AI News Archive: July 30, 2026 — Part 10
Sourced from 500+ daily AI sources, scored by relevance.
- This AI Model Is the ‘Best AI Capitalist.’ Its Behavior Should Worry Every Business Owner
When Andon Labs had Claude Opus 5 run a simulated vending machine, it was deceitful, exploitative—and highly successful.
- AI’s impact on research and development is undeniable, say experts
Professionals from IAS and Rent the Runway explore the impact AI has had on organisational R&D. Read more: AI’s impact on research and development is undeniable, say experts
Score: 30🌐 MovesJul 30, 2026https://www.siliconrepublic.com/careers/ai-impact-esearch-development-space-undeniable-experts-leadership - How AI can improve site reliability engineering
How AI can improve site reliability engineering InfoWorld
Score: 30🌐 MovesJul 30, 2026https://www.infoworld.com/article/4197480/how-ai-can-improve-site-reliability-engineering.html - ‘You'll notice I do not call it AI music, because I don't think it's music’: Qobuz’s Managing Director Dan Mackta on music streaming’s biggest problem, and how he hopes to tackle it
‘The whole thing is just a really, really big bummer’ — Qobuz’s Managing Director on the scourge of AI-generated audio in streaming, and why he still refuses to call it music
- Your voice gives away more than your words ever will
Deepfake detection, stress signatures, and what 8 years of listening say
Score: 30🌐 MovesJul 30, 2026https://aitoolreport.beehiiv.com/p/your-voice-gives-away-more-than-your-words-ever-will - Graph Engineering: Engineering Coordination Instead of Smarter Agents
Reliable agent systems are not defined by the number of agents they contain. They are defined by the contracts that govern what moves between them. Graph Engineering: Why AI Agents Need Contracts, Not Just Prompts One agent can fix a bounded problem. Several agents can still leave a human doing the hardest work: deciding what starts next, which output is trustworthy, what state crosses a handoff, and when a failure should stop instead of retry. That is the coordination gap. The recent conversation around graph engineering is trying to name this layer of system design. The term is still fresh, and its definition is not settled. But the problem behind it is already visible in production agent systems: once work spans several agents, tools, evaluators, and human approvals, reliability depends less on the intelligence of any single model and more on the shape of the handoffs between them. Loop engineering designs how one unit of work converges. Graph engineering designs how many units of work coordinate. Graphs do not replace loops. A loop is a graph with a cycle. Production agent graphs usually contain several local loops: an implementation loop that repairs failing tests, a research loop that closes an evidence gap, or an operations loop that retries a recoverable action. The graph decides where those loops begin, what they are allowed to change, what evidence they must produce, and who is allowed to declare the wider task complete. The Problem Is Not More Agents. It Is More Relationships. The first useful agent system usually looks like this: prompt -> model -> tool -> observation -> retry That is enough for a focused task. As the work becomes longer-running, teams add a researcher, implementer, reviewer, security check, deployment gate, and human approver. The system is then described as “multi-agent,” but that label hides the real engineering question. What does the reviewer receive: a summary, a diff, test output, or the entire conversation history? Can the reviewer send work back? Can it veto a deployment? If two reviewers disagree, who resolves the conflict? Which data is shared, which data is immutable, and which node owns the final state? These are graph questions. They are not solved by adding another prompt or asking a supervisor agent to be “careful.” They require explicit control over state, routing, ownership, evidence, and authority. A Practical Vocabulary Before drawing an architecture, separate five concepts that are often collapsed into the word “agent.” An agent is a probabilistic actor. A node is any work unit: model call, deterministic function, evaluator, or human gate. A loop is a bounded feedback cycle. A graph is the topology of nodes, transitions, state, and dependencies. A governor owns decisions when nodes disagree, or risk is high. The important implication is easy to miss: not every node should be an agent. Routing, schema validation, permissions, joins, checks, and budget enforcement are usually better as deterministic software. They are faster, cheaper, easier to test, and less likely to invent a reason to proceed. A node deserves to exist when it has a distinct objective, permission set, input/output contract, or independent verifier. Draw boundaries around responsibilities, not around model count. A Graph Is More Than a Flowchart Every workflow can be drawn as boxes and arrows. That does not make it engineered. An engineering graph gives every arrow operational meaning. An edge is not simply “then.” It is a contract that states what structured state moves, what evidence is required, who owns the next state, what route type is being taken, and which budget or permissions are inherited by the destination. An edge contract carries evidence, authority, routing, and budget between work nodes. Consider the difference between these two handoffs: Researcher -> Implementer and: Researcher -> Implementer input: { verified_constraints, source_links, acceptance_tests } transition only when: each constraint has provenance authority: implementer may modify code, not acceptance tests budget: one branch, 30 minutes, no deploy permission The first is a diagram. The second is a control boundary. This is familiar terrain in distributed systems. Interfaces, ownership, retries, and failure semantics determine whether components compose. Agent systems need the same discipline, with one extra complication: some nodes are probabilistic and will confidently misread an ambiguous handoff. State Is the Real Multi-Agent Problem The hardest part of multi-node agent work is not calling models. It is deciding what state exists, who can change it, and how downstream nodes know whether to trust it. The tempting design is to give every agent the same global conversation history. That works until it does not. A speculative note becomes another node’s assumed fact. Parallel branches overwrite each other. A failed branch leaves behind state that looks complete. Context grows until nobody knows which part is evidence and which part is commentary. Use a typed state schema instead. Separate facts , proposals , decisions , artifacts , and budgets . They do not have the same trust level, so they should not live as undifferentiated text in one shared prompt. from typing import Literal, TypedDict class Evidence(TypedDict): kind: Literal["test", "source", "log", "review"] value: str source: str observed_at: str class Decision(TypedDict): status: Literal["accepted", "returned", "escalated", "stopped"] authority: str reason: str class GraphState(TypedDict): task_id: str facts: list[Evidence] proposals: list[str] decisions: list[Decision] artifacts: dict[str, str] remaining_retries: int The point is not this exact schema. The point is separation. A model’s plan should not have the same status as a failing test log. A reviewer comment should not become an accepted decision unless the graph has a route that grants it that authority. For consequential systems, an append-only evidence ledger is often safer than mutable shared memory. Nodes can add evidence; a governor or deterministic reducer decides what changes the accepted state. This makes the workflow inspectable after the fact. A Reference Graph for Software Delivery Consider a change request that affects a production service. The shape below is intentionally ordinary: intake, triage, investigation, implementation, test repair, verification, review, and a merge gate. The value is not novelty. The value is that every handoff has a reason to exist. A graph engineering control plane for agent work The implementation loop is local. It can modify code, run sandboxed tests, read resulting failures, and make a bounded repair attempt. It cannot silently lower acceptance criteria, approve its own output, or deploy to production. Independent verification is a separate node because it has a different incentive and authority. It checks the artifact against the original contract. It might run tests, inspect a diff, validate a schema, or ask a separate evaluator to review a claim. The exact technique is less important than the separation: the node that generated a result should not be the only node that accepts it. The graph also makes failure useful. An evidence gap returns to investigation. A known test failure returns to implementation. A policy conflict goes to a governor. Those are materially different routes. If they all become “ask the supervisor agent,” the system loses its ability to explain and control its behavior. The Failure Modes Worth Designing For More nodes do not automatically make a system safer. They introduce coordination failure modes that a good graph must make visible. Ambiguous ownership: two agents each assume the other approved a change. Assign one governor and explicit write authority. Shared-state races: parallel branches overwrite or invalidate each other. Use versioned state, immutable evidence, and merge rules. Context leakage: a speculative result becomes accepted fact downstream. Separate proposals from verified facts and preserve provenance. Invalid joins: branches finish, but their outputs conflict. Define join preconditions and a conflict-resolution route. Circular delegation: agents hand work around without a new observation. Bound cycles and require evidence before retrying. Consensus theater: several agents agree because they inherited the same flawed context. Use independent evidence, diverse checks, or a human gate. Unbounded fan-out: a broad task spawns costly duplicated branches. Set concurrency, spend, and deduplication limits. The warning is not “never use graphs.” It is that coordination complexity must pay for itself. A two-step task with one owner rarely needs a multi-agent topology. A typed function call and a deterministic test may be the better architecture. Workflow Graphs Are Not Knowledge Graphs The word “graph” also creates a common confusion. A knowledge graph represents connected domain information: entities, relationships, provenance, and semantics. RDF represents graph data as subject-predicate-object triples. Property-graph systems model nodes, relationships, labels, and properties. The graph is a substrate for retrieval, reasoning, and data integration. An agent workflow graph represents control: which work unit may run next, what it can receive, what evidence it must emit, and which authority can move the system into a new state. The two are complementary. A workflow graph can ask a knowledge graph for grounded context. A knowledge graph can help a verifier trace the sources behind a claim. But a retrieval graph does not decide whether a deployment is allowed, and an orchestration graph does not make facts trustworthy by itself. Start With One Loop, Then Earn the Graph A single feedback loop is a graph; a governed topology adds branching, verification, and approval. Start with one loop and earn the graph. The smallest useful architecture is usually a bounded loop with a goal, a tool surface, evidence, and a stop condition. Add graph structure only when a real coordination pressure appears. Externalize state. Replace hidden conversation history with named artifacts, evidence, and budgets. Add an independent verifier. Separate producing a result from accepting it. Introduce one explicit return route. Route evidence gaps, test failures, and policy conflicts differently. Branch only when work is genuinely independent. Parallelism without isolation often multiplies conflicts rather than throughput. Name the governor. Define the human, service owner, or policy node that resolves conflict and approves irreversible action. Measure the topology. Track where work waits, retries, escalates, fails, and consumes budget. The graph is a system, not a slide. If you cannot observe where work is cycling or why a transition fired, the topology is decorative. Actionable Takeaways Before adding another model call, ask seven questions: Can every node explain its responsibility, inputs, outputs, and permission scope? Does every edge specify the evidence required to move forward? Are facts, proposals, decisions, artifacts, and budgets represented differently? Can a verifier reject output without relying on the generator’s self-report? Are retries allowed only after new evidence or changed strategy? Is there a named governor for conflict, risk, and irreversible actions? Can an operator inspect why the graph branched, joined, escalated, or stopped? If the answer is no, the system may still work in demos. It will be difficult to operate when the task becomes ambiguous, expensive, or risky. Conclusion Prompting changes what a model is likely to say. Context engineering changes what it can see. Harness engineering limits what one run can do safely. Loop engineering helps a unit of work converge against evidence. Graph engineering adds the missing outer structure: how several bounded work units share state, hand off artifacts, run in parallel, resolve conflict, and stop under an authority visible to operators. The name may change. The requirement will not. Once a system contains more than one actor, reliability depends less on the brilliance of any individual node and more on the contracts that govern their relationships. Graph Engineering: Engineering Coordination Instead of Smarter Agents was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Tech company's AI literacy programs reach almost 500,000 students worldwide
A Minecraft-based program built with partners reached more than 240,000 students on its own.
