the specialist for hard agent work, the all-rounder for general use, the small one for local, the long one for volume, and the clean-license one for enterprise” is a framework that will outlast any individual model on this list. Learn the shape of the toolbox, and you’ll always know roughly where to reach. If you have deployed any of these on real work, drop a comment with which you chose and what you used it for. The categories hold, but the best signal on any specific model is always someone who has actually run it on a task like yours. The 5 Open Models Worth Knowing in 2026, and Exactly What Each One Is Best At was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- What billions of AI predictions taught Expedia before the age of AI agents
There's an important distinction between AI that just works today, and AI that lasts at scale. Many companies optimize hard for the first one without ever asking whether they're building the second. Velocity without discipline and strategic direction is a liability, not an asset. The hardest part of building AI at scale isn't getting a model to work once. It's building systems that continue to work, scale beyond individual teams and use cases, and improve consistently over time. Today's AI systems do more than just predict and optimize. They converse, reason, and increasingly take action. An autonomous system making decisions on a traveler's behalf creates a very different set of expectations around reliability, governance, and accountability. As AI takes on more of those roles, the principles behind how these systems operate matter more than ever. We have spent years applying AI and machine learning (ML) across the traveler journey — from personalization, ranking, and recommendations, to fraud prevention, customer support, and, more recently, generative and agentic AI experiences. That depth of experience is what led us to develop a set of ML and AI principles to guide how we build, deploy, and evolve AI systems across our company. The goal is simple: Make sure the systems we build create real business value, scale, and operate safely. These principles define how we measure, design, govern, and operate our systems. From principles to practice Publishing principles is the easy part. The harder and more important work is turning them into operating mechanisms: Recommendations, requirements, tooling, and release processes that teams actually use. We have begun using 'Agentic Release' tollgates: A set of recommended and, in some cases, required checks before launching agentic AI features. These tollgates translate principles like clear ownership, risk-based governance, evaluation, safe rollout, and monitoring into concrete expectations for teams. Some of these recommendations and requirements are already being automated and integrated into the software development lifecycle (SDLC). Over time, the goal is for these expectations to become embedded in how we design, evaluate, approve, launch, and monitor AI systems from the start. Outcomes: Measuring what actually matters The first test for any model is whether it improves a business outcome and, ultimately, the traveler experience — not whether it just improves a technical metric. Align models to metrics with business impact: Every ML effort must tie directly to a key business outcome or traveler experience metric. Technical optimizations are useful midpoints, not end goals . Optimize for return on cost : The value a model creates has to justify what it costs to develop, train, and monitor, plus the operational complexity it adds. Favor solutions that deliver lasting impact relative to what they cost to run. Justify complexity against strong baselines: Complexity should be earned, not assumed. Start with a strong baseline: An existing general model, a simple heuristic, an off-the-shelf solution. Reach for specialized models or more complex architectures only when simpler options genuinely can't meet the bar. Require both offline and online evaluation : No model goes to broad deployment on offline validation alone or jumps straight to A/B testing. Every model must perform in both offline and online evaluations. Over time, our offline evaluations should reliably predict what we see online. Design: building systems that scale beyond the teams that build them Getting a model to work is one challenge. Making its value extend beyond a single team or use case is the harder one. Build on shared foundations; specialize only when justified: Favor shared, platform-wide foundations for core capabilities, data representations, and model building blocks. Specialization should build on those foundations, not spin up isolated stacks, so when the foundation improves, the gains flow across the organization. Treat data as a first-class product : A model's quality is bounded by the quality of its data. We need to maintain robust pipelines, clear lineage, reproducibility, and reusable features built with documented ownership, clear schemas, and SLAs that other teams can rely on. Prioritize generality over local optimization : When two approaches perform similarly, favor the one whose learnings, assets, and operating patterns can be reused across teams, brands, and use cases. We should optimize not just for local performance, but for how quickly improvements can diffuse across the company and compound over time. Minimize and sunset manual business rules: Manual rules are sometimes necessary for policy, safety, or compliance, but they should be explicit and reviewed regularly, never silent patches for weak models or a source of permanent maintenance debt. Reproducibility and traceability by default : Training data, features, configurations, evaluation results, deployment versions, and key decisions should all be documented and recoverable. That's what lets you debug a production issue months later and hand off ownership without losing institutional knowledge. Trust: ownership, governance, and operating responsibly at scale The bar for deploying AI isn't just "does it work?" It's "can we stand behind it?" Trust isn't something you add at the end; it's earned over time and maintained across the full lifecycle of every model we ship. Assign clear ownership and accountability: Every model needs defined ownership across its lifecycle — a business owner, a product owner, an AI owner, and an operational owner. These don't need to be four people, but the responsibilities must be explicit. Who's accountable for outcomes? Who responds if the model drifts? Who answers the incident at 2 a.m.? Without this in place, models become orphaned and problems surface with no one to own them. Adhere to standards and governance: AI and ML models must use approved platforms and comply with established company standards, release gates, and governance processes. Operating outside these guardrails requires a clear, defined path to remediation or deprecation, rather than an open-ended exception. Govern proportionally to risk : The level of review, evaluation rigor, and human oversight should scale with a model's impact. A customer-facing model that affects pricing or availability for millions of travelers demands a far higher bar than an internal tool used by a small team. For high-impact, safety-sensitive, or highly autonomous systems, human-in-the-loop checkpoints are built in from the start. Design for fairness, privacy, and transparency : We actively test for unintended bias, have strong data guardrails, and favor explainability when decisions meaningfully affect users. These are incorporated from the start, not added on. Design for safe rollout, rollback, and control : Deployments are progressive, with rollback paths, fallback mechanisms, and circuit breakers ready before launch. The ability to safely undo a deployment matters as much as the ability to ship it. Monitor continuously and adapt: Once live, teams must actively monitor quality, drift, latency, cost, and business performance and retrain or recalibrate when the data shifts. A team should always be able to explain how its model is performing now, not just how it performed when it launched. These principles do more than define how we build. They define what we're willing to ship and how we stand behind it. In a world where AI systems are increasingly consequential and make real decisions for real travelers and partners, these standards matter. Applied consistently, they build responsible AI that lasts. Xavi Amatriain is Chief AI and Data Officer at Expedia Group Xavier will share more details about Expedia's architecture during his session at VB Transform on July 14 at 11:10 am PT. He will discuss: "Expedia's blueprint for building autonomous agents for high-stakes transactional systems." Interested in attending VB Transform 2026? Register here . A select number of complimentary passes are also available to senior technology leaders. Contact us to get yours.
