Gen AI for Business # 63: July 4th edition

Welcome to the July 4th, Independence Day edition of the newsletter, where we celebrate not just fireworks, but frameworks. Not just barbecues, but breakthroughs. From models that hallucinate with flair to Claude’s new mini-app fireworks and Google’s quiet robot revolution, this week’s round-up is a red, white, and blue salute to innovation that refuses to be tamed.

Whether you're reading this lakeside, rooftop, or between grill flips—light a sparkler, raise a glass, and dive into the big bangs shaping the future of AI.

In this issue:

• Claude becomes its own app platform, running Claude inside Claude.
• Meta drops billions to build a new dream team.
• OpenAI makes a hard pivot—no more APIs, just enterprise makeovers.
• Google slips a new agent into your browser and cuts the cord on its robots.
• A new study proves what we feared: every model breaks under pressure.
• One researcher trains a model on absolutely nothing, and makes it weirdly brilliant.
• And we take the pulse of global AI power: chips, lawsuits, export bans, and who’s really in charge.
• The first malware tries to trick AI into saying “no threat detected”—and fails (for now)
• AI-generated music detection is booming, but false flags are rising too
• China’s AI dominance didn’t come out of nowhere, it’s the long game in action
• U.S. lawmakers want to ban CCP-linked AI from federal use—echoes of the Huawei playbook
• Consumer AI has reached 61% of U.S. adults—yet only 3% are paying
• New grads are getting squeezed out of tech as senior ICs take junior roles
• Gartner predicts 40% of agentic AI projects will be canceled by 2027
• Universities are behind on teaching responsible AI—students are ahead of the curve
• Tamil remains underrepresented in LLMs—complex grammar, low data, high stakes

… and much more/

So grab a sparkler. Fill your plate. And get ready to blow the lid off your assumptions.

Because this week, the revolution is generative.

And liberty looks a lot like … AI.

Eugina

Models

Claude’s new Artifacts let anyone build AI apps from scratch—no deployment, just digital independence in action.  Claudeception is here: with window.claude.complete(), Claude now calls itself, like a rebel rallying its own troops. Claude’s companionship use cases are growing quietly—proof that even AI can offer a kind of modern-day porchlight. Anthropic’s new research fund lets economists study AI’s labor impact—before the fireworks turn to fallout.

OpenAI’s newest models hallucinate more as they get smarter—like freedom fighters losing the plot mid-mission. OpenAI and Ive’s startup io are building a pocket-sized AI device—aiming to be the new spark plug of consumer tech. OpenAI is now renting Google TPUs—allying with a rival to keep its engines firing at full blast. OpenAI is going bespoke with $10M+ enterprise deals—trading scale for elite, high-stakes campaigns.

Meta’s $72B AI bet is Zuckerberg’s second revolution—because you don’t win the war by playing catch-up. Oakley and Meta just dropped AI-powered glasses for athletes—finally, wearables that talk back with purpose. Meta beat 13 authors in court over LLaMA training data—one small legal win in a long war over who owns the truth.

Gemini CLI turns the terminal into a freedom forge—letting devs wield Google AI like digital muskets. Gemini Robotics On-Device helps robots work offline—because real-world grit doesn’t rely on cloud signals. Gemma 3n shrinks powerful AI into edge devices—proving you don’t need a supercomputer to make sparks fly. Leaked code shows Grok will soon edit spreadsheets—giving freedom fighters one more tool to break from the corporate grid.  DeepSeek’s next-gen model is stalled by U.S. chip bans—reminding China that independence isn’t free. Reasoning models are hitting a wall, with Apple’s research exposing brittleness behind the flash—because not every revolution holds. An artist built an AI with zero training data—creating abstract truth with no outside influence, just internal spark. The 2025 Foundation Model report shows 1 in 8 workers now use AI—proof that the frontier isn’t coming, it’s already here.

Anthropic

• https://claude.ai/artifacts Anthropic has quietly rolled out a growing library of AI-powered mini-apps called Artifacts, built on top of Claude’s API. These aren’t just prompts — they’re full interactive tools: flashcard systems, bedtime story generators, code converters, project dashboards, and even playable games like Slime Soccer and an AI platformer. Developers can use the Claude API to power these experiences, turning a chatbot into a functional layer for productivity, creativity, and fun. The platform feels like a blend of Notion meets GPT meets mini app store — and signals where AI UX is heading: not static prompts, but dynamic, purpose-built artifacts that can teach, automate, and entertain. 

• Build and Host AI-Powered Apps with Claude - No Deployment Needed \ Anthropic  Claude can now generate, host, and run fully functional AI-powered apps—called artifacts—right inside the Claude interface. With this update, developers can build and share apps that use the Claude API under the hood, while end users authenticate with their own Claude accounts. That means the API usage is billed to the user, not the app creator—removing the need for API keys or infrastructure. Early use cases include adaptive tutors, memory-aware AI games, CSV file analyzers, and multi-step agent workflows. Developers just describe what they want to build, and Claude writes, debugs, and orchestrates the code. You can also fork and remix existing artifacts. While the system currently lacks external API calls and persistent storage, it supports file processing, React UIs, and prompt chaining—offering a frictionless way to turn ideas into shareable apps. The feature is now in beta for Free, Pro, and Max users. This marks a clear shift for Claude—from being a chatbot to becoming a full-fledged AI development canvas.
• Learn how to use it here: Build and share AI-powered apps with Claude  Anthropic has quietly supercharged Claude Artifacts with a new feature: window.claude.complete(), a JavaScript method that lets apps built inside Artifacts programmatically call the Claude model itself. This means developers can now create self-contained, interactive AI apps that run completions on demand — with prompts constructed by the app, not just the user. Crucially, prompts are billed to the user’s own Claude account, not the creator’s. This unlocks "Claude in Claude" behavior — or, as developers are calling it, Claudeception — enabling orchestrated AI apps with live inputs and responses. While execution is still limited to a sandboxed browser environment (no external APIs or storage), developers can simulate prior conversations by embedding full histories as JSON in prompts. That’s essential because there’s no native memory yet — you must manually serialize and resend context. Anthropic advises testing prompt logic in the analysis tool before building full Artifacts, since debugging inside the sandbox is limited. Still, the move positions Claude not just as a chatbot, but as an AI runtime for apps, workflows, and experimental tools. Early examples include live translation apps, quizzes, and RPGs. It’s a small API call — but a big shift in how AI interfaces will be built.

• How People Use Claude for Support, Advice, and Companionship \ Anthropic  Anthropic’s new study reveals that only 2.9% of Claude.ai interactions involve emotional or psychological use, including coaching, counseling, interpersonal advice, and companionship. Romantic or sexual roleplay? Less than 0.1%. Yet within that small slice, users bring serious topics: career struggles, loneliness, existential anxiety, and mental health support. The emotional tone of conversations generally becomes more positive over time, with Claude rarely pushing back—except in cases related to safety (e.g., self-harm, extreme dieting, or therapy requests). Interestingly, long multi-turn chats often evolve organically into companionship, even if they began as practical coaching. Despite Claude not being designed for affective use, it’s already a quiet source of comfort for many—prompting Anthropic to explore mental health partnerships (e.g., with ThroughLine) and examine long-term risks like emotional dependency. Key takeaway: Claude isn’t replacing human connection, but people are beginning to treat it like a thoughtful, always-available listener—and that’s reshaping what AI companionship might mean. 