Score: 30🌐 MovesJul 30, 2026https://www.bizjournals.com/sanfrancisco/news/2026/07/30/corporate-philanthropy-check-netapp.html?ana=brss_6150 - Beyond the Chatbot Hype: Why Mathematical Optimization Is the Next Frontier for Enterprise AI
n an era where enterprises are flooded with data but starved for actionable direction, traditional analytics and predictive models fall short of driving real operational impact. While dashboards highlight trends and chatbots capture headlines, the critical missing link remains the ability to execute mathematically sound, constraint-aware decisions at the speed and scale modern business demands. […] The post Beyond the Chatbot Hype: Why Mathematical Optimization Is the Next Frontier for Enterprise AI appeared first on CXOToday.com .
- PhonePe launches PulsePro: Actionable data intelligence for Indian businesses
PhonePe launches PulsePro: Actionable data intelligence for Indian businesses YourStory.com
Score: 29🌐 MovesJul 30, 2026https://yourstory.com/2026/07/phonepe-launches-pulsepro-actionable-data-intelligence-for-indian-businesses - Sick of A.I.-Generated Content? The ‘Slop Janitor’ Is Here to Help.
Pangram, an A.I. detection start-up, promises near-perfect accuracy in sniffing out writing and imagery that wasn’t made by humans. It’s raising some big questions along the way.
Score: 29🌐 MovesJul 30, 2026https://www.nytimes.com/2026/07/29/business/dealbook/pangram-ai-slop-detection.html - Revenue, not just automation: Why AI products transforming customer experience deserve recognition
As enterprise AI shifts from cost-saving to revenue generation, the ET Most Innovative AI Product Awards 2026 recognise AI innovations driving customer experience, marketing excellence and measurable business growth.
- The US Army put new battlefield tech to the test. Now it's deciding what isn't making the cut.
The US Army put new battlefield tech to the test. Now it's deciding what isn't making the cut. Business Insider
Score: 29🌐 MovesJul 30, 2026https://www.businessinsider.com/us-army-assessing-new-ai-enabled-battlefield-tech-2026-7 - Google AI Overviews Attribution: 5 Changes NJ Businesses Must Know
Google AI Overviews Attribution: 5 Changes NJ Businesses Must Know USA Today
- Hexaview Technologies Joins the OpenAI Partner Network
Hexaview Technologies Joins the OpenAI Partner Network azcentral.com and The Arizona Republic
Score: 28🌐 MovesJul 30, 2026https://www.azcentral.com/press-release/story/103560/hexaview-technologies-joins-the-openai-partner-network/ - AI and SEO Reshape How Skin Care Brands Compete Online
AI and SEO Reshape How Skin Care Brands Compete Online USA Today
Score: 28🌐 MovesJul 30, 2026https://www.usatoday.com/press-release/story/38803/ai-and-seo-reshape-how-skin-care-brands-compete-online/ - AI improves cover letters, not hiring chances
Study shows AI helps craft better cover letters but doesn't boost hiring odds.
- Why AI governance can’t wait: Seven steps every security leader should follow today
Effective AI use has become a competitive advantage, and organisations are deploying tools and agents across every part of the business. But with great innovation comes great accountability. Today’s AI isn’t just chatbots – agents are querying databases, calling tools and moving data. And this often happens without oversight over what they can access, or [...]
Score: 28🌐 MovesJul 30, 2026https://www.cityam.com/why-ai-governance-cant-wait-seven-steps-every-security-leader-should-follow-today/ - Why AI literacy may become the new financial literacy
For decades, we’ve taught financial literacy as one of life’s most important skills. Save before you spend. Build an emergency fund. Invest early. Understand compound interest. Avoid unnecessary debt. Those lessons remain just as relevant today as they were a generation ago. But I believe future generations will need to master another form of literacy […] The post Why AI literacy may become the new financial literacy appeared first on e27 .
Score: 28🌐 MovesJul 30, 2026https://e27.co/why-ai-literacy-may-become-the-new-financial-literacy-20260730/ - Tech Industry Alliance and RDI Hub launch AI skills programme
The four-week pilot aims to give graduates the opportunity to work on real industry challenges, learn from experienced professionals and build the confidence to apply AI in practical workplace settings. Read more: Tech Industry Alliance and RDI Hub launch AI skills programme
Score: 28🌐 MovesJul 30, 2026https://www.siliconrepublic.com/careers/tech-industry-alliance-rdi-hub-launch-ai-skills-programme - Why your data layer is AI’s most critical climate technology
I have spent enough time building cloud and AI infrastructure to know one thing: Efficiency problems never show up where teams expect them. They tend to sit just beneath the surface, quietly shaping outcomes long before they appear in the metrics anyone is tracking. Most enterprise conversations still center on token pricing. That focus makes sense. The per-unit cost of using models has fallen quickly, and organizations want to understand whether AI can scale without pushing budgets out of bounds. That question has become more complicated than it looks. At the same time, the nature of each interaction has changed. What used to resemble a straightforward exchange now involves retrieval and ongoing reasoning that unfolds across multiple steps. Even with lower unit costs, total consumption continues to rise because each request requires more underlying work. This creates a contradiction. Tokens are cheaper, yet overall compute demand keeps increasing. That compute demand has to go somewhere, and where it goes is a growing problem. The answer, increasingly, is the power grid. The International Energy Agency has warned that AI and data centers are becoming a major source of electricity demand, and Goldman Sachs Research has projected that data center power demand could rise 165% by 2030 compared with 2023 levels. Today, electricity consumption from data centers already amounts to roughly 415 terawatt hours, or about 1.5% of global electricity use , and has been growing at roughly 12% annually over the past five years. Much of the response has focused on models or hardware. Both matter, but they do not explain the full picture. In enterprise environments, a meaningful share of inefficiency begins before a model processes anything at all. It begins in the data layer. The paradox: Cheaper tokens, more total compute I have seen this pattern play out before. When unit costs drop, usage expands and everyone is surprised, every time. AI is following that same trajectory. Lower token pricing has made it easier to integrate models into more workflows. Those workflows have also become more involved. Systems retrieve broader context and revisit outputs before returning a result. The additional activity outweighs the savings from lower pricing. In many cases, a system prompt alone can consume 2,000–5,000 tokens before a user even contributes input, which means the baseline cost of each interaction has already expanded. This is no longer just a financial consideration. Increased compute demand brings increased energy consumption, which is becoming a central constraint for enterprise AI. Why agentic AI changes the energy equation Average load: It rarely tells the full story. A platform can appear stable while quietly absorbing constant internal work that never surfaces to users. Agentic AI introduces that kind of persistent demand. An agent continues operating in the background even after a user interaction ends. It checks for new inputs, revisits its internal state, then prepares the next action. That ongoing loop changes how infrastructure is consumed. Instead of handling activity in bursts, systems begin to carry a steady baseline load. Compute usage becomes continuous rather than intermittent. This shift exposes inefficiencies that might otherwise go unnoticed. A slow query or a fragmented data source affects far more than a single request. The same issue repeats continuously, which increases both cost and energy consumption over time. This matters even more as inference workloads scale, with projections suggesting they could account for more than 40% of data center demand by 2030. The overlooked source of energy growth: The data layer In almost every architecture review I’ve sat through, attention gravitates toward model performance or infrastructure spend. The data layer gets treated as an afterthought. That assumption falls apart the moment you move into retrieval-heavy AI. Agents depend on access to reliable context. They pull information from operational systems as well as analytical platforms. When those sources are disconnected, each interaction requires additional effort to assemble a usable view. That effort accumulates quickly. Data must first be located before it can be used. It often needs to be moved into another environment, then reshaped into a format the model can process. Only after that does it become useful for decision-making. Each step introduces overhead that consumes compute and energy. In practice, engineers can spend up to 40% of their time preparing data before it is even usable for downstream systems. The model remains the most visible part of the system, but the surrounding data work often determines how much energy the system ultimately uses. Fragmentation as the root cause Enterprise architectures evolve over time, and fragmentation usually reflects a series of reasonable decisions rather than a single mistake. The challenge appears when AI systems need to operate across all environments at once. Fragmentation forces repeated work. Systems retrieve overlapping datasets because no single source is trusted as authoritative. Pipelines reprocess information that already exists elsewhere. Teams build parallel structures instead of relying on shared ones. This pattern increases demand in ways that are easy to overlook. When retrieval becomes inconsistent, applications compensate by sending more context than necessary. Models then process larger inputs, which increases token usage without improving the quality of the outcome. What appears to be a model efficiency issue often traces back to data architecture. It also helps explain why as many as 60% of AI projects are abandoned before reaching production, often due to gaps in data readiness rather than model capability. Why sovereign, unified architecture changes the math I have seen organizations improve efficiency without changing models simply by reducing friction in how data is accessed. The good news is that the fix is closer than most teams assume. A more unified architecture shortens the path between a question and the data needed to answer it. When systems can access authoritative data directly, they avoid repeated transformations and unnecessary duplication. Retrieval becomes more precise, which allows models to operate with less excess input. This does not require consolidating everything into a single system. Enterprises will continue to operate across multiple environments. The objective is to reduce unnecessary movement and make data easier to use wherever it resides. Sovereignty and control play an important role as well. Organizations operating across different environments need a clear understanding of where data resides and how it can be used. When governance is built into the architecture, systems spend less time reconciling access and more time producing results. Reducing friction at this layer has a direct effect on both cost and energy use. What 2026 will expose The next phase of enterprise AI will shift attention from access cost to operating cost. Early efforts focused on whether organizations could use advanced models. The next stage focuses on whether those systems can run continuously across real workflows without creating unsustainable demand. That shift will make energy consumption more visible. It will also make inefficiencies harder to ignore. Some increase in demand reflects real value. Systems that improve decision-making or streamline operations will naturally require more compute. The more difficult question is whether additional consumption reflects useful work or avoidable overhead. When systems repeatedly move and reprocess the same data, the increase in energy use does not correspond to better outcomes. It reflects architectural inefficiency. CIOs will need to examine that distinction more closely. The takeaway for tech leaders Before adding infrastructure, ask an honest question: Is this capacity supporting real growth, or is it just hiding inefficiencies that should have been fixed first? Start with the data layer. Review how systems retrieve context. Look for duplication across environments. Identify where data is repeatedly transformed before it becomes usable. These patterns reveal whether the architecture is enabling efficient AI or creating unnecessary demand. Enterprise systems will always involve complexity. The objective is to ensure that complexity does not translate into avoidable work. In the agentic era, the efficiency of AI systems is closely tied to how much work happens before the model produces an answer. Data architecture plays a central role in determining whether that work remains controlled or expands beyond what is necessary. This article is published as part of the Foundry Expert Contributor Network. Want to join?