- Small businesses are adopting AI, with some snafus along the way
Small businesses are adopting AI, with some snafus along the way Business Insider
Score: 49🌐 MovesJul 6, 2026https://www.businessinsider.com/ai-transforms-small-businesses-but-challenges-persist-2026-7 - 6 ways to make AI accountability stick
As intelligent systems move into production environments and begin taking actions, organizations quickly discover that accountability becomes much harder . Unlike traditional enterprise software, these tools can produce unpredictable outcomes as they interact dynamically with data, APIs, and business workflows. “When something goes wrong with AI, it is generally assigned to whoever was closest to the pain point,” says David DuChene , manager of data and AI pre-sales at SHI International, which works with enterprises on AI deployments and governance. As these systems shift from advisor to actor within workflows, accountability becomes harder to enforce through policies alone. IT leaders must build it directly into the fabric of their operations through clear ownership, continuous observability, defined escalation paths, and infrastructure designed to make responsibility visible when things go wrong. Here are six ways to make AI accountability enforceable in production. 1. Assign direct ownership from the beginning Many enterprises still view AI accountability as a shared responsibility, but some experts argue that this is the first assumption to fail when systems enter production. “Shared accountability is not accountability,” says Joe Wilson , SVP and CIO of CSG, a customer experience, billing, and payments software provider. “You need a direct owner.” He says that at CSG, AI initiatives go through governance reviews involving executive leadership, and direct ownership is assigned at the start of projects. Wilson, who oversees the AI governance and deployment strategy for CSG, says the company also created “CIO reps” embedded inside business units and product groups to ensure accountability spans the entire lifecycle of AI initiatives. According to SHI’s DuChene, many enterprises still lack formalized accountability structures for those environments. “They may have responsible parties on paper, but once a system actually breaks down, everything gets relitigated,” he says. “It goes back to who’s closest to the pain point.” One diagnostic question, he argues, reveals whether organizations are truly prepared: “If your AI deployment generates a wrong answer and costs the business money tomorrow, who’s going to write the postmortem?” If leaders cannot answer that question quickly, accountability structures likely don’t yet exist in practice. 2. Build governance before scaling deployments In the past few years, many enterprises deployed AI systems before establishing the governance and operational foundations necessary to support them safely. “The biggest gap we see is a sequencing problem,” says DuChene. “We’ve gone around and built a bunch of houses where we’re standing up the walls before we’re pouring the foundations.” That sequencing problem creates expensive retrofitting efforts later. DuChene says teams frequently discover they lack data classification systems, AI-aware identity and access controls, lineage and provenance tracking, audit capabilities, and escalation channels for failures. According to Seth Dobrin , CEO of deterministic AI model maker Arya Labs and former global AI leader at IBM, governance often fails because organizations treat it as a policy layer rather than something embedded directly into operational workflows. “How do you integrate it into the workflow?” he asks. “If you don’t get that right, it’s going to fall apart.” Dobrin recalls working with an insurance company that spent 18 months building an intelligent system before legal teams blocked deployment entirely. The problem was not the technology itself, but the absence of governance early in the process. “They had to throw it away,” Dobrin says. “Had they started earlier, they would have steered it to a place where they could have gotten to yes.” Dobrin says governance should not slow projects down. Instead it should be integrated deeply enough into workflows that teams can move quickly without downstream compliance or operational failures. “The objective should never be to say no,” he says. “It should always be to figure out how to say yes.” Wilson at CSG makes a similar point, arguing that governance should help teams absorb complexity rather than simply restrict what they can do. He compares it to a vehicle suspension system rather than a braking mechanism. “Our intention is not to slow things down,” he says. “Our intention is to speed stuff up, but also when you get into rough terrain, to be able to navigate that terrain.” 3. Treat data governance as the foundation of accountability According to Wilson, CSG focused on governing its data before scaling AI initiatives across the business. Those efforts started with data synchronization and privacy impact assessments. “The foundation is data,” Wilson says. “If we don’t have clean, synchronized, and governed data across the board, we’re not going to win this battle.” Many organizations underestimate how difficult it becomes to maintain accountability once AI systems begin interacting with fragmented enterprise data environments, says Quais Taraki , CTO of EnterpriseDB, a company that works with enterprises on data infrastructure and governance. An AI assistant summarizing customer interactions, for example, may pull regulated or confidential data from systems that were never intended to feed generative AI tools. Strong data governance practices — including lineage, provenance tracking, classification systems, and access controls — not only help head off such problems but also create the foundation for accountability when something does go wrong. Otherwise, teams struggle to determine what data an AI system accessed, how outputs were generated, and whether sensitive information influenced a decision. “Without lineage and provenance, you can’t do root-cause analysis,” Taraki says. “You won’t know what to change, or how things mutated in ways you didn’t expect.” Taraki argues that accountability should follow governed data products rather than organizational silos. When ownership is split across infrastructure teams, data scientists, and application developers, responsibility can become difficult to establish after failures occur. Assigning clear ownership to the data products that feed AI systems helps create accountability throughout the AI lifecycle. 4. Build observability into (and beyond) AI systems Traditional enterprise monitoring systems were designed primarily to track uptime, infrastructure health, and application performance. AI introduces a different challenge: tracing reasoning paths, decision chains, and behavioral drift. Nik Kale , a member of the Coalition for Secure AI (CoSAI) and participant in AI security and agent identity standards efforts, describes this through what he calls an “Investigation Graph.” This is a reasoning trail showing what an AI system observed, what tools it accessed, what conclusions it reached, and what actions it ultimately took. “When something breaks, the first instinct is always to ask, ‘Why did the AI make that decision?’” Kale says. “Honestly, I think that’s the wrong question. The right question is, ‘What did the system actually do?’” That distinction is increasingly important because AI failures rarely originate from a model alone. Instead, they emerge from interactions between models, credentials, APIs, workflows, policies, and downstream systems. “The model didn’t act,” Kale says. “The system around the model acted.” That broader view of accountability is also changing how IT leaders think about observability. Rather than monitoring AI models in isolation, enterprises increasingly need visibility across the systems those models interact with, including data sources, APIs, applications, security controls, and downstream workflows. In practice, that starts with comprehensive logging of prompts, model outputs, tool calls, data access events, and agent actions. Combined with traditional application and infrastructure telemetry, those logs create an auditable record of how AI systems behaved and why decisions were made. That visibility becomes especially important when IT leaders try to identify unauthorized AI usage. While governance policies define which tools employees should use, observability helps reveal which tools they are actually using. Unusual data access patterns, unexpected API calls, traffic to external AI services, and unexplained movement of sensitive data can all be indicators of shadow AI . Even well-governed organizations can struggle when employees adopt unauthorized AI tools outside approved workflows. “If it’s shadow IT, we don’t even know it exists,” says DuChene. “We don’t know what data of ours is going into it, how it’s being used, or how it’s being distributed.” By extending observability beyond AI models to the broader enterprise environment, IT can detect those activities earlier, investigate them more quickly, and reduce the accountability gaps that shadow AI creates. 5. Create ‘escalate’ and ‘stop’ mechanisms The most important accountability question may not be what an AI system can see or do, but when it should stop and ask for help. According to Kale, that’s often the most underdeveloped part of enterprise AI deployments. “Most enterprises have figured out how to monitor their AI systems,” he says. “But nobody has really built the third piece, which is, when does the system actually stop and ask for help?” Kale argues that enterprises need explicit escalation paths, human decision points, and clearly defined stop mechanisms for systems operating in production. “You don’t want a rubber stamp — you want a human in the loop,” he says, adding that the human should be named and have the authority to say no. According to Wilson, incident response processes also need to evolve, because AI failures behave differently from traditional IT outages. “A traditional IT incident typically looks like it’s an up or down scenario,” he says. “AI failures are a little more subtle than that.” Models may drift gradually, outputs may degrade over time, or workflows may begin producing unexpected results without systems technically failing. The result, says Wilson, is a growing need for multidisciplinary response processes involving legal, communications, security, audit, business teams, and IT operations simultaneously. 6. Treat AI systems more like workers than software Some enterprises still govern AI like traditional applications. But according to Kale, AI systems behave more like workers and less like deterministic software. “You cannot just deploy once and be done,” he says. “Like workers, they need ongoing oversight.” That ongoing oversight is becoming a core accountability function. Employees are not hired, trained, and then left unsupervised indefinitely. Managers monitor performance, provide feedback, evaluate changing responsibilities, and intervene when behavior drifts from expectations. Kale argues that AI systems increasingly require similar treatment. Traditional software can often be reviewed and approved at release time because its behavior remains relatively stable between versions. AI systems are different. Models evolve, prompts change, retrieval systems are updated, and the information available to agents changes continuously. That challenge extends beyond internally developed systems. Enterprises must also monitor the third-party AI services they rely on. Not only do vendor models evolve on their own, but vendors also update software and capabilities behind the scenes. “The vendor we approved last quarter is functionally a different vendor this quarter,” Kale says. As a result, accountability cannot end when a system is deployed. Someone must remain responsible for monitoring performance, reviewing changes, assessing risk, and determining whether systems continue to operate within acceptable boundaries. Kale points to CoSAI’s AI Shared Responsibility Framework as one emerging effort to clarify those responsibilities across enterprises, software vendors, model providers, and infrastructure operators. The organizations making the most progress are discovering that accountability cannot be assigned on paper and forgotten. As AI systems become more autonomous, accountability is becoming an operational capability built into data governance, observability, escalation processes, and ongoing oversight. For IT leaders, the challenge is no longer defining responsibility. It is making responsibility enforceable. Related reading: CIOs plagued by growing AI accountability gap AI is spreading decision-making, but not accountability Who authorized the AI agent? Breaking the blame loop in agentic AI Why AI adoption keeps outrunning governance — and what to do about it 5 ways to curb AI sprawl without stifling innovation
Score: 49🌐 MovesJul 6, 2026https://www.computerworld.com/article/4184169/how-to-make-ai-accountability-stick.html - Meet Arpana Shahi, the woman entrepreneur building an AI-powered longevity platform with Gabit
Meet Arpana Shahi, the woman entrepreneur building an AI-powered longevity platform with Gabit YourStory.com
- Beyond the AI arms race: the cooperation we’re missing
Beyond the AI arms race: the cooperation we’re missing Bennett School of Public Policy
Score: 48🌐 MovesJul 6, 2026https://www.bennettschool.cam.ac.uk/blog/beyond-the-ai-arms-race-the-cooperation-were-missing/ - Beyond automation: Why AI infrastructure must be designed around business risk and resilience
By Amit Jaju, Senior Managing Director – India, Ankura Consulting AI has quietly shifted from being a set of experiments at the edge of the enterprise to becoming part of […] The post Beyond automation: Why AI infrastructure must be designed around business risk and resilience appeared first on Express Computer .
