Date: {{current_date_full_with_day}}

Hey {{first_name | AI Visionaries}},

In this 107th edition, we look at AI getting more practical across the stack with faster long-running agents and AI that helps build AI ay Anthropic.

Plus we have a new guest writer Venkitesh, who shares his experience about building AI systems at the edge. I am excited to feature him as we writes about an oft neglected field of AI in manufacturing.

The theme is simple: AI is moving from single-purpose tools into work systems businesses can actually use.

As usual, please send in your feedback and suggestions. It helps make the newsletter sharper for you.

-Renjit

PS: If you want to unleash the power of AI agents for personal and business productivity, you can setup time speak to me here »

Set up time to discuss:

NVIDIA released Nemotron 3 Ultra, an open model built for the kind of long-running agents that do more than answer one prompt. The model has 550 billion parameters, with 55 billion active at a time, and is aimed at workflows where agents plan, call tools, check results, delegate work, and keep going across many turns.

The useful part for founders is the cost and speed story. NVIDIA says Nemotron 3 Ultra can deliver up to 5x higher throughput than comparable open models in its class, and can lower cost to task completion by up to 30% on some agentic benchmarks. The model uses hybrid Mamba-Transformer layers, NVFP4 precision, LatentMoE routing, and multi-token prediction to keep long-context work moving.

SO WHAT?

NVIDIA also released recipes, weights, and training details, which matters for teams that need domain control instead of a closed black box. It could be a good starting model for businesses and startups who want to build out specialized use cases.

Read the source:

Guest Writer: Venkitesh Ramadas

The other day I sat in on an employee training session run by a customer, part of a companywide rollout of a popular enterprise AI assistant.

The trainer was teaching prompt engineering, showing the team why they need to pack in far more detail and context to get good answers. He literally demonstrated how a natural question looks, one simple sentence, against how an AI prompt is supposed to look, a block with lot more context.

Enterprises are buying these licenses by the thousands, yet we rarely see the real business impact. It reminded me of a report showing only 30 to 50 percent active AI assistant adoption at enterprises.

Manufacturing is a sector we care about, and our hypothesis is that adoption could be even lower here. The sector runs on a large frontline workforce, the everyday men and women working hard on factory floors, in warehouses, in logistics hubs and across supplier hubs to ship physical things to us.

So here is the question. Do we want our quality engineer to be a world-class quality innovator, or a prompt engineer who crafts perfect prompts just so the AI understands that the yield issue he is describing is First Pass Yield in production, not the bond yield a finance person might mean? The example might be extreme, but the point stands.

Context is everything. Expecting users to supply it every single time will not drive adoption. Context matters, but understanding intent is what puts it to work and makes AI truly magical.

We need intent-driven AI that makes interactions feel natural, where you barely notice AI is involved. It should feel like working with an expert who already understands your context, your priorities, your goals, and the cross-functional nature of solving real problems.

AI has to blend into the everyday life of an organization and its people, making work effortless, where no one sees the AI at all.

It reminds me of a line from Mark Weiser of Xerox PARC - "The most profound technologies are those that disappear. They weave themselves into the fabric of everyday life until they are indistinguishable from it."

For AI to create true organization-wide impact, including the frontline, we need intent-driven, context-aware, everyday AI.

———-

Anthropic published a clear warning and progress report on recursive self-improvement: the point where AI systems help design and build their own successors. The company says we are not there yet, and that the outcome is not inevitable, but its internal data shows the loop is already starting to tighten.

At Anthropic, engineers are now shipping far more code than they did in the 2021 to 2025 period, largely because Claude is doing more of the implementation work. The company says that as of May 2026, more than 80% of code merged into its codebase was authored by Claude. Claude is also being used to run experiments, review code, catch defects, and handle more open-ended engineering work.

The remaining gap is judgment. Humans still choose which problems matter, which results to trust, and when to stop. But the cost of doing the work is falling fast.

SO WHAT?