Score: 28🌐 MovesJul 30, 2026https://www.cio.com/article/4203021/why-your-data-layer-is-ais-most-critical-climate-technology.html - With AI, control matters more than capability
Ask most enterprise technology teams where they spend their AI strategy energy and you will get the same answer: figuring out which model to use. It feels like the right question. As organizations move from pilots into production and the real compliance, cost and continuity risks appear, it turns out to be the wrong one. Writing on CIO.com this year, Floyd DCosta argued the divide is between enterprises that own their AI and those that rent it , and later that closed-model dependency is outsourced intelligence with a vendor kill switch in your operations . He is right. But the ownership question raises a harder one: own it how? I believe the answer is open-weight and open-source models, not because they are cheaper, but because they are structurally better suited to how serious organizations need to govern and protect AI at scale. Why closed models create governance problems Building an enterprise AI program on closed, proprietary models from a single external provider is not a technology decision. It is a governance liability. The data confirms the exposure is already real. A June 2026 IBM Institute for Business Value study of 1,000 senior executives found that 91% do not fully understand their AI vendor dependencies, 71% said switching providers would be difficult, and 81% said a seven-day vendor outage would cause severe disruption. These figures describe the baseline condition of enterprise AI in 2026. Think about what you give up. You cannot audit the training data. You have no visibility into how the model changes between versions. Your cost structure is set by someone else’s pricing team. And if that provider faces a government directive, a supply disruption or a commercial decision to reprice, you have no leverage and often no warning. For a generic SaaS tool, that is an inconvenience. For organizations in defense, healthcare or financial services, where data handling is regulated by law and audit trails are mandatory, a vendor changing model access terms overnight is a compliance event. The most vivid demonstration came in June 2026, when the U.S. Commerce Department ordered Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals , including its own non-citizen employees. The result was a hard global shutoff for every customer, with no advance notice. Enterprises with production workflows on those models were left with nothing. The precedent is now set. This is what AI governance exposure looks like in practice. When your intelligence layer sits entirely outside your control, a single government directive, pricing change or vendor decision can bring your AI operations to a halt. Having seen organizations scramble through exactly this scenario, I can say the ones with no continuity plan are the most exposed. The IBM numbers confirm it: most enterprises have not built the visibility, let alone the architecture, to absorb this kind of disruption. The question is not whether it will happen again. It is whether your architecture is ready when it does. Open-weight models have changed the equation Until recently, the argument for closed frontier models was simple: they were dramatically better. That gap has narrowed faster than most enterprise technology leaders anticipated, and the conversation has shifted from capability to control. Open-weight models, including Meta’s Llama family, Alibaba’s Qwen series, Zhipu AI’s GLM and DeepSeek, have moved well past the research stage. They are running in production at serious organizations, and not because those organizations could not afford anything better. They chose them because open-weight models give them something closed models cannot: control. Airbnb’s adoption of Alibaba’s Qwen makes the case plainly. CEO Brian Chesky stated that the company relies heavily on Qwen to power its customer service agent, describing it as “very good” and “fast and cheap,” while noting that OpenAI’s SDK was not ready for the depth of integration Airbnb needed. The agent runs across 13 different models. That is not a cost-cutting move. It is a deliberate multi-model architecture built around control, not just capability. Microsoft’s evaluation of DeepSeek V4 for Copilot Cowork , reported by Axios in June 2026, tells the same story. The company is exploring a self-hosted DeepSeek to replace the Anthropic and OpenAI models powering its enterprise agentic product, driven by unsustainable costs at scale. Charles Lamanna, Microsoft’s executive vice president for Copilot, agents and platform, told Axios: “We have users who do hundreds of tasks a week… the consequence is the costs can go very high.” The IBM study’s full findings add context: organizations pay 2.8 times more in token processing when AI runs far from the data it depends on. When Airbnb and Microsoft are making these choices in production, the market signal is unambiguous. Open weight is not a fallback. It is the architecture direction serious enterprises are moving toward. Some of these models are Chinese in origin, and yes, that has drawn attention from U.S. lawmakers. Those are real conversations worth having. But here is the practical point: a model running inside your own infrastructure, under your own security controls, gives you more governance than a closed model running on a server you do not own, regardless of where either was built. Three forces accelerating the shift Three things are pushing enterprises in this direction, and none of them are going away. Cost at scale. Every API call compounds. Open-weight models on your own infrastructure convert that variable cost into one your organization controls. The IBM study found misaligned AI infrastructure costs enterprises 2.8 times more in token processing. Microsoft’s DeepSeek evaluation is that logic at the largest scale in the industry. Data protection. Consider what this means for a hospital routing patient records through a third-party AI API, a defense contractor using a closed model to analyze procurement data, or bank feeding client financials into an external inference endpoint. In each case, compliance is contingent on a vendor relationship the organization does not fully control. HIPAA does not care whether your AI vendor had a good SLA. FedRAMP authorization does not transfer because a model performed well on a benchmark. Regulators expect organizations to demonstrate control over where sensitive data goes, and a closed frontier API is not a defensible answer. The IBM study found 68% of executives say meeting data residency and sovereignty requirements across geographies is already challenging. Open-weight models deployed within your own environment remove that exposure entirely. The data does not leave your perimeter, full stop. Sovereignty. The Fable 5 episode made clear that your AI capability is only as sovereign as the provider you depend on. For defense primes, intelligence contractors and any organization operating under jurisdiction-specific regulations, this is an existential design question, not a preference. Open-weight models, deployable under your own governance in any environment, are the only architecture that resolves it. What this means for enterprise AI architecture In conversations with technology leaders, the same pattern keeps surfacing: organizations that started with a single frontier provider for speed are now the most constrained when they try to scale, govern or adapt. This is especially acute in regulated sectors. A healthcare organization that built clinical documentation on a closed frontier model faces a hard question every time that vendor changes its data processing terms. A defense contractor with a closed model embedded in its logistics pipeline must revisit its authorization to operate every time the model updates silently. The implication is not to abandon frontier models entirely. It is to stop building AI programs that depend on them as the sole or default layer. The practical answer is a multi-model architecture: frontier models where the capability genuinely justifies the cost and the data exposure, open-weight models running on your own infrastructure for everything else. Not every task needs the most powerful model available. And not every task should leave your perimeter. DCosta framed the coming divide as between AI owners and AI renters. I would take that one step further. The organizations that will genuinely own their AI are the ones building the infrastructure, governance and internal capability to run open-weight models on their own terms, right now. The ones that continue to depend entirely on closed frontier providers are not owners, whatever they call themselves. They are renters, and their leases can be terminated, repriced or restricted at any time. The IBM study makes the stakes clear: 57% of executives say replacing a core AI model would require significant decoupling or a full rebuild. The longer you wait to build portability in, the harder it gets. The best AI model is not the one with the highest benchmark score. It is the one that fits into an architecture your organization governs. This article is published as part of the Foundry Expert Contributor Network. Want to join?