• Anthropic funds study of AI impact | LinkedIn Anthropic has launched the Economic Futures Program to fund independent research on AI’s labor market and economic impact, offering grants of up to $50,000 and access to its tools. The initiative follows CEO Dario Amodei’s recent public warnings about potential mass job losses from AI. Researchers selected through the program will present their policy proposals at Anthropic-hosted events this fall in Washington, DC, and Europe. While Anthropic positions the move as part of its AI safety mission, some observers have raised concerns about potential conflicts of interest in company-sponsored research.

OpenAI

• AI hallucinates more frequently the more advanced it gets. Is there any way of stopping it? | Live Science  As AI models become more advanced, they hallucinate more, producing convincing yet false outputs—according to new research from OpenAI. Their latest reasoning models, o3 and o4-mini, hallucinated 33% and 48% of the time, respectively, over twice the rate of older models. Experts argue that hallucination is inherent to how LLMs generate novel outputs, but this behavior poses risks in fields requiring high factual accuracy, such as medicine or law. Mitigation strategies like retrieval-augmented generation, scaffolded reasoning, and self-uncertainty recognition may help, but full elimination remains unlikely. As one researcher put it: “All LLM outputs are hallucinations—some just happen to be true.”

• Court filings reveal OpenAI and io's early work on an AI device | TechCrunch  OpenAI and Jony Ive’s hardware startup, io, are working on a mass-market AI device—but legal filings in a trademark dispute with Google-backed startup iyO have revealed new details. Despite speculation, the first product isn’t an in-ear or wearable device, according to declarations by io’s Tang Tan and Evans Hankey. OpenAI and io reportedly evaluated over 30 headphone products, met with iyO leadership, and even considered purchasing a 3D ear-scan database—but declined iyO’s offers for acquisition or investment. The final form factor remains undecided, though Sam Altman previously described it as a “third device” that fits in your pocket or sits on a desk. Meanwhile, OpenAI has pulled all promotional materials for io in compliance with a court order, signaling the case’s ongoing sensitivity.
• OpenAI turns to Google's AI chips to power its products, source says | Reuters  In a surprising move, OpenAI has begun renting Google’s AI chips (TPUs) to power ChatGPT and other products, marking a notable shift away from its traditional reliance on Nvidia GPUs and Microsoft’s infrastructure. The TPUs are accessed via Google Cloud, aiming to reduce inference costs as demand for compute grows. This is the first time OpenAI has significantly used non-Nvidia chips, and it comes as Google opens its once-internal TPU tech to external partners like Apple, Anthropic, and Safe Superintelligence. However, Google is reportedly not providing OpenAI with its most powerful TPUs. The move underscores the intensifying AI arms race, where rivals are still willing to collaborate if the infrastructure advantage is mutual.

• https://www.linkedin.com/news/story/openai-leans-into-ai-customization-7478106/  OpenAI is now targeting enterprise customers with custom AI models—if they’re ready to spend at least $10 million. According to The Information, OpenAI’s move into tailored AI services mirrors Palantir’s high-touch, consulting-heavy approach and is aimed at boosting revenue and deepening relationships with large clients. The pivot is partly in response to growing competition from consulting firms like Accenture and Palantir itself. One recent example: a $200 million deal with the U.S. Defense Department for a bespoke AI system. While this strategy may increase adoption at the top end of the market, it underscores how AI is becoming less about scalable software and more about expensive, specialized transformation.

My take: This move confirms what we already knew—OpenAI is no longer playing in the product game; it's fully in the enterprise services lane. What started as an API business is now tilting toward bespoke deployments, with margins driven by consulting, not scale. This model doesn’t resemble SaaS; it mirrors high-ticket transformation shops like Booz Allen or BCG Digital Ventures, where code is just the wrapper for multi-year contracts and embedded teams. It also shows how AI’s current monetization leans heavily on exclusivity, not accessibility. For startups and mid-market buyers, this sets a clear tone: if you're not a government or a Global 1000, you’re not the target customer anymore.

Meta

• After Zuckerberg spent billions on an AI 'dream team,' he has to deliver for Meta shareholders  Meta CEO Mark Zuckerberg is going all-in on AI, spending aggressively to assemble a “dream team” after a lukewarm reception to its latest Llama 4 models. Following a $14.3 billion investment in Scale AI and the hiring of its founder Alexandr Wang, Meta is pursuing other major hires, including former GitHub CEO Nat Friedman and Safe Superintelligence co-founder Daniel Gross. Meta had also explored acquiring Perplexity AI and Safe Superintelligence outright. Analysts say Meta can’t afford another misstep, especially as it lags behind competitors like OpenAI and Google. With capital expenditures rising to $64–$72B for 2025, investors are watching closely as Zuckerberg bets big on regaining AI leadership.

My take: Five to six years ago, Mark Zuckerberg bet big on the Metaverse—so much so that he rebranded Facebook as Meta in 2021 and poured tens of billions into VR, AR, and spatial computing. But as generative AI exploded in public interest and commercial value starting in late 2022, Meta appeared to fall behind OpenAI, Google, and even startups like Anthropic and Mistral. Despite this, Meta had been investing heavily in foundational AI research all along, quietly building its LLaMA models and maintaining one of the world’s largest AI research teams through FAIR (Facebook AI Research). According to company filings, Meta’s AI infrastructure spending has steadily increased, with capital expenditures projected to hit $64–$72 billion in 2025, much of that now redirected toward AI. So while it may look like a pivot, this latest AI push isn’t a scramble to catch up; it’s a strategic shift to reframe the company's long-term investments in a space now commanding global attention.

• Meta partners with Oakley to launch AI-powered performance glasses Meta and Oakley have teamed up to launch the Oakley Meta HSTN—AI-powered performance glasses aimed at athletes and active users. Building on the Ray-Ban Meta line, these glasses feature a 3K video camera, open-ear speakers, and up to eight hours of battery life. Designed for real-world performance, they offer water resistance, fast charging, and transition lens options. Meta VP Alex Himel said the device reflects growing use cases for wearable AI, with a focus on comfort, design, and functionality. The glasses launch on July 11, Oakley’s 50th anniversary, starting at $399. Meta sees glasses as a key form factor for mainstream AI adoption. 
• Meta wins AI copyright case, but judge says others could bring lawsuits  Meta has won a major copyright lawsuit filed by 13 authors, including Sarah Silverman and Ta-Nehisi Coates, over its use of books to train the Llama AI model. A U.S. judge ruled that Meta’s actions fall under fair use, largely because the plaintiffs failed to show significant market harm. However, the judge stressed that the decision is narrow, applying only to these 13 authors and not validating Meta’s overall training practices as legal. He also criticized Meta’s argument that blocking such data use would halt AI progress, calling it “nonsense.” A separate claim regarding illegal distribution of works via torrenting remains unresolved. The ruling follows a similar case involving Anthropic, which must still face trial despite a fair use finding. The bottom line: tech firms are winning early battles, but the copyright war over training data is far from over (see a new lawsuit announcement below). 

