Get tomorrow's brief in your inbox
Today: Reliance Industries is embedding AI assistants directly into phone calls for 500+ million users in India. Subquadratic claims it solved the math bottleneck holding back LLMs, with third-party tests showing 12x longer context windows at a fraction of the cost. Qualcomm's CEO says AI agents will replace apps as we know them.
Reliance Industries is building AI into India's phone network, and the distribution advantage is enormous. The Indian conglomerate announced Jio Call Agent, an AI assistant that can join your phone calls to transcribe conversations, book cabs, order food, and make reservations when you say "Hey Jio." Instead of offering a standalone app, they're embedding it directly into the telecom network for 500+ million Jio users.
This is a fundamentally different distribution model than what Western AI companies are doing. You don't need to download anything or remember to open an app. The AI is native to the phone call itself. That removes all friction and could make AI assistance the default experience instead of an opt-in feature.
For small business owners, this points to where AI assistants are headed: ambient, always-on, and baked into the tools you already use instead of requiring you to switch contexts. If you're building AI tools or considering which platforms to bet on, watch for this shift from apps to infrastructure. The companies that win may not be the ones with the best standalone products, but the ones that make AI invisible by building it into the layer below. Source
AI agents are replacing apps, according to Qualcomm's CEO. Cristiano Amon laid out a vision where you wear smart glasses with a camera and just talk to an AI agent that coordinates across multiple apps to complete tasks like booking a restaurant. Instead of opening a booking app, a payment app, and checking email, the agent handles all three. Qualcomm is working on over 40 different AI gadget designs including jewelry, earbuds with cameras, pins, and watches. Amon says smart glasses could eventually be as big as smartphones (1.2 billion shipped last year). For now, this is future vision, not a product you can buy today, but it signals where device makers think this is going. If you run a service business with an app, start thinking about how your business works in an agent-first world where customers don't open your app directly. Source
Subquadratic says it cracked the LLM speed bottleneck, and independent tests back it up. The Miami startup came out of stealth last month claiming it solved a mathematical problem that's been holding back large language models for a decade. Skeptics called it "AI Theranos." But now third-party testing firm Appen has published results showing Subquadratic's SubQ model can process 12x more text at once than most models, runs faster and cheaper, and uses significantly less energy while matching the performance of Google, OpenAI, and Anthropic models on key tasks like coding. The catch: SubQ isn't widely available yet. If these claims hold up, this could mean dramatically lower AI costs for data-heavy tasks like analyzing entire code bases or hundreds of documents. Worth watching, but don't make purchasing decisions until broader access allows real-world validation. Source
Smartbird (formerly Allbirds) is now an AI infrastructure company with a new CEO and zero employees. The shoe company sold its footwear business for $43 million, raised $100 million, and hired former AWS exec Nadia Carlsten to build an AI compute service. The pitch: managed single-tenant AI infrastructure for companies that need data sovereignty or specialized model control. Think pharmaceutical companies, energy firms, and financial services that can't use public cloud for regulatory or business model reasons. Carlsten says customers will get hundreds to thousands of chips, not massive GPU clusters, optimized for agility and control instead of price. First deployments expected by year end. This is a real business model (Hewlett Packard and Equinix offer similar services), but it's not clear if it has the growth potential investors expect. File under "wait and see." Source
AI receptionists pay for themselves fast. If you run a service business that relies on phone bookings, AI receptionists can catch missed calls, answer questions, and free up staff to focus on customers who are already there. Inc. reports these tools are hitting an ROI inflection point where they're cheap enough to justify for small businesses. The math: if you're missing even a few bookings per week because nobody picks up the phone, an AI receptionist that costs $50 to $200/month pays for itself immediately. Look for tools that integrate with your existing booking system and can handle your specific industry's questions (salon vs. dental office vs. HVAC). Source
A founder tested AI-generated creative content and got a reality check. The numbers showed AI can generate content volume, but it still struggles to replicate human creativity that actually converts. Marketing's secret weapon remains human. If you're using AI for content, the pattern that works: AI for speed and scale on commodity content (product descriptions, FAQ answers, basic blog posts), humans for creative work that differentiates your brand (hero campaigns, video concepts, core messaging). Don't expect AI to replace creative strategy, but do use it to free up your team from repetitive tasks so they can focus on the work that actually drives revenue. Source
The US government forced Anthropic to pull its two newest models citing national security concerns. Fable 5 and Mythos 5 were yanked after Amazon researchers allegedly found a way to bypass Fable 5's guardrails. Cybersecurity researchers have since signed an open letter calling the move dangerous, and Anthropic noted the same jailbreaks exist in other models. This creates uncertainty for developers building on Anthropic's platform. If you're using Claude models in production, this is a reminder to have a fallback plan and not build your business on a single model provider. The episode also highlights the messy relationship between AI labs and government regulators. Expect more of this as AI capabilities advance. Source
MCP's real value is isolating auth outside the agent's context window. Developer Sean Lynch pointed out that the Model Context Protocol's most valuable capability isn't the tool ecosystem, but separating authentication flows from the AI agent's context. This prevents credentials from being exposed to the model or the harness running it. If you're building AI tools that need API access, this is the pattern to follow: keep secrets out of the context window entirely. Security-conscious businesses should look for tools that implement this separation by design. Source
Use AI to analyze customer support patterns, not replace support staff. If you're getting the same customer questions repeatedly, feed your support tickets into an AI model and ask it to cluster common issues and suggest knowledge base articles or process improvements. This is lower risk than deploying an AI chatbot and often reveals problems in your product or onboarding flow. Start with a simple export from your support system, paste into Claude or ChatGPT, and ask: "What are the top 5 recurring issues and what would prevent them?"
The shift from AI apps to AI agents is accelerating. Reliance's phone network integration and Qualcomm's vision both point to the same future: AI that's embedded in infrastructure instead of living in standalone apps. For businesses, this means the tools you build today need to work with agents tomorrow. The companies winning on AI infrastructure aren't necessarily the ones with the flashiest demos, but the ones solving unglamorous problems like context window limitations (Subquadratic) and data sovereignty (Smartbird). Watch for consolidation as businesses realize they can't afford to maintain integrations with 50 different AI tools and start demanding unified platforms.