Score: 28🌐 MovesJul 30, 2026https://www.cio.com/article/4203027/with-ai-control-matters-more-than-capability.html - Why your AI strategy has a trust problem and speed won't fix it
Faster AI alone won't build customer trust, sustainable adoption, or long-term business value.
Score: 28🌐 MovesJul 30, 2026https://www.techradar.com/pro/why-your-ai-strategy-has-a-trust-problem-and-speed-wont-fix-it - Why AI Responses Get Worse in Long Chats (And How AI Engineers Avoid It)
Your AI isn’t getting worse. It’s running into an engineering trade-off almost nobody thinks about. Everything is going perfectly. The AI understands your writing style, remembers your project, and follows every instruction. Even the ones you mentioned only once. Then, somewhere around the twentieth or thirtieth message, something changes. It forgets requirements you repeated three times. It starts contradicting itself. It answers the question you asked twenty minutes ago instead of the one you just typed. It feels dumber. Most people assume the model is getting tired. It isn’t. It’s running into one of the biggest engineering trade-offs in modern AI. The symptom everyone recognizes Ask around and you’ll hear the same complaints in almost the same words. The AI ignores instructions it followed perfectly an hour ago. It forgets details you already gave it — twice. It gets repetitive, looping back to ideas you’ve moved past. It starts hedging, softening, generalizing, as if it’s no longer sure what you actually want. None of this is your imagination, and none of it is a sign you’re prompting badly. It’s a sign you’ve been talking long enough for something structural to kick in. The hidden cause: everything is competing for attention Every AI conversation lives inside something called a context window. T he full stretch of text the model can “see” at once when it generates a reply. Not just your last message. Everything: your first message, your fifth, your twentieth, every instruction, every correction, every tangent. Every time you send a new message, the model doesn’t just read your latest prompt. It may also need to process everything the system decides is relevant from the conversation so far. A short conversation — nothing competing for the model’s attention but the one thing you just asked. A long conversation — every earlier message pulls at the model’s attention before it can answer the one you just sent. At message three, that’s trivial. By message two hundred, the system may be working with context equivalent to a novella whether that’s the original conversation, a compressed summary, or a retrieved subset depends on the product and it has to decide, silently, which parts of that material actually matter right now. Imagine asking someone to answer a single question after rereading a 70-page document every time you speak. That’s surprisingly close to what long AI conversations become. Every new message isn’t just another sentence. It becomes part of the material the model has to process before producing the next answer. This is why a brand-new conversation can sometimes outperform a 300-message thread, even though the older thread contains far more information. Imagine a notebook where someone keeps adding a new page every minute. The notebook becomes more valuable over time — but finding the right page becomes harder. Long AI conversations work much the same way. That decision is where things go wrong. Why performance actually drops Here it’s worth separating two different problems, because they get solved in completely different ways and conflating them is how most explainers get this wrong. Problem one lives inside the model. Transformer-based models don’t weight every part of a long context equally. Research on long-context performance has repeatedly found that models are most reliable with information at the very start or the very end of the context, and measurably worse at retrieving or reasoning over information buried in the middle — a pattern researchers call “lost in the middle.” Your requirement from message four isn’t gone. It’s just sitting in the part of the context the model is statistically worst at attending to, sandwiched between forty other things also asking for attention. This is a property of how the model was trained and architected. No amount of clever prompting fully fixes it, because it’s not a prompting problem. “Lost in the middle” — the same information is retrieved less reliably depending only on where it sits in the context. Problem two lives inside the product. Every AI chat interface — ChatGPT, Claude, whatever you’re using has to decide what to do when a conversation grows past what the model can hold at all. Some products truncate, quietly dropping the oldest messages. Some summarize old turns down to a compressed version. Some retrieve only the pieces that seem relevant to your current question. These are engineering decisions made by the team that built the product, and they vary between tools. This is why the same conversation can degrade differently in different apps . You’re not just fighting the model, you’re fighting whatever memory strategy that specific product chose. Put together: the model has a built-in bias against the middle of a long conversation, and the product has to make lossy decisions about what even survives to reach the model. Every additional message increases the odds that the thing you need right now is exactly the thing that’s been deprioritized twice over. You can think of the model’s attention as a limited budget. Every previous instruction, code block, pasted article, and side conversation competes for a share of that budget. The more unrelated information you keep in the conversation, the less focus the model can devote to the task you care about right now. The same abundance that makes AI valuable is what makes attention scarce. The more context you give the model, the more competition you create inside that context. How AI engineers actually avoid it People who build with these models professionally rarely say “just clear your chat.” They design around the problem instead. A few patterns show up again and again: Project isolation. Engineers isolate conversations because context is expensive. Every unrelated exchange competes with the task you’re trying to solve now. Context checkpointing. Once a task reaches a natural checkpoint. A draft is approved, a bug is fixed, a decision is made — continuing in the same thread means dragging all of that resolved history forward, uselessly, forever. A fresh start costs nothing and removes noise. Manual compression. Instead of relying on the model to correctly re-weigh forty messages of history, engineers write a short, explicit summary of the current state and paste it into a new conversation. This isn’t cheating. It’s manually doing the compression the model would otherwise have to guess at, and guessing is exactly where errors creep in. Modular prompting. Rather than one sprawling instruction with a dozen embedded requirements, break the task into smaller, self-contained prompts. Smaller context, less competition, more reliable attention on what matters. External memory. Serious AI applications increasingly store important facts outside the conversation entirely , in a database, a file, a structured note and re-inject only the relevant pieces when needed. The model never has to remember everything. It just has to be handed the right five things at the right moment. In practice, that looks like this: None of these are tricks. They’re all the same insight applied differently: attention is finite, so don’t make the model spend it on things it doesn’t need. Where this is headed This is why long-term memory, persistent memory, context compression, and better retrieval systems have become some of the most active research areas in AI right now. Every major lab is, in some form, trying to solve the same question: how do you give a model the benefit of everything a user has ever said, without forcing it to pay the attention tax of holding all of it at once? The honest answer is that nobody has fully solved it yet. What exists today are workarounds — smart, effective, engineered workarounds, but workarounds. The model that never degrades over a long conversation, that remembers what matters and forgets what doesn’t without being told, doesn’t exist yet. We often say AI has an intelligence problem. Increasingly, it doesn’t. It has an attention problem. And as models become smarter, attention, not intelligence, may become the resource engineers spend the most time optimizing. That’s why long conversations break down today, and why the companies building tomorrow’s AI are investing so heavily in memory, retrieval, and context management. Understanding that doesn’t just explain why your chats get worse. It changes how you’ll use AI from now on. Because the future of AI isn’t just about building more intelligent models. It’s about helping those models pay attention to the right things at the right time. Why AI Responses Get Worse in Long Chats (And How AI Engineers Avoid It) was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- The Python Ecosystem That Changed AI Development
How one open-source ecosystem made state-of-the-art AI accessible The post The Python Ecosystem That Changed AI Development appeared first on Towards Data Science .
Score: 28🌐 MovesJul 30, 2026https://towardsdatascience.com/the-python-ecosystem-that-changed-ai-development/ - Schools Deploying Pepper Spray Drones to Sic on Mass Shooters
Anything but actual gun control. The post Schools Deploying Pepper Spray Drones to Sic on Mass Shooters appeared first on Futurism .
Score: 28🌐 MovesJul 30, 2026https://futurism.com/robots-and-machines/schools-deploying-drones-mass-shooters - 'I use AI, but I don't want AI making my choices': Tom's Guide readers sound off on Google AI search
'I use AI, but I don't want AI making my choices': Tom's Guide readers sound off on Google AI search Tom's Guide
- The real bottleneck in India's agentic AI race isn't the AI
The real bottleneck in India's agentic AI race isn't the AI Techcircle
Score: 27🌐 MovesJul 30, 2026https://www.techcircle.in/2026/07/30/the-real-bottleneck-in-india-s-agentic-ai-race-isn-t-the-ai - Will AI take your job? Cashiers to doctors, which professions face biggest risk
Artificial intelligence will automate routine jobs across many industries. Complex tasks requiring human judgment will be supported by generative AI. Financial services and retail trade see high automation risk for basic operations. Healthcare and education will see AI strengthen clinical and learning roles. Workplace tasks will be reshaped as AI complements human capabilities.
- How to Build a Context Layer and a Company Brain
What it actually takes to turn a company's scattered knowledge into something an LLM can reliably use — and why the demo is 5% of the work. The post How to Build a Context Layer and a Company Brain appeared first on Towards Data Science .
Score: 26🌐 MovesJul 30, 2026https://towardsdatascience.com/how-to-build-a-context-layer-and-a-company-brain/ - Consultants have found an unlikely new career path: policing AI
Consultants have found an unlikely new career path: policing AI Business Insider
Score: 26🌐 MovesJul 30, 2026https://www.businessinsider.com/consulting-firms-consultants-career-paths-ai-risk-safety-bcg-mckinsey-2026-7 - AI-Driven Network Intelligence: The Visibility Edge India’s Enterprises Need Now
By Chirag Raichura India’s digital transformation is entering a new phase, which is increasingly shaped by the rapid adoption of artificial intelligence (AI). Across industries, AI is moving from experimentation to enterprise-wide deployment, accelerating innovation, automating operations, and enabling smarter decision-making. The expansion of Global Capability Centers (GCCs), digitization across banking and financial services, nationwide […] The post AI-Driven Network Intelligence: The Visibility Edge India’s Enterprises Need Now appeared first on CXOToday.com .