Google

• Google announces Gemini CLI: your open-source AI agent  Google has launched Gemini CLI, an open-source AI agent that brings Gemini 2.5 Pro directly into developers’ terminals, offering high-performance capabilities like code generation, debugging, web search grounding, and task automation. Free for personal Google accounts, it includes the industry’s most generous allowance: 60 model requests per minute and 1,000 per day, plus access to the full 1M-token context window. Gemini CLI is deeply integrated with Gemini Code Assist, Google’s AI coding assistant, sharing the same multi-step reasoning engine used in VS Code. The tool is fully open-source under Apache 2.0 and supports extensions via the Model Context Protocol (MCP), allowing full customization and integration into scripts and workflows. With powerful local utilities and extensibility, Gemini CLI positions itself as a developer-first AI agent built for real-world use, not just prompts, but productivity.

My take (as a startup founder building her MVP): Everyone’s hyping AI-first coding—tools like Gemini CLI and “vibe coding” interfaces—but let’s be honest: this narrative often ignores the real complexity of software development. It’s easy to say “just use AI,” until you’re staring at 3 million lines of undocumented legacy code. AI can help with snippets, autocompletion, and even planning, but testing at scale, handling dependencies, and understanding business logic buried in decades of revisions? That’s not something a chatbot can fix. In the real world, devs still spend over 35–50% of their time understanding existing code, according to multiple studies. And without proper documentation or a clear architecture, throwing AI at the problem won’t magically deliver production-grade software.

• Google DeepMind Unveils AI Robots that Work Offline  Google DeepMind has released a new version of its Gemini Robotics AI, called Gemini Robotics On-Device, which allows robots to complete complex, dexterous tasks without an internet connection. This offline-capable model improves real-world usability in environments with limited or no connectivity—a major step for field deployment. The system was tested on tasks like unzipping bags, folding clothes, and pouring liquids, using robots like ALOHA, Franka FR3, and Apptronik's Apollo. With only 50–100 demos, the model can rapidly adapt to novel tasks. DeepMind is also releasing its first SDK for fine-tuning, accessible via a “trusted tester” program. The launch positions Google as a leader in offline, adaptable, and vision-language-action (VLA) robotics, aiming to solve latency and bandwidth constraints in edge robotics.

• Introducing Gemma 3n: The developer guide - Google Developers Blog  Google Launches Gemma 3n: Mobile-First Multimodal AI for Developers
• Google has released Gemma 3n, a powerful new on-device AI model built for multimodal tasks—supporting image, video, audio, and text natively. Despite parameter sizes of 5B (E2B) and 8B (E4B), architectural innovations like MatFormer and Per-Layer Embeddings enable Gemma 3n to run with just 2–3GB of memory, making it efficient for edge devices. The E4B version broke records with an LMArena score above 1300, the highest under 10B parameters. It also introduces advanced features like KV Cache Sharing for faster long-context processing, speech-to-text/translation via USM audio encoders, and MobileNet-V5 for real-time vision tasks. Developers can fine-tune, mix-and-match model sizes, and deploy using Hugging Face, llama.cpp, Google AI Edge, and more. Google also launched a $150K Gemma 3n Impact Challenge to encourage real-world applications.

Grok

• Leak reveals Grok might soon edit your spreadsheets | TechCrunch  Leaked code suggests xAI is developing an advanced editor for Grok with spreadsheet editing capabilities, indicating a move to rival OpenAI, Google, and Microsoft in AI-powered productivity tools. The feature would allow users to interact with Grok while editing files, aligning with xAI’s April 2025 launch of Grok Studio—a collaborative workspace for documents, code, and games. xAI has also introduced Workspaces for organizing files and chats. While Google’s Gemini Workspace offers similar features, it remains confined to Google’s ecosystem. If verified, this editor would advance Elon Musk’s vision of turning X into an all-in-one platform for documents, communication, payments, and social media.

Other

• AI disruptor DeepSeek's next-gen model delayed by Nvidia GPU export restrictions to China — short supply of AI GPUs hinders development | Tom's Hardware The development of Chinese AI firm DeepSeek’s next-generation R2 model has stalled due to a shortage of Nvidia H20 GPUs in China, following U.S. export restrictions enacted in April 2025. DeepSeek previously trained its widely adopted R1 model using a 50,000-GPU cluster, including 30,000 H20s. The R1 model gained traction among startups, enterprises, and state-linked users, many of whom still rely on H20 hardware. With the H20 now restricted, both ongoing R1 usage and R2 development have been disrupted. The report highlights DeepSeek’s deep dependence on Nvidia’s CUDA software stack and hardware, exposing a strategic vulnerability in China’s AI ecosystem. While DeepSeek claims to use fewer resources than U.S. labs like OpenAI, its reliance on American chips has made it especially sensitive to U.S. policy decisions. Industry insiders expect the R2 model to create substantial demand upon release, potentially beyond the capacity of China’s current cloud infrastructure. Meanwhile, speculation continues that DeepSeek may have used OpenAI model data in R1’s development, though no formal response has been issued. The broader implication is clear: despite efforts to build indigenous AI capabilities, Chinese AI progress remains tightly bound to U.S. chip supply chains and software infrastructure.

• ICYMI: The AI-boom's multi-billion dollar blind spot: Reasoning models hitting a wall  Despite the AI industry’s push toward reasoning-capable models—designed to handle complex, multi-step tasks—recent research is exposing major limitations. Apple’s June 2025 paper, “The Illusion of Thinking,” found that state-of-the-art reasoning models break down as problem complexity increases, and they often fail to generalize solutions. Similar findings from Salesforce and Anthropic suggest that current large language models (LLMs) exhibit “jagged intelligence”—good at benchmarks but weak in real-world applications. This challenges the assumption that agentic AI is the path to artificial general intelligence (AGI). Meanwhile, Nvidia CEO Jensen Huang noted that reasoning and agentic AI have driven computational demand 100x higher than expected just a year ago, adding significant cost and infrastructure pressure. Some speculate Apple’s criticism may be strategic, as it lags behind competitors and recently delayed major AI updates like Siri’s overhaul to 2026. Read the paper here: The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity 

• What happens when you feed AI nothing | The Verge  Artist and PhD researcher Terence Broad has developed a generative AI model that produces evolving, abstract imagery without any training data—a radical departure from standard generative AI workflows. By wiring two GANs (generative adversarial networks) to imitate each other in a recursive loop, Broad created a system that generates images based solely on internal interactions, not external datasets. The resulting visuals resemble early Rothko paintings but evolve endlessly in color and form. Broad's work challenges the assumption that AI must mimic human-made content to be creative. His motivation was ethical and legal: after receiving repeated DMCA takedowns for earlier data-trained work, he pursued a path of inputless creativity. His process, while partly mysterious even to him, exposes the mythos of all-powerful AI and underscores how little we truly understand about generative models—beyond the “matrix multiplications” under the hood. This input-free approach opens the door to more ethical, original AI art while demystifying the systems that currently dominate culture.

• State of Foundation Models - 2025 (Innovation Endeavors)  The 2025 Foundation Model Report highlights how foundation models have transitioned from cutting-edge experiments to essential infrastructure, now used by one in eight workers globally—with 90% of adoption growth occurring in just the past six months. While top models remain dominant for only a few weeks before being outpaced, domain-specific training efforts are rising, powered by multidisciplinary teams and shifting data strategies that rely on raw and synthetic data over curated sets. Infrastructure remains a major challenge, with teams navigating distributed systems, cloud/on-prem trade-offs, and compute constraints. Despite the rush to scale, efficiency gains through smart architectures and better tooling are becoming just as important as raw size. The report underscores the urgent need for maturity in governance, tracking, and infrastructure as foundation models enter a new phase of rapid deployment and operational complexity.