- If you use Google, you’re training its AI. Here’s how to opt out.
Consider this a belated PSA: A recent change to Google’s privacy settings is allowing the company to store more of your data, including media such as “images, files, and audio and video recordings,” to improve its AI models.
Score: 47🌐 MovesJul 6, 2026https://techcrunch.com/2026/07/06/if-you-use-google-youre-training-its-ai-heres-how-to-opt-out/ - Path-Constrained Mixture-of-Experts
Sparse Mixture-of-Experts (MoE) architectures route each token through a subset of experts at each layer independently. We propose viewing MoE computation through the lens of expert paths—the sequence of expert selections a token makes across all layers. This perspective reveals that, despite N^L possible paths for N experts across L layers, tokens in practice cluster into a small fraction of paths that align with linguistic function, yet the vast majority of paths remain unexplored, representing a statistical inefficiency. This motivates architectures that constrain the effective path space…
Score: 47🌐 MovesJul 6, 2026https://machinelearning.apple.com/research/path-constrained-mixture-experts - The Hidden Psychological Cost Of Having AI As Your Boss
As AI takes on more management decisions, leaders face an unexpected challenge: preventing technology from weakening trust, relationships, and employee engagement.
- Revisiting ASR Error Correction with Specialized Models
Language models play a central role in automatic speech recognition (ASR), yet most methods rely on text-only models unaware of ASR error patterns. Recently, large language models (LLMs) have been applied to ASR correction, but introduce latency and hallucination concerns. We revisit ASR error correction with compact seq2seq models, trained on ASR errors from real and synthetic audio. To scale training, we construct synthetic corpora via cascaded TTS and ASR, finding that matching the diversity of realistic error distributions is key. We propose correction-first decoding, where the correction…
- [Framework] The Asymmetric Key-Value Cache Compression
Why hardware pipelines hate irregular memory layouts and what you should do instead. A tactile physical installation illustrating the asymmetric decoupling of the key-value cache to optimize memory bandwidth in hardware pipelines. I remember standing in the glass-walled server room of a major consumer internet company, watching a tier-one systems architect stare at a dashboard flashing deep, angry amber. We were attempting to deploy a state-of-the-art 175-billion parameter model with a newly expanded context window, expecting a triumph of modern engineering. Instead, we were witnessing a physical crisis of memory bandwidth, a silent strangulation where the model was bottlenecked not by its static weights, but by the sheer volume of its historical conversational state. Every single user token required us to store, retrieve, and process keys and values, turning our elegant silicon architecture into a sluggish, overstuffed filing cabinet. In that moment, it became blindingly obvious that the race for million-token context windows is not an algorithmic victory, but a physical war against modern physics. 📊 Executive Summary: Scaling LLM contexts creates a massive memory wall, forcing the KV cache of a 175B model to balloon to 1.2TB. This framework addresses this bottleneck by decoupling Key (per-channel) and Value (per-token) quantization while neutralizing Token Norm Imbalance via Omni-Scaled Canalized Rotation (OScaR). Compressing caches to sub-1-bit limits (0.81 bits/activation) yields a 5.3x memory footprint reduction and a 3.0x decoding speedup with zero linguistic degradation. I. The 1.2 Terabyte Memory Wall: Why Context Expansion is Breaking Modern Silicon The contemporary gold rush toward million-token context windows is celebrated as a cognitive leap for machine intelligence, but behind the scenes, datacenter operators are quietly drowning in hardware bills. Autoregressive decoding is inherently bound by memory bandwidth rather than raw arithmetic capability, forcing memory architectures to constantly stream historical data to the processor cores (Wang et al., 2024). When a standard 175-billion parameter model processes a high-concurrency batch, its dynamic Key-Value (KV) cache can swell to an astronomical 1.2 Terabytes (Liu et al., 2024; Su et al., 2026). This colossal footprint completely eclipses the static weights of the model itself, forcing standard systems to spill their cache lines over into slow off-chip memory and stalling computation (Liu et al., 2024). To survive this memory wall, we must abandon the long-held industry assumption that symmetric post-training quantization is sufficient for extreme compression. Traditional low-bit quantization schemes apply uniform reduction across weights, activations, and caches, but this blunt approach triggers severe representation loss at the context boundaries (Dettmers et al., 2022; Su et al., 2026). This article introduces The Asymmetric Key-Value Cache Compression Framework, a paradigm shift that decouples Key and Value quantization matrices while mathematically neutralizing Token Norm Imbalance (Liu et al., 2024; Su et al., 2026). By tailoring the quantization strategy to the unique mathematical topologies of Keys and Values, this framework compresses the KV cache to sub-1-bit levels (0.81 bits per activation) while fully preserving the model’s complex reasoning capacity (Liu et al., 2024; Zhang et al., 2024). Our journey through this architecture will trace the systemic failures of naive symmetric quantization, explore the non-overlapping mathematical structures of Keys and Values, and examine the rotation techniques that eliminate token-level variance. Finally, we will outline a complete, co-designed pipeline that unites sub-1-bit caches with ultra-low-bit weights, creating an engine optimized for next-generation hardware. A physical scale model depicting a massive concrete wall blocking high-bandwidth data pathways, visualizing the 1.2 Terabyte memory wall bottleneck. “We shape the silicon, but the memory shapes the model.” — Mohit Sewak, Ph.D. II. The Symmetrical Failure: Why Naive Quantization Triggers Representational Collapse When machine learning engineers attempt to shoehorn a massive model’s activations into low-bit spaces using traditional absolute maximum (absmax) scaling, the result is almost always catastrophic. Standard absmax quantization relies on a simple scaling factor, calculated by dividing the target discrete integer range by the maximum absolute value present in the tensor: S = target_integer_range / max(|X|) While this formula works perfectly for stable, bounded distributions, it disintegrates when applied to the activation states of modern deep transformers (Dettmers et al., 2022). The culprit is a highly systematic structural anomaly: the spontaneous emergence of massive activation outliers (Dettmers et al., 2022). In models exceeding several billion parameters, specific hidden dimensions develop activation values that can exceed magnitudes of 1⁰⁵, while the surrounding features maintain a quiet median value of roughly 0.1 (Su et al., 2026). When absmax scales this tensor, the enormous magnitude of the outlier forces the scaling factor S to become incredibly small. As a consequence, the normal features that carry the vast majority of the model’s linguistic nuances are rounded down to absolute zero, causing extreme quantization degradation and an immediate drop in zero-shot accuracy (Dettmers et al., 2022). Naive Absmax Quantization Squeezing: [ 0.1, 0.12, 0.08, 100000.0, 0.11 ] <-- Raw Activations with Outlier (10⁵) ↓ (Scaled by S = 127 / 100000) [ 0, 0, 0, 127, 0 ] <-- Normal features obliterated