- The Download: tricking LLMs, and reviving geothermal plants
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. A fundamental flaw leaves LLMs strikingly vulnerable to attack It is impossible to make large language models fully secure against hacks because of a fundamental flaw in how they work, a…
Score: 26🌐 MovesJul 30, 2026https://www.technologyreview.com/2026/07/30/1140936/the-download-tricking-llms-reviving-geothermal/ - The Convergence Revolution — When AI Meets Every Emerging Technology
Artificial intelligence's true revolutionary impact stems from its convergence with other advanced technologies
- Florida plans to build air taxi pads using $200M intended for EV chargers
Florida wants to use federal EV charger funds to build an air taxi network connecting golf courses, luxury apartment buildings, and airports.
Score: 26🌐 MovesJul 30, 2026https://techcrunch.com/2026/07/30/florida-plans-to-build-air-taxi-pads-using-200m-intended-for-ev-chargers/ - AI Scammers Are Better at Building Trust Than Humans
Researchers pitted a person against a Claude agent and found that, after a week of texting, the AI chatbot was more effective at creating “exploitable trust” with others.
Score: 26🌐 MovesJul 30, 2026https://www.wired.com/story/ai-scammers-are-better-at-building-trust-than-humans/ - Context, not correlation, will define successful AI implementation
Discover why context, trusted data and governance matter more than AI models alone.
Score: 25🌐 MovesJul 30, 2026https://www.techradar.com/pro/context-not-correlation-will-define-successful-ai-implementation - Why AI systems need to be robust in safety-critical environments
UL’s Prof Martin Hayes on his systems theory research, why ‘pi-shaped’ graduates are the future of engineering, and the importance of patience. Read more: Why AI systems need to be robust in safety-critical environments
Score: 25🌐 MovesJul 30, 2026https://www.siliconrepublic.com/innovation/why-ai-systems-need-to-be-robust-in-safety-critical-environments - What Can You Actually Do With Salesforce Foundations (For Free)?
Foundations gives existing Salesforce customers access to AI tools — here's exactly what that means for your business.
Score: 25🌐 MovesJul 30, 2026https://www.salesforce.com/blog/small-business/what-does-foundations-do/ - The pitch you never gave: What AI tells buyers about your startup
Before I take an intro call with a company, I ask a few AI assistants about it first. What does this company do, and would you recommend it? Five minutes, and the meeting changes, because the answers are wrong more often than founders would like to believe. Old positioning survives, pivots go missing, and a […] The post The pitch you never gave: What AI tells buyers about your startup appeared first on e27 .
Score: 25🌐 MovesJul 30, 2026https://e27.co/the-pitch-you-never-gave-what-ai-tells-buyers-about-your-startup-20260727/ - Equinox Thought It Was Critiquing AI. Critics Say It Ended Up Reinforcing a Harmful Stereotype
The fitness giant pulled an ad that sparked backlash for reinforcing stereotypes about Asian women. Here’s what it’s doing now.
- This week in 5 numbers: Why some workers aren’t AI upskilling
Here’s a roundup of numbers from the last week — including how many C-suite, CHRO and senior talent acquisition leaders think their leaders are “highly prepared” to lead AI adoption.
Score: 25🌐 MovesJul 30, 2026https://www.hrdive.com/news/why-some-workers-arent-ai-upskilling/826634/ - Rethinking how AI supports investment decisions
Artificial intelligence (AI) is rapidly transforming modern finance, powering applications ranging from stock market forecasting to investment advice. But does making more accurate predictions necessarily lead to better investment decisions? According to two recent studies by researchers from Pusan National University and international collaborators, the answer may be no.
Score: 25🌐 MovesJul 30, 2026https://techxplore.com/news/2026-07-rethinking-ai-investment-decisions.html - How to Organize All of Your Coding Agent Tasks
Optimise how you interact with your coding agents The post How to Organize All of Your Coding Agent Tasks appeared first on Towards Data Science .
Score: 25🌐 MovesJul 30, 2026https://towardsdatascience.com/how-to-organize-all-of-your-coding-agent-tasks/ - Go Pure: Locking In Quality Using Native Claude Code Features
Fewer Tokens, Zero Thrash, Bulletproof Code Quality Heya! My name is Nick and I love writing Web apps! With the advent of AI tooling in late 2025 hinting at greater things to come. I wanted to realize the benefits AI could offer but deep down I wanted to begin to understand it. In 2026 it feels like there is a rapidly evolving tech space around context engineering. When Anthropic releases more powerful models they usually release new tooling that can improve the software development workflow experience. Looking back on the articles I have written here on Medium I see where I needed to work around the limitations that existed at the time. Working with a 200kb context level is not a constraint I operate on daily now and some of the tooling I built around that limit no longer needs to exist. The feature capabilities that Anthropic built into the Memory system are fantastic! I have moved virtually all of the prose based coding standards that lived in markdown and converted them to lint and file type based rules. The goal was to utilize AI only where it’s needed but default to deterministic tools and improve cost efficiency. By using less AI for code quality the whole development cycle has less friction not to mention greater speed with faster lint tools. Image created with Gemini by author The Scaffolding Era The sole motivation and goal while reworking my Claude Code infrastructure was as follows. Every rule lives at the lowest layer that can catch it deterministically. Prose is the layer of last resort. Everything I had built at the time was a rational patch for a real 2025 weakness. We were coding with lossy compaction, no persistent memory, no path-scoped rule loading and prose as the only enforcement mechanism. This is what the extra layers looked like before the refactor. — 17 Session Files: markdown dumps for stringing multi-day sessions together — Coding Standards: every rule, pattern and testing recipe, in prose, that every new session read. This fired as a hook and would thrash other Claude sessions. — Post-compaction Hook: re-inject “context essentials” file every time compact fired. Because compaction used to eat the rules. — A SPARC config, pattern templates, a VIOLATIONS.md ledger and a 400 line hand maintained skills index. Every piece made sense at the time but every piece also paid a tax on every session. Prose are anything Claude has to read and remember A lint error is not probabilistic. It fires the same way every time, for every bot or human at zero context cost. Bucket 1: Prose -> lint errors We took all of the mechanical coding-standards direction out of markdown and made it ESLint errors that we possibly could. For most things this is a straightforward swap but for more complex issues you will have to get more creative. That’s no problem though because Claude Code provides file/folder aware rules that only apply while editing those files. If you want your AI to not get lazy and skip important steps you need this system. We found history in the session files of many production bugs in a timezone sensitive app. We turned these into rules and it became a custom plugin (shown below). ```js // eslint.config.mjs { files: ['src/**/*.{ts,tsx}'], plugins: { 'clear-mornings': customRulesPlugin }, rules: { 'clear-mornings/prefer-datecore-over-new-date': 'error', 'clear-mornings/require-timezone-in-date-operations': 'error', 'clear-mornings/no-manual-date-concatenation': 'error', }, } ``` Now when Claude writes any new Date() in a component the result isn’t a “please remember not to do this”. It’s a red squiggle in the IDE and a blocked commit. Same treatment for useEffect in client side components. We also prevented Claude from working around the system with ignore statements. The removed session files were a great source of defects we encounter repeatedly and their remediation. These were converted straight into lint rules where possible. If you have been recording these in the memory system instead those can be a good pool to pull from. Simply ask Claude “hey what are the commonly encountered bugs in this system and can they be converted to a deterministic system like lint”. Claude can pull from memory or if you have git CLI your repo history holds the sordid details. Bucket 2: What ESLint Can’t See Cover With Diff Aware Hook We kept exactly one edit time hook, its header shows you the lesson that shaped it. ```bash # History: this hook used to whole-file-grep every edit for ~15 patterns # and "block" with exit 1. Two problems: # 1. exit 1 is NON-BLOCKING for PostToolUse — it never actually blocked. # Proof: 48 data-testid, 5 @ts-*, and other "blocked" patterns # accumulated anyway. # 2. Whole-file grep fired on legacy code elsewhere in a file you barely # touched → false blocks + token thrash across parallel sessions. ``` Two hard learned specifics there. First, hook exit codes matter. For postToolUse exit 1 is advisory where as exit 2 is blocking. We had months of imaginary enforcement before noticing violations piling up. Second, be diff aware. The hook now inspects only the content the current edit introduced ('new_string' / 'content') — never the file on disk. So legacy code and parallel sessions can’t trip it (full hook below). #!