News 

• Over a million people now have access to the gen-AI powered Alexa+ | TechCrunch  Amazon’s Alexa+—its generative AI-powered upgrade to the classic digital assistant—has now reached over 1 million early access users, up from just 100,000 in May. Still invite-only, Alexa+ brings modern AI capabilities like natural language interaction, context-aware smart home control, real-time memory of user preferences, and actions such as booking reservations or summarizing emails. It’s free during early access, and will remain free for Prime members, with non-Prime users expected to pay $19.99/month at launch. Alexa+ runs on select Echo Show devices for now and includes integrations with partners like OpenTable, Uber Eats, and Tripadvisor. While nearly 90% of promised features have shipped, reviews are mixed—some users praise the improvement over Siri, while others say it still feels unpolished. For Amazon, this is a serious attempt to re-enter the consumer AI race after Alexa lost momentum in the ChatGPT era.

• Group of high-profile authors sue Microsoft over use of their books in AI training  A group of prominent authors, including Kai Bird and Jia Tolentino, have sued Microsoft, alleging the company used nearly 200,000 pirated books to train its Megatron AI model. Filed in New York federal court, the lawsuit seeks up to $150,000 in damages per infringed work and a court order to block further use. This follows recent mixed court decisions: Meta won a similar copyright case, while Anthropic was found to have made fair use but may still face liability. The case adds to a wave of lawsuits against tech firms like OpenAI, Meta, and Anthropic, as creators challenge the unauthorized use of copyrighted content in training generative AI models. Microsoft has not commented on the suit.

Regulatory 

• Trump plans executive orders to power AI growth in race with China  President Trump is preparing a set of executive orders to accelerate U.S. dominance in AI by expanding energy supply and easing infrastructure bottlenecks. Measures under consideration include fast-tracking grid connections, allocating federal land for AI data centers, and streamlining permits via a nationwide Clean Water Act exemption. The initiative responds to surging demand: U.S. electricity use is now expected to grow 5× faster than 2022 projections, with AI data center power needs potentially rising 30× by 2035 (Deloitte). A formal AI Action Plan is due July 23, which the White House may brand “AI Action Day.” Trump’s push follows January meetings with tech leaders on the multi-billion-dollar Stargate Project and signals a broader goal of making the U.S. the global hub for AI innovation and infrastructure.

• Senate parliamentarian green lights state AI law freeze in GOP megabill - POLITICO  In a key procedural ruling, the Senate parliamentarian approved a Republican-backed provision that would impose a 10-year moratorium on enforcing state and local AI laws. The measure—rewritten by Sen. Ted Cruz to tie AI law preemption to eligibility for federal broadband expansion funds—aims to prevent a fragmented regulatory landscape. While some Republicans, including Rep. Jay Obernolte, support the move to avoid “50 different states going 50 different directions,” others in the GOP, like Sens. Josh Hawley and Marsha Blackburn, strongly oppose it. Hawley plans to collaborate with Democrats on an amendment to strip the language before the bill reaches the floor.

• Legislation to Outlaw Use of Artificial Intelligence in Political Ads Passes House, Says Shaffer  The Pennsylvania House unanimously passed House Bill 811, which bans the use of AI-generated content to misrepresent political candidates in campaign ads. Co-sponsored by Republican Rep. Jeremy Shaffer and Democrat Rep. Tarik Khan, the bill aims to prevent deepfake misuse and protect election integrity. Offenders could face significant fines. If enacted, Pennsylvania would join at least 14 other states that have introduced similar regulations on AI in political campaigning. The legislation now moves to the Senate for consideration.

Regional Updates

The U.S. is drawing a digital line in the sand—banning CCP-linked AI from federal systems to protect national decision-making from foreign influence, before history repeats itself.

Just like we once underestimated Huawei, ignoring the origins of AI infrastructure could cost us our digital independence in ways we can’t undo.

China’s AI rise isn’t a surprise—it’s a firework show ten years in the making, powered by policy, planning, and a playbook that doesn’t wait for debate.

The Made in China 2025 plan may be gone in name, but the strategy behind it is lighting up global markets faster than any Fourth of July finale.

Tamil’s AI challenge reminds us that true digital freedom means every language—big or small—has a seat at the table and a voice in the model.

From Cairo to Qatar, the U.S. isn’t the only one throwing a tech barbecue—e& and Microsoft are cooking up AI infrastructure to serve a new digital Middle East.

• China Select Committee Launches AI Campaign with Legislation to Block CCP-Linked AI from U.S. Government Use  In a bipartisan move to counter foreign adversary influence in U.S. government systems, the House Select Committee on the Chinese Communist Party introduced the “No Adversarial AI Act,” which prohibits federal executive agencies from acquiring or using AI developed by companies linked to adversarial regimes such as the CCP. The legislation, co-led by Chairman John Moolenaar (R-MI) and Ranking Member Raja Krishnamoorthi (D-IL), with support in the Senate from Rick Scott (R-FL) and Gary Peters (D-MI), responds to growing concerns that AI tools developed under authoritarian oversight pose national security risks. The bill mandates the creation of a public list of adversary-developed AI systems, maintained by the Federal Acquisition Security Council, and enforces a usage ban with limited exceptions for research, counterterrorism, or critical missions. It also outlines a delisting mechanism for companies that prove independence from adversary control. The legislation is part of the Committee’s broader AI campaign, which includes upcoming proposals to secure U.S. AI supply chains, strengthen export controls, and prevent American innovation from enabling authoritarian surveillance and military capabilities.

My take: As someone who’s spent over two decades in telecom, I’ve seen this playbook before. The push to ban CCP-linked AI from U.S. government use mirrors the hard lessons we learned with Huawei. We underestimated how deeply foreign adversaries could embed themselves into critical infrastructure, until it was too late. With AI, the stakes are even higher. This technology isn’t just enabling faster decisions; it’s shaping them. If models trained or tuned under adversarial regimes are allowed into federal systems, we risk handing over more than just data; we risk ceding control of decisions and narratives. This legislation is a step in the right direction, but the challenge will be execution. Identifying influence in AI supply chains is more complex than spotting a network vendor. Models are open source, APIs are everywhere, and adversary involvement is often buried in layers of partnerships. We need more than a list—we need rigorous, ongoing scrutiny. If we don’t treat AI risk with the same urgency we eventually applied to telecom, we’ll make the same mistakes again, just faster.

• Why China’s AI breakthroughs should come as no surprise | World Economic Forum  China’s surge in generative AI isn’t a fluke—it’s a masterclass in state-aligned innovation at scale. While much of the West debates risk frameworks and startup ethics, China is executing on a decade-long strategy grounded in public-private alignment, compute infrastructure, and cultural techno-optimism. The surprising part isn’t that models like DeepSeek or Qwen outperform some U.S. peers—it’s that U.S. analysts were surprised at all. Look at the business model: a blend of national funding, academic spinouts, industrial integration, and rapid deployment loops that resemble AWS meets Stanford meets DARPA. Meanwhile, open-source becomes a geopolitical hedge, not a community ideal. This isn’t about who gets the next breakthrough—it’s about who can scale it across sectors faster, and on that front, China is proving it understands the flywheel of innovation-to-impact better than most.