to zero! This structural squashing triggers what researchers term representational collapse — a state where the model’s parameters lose their geometric diversity and information capacity (Huang et al., 2026; Oberländer et al., 2026). During Quantization-Aware Training (QAT), this manifests as weight trapping (Huang et al., 2026). Because the forward pass uses discrete values but the backward pass relies on real-valued gradients calculated via Straight-Through Estimators (STE), the lack of gradient variance leaves weights trapped in local, sub-optimal regions (Huang et al., 2026). In Post-Training Quantization (PTQ), the failure mode is even more immediate (Oberländer et al., 2026). Quantization errors do not remain isolated; instead, they compound exponentially as they propagate through the deep layers of the model (Oberländer et al., 2026). Without corrective measures, this cumulative quantization bias leads to a total collapse of the model’s expressive capabilities, resulting in a perplexity explosion that can exceed 1⁰⁸ in uncorrected 1-bit baselines (Oberländer et al., 2026). A physical mechanical metaphor demonstrating representational collapse, where a massive steel outlier forces an absmax clamp to crush normal glass activations to zero. 🔍 Fact Check: While developers expect quantization noise to propagate linearly, forensic analyses of PTQ baselines reveal that uncorrected 1-bit quantization causes an immediate exponential amplification of error across deep layers, triggering a catastrophic perplexity explosion exceeding 1⁰⁸ on standard language evaluation benchmarks. This issue highlights a fundamental software-hardware gap (Dettmers et al., 2022). Unlike weight outliers, which are static, offline properties of a fully trained model, activation and KV cache outliers are dynamic, runtime phenomena generated in response to incoming user tokens (Dettmers et al., 2022). Since activations cannot be pre-processed offline, they represent the single most stubborn bottleneck in extreme low-bit serving (Dettmers et al., 2022). III. Pillar I: Decoupling Key and Value Asymmetry via KIVI Partitioning The foundational breakthrough of the KIVI (Key-Value Quantization) framework was the discovery that Key and Value caches exhibit completely divergent, non-overlapping statistical topologies (Liu et al., 2024). Key and Value matrices serve fundamentally different roles in the attention mechanism, and their activation distributions reflect this functional division (Liu et al., 2024). In the Key cache (K), activation outliers are strictly channel-wise (Liu et al., 2024). Across the entire sequence dimension, specific, predictable channels consistently exhibit massive activation magnitudes (Liu et al., 2024). Conversely, the Value cache (V) acts as an attention output mixer and displays no such channel-wise anomalies (Liu et al., 2024). Instead, variations in the Value cache are stochastically distributed per-token, meaning the magnitude spikes are tied to individual sequence steps rather than specific channels (Liu et al., 2024). Statistical Topologies: Key Cache (K): [Token Dim] × [Channel Dim (Outliers Concentrated in Specific Columns)] Value Cache (V): [Token Dim (Outliers Concentrated in Specific Rows)] × [Channel Dim] When engineers apply a symmetric, uniform quantization scheme across both caches, they force a lose-lose compromise (Liu et al., 2024). If they quantize per-token, the Key cache’s massive channel outliers dictate the scaling factor for the entire token vector, obliterating the normal channels (Liu et al., 2024). If they quantize per-channel, they fail to capture the token-level variations of the Value cache, introducing significant reconstruction errors (Liu et al., 2024). A physical KIVI sorting rig illustrating asymmetric partitioning, separating Key activations per-channel and Value activations per-token. The KIVI framework elegantly resolves this dilemma through a tuning-free, asymmetric quantization pipeline (Liu et al., 2024): ┌───────────────┐ │ Raw KV Cache │ └───────┬───────┘ │ ┌────────────┴────────────┐ ▼ ▼ ┌───────────────┐ ┌───────────────┐ │ Key Cache (K) │ │Value Cache (V)│ └───────┬───────┘ └───────┬───────┘ │ │ ▼ ▼ [Per-Channel Quant] [Per-Token Quant] │ │ └────────────┬────────────┘ ▼ ┌───────────────┐ │ FP16 Sliding │ (Most recent 32 tokens) │ Window │ └───────────────┘ Quantize the Key cache (K) strictly per-channel, confining the outlier variance to its native dimension and preserving normal channels (Liu et al., 2024). Quantize the Value cache (V) strictly per-token, capturing its stochastic, sequence-dimensional variations (Liu et al., 2024). Maintain a small, full-precision (FP16) sliding window for the most local relevant sequence tokens (e.g., the last 32 tokens) to guarantee high-fidelity local context processing (Liu et al., 2024). 💡 ProTip: When implementing asymmetric KV cache partitioning, never quantize the local sliding-window buffer. Pinning the most recent 32 tokens in FP16 acts as an anchor for local syntactic dependencies, ensuring that the cumulative quantization error of historical tokens does not poison the immediate next-token generation pipeline. By splitting the quantization strategy along these asymmetric lines, KIVI enables stable 2-bit KV cache quantization without requiring any expensive fine-tuning or calibration (Liu et al., 2024). In production environments, this simple partition reduces memory overhead so dramatically that it delivers up to a 3.47x higher inference throughput (Liu et al., 2024). IV. Pillar II: Eradicating Token Norm Imbalance with Omni-Scaled Canalized Rotation (OScaR) While KIVI’s per-channel and per-token split operates beautifully at 4-bit and 2-bit levels, pushing the boundaries to extreme 1-bit or sub-2-bit limits reveals a deeper structural bottleneck: Token Norm Imbalance (TNI) (Su et al., 2026). Token Norm Imbalance describes a phenomenon where the L₂ norm of individual token vectors varies wildly across the sequence dimension (Su et al., 2026). Highly attended tokens — such as the initial “attention sink” tokens — possess massive activation energy, while normal tokens in the middle of the sequence carry negligible norm (Su et al., 2026). Under standard per-channel quantization, a single shared scaling parameter is calculated across the sequence dimension for each channel (Su et al., 2026). Because this parameter must span both high-norm and low-norm tokens, it systematically amplifies the reconstruction error of the low-norm tokens (Su et al., 2026). When pushed to the 1-bit limit, this imbalance causes a total collapse in the model’s linguistic fidelity (Su et al., 2026). To solve this physical limitation, the OScaR (Omni-Scaled Canalized Rotation) framework introduces a dual-stage mathematical pipeline that flattens this imbalance with near-zero computational overhead (Su et al., 2026). Token Norm Imbalance (TNI) Mitigation Pipeline: [Raw Key Cache X] ──► [Stage 1: Canalized Rotation (Walsh-Hadamard Y = X × H)] ──► [Stage 2: Omni-Token Scaling] ──► [Smooth Gaussian Distribution] Stage 1: Canalized Rotation (The Walsh-Hadamard Transform) First, the framework applies an online orthogonal transformation to rotate the activation space, using the formula: Y = X × H where H is an N × N orthogonal matrix where H_i,j ∈ {+1, -1} (Wang et al., 2025). This rotation distributes the concentrated energy of outlier spikes evenly across all dimensions, transforming a heavily skewed, heavy-tailed distribution into a neat, quantizable Gaussian curve (Wang et al., 2025). Because the Fast Walsh-Hadamard Transform can be executed in O(N log N) runtime complexity, it avoids expensive matrix multiplications and runs with near-zero latency overhead (Wang et al., 2025). 