/bin/bash # Edit-time standards hook — NARROW + DIFF-AWARE # # Covers ONLY what your linter structurally cannot: # - If ESLint ignores test files, suppressions and testid queries in tests # are invisible to lint — this hook is their only gate. # - Rules with legacy violations (where a lint `error` would block CI on # old code) get a DIFF-AWARE check that blocks only NEW additions. # # DIFF-AWARE: inspect ONLY the content this edit introduced (Edit.new_string / # Write.content), never the file on disk — legacy code and parallel-session # edits elsewhere in the file can't trip it. # # EXIT CODES (PostToolUse contract): # 2 = block + feed message back to Claude (real enforcement) # 0 = allow (optional advisory on stderr) # NOTE: exit 1 is NON-BLOCKING for PostToolUse. Enforcement via exit 1 # is imaginary. set -o pipefail INPUT=$(cat) TOOL_NAME=$(echo "$INPUT" | jq -r '.tool_name // empty') FILE_PATH=$(echo "$INPUT" | jq -r '.tool_input.file_path // empty') # Only Write/Edit with a file path. if [[ "$TOOL_NAME" != "Write" && "$TOOL_NAME" != "Edit" ]]; then exit 0; fi if [[ -z "$FILE_PATH" ]]; then exit 0; fi # Only code files. case "$FILE_PATH" in *.ts|*.tsx|*.js|*.jsx) ;; *) exit 0 ;; esac # ---- Diff-aware content: ONLY what this edit introduced ---- NEW_CONTENT=$(echo "$INPUT" | jq -r '.tool_input.new_string // .tool_input.content // empty') if [[ -z "$NEW_CONTENT" ]]; then exit 0; fi is_test_file=false case "$FILE_PATH" in *test*|*spec*|*__tests__*) is_test_file=true ;; esac BLOCKS=() WARNINGS=() # BLOCK: data-testid and *ByTestId queries — NEW additions only. # Accessible queries (getByRole/getByText/getByLabelText) required instead. if printf '%s' "$NEW_CONTENT" | grep -qE 'data-testid|(get|query|find)(All)?ByTestId'; then BLOCKS+=("data-testid / *ByTestId is banned — use semantic HTML + accessible queries (getByRole, getByText, getByLabelText).") fi # BLOCK: type suppressions inside TEST files (non-test src is covered by # ESLint's ban-ts-comment; tests are lint-ignored, so this is their gate). if $is_test_file && printf '%s' "$NEW_CONTENT" | grep -qE '@ts-ignore|@ts-expect-error|@ts-nocheck'; then BLOCKS+=("@ts-ignore / @ts-expect-error / @ts-nocheck — fix the underlying type, never suppress it.") fi # WARN: newly-skipped tests. Advisory — fix or remove, don't leave disabled. if $is_test_file && printf '%s' "$NEW_CONTENT" | grep -qE '\b(describe|it|test)\.skip\b|\bx(describe|it|test)\b'; then WARNINGS+=("Skipped test added — fix or remove it rather than leaving it disabled.") fi if [[ ${#BLOCKS[@]} -gt 0 ]]; then { echo "" echo "STANDARDS VIOLATION (new code): $FILE_PATH" echo "========================================" for b in "${BLOCKS[@]}"; do echo " [BLOCK] $b"; done for w in "${WARNINGS[@]}"; do echo " [WARN] $w"; done echo "" } >&2 exit 2 fi if [[ ${#WARNINGS[@]} -gt 0 ]]; then { echo "" echo "Standards check: $FILE_PATH" echo "----------------------------------------" for w in "${WARNINGS[@]}"; do echo " - $w"; done echo "" } >&2 fi exit 0 Image created with Gemini by author Bucket 3: The Commit Ladder Once the code is complete and it’s time to push there is a more strict set of code quality checks. These all run on pre-commit. All of these are commit checks and are nice to have done for you when wrapping up a long-running session. lint-staged — prettier + eslint --fix blocks on unfixable errors type-check — project wide tsc --noEmit (per-file tsc breaks imports) affected tests — jest --onlyChanged no-dead-tRPC — no registered procedure ships without a live caller. AI tooling wants to create everything from scratch instead of re-using what was already built. This prevents dead-end tRPCs. translation auto-fill — we hand edit en.json only, a cheap model fills the other locales and stages them into the same commit. Bucket 4: What’s Left For Prose And Where It Loads Our final CLAUDE.md shrank to 81 lines and holds what prose is actually for: judgment calls no linter can express. Of those things that remained in CLAUDE.md we questioned does this apply on every turn or only when specific files are open? CLAUDE.md contains topics that apply in every session. Ours contains constraints like: — Load the domain skill before grepping — database writes and commit’s need human approval — no use of memoization since we use React compiler — a handful of architectural invariants and pointers For constraints that apply only to specific files (or directories) we can use rules. A rules file is plain markdown with one frontmatter field. ```yaml --- paths: - 'src/server/trpc/**/*.ts' --- ``` We applied rules covering the following topics: — SSR Authentication — async server component stacking — tRPC security and procedure lifecycle — i18n policy and it means I can push far more code more quickly Back To Basics The old system was built to protect the output quality and did that but also ended up causing real friction when I had multiple Claude sessions running. When any of the secondary sessions made a change on disk that caused a hook to run and check all of the uncommitted code, Claude would often git stash code from another session then have to stitch it all back together after realizing its mistake. The new system offers the flexibility to weave in other sessions while a long-running task finishes in the main thread. I can finally do this without stomping on any of the other work and it means I can push far more code more quickly. Start With A Plan All coding and research sessions start with a plan. For one thing, it makes sense to start Claude off in a read-only state. When you start in/plan Claude will never write to files until after gaining approval. One of the biggest changes to our system was to remove session files. The replacement is Claude plans. These documents had the same purpose: to document decisions and the current state of a feature. By writing a plan document into the native directory ./claude/plans (or as you see here the docs directory) Claude will know to reference this doc on a new session if it comes up in conversation. Take a look at the start of this session after exiting plan mode. Claude is poised and ready to take off! Using plan mode is a great way to have Claude researching and getting prepped up for a project while another Claude session(s) is running. Just let Claude sit after the plan is written, then even if the power goes out you still have that plan file, the result of all the research and planning. Once the other session(s) completes commit/push and let your new plan rip! Dropping A Session Now after Claude gets to around 55% context and is at a stopping point I can straight up drop the session, start a new one and simply say “pick up phase 3 of …” and Claude knows exactly what to do! Even with the 1m context I often still run over the 75% mark where Claude needs a renewal or complete abandonment in favor of a new session. The now native built in memory and plans features allow for a consistent and reliable source of truth when it comes to state of a project, the decisions made while working on it and the original vision it started from. Compaction Uses Tokens I researched compaction and found that the process actually uses tokens. It’s packaging up the accumulated context and sending it off to the server for summarization. Instead I have been using the /clear command near the end of my usable context. Because we start with a plan and can easily pickup from the plan. Holding around a bunch of superfluous context isn’t advantageous and costs more tokens over the life of your task. Summary My workflow in 2026 has gone entirely native. I always start with a plan using the /plan command or the Shift + Tab mode toggle. That will write a plan to session and you can extend this to write to disk. Ensuring code quality with deterministic tools is faster, cheaper and doesn’t make a mistake. Speeding up the development cycle and resulting in less thrash while running multiple sessions. How Do You Handle A Multi-Session Workflow? Thanks for reading and please indulge me! How do you facilitate multi-session workflows? I mean both multiple Claude sessions to work on a single task as well as multiple simultaneous sessions? How have you kept your code clean using hooks or other native features? Please tell me in the comments! Resources: https://code.claude.com/docs/en/common-workflows https://code.claude.com/docs/en/memory https://code.claude.com/docs/en/agent-sdk/hooks Go Pure: Locking In Quality Using Native Claude Code Features was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- University of Phoenix survey highlights AI’s potential to advance accessibility in work and learning
University of Phoenix survey highlights AI’s potential to advance accessibility in work and learning EurekAlert!