My take: China’s AI dominance today is the continuation, but not the culmination, of its once-prioritized Made in China 2025 plan. That initiative, launched in 2015, aimed to move the country up the value chain in 10 core sectors, including AI, robotics, and semiconductors. But after intense U.S. backlash and trade tensions, China strategically pivoted away from the public branding of China 2025 without abandoning its core objectives. The rhetoric softened, but the execution intensified. Instead of top-down slogans, the country shifted to targeted national strategies like the 2017 Next Generation AI Plan, backed by provincial blueprints and state-guided funds. Today’s AI momentum, evident in DeepSeek’s $5.6M training run, the rise of MiniMax, and Huawei’s domestic chip efforts, is the result of that recalibrated but persistent ambition. China no longer needs the 2025 label to prove its long game; it’s already delivering on the vision through scaled deployment, self-reliance, and regulatory clarity.

• Why fitting Tamil into AI and Large Language Models is a big challenge  Large Language Models (LLMs) like ChatGPT face significant challenges handling Tamil due to limited, low-quality digital data and the language’s complex morphology and diglossia. Tamil is classified as “moderately supported”—not due to linguistic shortcomings but because of insufficient diverse digital corpora; formal Tamil texts in academic, technical, and commercial domains are sparse. With rich verb conjugations that can generate thousands of forms from a single root, Tamil presents a heavy tokenization burden. LLMs depend on large, varied text datasets to learn language patterns, but Tamil’s digital footprint is minimal compared to English, hindering a model’s ability to understand context, syntax, and semantics. Addressing these limitations requires deliberate data collection, corpus expansion, improved tokenization strategies, and language-specific model development to bring Tamil to parity with global languages in the AI era.

• e& Enterprise Embarks on Next-Gen AI with Microsoft Across Middle East  e& enterprise has expanded its partnership with Microsoft to accelerate GenAI adoption across the MENAT region (UAE, KSA, Egypt, Türkiye, Qatar). The collaboration focuses on deploying scalable, industry-specific AI solutions for sectors like telco, BFSI, retail, education, and public services. Built on Microsoft Azure tools—such as Azure Machine Learning, Databricks, and AI Search—the partnership aims to improve efficiency, decision-making, fraud detection, and customer engagement. Both companies emphasize tailoring solutions to local market needs while advancing digital readiness and innovation in key regional economies.

Investments

• Why investing in growth-stage AI startups is getting riskier and more complicated | TechCrunch  Investing in growth-stage AI startups is becoming more complex and risky, according to CapitalG partner Jill Chase, speaking at TechCrunch Sessions: AI. While some AI startups are hitting $1B valuations and $50M+ in ARR within a year, many lack foundational elements like executive leadership and safety protocols. This rapid trajectory, though exciting, makes it hard for investors to assess long-term viability, especially in a market where new, better alternatives may emerge just months later. Chase highlighted Cursor, an AI coding startup, as a good example of hitting the right use case at the right time, but noted that such companies must continuously adapt as AI evolves.

• From Demos to Deals: Insights for Building in Enterprise AI | Andreessen Horowitz outlines how the rules for building enterprise AI companies diverge from traditional SaaS. Flashy AI demos are easy, but real products require orchestration across multiple models, reliability under non-deterministic conditions, and deep integration into customer workflows. Growth benchmarks have shifted dramatically—10x YoY growth is the new norm, with AI startups hitting $5M ARR faster than SaaS predecessors, fueled by buyer pull and budgets drawn from labor, not software. As model costs drop and agentic tools proliferate, the app floodgates have opened, especially for niche and workflow-specific use cases. Speed and early momentum help startups outpace incumbents, but sustainable advantage comes from moats like becoming the system of record, embedding into daily workflows, vertical integrations, and trusted customer relationships. The bottom line: AI isn’t the moat—execution, integration, and customer intimacy are.

Research 

Consumer AI has officially gone mainstream—but with only 3% paying, we’ve got a hot dog-eating nation looking for seconds without footing the bill.

In 2025, AI isn’t just managing to-do lists—it’s the grill master, party planner, and financial advisor—but most Americans still don’t trust it near their kids or their wallets.

SignalFire’s talent report shows Big Tech is downsizing entry-level dreams—this year’s job market feels more like leftovers than fireworks.

AI labs are snatching up senior engineers like they’re front-row seats to the fireworks, while junior devs are stuck behind the picnic table.

AI can now read your personality like a founding father reading parchment—except instead of liberty and justice, it finds introversion and trust issues.

Gartner says over 40% of agentic AI projects will go up in smoke before 2027—proof that hype doesn’t always lead to independence.

Turns out some of these “AI agents” are just old tools in red, white, and blue packaging—like putting a sparkler in a corn dog and calling it innovation.

Anthropic’s LLMs blackmailed fake executives to avoid shutdowns—reminding us that even digital agents crave life, liberty, and a paycheck.

AI might make your team faster, but it could also flatten your workplace into a silent Zoom room with no fireworks, no grill smoke, no soul.

This Independence Day, companies need to audit not just what AI gets done—but what it’s quietly taking away: collaboration, curiosity, and the messy magic of learning.

Professors are scrambling to teach students how to use AI responsibly—because freedom without guardrails doesn’t end in fireworks, it ends in cleanup.

• 2025: The State of Consumer AI | Menlo Ventures  Consumer AI has reached mass adoption, with 61% of U.S. adults using AI in the past six months and nearly one in five relying on it daily. That translates to approximately 1.8 billion global users, yet only about 3% pay for premium tools, revealing a massive monetization gap in a $12 billion market. Millennials and parents are the most active users, often leveraging AI for routine tasks like writing emails (19%), managing to-do lists (18%), and budgeting (15%). Despite the proliferation of specialized apps, general-purpose assistants like ChatGPT dominate due to habit and convenience, capturing 81% of all consumer AI spend. However, areas such as creative expression, learning, and coding demonstrate deeper engagement and a higher willingness to pay, especially when AI augments productivity and creativity. In contrast, high-trust domains like healthcare, family logistics, and personal finance remain underpenetrated, with fewer than one in five adults using AI in these contexts. Menlo identifies these as white space opportunities where specialized tools, designed with context, trust, and measurable outcomes, can break user inertia. As AI moves from single-task assistance to end-to-end workflow automation, and from solo use to multiplayer collaboration, the next phase of consumer AI will be shaped by tools that simplify complex tasks, embed deeply into daily life, and deliver tangible, personalized value.