💡 ProTip: To implement the Walsh-Hadamard Transform without bottlenecking the decoding pass, fuse the rotation directly into the CUDA kernel of the preceding projection layer. Performing the rotation in-place bypasses the massive overhead of reading and writing unrotated activation tensors to High Bandwidth Memory. Stage 2: Omni-Token Scaling Immediately following the rotation, the framework applies token-level L₂ norm normalization (Su et al., 2026). This step mathematically equalizes the norms across the sequence dimension, effectively neutralizing Token Norm Imbalance (Su et al., 2026). By implementing this dual rotation-and-scaling sequence during Key matrix construction, system architects can completely decouple token magnitude from channel constraints (Su et al., 2026). The hardware benefits of this approach are profound. When benchmarked against BF16 FlashDecoding-v2, the OScaR framework delivers a 5.3x reduction in memory footprint and up to a 3.0x speedup in decoding, establishing a new Pareto frontier for extreme context serving (Su et al., 2026). V. Pillar III: Transcending the 1-Bit Barrier with Coupled Quantization (CQ) Even at 2 bits, a million-token cache demands significant memory bandwidth. To break through the 1-bit barrier, we must challenge the traditional assumption that individual activation channels are independent variables (Zhang et al., 2024). In reality, adjacent channels within a transformer’s embedding space share significant mutual information and deep interdependencies (Zhang et al., 2024). The Coupled Quantization (CQ) framework exploits these correlations to compress the cache to sub-1-bit levels without triggering accuracy collapse (Zhang et al., 2024). Instead of compressing each channel as an isolated scalar stream, CQ couples multiple channels together and encodes them as a unified vector (Zhang et al., 2024). Coupled Quantization (CQ) Grouping: Raw Channels: [ Ch 1, Ch 2, Ch 3, ... Ch 16 ] (High mutual information) │ │ (Grouped into blocks) ▼ CQ Joint Encoder: [ 16-Channel Contiguous Block ] │ │ (Fisher-Guided Centroid Search) ▼ Encoded Output: [ 12-bit Codebook Centroid ] <-- Average of 0.81 bits/activation! This vector quantization is implemented via a highly specialized three-step pipeline (Zhang et al., 2024): Step 1: Channel Partitioning. Partition the activation channels into non-overlapping, contiguous groups, typically consisting of 16 channels per block (Zhang et al., 2024). Step 2: Joint Codebook Encoding. Jointly encode these grouped channels using multi-dimensional centroids learned offline (Zhang et al., 2024). Step 3: Fisher-Guided Selection. To preserve high-impact activations, discard standard K-means clustering. Instead, use second-order Fisher Information matrices during centroid training to prioritize the preservation of channels that contribute most to the model’s loss function (Zhang et al., 2024). This channel-coupling paradigm allows for extreme codebook sharing. Mapping a 12-bit code to a group of 16 coupled channels yields an average of only 0.81 bits per activation (Zhang et al., 2024). 🔍 Fact Check: While 1-bit is widely regarded as the absolute minimum representation limit, the Coupled Quantization (CQ-16c12b) framework achieves a sub-1-bit average of 0.81 bits per activation by exploiting the shared mutual information across 16 contiguous channels, enabling a 15x increase in concurrent serving capacity. For system architects, the optimal deployment strategy combines this CQ-16c12b configuration with a tiny, high-precision sliding window of 32 to 128 FP16 tokens for localized, high-fidelity context processing (Zhang et al., 2024). This sub-1-bit strategy allows data centers to scale concurrent batch sizes by up to 15x compared to standard FP16 baselines, maximizing hardware utilization (Zhang et al., 2024). VI. Pillar IV: Co-Designing the Complete Pipeline — W1A4 Weights and H-BitLinear Integration Compressing the KV cache to extreme low-bits solves the memory capacity issue, but if the rest of the compute pipeline is poorly integrated, the system will still stall (Wang et al., 2024; Wang et al., 2025). For instance, if weights are compressed to 1.58 bits (as in BitNet b1.58) but activations remain in FP16 or INT8 format, the hardware cannot leverage fast INT4 or binary tensor cores (Dettmers et al., 2022; Ma et al., 2024). The entire inference engine is dragged down to the speed of the slowest common denominator (Dettmers et al., 2022). Standard BitNet b1.58 vs. H-BitLinear Co-Design: ┌──────────────────────────┐ ┌──────────────────────────┐ │ BitNet b1.58 (W1A8) │ │ BitNet v2 (W1A4KV2) │ ├──────────────────────────┤ ├──────────────────────────┤ │ Weights: 1.58-bit │ │ Weights: 1.58-bit │ │ Activations: 8-bit │ │ Activations: 4-bit (Rot) │ │ KV Cache: 16-bit │ │ KV Cache: 2-bit/Sub-1 │ └──────────────────────────┘ └──────────────────────────┘ (Stalls on mixed pipelines) (Fully integer-native speed) To resolve this, modern deployment architectures must co-design weights, activations, and caches (Wang et al., 2024). The most elegant solution utilizes BitNet v2 (H-BitLinear), which enables native 4-bit activations for 1-bit models (Wang et al., 2025). The H-BitLinear module leverages online Hadamard rotations prior to activation quantization (Wang et al., 2025). By rotating the activations, the module smooths heavy-tailed outlier distributions, allowing them to fit neatly into narrow 4-bit integer grids with near-zero latency overhead (Wang et al., 2025). This achieves the W1A4 standard, unlocking fully integer-native execution (Wang et al., 2025). An alternative approach is hybrid sparsification, exemplified by BitNet a4.8 (Wang et al., 2024). This architecture recognizes that different components of a transformer exhibit different sensitivities to precision reduction (Wang et al., 2024). It applies 4-bit quantization to the inputs of the Multi-Head Attention (MHA) and Feed-Forward Network (FFN) layers, where representations are stable, but retains an 8-bit pathway for highly sensitive intermediate states (Wang et4.8; Wang et al., 2024). To offset the memory cost of the 8-bit path, BitNet a4.8 employs aggressive Top-K sparsification via a squared ReLU gating mechanism, activating only 55% of the total parameters during inference (Wang et al., 2024). By combining asymmetric KV cache compression (such as OScaR or KIVI) with these W1A4 weight architectures, system architects can build fully integer-native pipelines (Ma et al., 2024). These pipelines bypass floating-point matrix multiplications entirely, executing operations as simple, energy-efficient additions and bit-shifts (Ma et al., 2024). VII. The W1A4KV2 Dawn: Transitioning from High-Precision Monoliths to Ultra-Compressed Edge Engines The evolution of LLM quantization has reached a pivotal turning point. The industry is rapidly moving away from fragile post-training hacks on FP16 models and embracing holistic hardware-software co-design (Su et al., 2026). This paradigm shift culminates in the W1A4KV2 architecture (1-bit weights, 4-bit activations, and 2-bit or sub-1-bit KV caches) as the new gold standard for high-performance inference (Su et al., 2026). ┌────────────────────────────────────────┐ │ The W1A4KV2 Gold Standard │ ├────────────────────────────────────────┤ │ Weights: 1-bit Ternary │ │ Activations: 4-bit Smooth Rotated │ │ KV Cache: 2-bit Asymmetric/Sub-1 │ └────────────────────────────────────────┘ Looking further ahead, the next frontier points toward event-driven neuromorphic architectures, such as Spiking Neural Networks (SNNs) (Zhu et al., 2024). Frameworks like SpikeGPT and Dual-Path SparseTCAM bypass continuous algebraic calculations entirely, representing information as discrete binary {0, 1} spikes