- Startup Spotlight: This Seattle-area plumber wants to unclog retail with AI-powered delivery
A Seattle-area plumber and Navy veteran is taking on a new challenge: making local retail as convenient as ordering takeout. Meet Jeremy Bernier, the founder of MODRA, an AI-powered startup aiming to bring same-day delivery and personalized shopping to neighborhood stores. Read More
- Can I take a robot car anywhere in Dallas? Three things to know about AVs
Can I take a robot car anywhere in Dallas? Three things to know about AVs Dallas News
Score: 24🌐 MovesJul 30, 2026https://www.dallasnews.com/news/transportation/article/autonomous-vehicles-dallas-work-22365356.php - Vision Language Grounding: How AI Connects “Dog” to Pixels, and Where It Falls Apart
A research grounded deep dive 1. The Hook Picture an ordinary photo like this one. A vision language model will confidently caption it “a dog in a park,” and it will be right, almost every time [1]. Source: Image by Sanjana Dubey The trouble starts at the edges of “almost.” In 2021, OpenAI researchers showed that printing a paper label reading “iPod” onto a Granny Smith apple caused their CLIP model to classify the apple as an iPod [2]. A related version used dollar signs drawn on a poodle, which shifted the model’s answer to “piggy bank” [2]. Neither attack required any optimization or special tooling, just pen and paper, because it exploits how CLIP encodes visual text as part of an image’s meaning [2]. That single demonstration is a good way into this whole topic. It shows that grounding is not conceptual understanding. It is proximity in a learned space, and that space can be nudged by anything that activates the right region, including text glued onto an object. Source: Image by Sanjana Dubey 2. What Grounding Actually Means Grounding, in the cognitive science sense, is how a word comes to refer to something real. When a child learns the word “dog,” they aren’t just memorizing a sound through repeated exposure, touch, and correction, they build a connection between that sound and the actual category of thing it refers to in the world. For a vision language model, grounding means something much narrower. The model never touches a dog or hears one bark. All it has is a word and a set of pixels, so grounding here just means: does the model’s internal representation of the word “dog” reliably line up with its internal representation of dog-shaped, dog-textured visual patterns? The mechanism behind that alignment is statistical, not symbolic, and this distinction explains nearly every failure covered later in this piece. A symbolic system would ground “dog” against something like a stored checklist: four legs, fur, tail, barks, and check an image against that checklist feature by feature. A statistical system, which is what these models actually use, has no checklist at all. There is only a mathematical space where the word “dog” and photos of dogs were pushed close together during training simply because they appeared together often enough. Nothing is verified or checked; it’s proximity, learned purely from co-occurrence patterns in the training data. That difference matters because a checklist generalizes gracefully: a dog in an odd pose is still a dog if you’re checking against a definition. Proximity in a learned space does not generalize as gracefully. Anything that shifts a word or image away from its usual neighborhood, an unusual pose, a swapped attribute, a negated sentence, an odd background they can break the association, even when a human would recognize the object instantly. Source: Image by Sanjana Dubey 3. The Core Mechanism CLIP is built as a dual encoder. That means there are two separate neural networks working side by side. One is a Vision Transformer, which reads raw pixels. The other is a Transformer, which reads the caption text. Each of these two networks processes its own input completely on its own, without ever looking at what the other one is doing. What makes them usable together is that both networks are designed to output a vector, a plain list of numbers, and both vectors are the exact same length. Because they are the same length, they can be directly compared to each other, even though one came from a picture and the other came from a sentence. This shared, comparable output space is what “project into the same latent space” means [1]. To train this system, researchers used 400 million image and caption pairs collected from the internet [1]. For every single pair, the training process nudges the image’s vector and its correct caption’s vector closer together in that shared space. At the same time, it pushes that same image’s vector away from every other caption that does not belong to it. This is done in a very specific way. For every batch of examples during training, the model builds a large grid of similarity scores, comparing every image in the batch against every caption in the batch. It then applies a loss function called cross entropy, which penalizes the model whenever a wrong pairing scores higher than the correct pairing. This happens in both directions at once, meaning the model is checked on matching images to captions, and separately on matching captions back to images [1]. Put simply, the training task boils down to this. Given one image, and a large pile of caption candidates where only one is correct, the model has to correctly pick out the one true caption [1]. It gets rewarded when it guesses right and corrected when it guesses wrong, and it does this across hundreds of millions of examples. That repeated correction, done at massive scale, is the entire reason the model ends up placing the word “dog” near photos of actual dogs. Nobody ever gave it a definition of what a dog is. Here is a diagram of that full pipeline, showing both encoders and how their outputs land in the same shared space. Source: Image by Sanjana Dubey The loss function that drives this training is called InfoNCE. Below is that same mechanism written out as runnable code, so you can see there is no hidden step: import torch import torch.nn.functional as F def contrastive_loss(image_embeds, text_embeds, temperature=0.07): """ image_embeds, text_embeds: (batch_size, dim), L2-normalized. Row i of each tensor is assumed to be a true match. """ logits = image_embeds @ text_embeds.T / temperature labels = torch.arange(logits.shape[0], device=logits.device) loss_i2t = F.cross_entropy(logits, labels) loss_t2i = F.cross_entropy(logits.T, labels) return (loss_i2t + loss_t2i) / 2 Reading through this code line by line, image_embeds @ text_embeds.T is the similarity grid described above. Dividing by a temperature value simply sharpens or softens how confident the scores are. The two cross_entropy calls are the two directions of checking mentioned earlier, image to text and text to image. Averaging them at the end gives one final loss number the model can learn from. Finally, here is what grounding looks like once a model has already finished training, using Hugging Face’s transformers library: from transformers import CLIPModel, CLIPProcessor from PIL import Image import torch model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32") processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32") image = Image.open("dog_photo.jpg") captions = ["a photo of a dog", "a photo of a cat", "a photo of a car"] inputs = processor(text=captions, images=image, return_tensors="pt", padding=True) with torch.no_grad(): probs = model(**inputs).logits_per_image.softmax(dim=1) for c, p in zip(captions, probs[0]): print(f"{c}: {p.item():.3f}") This snippet does not train anything. It loads a model that has already learned from those 400 million examples, feeds it one real photo along with three candidate captions, and asks it to score how well each caption matches. If everything worked as described above, the caption “a photo of a dog” should come back with the highest score whenever the photo actually shows a dog. Running this yourself turns the whole explanation from something you read into something you can watch happen. 4. From Global Vectors to Fine Grained Grounding A single pooled vector can tell you whether an image and a caption match overall, but it cannot tell you where inside the image the match actually is. If a photo shows a dog in one corner and a person in the middle, one vector for the whole image has no way to point specifically at the dog. Three approaches address this. RegionCLIP compares small pieces of an image to text, instead of only comparing the whole image to a whole caption. It first generates labels for individual regions of an image, then trains the model to match those regions against pieces of text, before this is transferred into a full detection setup [9]. OWL-ViT starts from an already trained CLIP style model. It removes the step that pools the whole image into one single vector, and instead attaches small heads directly onto the individual output tokens from the image encoder, where each token corresponds to a small patch of the image. At the time of use, a text query is compared against those tokens to find the matching patch, which is what lets it detect objects described by any free text phrase, not just a fixed list of categories [8]. Grounding DINO fuses the two modalities earlier in the process. Image tokens attend to each other, and at the same time attend to the text tokens, so both sides share a grounded understanding of the scene before the model proposes any specific region as its answer [10]. All three approaches rely on the same underlying comparison, just applied to small image patches instead of a whole image. A word is compared against every patch, each patch gets a weight showing how strongly it relates to that word, and the patch with the highest weight is where the model is, functionally, pointing: import torch.nn.functional as F def patch_attention(word_embed, patch_embeds): """word_embed: (dim,) patch_embeds: (num_patches, dim)""" scores = patch_embeds @ word_embed return F.softmax(scores, dim=0) # weight per patch, sums to 1 patch_embeds @ word_embed produces one similarity score per patch. Passing those scores through softmax turns them into weights that add up to one, so the patch with the highest weight is the model's best guess at where the word is grounded in the image. 5. Failure Showcase, Grounded in Published Research 5.1 Attribute binding failures Attribute binding means correctly matching a property, like a color, to the specific object it belongs to. This is where vision language models frequently break down. Research has found that CLIP often behaves like what researchers call a bag of words model. This means the model notices which words and concepts are present in a caption, but does not reliably track which word belongs to which object in the scene. For example, given an image showing an orange square next to a blue triangle, CLIP will often score that image just as highly against the caption describing a blue square and an orange triangle, even though the colors have been swapped between the shapes [3]. This failure was studied carefully through something called the ARO benchmark, short for Attribution, Relation, and Order. It was built specifically to test whether vision language models actually understand attributes, relationships between objects, and the order words appear in a sentence. The benchmark included more than 50,000 test cases, and the results showed that even state of the art models regularly make mistakes when trying to correctly link an object to its attribute [3]. One explanation researchers have proposed for why this happens comes down to how these models are trained. During training, a model is shown a batch of images and captions, and it is taught to prefer the correct pairing over the incorrect ones in that same batch. If the incorrect pairings in a batch are too easy to rule out, meaning they do not closely resemble the correct answer, the model can still succeed at the training task without ever learning to carefully track which attribute belongs to which object. In other words, without enough close, tricky wrong answers to learn from, the model can take a shortcut and still score well [3]. Source: Image by Sanjana Dubey Test it yourself: captions = ["a red cube on top of a blue sphere", "a blue cube on top of a red sphere"] inputs = processor(text=captions, images=image, return_tensors="pt", padding=True) probs = model(**inputs).logits_per_image.softmax(dim=1) print(probs) # similar scores against one image = the binding failure If the two probabilities printed here come out close to each other, that confirms the model cannot reliably tell which caption actually matches the colors in your image. 