The bottleneck is shifting from execution to judgment, so founders need better taste, review systems, and governance, not just more automation. My skeptical take is that Anthropic is following the “regulatory capture” playbook, where you raise the regulatory hurdles for your competitor AFTER you are at the front of the AI race.

Read the source:

Generalist AI announced $400 million in new funding, bringing its total raised to more than half a billion dollars. The company is building what it calls physical AGI: intelligence that can understand and act in the real world, across factories, warehouses, labs, farms, homes, restaurants, and even space. This is our second article touching upon the “real” world after the article by Venkitesh.

The funding round includes Radical Ventures as lead investor, plus 8VC, Union Square Ventures, Norwest, NVIDIA, Spark Capital, Bezos Expeditions, and angel investors including Fei-Fei Li and Naval Ravikant. That is pretty much most of the big names in AI funding!

Generalist says its earlier GEN-0 work showed scaling laws in robotics, while GEN-1 moved closer to commercial use with 99% reliability on diverse tasks, execution up to 3x faster than prior state of the art, and the ability to learn complex physical skills.

The company plans to use the money to scale robot learning, data, compute, infrastructure, and deployment partnerships.

SO WHAT?

Robotics is starting to look like frontier AI in 2023: capital, data, model scale, and deployment are beginning to compound together.

Read the source:

9to5Mac previewed five iOS 27 features ahead of Apple’s June 8 WWDC reveal, and the list shows where consumer AI may become more practical. The biggest item is a rumored Siri overhaul, with a Google Gemini-based foundation and upgrades such as a chatbot-like interface, Dynamic Island integration, better world knowledge, multi-action requests, and deeper personal context.

The other features are smaller but useful. iOS 27 is rumored to add nutrition-label scanning into Health, more nutrition features, Liquid Glass refinements, and a stronger Add to Notes action that lets Siri save useful information into Apple Notes. I will be very happy with a more intelligent Siri! One confirmed feature is generated subtitles for videos, even when captions are not already available.

For users, these may feel like conveniences. For builders, they show Apple pushing AI into daily phone behavior instead of keeping it inside a separate chatbot app.

SO WHAT?

If Apple turns AI into default phone behavior, startups should expect users to demand smarter, context-aware experiences everywhere.

Read the source:

OpenAI announced new Codex features aimed at making the tool useful across more roles, not only software teams. More than 5 million people now use Codex every week, and OpenAI says non-developers already make up about 20% of overall Codex users, growing more than 3x as fast as developers.

The update introduces role-specific plugins, annotations, and a preview of shareable interactive sites and apps. The six new plugins cover data analytics, creative production, sales, product design, public equity investing, and investment banking. Reminds me of the approach that Anthropic took by launching skills and plugins in Claude Co-work. Makes it more useful for real business use cases.

Together, they include 62 popular apps and 110 skills, connecting Codex to tools like Snowflake, Tableau, Figma, Canva, Salesforce, HubSpot, PitchBook, and others.

OpenAI also shared examples from its own teams, Zapier, and NVIDIA, where Codex is being used to build internal apps, prepare materials, create dashboards, turn tool context into plans, and speed research workflows.

SO WHAT?

The next phase of AI work may be less about coding assistants and more about role-specific operating layers for every business function.

Read the source:

Reve launched Reve 2.0, its newest image model, with a pitch that goes beyond prettier pictures. The company says the model generates native 4K images and is built around precise layouts, so users can shape and edit parts of an image instead of repeatedly rewriting prompts and hoping the next output lands closer.

The official launch post says Reve 2.0 debuted at number two on Arena’s text-to-image leaderboard, jumped 125 Elo over Reve 1.5, and trained on 10x fewer GPUs than some competitors. The bigger product idea is layout control. Every image is built from a structured map of objects, text, and regions, which means a user can move a subject, rewrite a sign, swap a background, or stack more reference images into one composition.

Reve also rebuilt its editor so object controls appear directly on the canvas and results stream as the model works.

SO WHAT?

Image AI is moving from prompt lottery to controllable creative software, which is better for teams that need brand consistency and repeatable production.

See the source on X:

You currently have X referrals.

Click to Share