that propagate through temporal membrane dynamics (Zhu et al., 2024). While still in their infancy, these neuromorphic designs promise to reduce the power envelope of language models by orders of magnitude (Zhu et al., 2024). For engineering teams operating in the here and now, the path forward is clear. Transitioning to asymmetric KV cache compression and low-bit weight co-design is the only viable way to escape the physical constraints of the memory wall. If your team is ready to scale context windows without breaking your hardware budget, download our open-source, hardware-aligned packing kernels for asymmetric KV caching, or subscribe to our Deep Tech Infrastructure newsletter to receive our next code-level implementation guide of the OScaR rotation pipeline. References & Further Reading Core Quantization & Weight Representation Concepts Dettmers, T., Lewis, M., Belkada, Y., & Zettlemoyer, L. (2022). LLM.int 8(): 8-bit Matrix Multiplication for Transformers at Scale. Advances in Neural Information Processing Systems , 35, 30336–30349. https://doi.org/10.48550/arXiv.2208.07339 Huang, H., Wu, D., Hu, Q., Yu, G., Yang, J., Zhu, J., Liu, X., & Wu, D. (2026). Sherry: Hardware-efficient 1.25-bit ternary quantization via fine-grained sparsification. 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) . https://doi.org/10.48550/arXiv.2601.07892 Ma, S., Wang, H., Ma, L., Wang, L., Wang, W., Huang, S., Dong, L., Wang, R., Xue, J., & Wei, F. (2024). The era of 1-bit LLMs: All large language models are in 1.58 bits. arXiv preprint arXiv:2402.17764 . https://doi.org/10.48550/arXiv.2402.17764 Oberländer, J., Finkbeiner, J., Schöfmann, C. M., & Neftci, E. (2026). GRINQH: Graded input-based quantization hierarchy for efficient LLM generation. arXiv preprint arXiv:2606.23419 . https://doi.org/10.48550/arXiv.2606.23419 Advanced Activation Scaling & Co-Design Paradigms Wang, H., Ma, S., & Wei, F. (2024). BitNet a4.8: 4-bit activations for 1-bit LLMs. arXiv preprint arXiv:2411.04965 . https://doi.org/10.48550/arXiv.2411.04965 Wang, H., Ma, S., & Wei, F. (2025). BitNet v2: Native 4-bit activations with Hadamard transformation for 1-bit LLMs. arXiv preprint arXiv:2504.18415 . https://doi.org/10.48550/arXiv.2504.18415 Next-Generation KV Cache & Neuromorphic Frontiers Liu, Z., Yuan, J., Jin, H., Zhong, S., Xu, Z., Braverman, V., Chen, B., & Hu, X. (2024). KIVI: A tuning-free asymmetric 2bit quantization for KV cache. International Conference on Machine Learning (ICML 2024) . https://doi.org/10.48550/arXiv.2402.02750 Su, Z., Yang, R., Zhang, C., Liu, Y., Zhang, Y., Wu, W., Xiong, J., Du, D., Zhuang, X., & Wong, N. (2026). OScaR: The Occam’s razor for extreme KV cache quantization in LLMs and beyond. arXiv preprint arXiv:2605.19660 . https://doi.org/10.48550/arXiv.2605.19660 Zhang, T., Yi, J., Xu, Z., & Shrivastava, A. (2024). KV cache is 1 bit per channel: Efficient large language model inference with coupled quantization. Advances in Neural Information Processing Systems , 37, 3304–3331. https://doi.org/10.48550/arXiv.2405.03917 Zhu, R.-J., Zhao, Q., Li, G., & Eshraghian, J. K. (2024). SpikeGPT: Generative pre-trained language model with spiking neural networks. Transactions on Machine Learning Research . https://doi.org/10.48550/arXiv.2302.13939 Disclaimer: The views and opinions expressed in this article are personal and do not necessarily reflect the official policy or position of any associated agencies, organizations, or the India AI Mission. AI assistance was utilized in the research, drafting, and ideation of this article. Licensed under CC BY-ND 4.0. [Framework] The Asymmetric Key-Value Cache Compression was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Scaling Properties of Continuous Diffusion Spoken Language Models
Speech-only spoken language models (SLMs) lag behind text and text-speech models in performance, with recent discrete autoregressive (AR) SLMs indicating significant computational and data demands to match text models. Since discretizing continuous speech for AR creates bottlenecks, we explore whether continuous diffusion (CD) SLM is more viable. To quantify the SLMs linguistic quality, we introduce the phoneme Jensen-Shannon divergence (pJSD) metric. Our analysis reveals CD SLMs, mirroring AR behavior, exhibit scaling laws for validation loss and pJSD, and show optimal token-to-parameter…
Score: 45🌐 MovesJul 6, 2026https://machinelearning.apple.com/research/scaling-properties-continuous-diffusion - Tracking AI in the supply chain
Supply chains networks span manufacturers, warehouses, shipping lines, trucking fleets, retailers and customers. Recent disruptions ranging from Covid-19 and extreme weather events to geopolitical conflicts in the Middle East and Eastern Europe have exposed the limitations of traditional planning tools. In response, companies are increasingly deploying AI to improve visibility, predict disruptions and, in some cases, make operational decisions without human intervention. Increasingly, however, the ambition extends beyond prediction and analysis . Some now believe AI can actively manage and optimise supply chains in real time, learning from outcomes and continuously improving performance. Few organisations have embraced that vision with greater intent than Minnesota-based logistics giant CH Robinson. Founded in 1905, the company now boasts 75,000 customers and 450,000 contract carriers, managing around 37 million shipments, or $23bn worth of freight annually. Last month, the company launched Lean AI Engineer, which builds on its existing Lean AI Planner to create what executives describe as an “agentic supply chain” – an AI ecosystem capable of continuously learning, adapting and acting across one of the world’s most complex logistics networks. Jordan Kass, CH Robinson vice-president of managed solutions, tells Computer Weekly that the Lean AI Engineer now effectively closes the loop: “It will run continuously, improve the operation it’s running and heal itself when something breaks – without an alert or a human noticing a problem first. The Lean AI Planner executes in real time while the Lean AI Engineer studies the results, identifies patterns, adapts logic and influences future decisions.” Kass explains that the company oversees a network connecting trucking operators, ocean shipping companies, airlines, rail and road freight providers spanning manufacturing, distribution, retail and customer delivery centres. Add in a constantly shifting mix of carbon emissions requirements, customs rules and regulatory obligations across hundreds of jurisdictions and the complexity deepens even further. It is precisely the sort of sprawling, interconnected network that would be almost impossible for humans alone to continuously optimise. Kass says the technology effectively ends the need for separate supply chain intelligence and orchestration tools. “It’s what businesses with complex logistics have wanted for decades.” The technology now handles 92% of fourth-party logistics shipments globally across trucking, ocean, air and rail, from the moment an order is created through tendering, routing, delivery, exceptions and carrier payment. “Now we’ve reached the point where our customers have an agentic supply chain – an entire AI ecosystem that continuously thinks, learns, adapts and acts,” CH Robinson CTO Mike Neill tells Computer Weekly. AI scaling human talent Sounding a voice of reason amid growing concerns AI is coming for people’s jobs, Kass stresses that CH Robinson’s approach is ultimately about scaling human talent. “This level of premium logistics service has traditionally depended on talented people to manage complexity, make smart decisions day to day and intervene during disruption,” he says. “The