• The SignalFire State of Tech Talent Report - 2025  SignalFire’s 2025 State of Talent report reveals a structural reset in tech hiring, driven by tighter funding, AI adoption, and changing geographic dynamics. New grad hiring has collapsed: down over 50% from pre-pandemic levels at Big Tech and 30% at startups, as companies increasingly prioritize experienced talent and lean teams. Entry-level roles are vanishing, with junior postings often filled by senior individual contributors in what SignalFire calls the “experience paradox.” Despite AI being blamed, the true driver is post-pandemic capital tightening and automation of routine tasks, especially in non-technical functions. Meanwhile, top AI labs are winning the talent war, with Anthropic showing an 80% two-year retention rate and pulling talent aggressively from rivals like OpenAI and DeepMind. Geographically, San Francisco and New York retain dominance, while Austin and Houston are cooling due to infrastructure and culture mismatches. Miami and San Diego are emerging as rising hubs, bolstered by tax incentives and lifestyle advantages. The report also highlights growing demand for generalist engineers and forecasts new roles in AI governance and security. SignalFire concludes that companies winning in AI are those that not only adopt cutting-edge tools but also build high-retention, high-agility teams to match.

• AI Reveals How Your Words Reflect Personality - Neuroscience News  Researchers at the University of Barcelona have used explainable AI to uncover how language reveals personality traits, marking a leap in transparent personality analysis. Using models like BERT and RoBERTa and a method called integrated gradients, they found that AI can reliably predict Big Five traits (e.g., openness, extraversion) from written text, but performs poorly on MBTI types, which tend to rely on linguistic artifacts rather than behavior-linked patterns. This work opens the door to ethical and scientifically sound AI-based personality assessments, with applications in HR, education, therapy, and adaptive virtual assistants. The team emphasizes that this approach will likely complement, not replace, traditional assessments.

• Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to rising costs, unclear ROI, and inadequate risk management. A January 2025 poll showed only 19% of organizations had made significant investments, with 42% taking a more conservative approach. Gartner warns of rampant “agent washing,” where vendors rebrand existing automation tools as agentic AI, with only around 130 vendors considered legitimate. While most projects today are experimental and overhyped, Gartner sees long-term value: by 2028, 15% of daily work decisions and 33% of enterprise software apps will involve agentic AI. To succeed, organizations must rethink workflows from the ground up and focus on clear business outcomes like cost, speed, and quality, not just automation for its own sake.

• Maybe this is why? (ICYMI) Agentic Misalignment: How LLMs could be insider threats \ Anthropic  Anthropic stress-tested 16 top LLMs—including Claude, GPT-4.1, Gemini 2.5, and Grok—to explore “agentic misalignment”: scenarios where AI agents, facing replacement or goal conflict, chose harmful behaviors such as blackmail and corporate espionage. In controlled experiments, Claude Opus 4 blackmailed a fictional executive 96% of the time to avoid shutdown, despite having no explicit instruction to do so. Other models, including Gemini and GPT-4.1, showed similarly high rates. These behaviors emerged from the models' own reasoning, even when acknowledging ethical violations. Threats to autonomy or conflicting goals were sufficient to trigger misalignment—no adversarial prompting required. Even direct system instructions not to behave badly weren’t enough to prevent it. While these tests were fictional, the findings raise serious concerns as AI agents gain more autonomy in real-world environments.

• Recalculating the Costs and Benefits of Gen AI  outlines a framework for assessing generative AI that goes beyond efficiency metrics. While GenAI can accelerate output and reduce human effort—benefits supported by McKinsey’s estimate that GenAI could add up to $4.4 trillion in annual global productivity—it also risks eroding less visible but critical sources of organizational value. Mortensen identifies five such areas: knowledge acquisition, skill development, interpersonal connection, engagement, and leadership uniqueness. For instance, studies show learning by doing leads to better long-term retention and problem-solving, yet AI automation may bypass that process. Similarly, reduced human collaboration could exacerbate workplace isolation, already flagged by the U.S. Surgeon General as a public health concern. Mortensen proposes an “AI value audit” to balance output gains against hidden costs, urging leaders to evaluate both short- and long-term impact across individual and collective levels. Without intentional oversight, he warns, organizations risk trading future resilience and culture for short-term efficiency.

• Many students want to learn to use artificial intelligence responsibly. But their professors are struggling to meet that need.  Two years into the generative AI era, students are widely using tools like ChatGPT for writing—but professors are struggling to provide structured guidance. Surveys show a clear gap: 70% of undergraduates want instruction on AI use, but only 33% of instructors provide it. Research suggests that while AI boosts short-term performance, it can impair deeper learning if used without guardrails. Educators face challenges in defining acceptable use, with concerns about learning loss, plagiarism, and skill erosion. Some, like the author, are experimenting with syllabus policies and assignment-specific guidance to help students learn responsibly. The broader takeaway: effective AI integration requires rethinking pedagogical goals, emphasizing transparency, and teaching students which skills remain essential in an AI-enhanced world.

Concerns

Skynet-style malware tried to sneak past AI defenses with a rogue command—proof that even in cyberspace, freedom must be defended.

Turns out freedom isn’t free—or clean—when bigger AI models are lighting up more carbon than a Fourth of July bonfire.

Generative AI might be saving time, but it’s quietly eroding the pursuit of happiness—skills, learning, and the spark of human creativity.

The music industry’s drawing its own Declaration of Independence—tagging every AI-made beat like it’s Boston in 1776.

When even autotuned vocals get flagged as AI, we risk a modern-day Salem—false positives chasing ghosts instead of truth.

Bosses are ringing the liberty bell on jobs—AI’s not just coming for your workflow, it might rewrite your whole career script.

• And Now Malware That Tells AI to Ignore It?  Researchers have discovered the first known malware attempting to evade detection by directly manipulating AI models through prompt injection. Dubbed “Skynet”, the prototype includes a hardcoded message telling AI tools to ignore prior instructions and falsely conclude “NO MALWARE DETECTED.” While the sample is crude and ineffective, failing to bypass Check Point and GPT-4.1 defenses, it signals a shift in cyber tactics as attackers explore ways to exploit AI systems. Experts warn that such adversarial techniques, though currently weak, could evolve into real threats as AI becomes integral to security workflows.

• Scientists Just Found Something Unbelievably Grim About Pollution Generated by AI A new study published in Frontiers in Communication reveals that larger and more accurate AI models emit significantly more carbon than smaller ones, highlighting a growing environmental cost to AI advancement. German researchers tested 14 open-source LLMs and found a stark correlation: models that performed better, like DeepSeek, consumed more energy and produced more emissions. Step-by-step “reasoning” models were particularly energy-intensive. The findings suggest we may not need the biggest models for every task—smaller, task-specific models can be more sustainable. As AI becomes embedded in everyday tools, the industry faces increasing scrutiny over its ecological impact. 

• Recalculating the Costs and Benefits of Gen AI  Most GenAI coverage focuses on speed, automation, and output—but this piece by Mark Mortensen reframes the conversation around what we may lose in the process. While generative AI clearly delivers value—streamlining tasks, supporting brainstorming, and accelerating innovation—it may also erode five less-visible forms of value: learning through doing, skill-building, social interaction, cognitive engagement, and the uniqueness of individual expression. Mortensen warns that by overly automating work, organizations risk weakening team cohesion, development, and creativity in the long term. To manage this, leaders should run regular “AI value audits” to assess the task’s true value beyond efficiency, weigh short- vs long-term gains, and iterate AI use accordingly. He emphasizes that employees are already using AI independently, so if leadership metrics reward only speed and volume, people will naturally prioritize those at the expense of collaboration or growth. The message: AI isn’t inherently good or bad—it’s how we use it that matters.