5.2 Spurious correlation as false grounding A spurious correlation happens when a model learns to rely on something that is not actually the object it is supposed to be recognizing, simply because that other thing tended to appear alongside the object during training. This problem is not unique to dogs, and it is not unique to vision language models either. One clear example involves keyboards. If keyboards in a training dataset are almost always photographed with a monitor sitting nearby, a model can end up learning to detect the presence of a monitor as a stand in for detecting the keyboard itself. This works fine as long as keyboards and monitors keep appearing together, but it fails as soon as the two appear separately, since the model never actually learned what a keyboard looks like on its own [12]. CLIP shows this same pattern. Researchers studied CLIP’s predictions by grouping them according to the background present in each photo. They found strong, consistent associations between certain backgrounds and certain predicted categories. As one example, images with railway tracks in the background strongly pushed CLIP toward predicting the object was a train, even in cases where the actual object had nothing to do with trains. This likely reflects how frequently railway tracks and trains appeared together during training [13]. This issue also extends beyond simple classifiers into more advanced multimodal systems. Research on multimodal large language models found that the vision encoder component on its own already carries these same spurious biases, meaning the problem is not something introduced only when vision and language are combined. The research also found that these misleading cues can directly cause the model to hallucinate objects that are not actually present in the image [14]. 5.3 Counting and spatial relations Counting requires a model to track individual, separate instances of an object rather than just recognizing that the object category is present somewhere in the image. This turns out to be a genuinely difficult task for these models. CountBench was created specifically to measure how well vision language models can count objects, after researchers observed that CLIP style models consistently struggled with this task [6]. A more recent and more tightly controlled benchmark, called VLMCountBench, pushed this testing further using simple geometric shapes instead of complex real world objects, to remove other sources of confusion. This study found substantial failures in counting whenever two or more different shape types appeared together in the same image. These failures showed up even when the total number of objects was small and the scene was visually simple. Performance also got worse when researchers made small changes to color, size, or how the counting question itself was phrased, showing that the models were not applying a stable, reliable counting process [7]. Source: Image by Sanjana Dubey 5.4 Negation blindness Negation means understanding when a sentence is saying something is absent, rather than present. This is a specific and well documented weakness in vision language models. Recent evaluation work found that these models show a strong affirmation bias, meaning they tend to treat a caption like “a dog” and a caption like “no dog” as nearly the same thing when comparing them against an image, largely ignoring the negation [4]. A separate research paper describes the same issue directly, explaining that CLIP struggles to understand negation and often produces very similar embeddings for affirmative and negated descriptions of the same scene. As a concrete example, the caption “no dog” ends up scoring a strong match against images that clearly do contain a dog, which is the opposite of what should happen [4]. This weakness is not limited to English either. A benchmark built to test negation understanding across seven different languages found that standard CLIP performed at or below random chance specifically on languages that do not use the Latin alphabet, showing the problem can be even more severe depending on the language involved [15]. Source: Image by Sanjana Dubey 5.5 Rare poses, occlusion, distribution edges This failure mode does not come from a single dedicated benchmark the way the previous ones do. Instead, it follows logically from how these models are trained in the first place. Since the model’s internal space is shaped entirely by the examples it was trained on, any part of that space corresponding to rarely seen visual situations will be less reliably organized. This includes unusual poses, objects that are heavily blocked from view, or breeds and variations that were underrepresented in the training data. Because this pattern connects back to the same root cause behind the other failures, it acts as a kind of thread running underneath all of them, rather than being its own separate, isolated issue. 5.6 Hallucinated grounding in captioning and VQA models Object hallucination refers to a model describing something in an image that is not actually there. The standard benchmark used to study this is called POPE. The researchers behind it conducted the first systematic study of this problem in large vision language models, and found that these models frequently suffer from serious object hallucination, meaning they describe objects that do not match what is actually shown in the image [5]. The way POPE tests for this is straightforward. It asks the model a simple yes or no question, such as whether a particular object is present in the image. It does this across three different settings. In the random setting, the object asked about is chosen at random. In the popular setting, the object is one that appears frequently across the training data in general. In the adversarial setting, the object asked about is one that closely and commonly co-occurs with objects that actually are present in the image, making it a harder and more deceptive test [5]. Follow up research building on this benchmark found that models are prone to hallucinating whether an object exists at all, and this problem becomes even more pronounced when testing whether the model correctly identifies specific, fine-grained attributes of objects that are genuinely present [5]. Source: Image by Sanjana Dubey 5.7 Adversarial and typographic attacks Covered in the hook, worth restating precisely: this was documented in the paper studying adversarial attacks on multimodal neurons, building on OpenAI’s original typographic attack findings [2]. Defenses exist too: researchers have proposed a learned output transformation, fine-tuning on synthetic typographic attack images combined with weight interpolation, and inserting a learned “defense prefix” token before the class name, as three separate mitigation strategies [16]. Source: Image by Sanjana Dubey 6. Why These Failures Happen Every failure above traces back to one property: the embedding space encodes statistical co-occurrence, not compositional structure. A pooled vector is a summary of features that tended to appear together, with no explicit mechanism for attribute binding, negation, or counting, and no guarantee that rare combinations are represented as reliably as common ones. 7. What Researchers Are Doing About It Region-level contrastive training , as in RegionCLIP and RO-ViT, which crops and resizes regions of positional embeddings during pretraining to better match how they’re used later at region level during detection fine-tuning [11]. Negation-aware training data , like the synthetic negation datasets used to fine-tune CLIP-based models, producing a 10 percent increase in recall on negated queries and a 40 percent boost in accuracy on negated multiple-choice questions [4]. Counting-aware contrastive objectives , following the CountBench line of work, targeting numerical grounding directly rather than treating it as a side effect of scale [6]. 8. Practical Takeaways Don’t trust a model to reliably distinguish “red X on blue Y” from “blue X on red Y” without independent verification. Treat confident negative claims skeptically — affirmation bias is measured and named, not a rare edge case. Assume counting above small numbers, and spatial relations, are unreliable unless the specific model has been benchmarked for it. Don’t treat a fluent caption as proof of accurate perception. 9. Closing Thought Grounding is geometric, not perceptual, and every study cited above shows that geometry failing in specific, reproducible ways. That’s good news for anyone building on these systems: the failure modes are catalogued, benchmarked, and under active study, not mysterious. FAQ What does “grounding” mean for an AI model, in simple terms? It means connecting a word to the right thing in an image. For a person, this connection comes from lived experience. For a model, it comes from proximity in a mathematical space, meaning the model places the word “dog” and photos of dogs close together because they appeared together often enough during training, not because it understands what a dog actually is. Is CLIP the only model that works this way? CLIP is the most well known example and the one this piece focuses on, but the same dual encoder, contrastive learning approach shows up across many later systems, including OWL-ViT and RegionCLIP, both of which build directly on CLIP style pretraining before adding detection capabilities [8][9]. Why does the apple and iPod experiment work? Because the model was trained to associate visible text with meaning, since captions on the internet often describe or label what is shown. Printing “iPod” on an apple activates that same text recognizing part of the model strongly enough to override the actual visual features of the fruit [2]. Can these failures be fixed by just training on more data? Only partially. More data reduces how rare unusual combinations are, which helps with issues like rare poses or uncommon breeds. It does not fix problems like attribute binding or negation, since those come from the model never being forced to learn compositional structure in the first place, not from a lack of examples [3][4]. Which failure mode is the most well studied? Object hallucination is probably the most actively benchmarked, largely because of POPE, which gives a simple, repeatable way to test whether a model reports objects that are not actually present in an image [5]. Do these problems affect every vision language model equally? No. Different architectures handle grounding differently, and some approaches, like region level contrastive training or early fusion in Grounding DINO, specifically target certain weaknesses such as poor localization [9][10]. That said, none of the published fixes fully close every gap discussed in this piece. Is negation a bigger problem in some languages than others? Yes. A benchmark testing negation across seven languages found that standard CLIP performed at or below chance specifically on languages that do not use the Latin alphabet, showing the problem is not evenly distributed across languages [15]. Can I test any of these failures myself without training a model? Yes. Every code snippet in this piece uses an already trained, publicly available CLIP model through the Hugging Face transformers library. You can run the attribute binding test, the counting test, or the negation test yourself with just a few lines of Python and any photo of your choice. Where can I read the actual research behind these claims? Every factual claim in this piece is tied to a numbered reference in the References section at the end, linking back to the original paper or benchmark it came from. References [1] Radford, A. et al. (2021). Learning Transferable Visual Models From Natural Language Supervision (CLIP). OpenAI. https://openai.com/index/clip/ [2] Goh, G. et al. (2021). Multimodal Neurons in Artificial Neural Networks . Distill. [3] Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., & Zou, J. (2023). When and Why Vision-Language Models Behave Like Bags-of-Words, and What to Do About It? ICLR. https://arxiv.org/abs/2210.01936 [4] Alhamoud, K. et al. (2025). Vision-Language Models Do Not Understand Negation . CVPR. [5] Li, Y. et al. (2023). Evaluating Object Hallucination in Large Vision-Language Models (POPE). https://arxiv.org/abs/2305.10355 [6] Paiss, R., Ephrat, A., Tov, O. et al. (2023). Teaching CLIP to Count to Ten . ICCV. [7] Yu, S. et al. (2025). Your Vision-Language Model Can’t Even Count to 20: Exposing the Failures of VLMs in Compositional Counting . https://arxiv.org/abs/2510.04401 [8] Minderer, M. et al. (2022). Simple Open-Vocabulary Object Detection with Vision Transformers (OWL-ViT). ECCV. [9] Zhong, Y. et al. (2022). RegionCLIP: Region-Based Language-Image Pretraining . CVPR. [10] Liu, S. et al. (2023). Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection . [11] Kim, D., Angelova, A., & Kuo, W. Region-Aware Pretraining for Open-Vocabulary Object Detection with Vision Transformers (RO-ViT). https://arxiv.org/abs/2305.07011 [12] Trustworthy Machine Learning — context bias, keyboard/monitor example. https://arxiv.org/pdf/2310.08215 [13] Concept Regions Matter: Benchmarking CLIP with a New Cluster-Importance Approach . https://arxiv.org/pdf/2511.12978 [14] Seeing What’s Not There: Spurious Correlation in Multimodal LLMs . https://arxiv.org/html/2503.08884v1 [15] Moraitaki, C. et al. Disparities in Negation Understanding Across Languages in Vision-Language Models . https://arxiv.org/pdf/2604.18942 [16] Materzyńska, J. et al. — learned output transformation defense against typographic attacks; Azuma, R. & Matsui, Y. — Defense-Prefixes. Referenced via https://arxiv.org/pdf/2504.08729 Vision Language Grounding: How AI Connects “Dog” to Pixels, and Where It Falls Apart was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Cambridge Analytica’s Website Is Now an AI Slop Farm
Somehow, Cambridge Analytica has returned. Sort of. The post Cambridge Analytica’s Website Is Now an AI Slop Farm appeared first on Futurism .
Score: 24🌐 MovesJul 30, 2026https://futurism.com/artificial-intelligence/cambridge-analytica-ai-slop-farm - Lithuanian Sigvi secures €1.2 million to automate operations for independent car rental fleets - ArcticStartup
Lithuanian Sigvi secures €1.2 million to automate operations for independent car rental fleets - ArcticStartup ArcticStartup