problem was that talent didn’t scale.” The company has changed this by encoding expertise – some 120 years’ worth – into the technology itself. This means shippers can access the same expertise consistently across every shipment, regardless of who is available, what time zone they operate in or how dramatically shipping volumes grow or spike. “Their team and our team can focus on strategic priorities and driving the best business results,” adds Kass. In a sign of what many enterprises – not only logistics companies – may look like in years to come, CH Robinson currently employs some 450 data scientists and software engineers. Inflated expectations Not everyone believes the path forward will be straightforward. Thomas O’Connor, Gartner vice-president for logistics and planning for APAC, says AI is currently at the “peak of inflated expectations”. He believes technology leaders across logistics and other industries are under growing pressure from the C-suite to deploy AI, even as many organisations continue to struggle with their definitions. “What kind of AI are people talking about?,” he adds. “There’s a desire for productivity improvements, yet a lack of clarity in terms of outcomes.” Chief supply chain officers…need to work with different ecosystem partners to ensure data provided into the data pool is accurate and safe Thomas O’Connor, Gartner Echoing remarks from CH Robinson’s Kass, O’Connor says supply chain organisations need to embrace a two-track approach focused on both “exploitation” and “exploration”, with the latter representing the most dramatic shift. He says the challenges facing technology leaders in logistics and supply chain are not fundamentally different from those confronting other sectors. “Chief supply chain officers need to have clarity in terms of data,” O’Connor says. This means understanding ownership, accurately identifying input sources and knowing where data is actually coming from. “They need to work with different ecosystem partners to ensure data provided into the data pool is accurate and safe.” Achieving all of this demands robust data governance frameworks. Managing uncertainty Deloitte Asia Pacific CEO Rob Hillard says AI is already transforming the supply chain and logistics sector, where there is a growing need to better manage what he terms “ambiguous exception management”. This has become a major priority for supply chain leaders as Covid-19, natural disasters and geopolitical conflicts have introduced unprecedented uncertainty into global logistics networks. Hillard says AI is expected to empower smaller manufacturers through lower-risk, data-driven experimentation, including product launches and expansion into markets directly aligned with supply chain realities and costs. He points to additive manufacturing and 3D printing as examples likely to benefit. “This could allow smaller businesses to create specialist products and distribute them more effectively,” he adds. Hillard says that digital leaders across supply chain ecosystems must also ensure systems can integrate and communicate if efficiencies are to be realised. At the same time, AI is creating a major push towards integration and interoperability across supply chains, manufacturers, logistics providers and technology platforms. 100 trillion data points CH Robinson says Lean AI Engineer can assess an entire supply chain in 25 to 30 minutes and determine improvements before performance is affected, compared with traditional supply chain assessments that can take up to four weeks and often focus on what has happened rather than what should happen next. While Lean AI Engineer delivers intelligence, Lean AI Planner manages shipments through hundreds of interconnected AI agents and in turn feeds more data back to Lean AI Engineer to develop even smarter refinements. As with all AI, success depends on managing and contextualising massive amounts of data. The company claims to now be managing 100 trillion data points across its global network. Kass explains that its Lean AI systems are able to understand customers’ supply chains from the inside out as they leverage data end-to-end across every step of the shipping process, above and beyond parts visible to disparate tools. The company’s 450 technology specialists play a key role in capturing and organising historical data reaching back to its earliest digital systems, while simultaneously collecting massive amounts of information about each client’s business and operating environment, eliminating generic or theoretical assumptions. One early-adopter customer realised annual savings of more than $1m by shifting from a variable shipping schedule to a weekly model. Another found that having one pickup serve three delivery locations cut loads by 81% and generated savings of around 40%. The robots are coming While robotics have long been integral to manufacturing and supply ecosystems, expect to see a sharp uplift in innovation and capability as AI seeps into this evolving industrial DNA. “Production lines are increasingly being integrated with robots,” Hillard notes. He cites Deloitte’s State of AI survey 2026 report , which observed that 2025 was the year physical AI – the merger of physical systems with AI – emerged from the realms of science fiction into mainstream business consciousness. Notably, while only 5% of surveyed organisations believe physical AI is transforming their industry today, more than 40% expect it will transform their industry within the next three years. Meanwhile, robots themselves are becoming vastly more intelligent and physically capable, in no small part due to advances in China, while the evolution of IoT and communications networks – from 5G and eventually 6G through to ubiquitous satellite connectivity – is opening new possibilities for smarter supply chains, logistics and manufacturing. This has seen the emergence of new players and innovations across AI, digital twins, robotics and industrial automation. Nvidia has emerged as a major force through its investments in physical AI, digital twins and industrial simulation platforms, while Siemens, Schneider Electric, ABB and a growing ecosystem of startups are developing technologies that connect AI-driven decision making with real-world operations. Meanwhile, software giants including SAP, Oracle, Microsoft and Salesforce are embedding generative and agentic AI capabilities into supply chain platforms. And specialists like Kinaxis, Blue Yonder, o9 Solutions, Manhattan Associates and Coupa now target everything from demand forecasting and inventory optimisation through to procurement, warehouse operations and transport logistics. If the first wave of AI in the supply chain was about analysing data, the next is most certainly all about acting on it at unprecedented scale with humans and machines working closely together to learn from the past, optimise the present, and, if to not actually predict, at least be better prepared for the future. If CH Robinson’s experience is anything to go by, the future may have already arrived. Read more stories about supply chain optimisation How AI can help to optimise supply chains under pressure: Globalisation boosted trade for decades, but crises, tariffs and climate change have made supply chains more fragile. Companies are responding with AI, nearshoring and planning. ‘Don’t break the business’ – Lessons from Ann Summers’ ESB transformation : Ann Summers’ technology and supply chain director Jeannette Copeland talks through lessons learned during the retailer’s recent ESB overhaul.
- TeraWulf shares soar after Anthropic leases data center in Kentucky
TeraWulf is a crytpo mining company that has pivoted to AI data center infrastructure. Its stock is up more than 80% this year.