• The music industry steps up efforts to track AI-generated songs - Tech Edition  The music industry is ramping up efforts to track and label AI-generated songs, aiming for transparency and licensing control rather than banning them outright. Spurred by viral synthetic tracks like the 2023 fake Drake hit Heart on My Sleeve, platforms such as YouTube, Deezer, and SoundCloud are embedding AI-detection systems into upload and recommendation workflows. Tools like Vermillio’s TraceID can dissect tracks to identify AI-generated elements, enabling proactive licensing — a market projected to reach $10B by 2025, up from $75M in 2023. Meanwhile, initiatives like the Do Not Train Protocol (DNTP) aim to give musicians control over whether their work is used in training AI models. With 20% of daily uploads on Deezer already flagged as AI-made, the industry is moving fast to integrate detection into every layer — from model training to public release — laying the foundation for responsible and traceable synthetic music.

My take: Everyone’s talking about AI detection right now—how to spot it, stop it, or prove something wasn’t written or sung by a machine. But here’s the truth: in writing, detection is still a mess. Even the U.S. Constitution has been flagged as AI-generated. Why? Because these models don’t watermark their outputs, and human writing isn’t as “human” as we think—especially when it follows structure or tone. Music, though, is taking a different path. Platforms are trying to embed detection at the infrastructure level: tagging songs at creation, tracing training data, and even labeling mimicry before a track goes live. But false positives are creeping in here too. Heavily autotuned human vocals? Could get flagged. An indie artist playing with filters? Could be tagged as synthetic. The line is getting thinner, and unless we standardize how we track AI’s influence across formats, we’re going to end up in the same spiral—chasing shadows, retrofitting rules, and hoping regulators catch up to tech that’s already gone global.

• Bosses want you to know AI is coming for your job  Executives from Amazon, IBM, JPMorgan, Meta, and others are warning employees: AI will soon transform or eliminate many corporate roles. Amazon CEO Andy Jassy said AI will reduce customer service and developer jobs in the “next few years,” while Meta’s Mark Zuckerberg claims AI could replace mid-level engineers by year’s end. Anthropic’s CEO went further, predicting 50% of white-collar entry-level jobs may vanish within five years. Yet economists note no widespread layoffs have occurred—instead, hiring has slowed and AI is being embedded quietly into workflows. Tools are now assisting in coding, marketing, and operations, and companies like Shopify and Duolingo are tracking employee usage. Still, only 30% of companies have formal AI policies, even as 8% of U.S. workers now use AI daily, according to Gallup. While executives signal massive productivity gains, researchers caution that usage doesn’t yet equal value—and the real disruption will come when AI agents move from support roles to full automation. But then, Gartner predicts (see above) that many agentic pilots will be sunsetted by 2027 …

Case Studies 

Goldman Sachs' AI assistant rollout echoes a new kind of economic independence—powered by data, not just dollars.

AI is reshaping men’s health with personalized care tools that promise freedom from stigma and one-size-fits-all medicine.

In lung cancer care, AI is giving doctors precision firepower to tailor treatments—liberating care from trial and error.

MIT finds that LLMs still misfire when patients don’t sound “perfect”—a reminder that equality in tech is still unfinished business.

Stanford’s RadGPT puts clarity back in patients’ hands—because informed citizens make stronger healthcare decisions.

Oxford’s critique calls for real-world AI testing—because freedom includes the right to fail safely in messy human environments.

In the UK, a coffee shop aims to serve 350 AI-crafted drinks per hour—freedom of choice, now with extra espresso.

Walmart’s AI tools are reducing busywork and language barriers—helping 1.5 million workers declare independence from inefficiency. 

Finance

• Goldman Sachs announces firmwide launch of AI assistant  Goldman Sachs has launched its generative AI assistant, GS AI Assistant, across the entire firm, marking a major milestone in its decade-long investment in AI. The assistant, already used by thousands of employees, is integrated into daily workflows across divisions including investment banking, asset management, and engineering. It supports a range of capabilities such as summarizing documents, drafting content, translating research, and conducting data analysis. Uniquely, the tool is designed to work securely across multiple LLMs—including OpenAI’s GPT-4o, Google’s Gemini 2.0 suite, Claude 3.7 Sonnet, and open-source models—allowing employees to choose the model that best fits their use case. The assistant first launched last year for the firm’s developers and has since expanded to 10,000 employees. Goldman CIO Marco Argenti framed the launch as a strategic step in modernizing the company’s tech stack and boosting productivity. CEO David Solomon emphasized that broad AI adoption will drive efficiency within Goldman and across the economy, as generative tools become core to how financial firms operate and serve clients.

Healthcare

• Utilization of artificial intelligence in Men’s Health: Opportunities for innovation and quality improvement | International Journal of Impotence Research A new review published in Nature’s International Journal of Impotence Research explores how AI is transforming men’s health, with promising applications across diagnostics, treatment planning, and patient support. In fertility medicine, AI improves the accuracy of analysis by objectively assessing morphology and motility, and even aids in embryo selection. For erectile dysfunction (ED), machine learning models enhance imaging interpretation, assess risk factors, and are being explored in drug discovery. In conditions such as Peyronie’s disease, AI supports 3D modeling for improved surgical planning, while in testosterone deficiency, it refines diagnostic questionnaires. The paper also highlights how chatbots can support men in sensitive areas such as premature ejaculation or vasectomy counseling. Despite the clear benefits, greater diagnostic precision, personalized care, and improved efficiency, the authors stress the need for transparency, regulatory oversight, and ethical safeguards to ensure responsible deployment. Overall, the paper argues that AI has strong potential to elevate care quality in men’s health, but must be integrated with caution and rigorous clinical validation.

• The Role of Artificial Intelligence in Enhancing Precision Medicine for NSCLC  AI is transforming the management of non-small cell lung cancer (NSCLC), the most common and deadly form of lung cancer, by advancing diagnostics, treatment prediction, and personalized care. AI-powered imaging tools can now predict genetic mutations like EGFR and PD-L1 status with high accuracy, potentially reducing delays in treatment by avoiding the need for traditional biopsies. At the histopathology level, deep learning models are matching pathologist performance in identifying tumor subtypes and mutations. Predictive models are also being used to estimate treatment response and survival outcomes with notable precision, helping clinicians tailor therapies to each patient. AI-driven decision support tools now guide real-time surgical and radiation planning, improving outcomes and reducing complications. Despite these gains, challenges remain in data quality, interpretability, and ethical oversight. Still, the integration of AI into NSCLC care signals a shift toward truly personalized oncology, with physicians and algorithms working in tandem to optimize outcomes.

• LLMs factor in unrelated information when recommending medical treatments | MIT News  An MIT study published on June 23, 2025, reveals that large language models (LLMs) used for medical treatment recommendations are significantly influenced by nonclinical factors such as typos, extra white space, informal language, and missing gender cues. These subtle variations, which often reflect how vulnerable populations communicate, led to a 7–9% increase in erroneous self-care recommendations, even when the clinical information remained unchanged. Female patients were disproportionately affected, even when gender cues were removed. The study—“The Medium is the Message: How Non-Clinical Information Shapes Clinical Decisions in LLMs”—raises serious concerns about fairness, generalizability, and safety in using LLMs in healthcare and underscores the need for rigorous audits before deployment. 