- AI doesn’t eliminate inefficiency. It amplifies it
Over the past two years, I have spent a significant amount of time discussing artificial intelligence with technology leaders, business executives and teams across my own organization. Most conversations begin with questions about the use cases, tools, governance and return on investment. Leaders want to know which technologies are creating the most value, where to invest next and how quickly they should scale adoption. Those are important questions, but I have noticed another pattern emerging as organizations move beyond experimentation and begin embedding AI into everyday work. In many cases, the technology itself is not the primary obstacle to success. Instead, AI is exposing organizational challenges that have existed for years. Processes that were already inefficient become more visible. Ambiguous decision-making structures become harder to ignore. Accountability gaps that once slowed projects quietly now become more apparent as work accelerates. This has led me to a simple conclusion: AI does not eliminate inefficiency. It amplifies it. That observation should not be interpreted as a criticism of AI. In fact, it highlights just how powerful the technology can be. AI accelerates workflows, shortens analysis cycles, improves access to information and increases employee productivity. However, because it accelerates the way work gets done, it also magnifies the strengths and weaknesses of the operating environment in which it is deployed. Organizations with strong processes and clear accountability often realize value quickly. Organizations with operational complexity frequently discover that technology alone cannot overcome management challenges. AI accelerates existing operating models Many organizations approach AI as a technology initiative. They evaluate platforms, launch pilots and identify tasks that can be automated. While those activities are important, they can also create a false impression that AI itself is the primary driver of transformation. In my experience, the greatest value comes not from the technology alone but from the willingness to rethink how work gets done. AI can automate tasks, but it cannot redesign a broken workflow. If a process contains unnecessary approvals, duplicate activities, conflicting priorities or poorly defined handoffs, those issues remain regardless of how sophisticated the technology becomes. This idea is consistent with a broader lesson I explore in my latest book, Digital Inside Out : digital transformation succeeds when organizations focus first on how work gets done, how decisions are made and how accountability is established. Technology can accelerate performance, but it rarely compensates for weaknesses in the underlying operating model. In many cases, new technologies simply make those weaknesses more visible. Researchers at the MIT Sloan School of Management have reached a similar conclusion. Their work suggests that organizations generate the greatest value from AI when they redesign workflows rather than simply automate individual tasks. In other words, the most significant gains come from rethinking how work flows through the organization rather than accelerating isolated activities. I have seen this pattern repeatedly throughout my career. Enterprise systems did not fix poor business processes. Collaboration platforms did not automatically improve communication. Analytics tools did not create accountability. Each technology delivered substantial benefits, but only when accompanied by process redesign, governance improvements and leadership commitment. AI follows the same pattern. Organizations that simply layer AI on top of existing complexity often find themselves completing inefficient work faster. Employees may generate reports in minutes instead of hours, produce presentations more quickly and analyze larger volumes of information. Yet the underlying process may still contain the same bottlenecks that limited performance before AI was introduced. The technology increases speed, but it does not automatically improve effectiveness. Why decision-making becomes the new bottleneck One of the most interesting effects of AI is how it changes the nature of organizational constraints. Historically, many companies struggled because information was difficult to access. Data was fragmented across systems, reporting cycles were slow and analysis required significant manual effort. Leaders frequently spent considerable time gathering information before they could make decisions. AI is rapidly reducing those barriers. Teams can now summarize large volumes of information, identify patterns, generate recommendations and produce insights in a fraction of the time previously required. Access to information is becoming less of a competitive differentiator because the effort required to generate it continues to decline. As this happens, another challenge becomes more visible. Many organizations discover that their greatest constraint is no longer information. It is decision-making. When ownership is unclear, faster insights do not necessarily produce faster outcomes. Teams may have access to excellent recommendations yet still struggle to determine who is responsible for acting on them. Multiple stakeholders may believe they have authority over a decision. Escalations become more common. Consensus-driven cultures can become overwhelmed by the volume of information being generated. Some of the most difficult conversations I have encountered in AI initiatives have had little to do with models, prompts or technical architecture. Instead, they involve governance, ownership, accountability and decision rights. These challenges existed before AI, but the technology makes them more visible because it removes many of the delays previously associated with gathering and analyzing information. This trend is likely to become even more pronounced as organizations adopt AI agents capable of executing tasks and workflows. While technology can automate actions, accountability remains a leadership responsibility. Leaders must still determine who owns outcomes, who approves actions and who is responsible when decisions create unintended consequences. What leaders should fix before scaling AI Deloitte’s annual State of AI in the Enterprise research highlights the challenges organizations face when attempting to scale AI beyond pilots and isolated use cases. This finding reinforces a lesson many leaders are learning firsthand: realizing value from AI requires organizational change, process redesign and strong leadership, not just new technology For CIOs and business leaders, one of the most important priorities should be simplifying processes before automating them. AI can reduce manual effort, but it rarely eliminates complexity that has been embedded into a process over many years. Organizations often achieve greater value by removing unnecessary steps before introducing automation. As Jon McNeill writes in his book, The Algorithm, “No need to waste time speeding up the old process. Instead, design, simplify, optimize and begin to work your new process. Then speed it up.” Leaders should also establish clear decision rights before scaling AI-enabled workflows. As information becomes easier to generate, organizations need clarity regarding who is accountable for making decisions and driving action. Without that clarity, AI can create more recommendations than the organization is capable of acting upon. Another important consideration is measurement. Many organizations continue to evaluate AI success through adoption rates, license utilization or employee engagement metrics. While these measures provide useful signals, they do not necessarily reflect business value. Leaders should focus on outcomes such as productivity improvements, revenue growth, cost reduction, customer experience enhancements and risk mitigation. Most importantly, leaders should recognize that AI adoption is fundamentally a leadership challenge. Technology can accelerate work, but leaders determine how work is organized, governed, measured and improved. Organizations that treat AI solely as a technology initiative often struggle to move beyond experimentation. Organizations that use AI as an opportunity to improve processes, clarify accountability and modernize operating models are more likely to achieve sustainable results. As AI adoption continues to accelerate, I believe the organizations that realize the greatest value will not necessarily be those with the largest investments or the most advanced models. They will be the organizations willing to address the management and operational issues that AI brings into focus. In many cases, AI is not creating new problems. It is revealing existing ones with greater speed and clarity. That may be one of the most valuable contributions AI can make. By exposing inefficiencies that organizations have learned to tolerate, it creates an opportunity for leaders to address them directly. The companies that seize that opportunity will be better positioned not only to benefit from AI, but also to improve the way their organizations operate long after the current wave of innovation has passed. This article is published as part of the Foundry Expert Contributor Network. Want to join?
Score: 45🌐 MovesJul 6, 2026https://www.cio.com/article/4192406/ai-doesnt-eliminate-inefficiency-it-amplifies-it.html - BCSSL deploys AI-powered multilingual FIR app with Hyderabad City Police
Blue Cloud Softech Solutions Limited (BCSSL) has announced the deployment of AI-CopWriter, an AI-powered multilingual complaint recording application developed in collaboration with the IT Cell of the Hyderabad City Police. […] The post BCSSL deploys AI-powered multilingual FIR app with Hyderabad City Police appeared first on Express Computer .
- Dublin’s Everhaze secures €450k as AI assistant Lú launches in UK
The funding lands as Everhaze rolls out Lú, a conversational AI built for PR professionals, across Ireland and the UK from today. Read more: Dublin’s Everhaze secures €450k as AI assistant Lú launches in UK
- From shampoo to cookies, consumer products get an AI makeover
From shampoo to cookies, consumer products get an AI makeover Reuters
Score: 44🌐 MovesJul 6, 2026https://www.reuters.com/business/shampoo-cookies-consumer-products-get-an-ai-makeover-2026-07-06/ - TopoPrimer: The Missing Topological Context in Forecasting Models
We introduce TopoPrimer, a framework that makes the global topological structure of the series population an explicit input to any forecasting model. TopoPrimer improves accuracy across diverse domains, stabilizes forecasts under seasonal demand spikes, and closes the cold-start gap. Precomputed once per domain via persistent homology and spectral sheaf coordinates, TopoPrimer deploys per token for fully-trained models and as a lightweight adapter for pre-trained backbones. Of these two components, sheaf coordinates are the primary accuracy driver. Across four public benchmarks on Chronos and…
Score: 44🌐 MovesJul 6, 2026https://machinelearning.apple.com/research/topoprimer-topological-context