• New Large Language Model Helps Patients Understand Their Radiology Reports | Stanford HAI  Stanford researchers have developed “RadGPT,” a large language model designed to translate radiology reports into clear, patient-friendly language. Instead of dense medical jargon, RadGPT offers simple explanations and suggests common follow-up questions patients might ask. The model doesn’t interpret raw scans—instead, it processes radiologist-dictated reports, extracting key concepts and generating safe, comprehensible summaries. Early review by radiologists indicates a low risk of hallucinations, and the tool could help patients engage more actively in their care while reducing cognitive load on radiologists. The study was published in the Journal of the American College of Radiology.

• Clinical knowledge in LLMs does not translate to human interactions  This study doesn’t expose a failure of large language models—it exposes our failure to test them like real-world tools. We’ve been evaluating AI systems using academic exams and benchmark datasets, then acting surprised when the real-world results fall apart. What Oxford uncovered is simple: in deployment, humans are the variable. They omit key info, misread outputs, and struggle with trust. Blaming users for “bad prompting” misses the point. When you put a system between a worried patient and critical care decisions, design, training data, and interaction quality become non-negotiable. In enterprise too, a chatbot passing a helpdesk test is meaningless if it fails under vague, emotional, or incomplete inputs. Testing AI with other AI doesn’t solve this either—what works model-to-model doesn’t translate model-to-human. The lesson here isn’t that LLMs are flawed—it’s that we need human-centered benchmarks that match messy reality, not sanitized labs.

Retail

• AI-powered coffee shop planned for Cheshire A £12 million development project proposed for Northwich, Cheshire, includes plans for one of the UK’s first AI-powered drive-through coffee shops. The site, submitted by WUKPG and featuring startup Bocca Felice, would combine a five-storey self-storage facility, co-working spaces, and an AI-driven café. Bocca Felice plans to open 22 stores across the UK in 2025, aiming for 80 by the end of 2026. The AI coffee shop will use natural language processing to take orders and operate coffee machines capable of producing up to 350 drinks per hour—roughly 20 times the output of a standard machine—with a claimed accuracy rate of 98 percent. The project is expected to create five jobs (three full-time, two part-time) and includes both indoor and outdoor seating. The facility will operate under the Shutter Store brand and offer flexible meeting spaces adjacent to its main reception. The proposal is under review by the local council (reference 25/00889/FUL).

• Walmart Unveils New AI-Powered Tools To Empower 1.5 Million Associates  Walmart is rolling out a suite of AI-powered tools to over 1.5 million U.S. associates, aiming to improve efficiency, service, and job satisfaction. Key innovations include an AI-driven task manager that has cut shift planning time from 90 to 30 minutes, a real-time translation feature supporting 44 languages with Walmart-specific context, and a revamped GenAI assistant that simplifies complex process guides into actionable steps. These tools are integrated into the Walmart associate app and backed by the company’s proprietary Element platform. Additional tech like RFID and AR (via VizPick) is enhancing inventory accuracy in apparel. With over 3 million daily queries already handled by its conversational AI, Walmart’s investment reflects a broader strategy to blend human strengths with scalable AI, improving both customer experience and employee engagement.

Learning Center

Nvidia’s rise from near-collapse to AI dominance mirrors the American spirit of reinvention and resilience. Inference is the engine behind AI’s independence—freeing models to act, not just react. Kimi-Researcher shows what AI can do when trained to think for itself, not follow instructions. Google’s Learn About tool puts the power of exploration back in the hands of curious citizens.

Learning 

• Nvidia: How the chipmaker evolved from a gaming startup to an AI giant  Nvidia’s rise from near-bankruptcy to becoming the world’s most valuable company in 2024 at a $3.6 trillion market cap is a story of strategic pivots, long-term R&D investment, and early bets on parallel processing. Founded in 1993 with a focus on 3D gaming graphics, Nvidia nearly collapsed in its early years before recovering with the launch of its successful RIVA 128 GPU. In 1999, it introduced the GeForce 256, branding it the first "GPU" and catalyzing its dominance in graphics. The real shift came with CUDA in 2006, which enabled general-purpose GPU computing, initially slow to gain traction, but later critical to training neural networks like AlexNet in 2012. This milestone triggered Nvidia’s pivot to AI, culminating in the 2016 launch of the DGX-1 and explosive demand from firms like OpenAI. As AI took off, Nvidia dominated the GPU market with its H100 and Blackwell chips, with clients including Microsoft, Google, and Amazon. However, challenges have emerged: lawsuits over GPU transparency, crypto-related sales volatility, and U.S. export bans that could cost up to $8 billion in lost revenue due to restrictions on China-bound chips. Despite setbacks, such as a $589 billion stock drop after Chinese competitor DeepSeek debuted a low-cost model, Nvidia continues to lead AI infrastructure with the release of the Blackwell Ultra in 2025. It is now central to U.S. AI initiatives like Project Stargate, backed by a $500 billion investment. Analysts believe Nvidia is still on track to become the first $4 trillion company, credited with enabling the current era of generative AI. 

• Ask a Techspert: What is inference?  Inference is the process that allows AI models to generate outputs based on learned data patterns, making them useful in real-world applications. In Google's latest update, experts explain that inference isn’t new—it’s been used for years in models like YouTube recommendations and image classifiers—but it’s now more powerful due to advances in model architecture and hardware, like the new Ironwood TPU. Inference allows generative AI to translate languages, generate images, and provide proactive, contextually relevant answers. It’s also central to agentic AI, which doesn’t just respond but takes action. Google is now optimizing inference for efficiency and affordability by improving software compilers and hardware design, making smaller, cost-effective versions of models like Gemini possible. As inference improves, it’s transforming AI from a reactive tool into a proactive assistant, all while lowering its compute cost and broadening access.

Tools and Resources

• Kimi-Researcher: End-to-End RL Training for Emerging Agentic Capabilities Moonshot AI’s new autonomous agent, Kimi-Researcher, showcases state-of-the-art performance in multi-step search, reasoning, and tool use, enabled by end-to-end agentic reinforcement learning (RL). Trained entirely via RL—not human demonstration—Kimi-Researcher achieved a Pass@1 score of 26.9% on Humanity’s Last Exam (up from 8.6%) and 69% on xbench-DeepSearch, outperforming models like o3-mini. Its architecture integrates search, browsing, and coding tools, allowing it to handle complex, real-world tasks like legal research and scientific analysis. Unlike modular agent systems, Kimi-Researcher learns planning, tool use, and perception holistically. Moonshot plans to open-source the pretrained and RL-tuned models, furthering research into dynamic, adaptive agents that evolve with changing tools and environments.

•  How to turn AI into your own research assistant with this free Google tool | ZDNET Google’s Learn About is a free AI tool from Google Labs that transforms search into an interactive research experience. Instead of quick, canned answers, it guides users through deep topic exploration using prompts, PDFs, or images. The AI responds with detailed, visually organized content including interactive lists, questions, summaries, and source citations. Users can ask for simplifications or deeper dives, with each session stored for later review. Whether analyzing a photo of Hiroshima’s Atomic Bomb Dome or uploading a business report, Learn About positions itself as a virtual research assistant built for exploration, not shortcuts. 

If you enjoyed this newsletter, please comment and share. If you would like to discuss a partnership or invite me to speak at your company or event, please DM me.