Welcome to Issue #58

Coming Up:

Zuck Bucks: Can Meta Buy an AI Breakthrough?

MCP and the End of App Lock-In

From Prompts to Context: A Better Way to Work With Agents

Plus, we discuss simulating evolution with AI, the human role of debugging, the role of AI in the church sermon, and all the news we found interesting this week.

It’s Sunday morning!

We are officially closer to 2050 than the year 2000 and AI is just finishing its warm up stretches. The next 25 years should be real interesting.

I’m sure we will cover it all in issue #1347.

But until then, let’s get into issue #58.

The DAS Crew - Andy, Beth, Brian, Eran, Jyunmi, and Karl

Zuck Bucks: Can Meta Buy an AI Breakthrough?

Meta has launched a bold new talent raid, with Mark Zuckerberg personally reaching out to top AI researchers and reportedly offering packages reaching into nine figures. Dubbed “Zuck Bucks,” these offers have already lured key names away from OpenAI and other labs. It is sparking debate about whether you can buy an AI future or just disrupt your competitors along the way.

The timing is strategic. Meta’s LLaMA 4 model has not performed on par with its level of training investment and compute budget. Others like Gemini 2.5 Pro and OpenAI’s GPT-4 series continue to dominate the narrative. With competition heating up, and new models spawning from the leaders, Zuckerberg is pushing Meta’s AI ambitions into high gear, shifting from open academic vibes to a closed, high-stakes race for leadership, guided by battle-proven talent.

The stakes are bigger than training the next large language model. Meta’s core business is not immediately threatened by AI, giving it room to invest aggressively while others navigate existential risks. Meanwhile, these high-dollar offers reset the market, raising the price of top talent and making it harder for smaller startups to compete.

But there is tension. Meta’s top AI scientist, Yann LeCun, has publicly criticized the current LLM arms race, calling it a “dead end.” Meta’s culture, known for speed and centralized control, may not easily fit the mission-driven researchers it’s hiring. And it’s unclear what the endgame looks like: Is it about superintelligence? Is it about making better AR glasses and social products? Or is it about making sure Meta doesn’t get left behind?

WHY IT MATTERS

The Talent Market Just Changed: If even a fraction of the reported sums are true, the bar for attracting top AI talent just got higher, squeezing smaller players.

Product vs. Mission: Meta’s business is not threatened by AI, allowing it to play long. But researchers motivated by safety and alignment may hesitate to become mercenaries with a company whose history raises trust questions.

LLMs Are Not the Only Game: Meta’s AR, VR, and social media data ambitions intersect with AI. Buying talent could mean preparing for a future where assistants and embodied AI are central to its broader businesses.

Culture Still Matters: Talent is only part of the puzzle. Meta will need to build the right operating culture to overcome deficiencies evidenced by its development pipeline, and transform those big salaries into breakthrough products.

This Is a Power Play: Beyond technology, the Zuck Bucks strategy is about weakening competitors, raising the floor for talent costs, and shaping the landscape in Meta’s favor.

MCP and the End of App Lock-In

If you’ve been following the Agentic web closely or listening to our show, you already know MCP servers let agents plug into any app, tool, or legacy system. The real story now is what happens next.

SaaS companies will face hard choices. If your platform doesn’t expose clean, agent-ready MCP endpoints, your subscribers’ agents will route around you. But the deeper disruption hits your business model. Once MCP normalizes agent-to-app connections, the moat shifts from the SaaS tool to the data store within and the customer relationship. That’s why leaders like Satya Nadella say SaaS must adapt or risk becoming an overpriced backend in an agent-driven world.

For enterprises, MCP unlocks a path to building lightweight, agentic workflows without brittle integrations or endless Zapier chains. Instead of maintaining fragile RPA bots or custom API scripts, you give your AI agents goals, and they figure out the requisite tools and use MCP to manipulate them. Quarterly report generation, revenue pipeline prep, meeting coordination, or pulling cross-platform marketing analytics, MCP makes these composable, agent-driven, and much cheaper to maintain.

What could be a real wild card is vibe coding. If the heavy lifting of app-building is wiring together systems, MCP wipes that out. Suddenly, small teams and even solo operators can stitch together mini-apps or task-specific agents that pull exactly what they need from systems like Salesforce, Asana, or legacy ERPs without a dev team or a venture-scale SaaS startup. You don’t need to build a SaaS business, you just need a micro-app that solves your exact workflow. The business case for many SaaS platforms starts to wobble under that reality.

MCP may also fuel a secondary marketplace that is a “Spotify for mini-apps”, a place where businesses can grab lightweight MCP-powered templates, vibe-code their functions for internal needs, and move on, without locking into a full vendor. It’s a path toward radical AI customizations where what once was “software eats the world” flips to become “every business eats just the software it needs.”

WHY IT MATTERS

SaaS is on Notice: MCP forces SaaS platforms to rethink how they provide value. If agents can access any micro-service or feature through a universal handshake, your pricing and stickiness come down to data and outcomes, not gated features.

Agents Will Handle the Tedious Work: Complex manual reporting, syncing integrations between systems, and repetitive operational tasks will increasingly be handled by AI agents with MCP, freeing teams to focus on higher-leverage work.

Vibe Coding Goes Enterprise: MCP makes it viable for teams to build micro-tools for internal needs without becoming a software company, shifting power back into the hands of operators.

The Data Moat Becomes Everything: When every app is a plugin, aggregating and protecting a private, proprietary datastore becomes the primary competitive defense.

Legacy Systems Don’t Need to Die: MCP wrappers can be used to give ancient ERP and obsolescent niche tools a new lease on life, allowing agents to extract value from hard-to-abandon legacy platforms without ripping them out and replacing the software.

From Prompts to Context: A Better Way to Work With Agents

Context engineering is quickly becoming the framework that separates people who get results using Agentic AI from those who are mired in the chatbot paradigm where non-reasoning inference runs require extensive specification and instructions. It moves beyond prompt engineering, which tried to stuff everything into a single request, and shifts toward designing the resource environment in which a self-directed multi-turn agent can reason, plan, and act.

When using the prompt engineering approach, you give the model instructions and fine those instructions until it returns what you need. When you use context engineering, you give the reasoning agent an objective, set the right boundaries, provide the right data, and let the agent decide the steps it needs to take. The agent pulls in tools, searches for new data, and adapts as it learns, using memory to improve results across steps.

Agentic pioneers are already using this approach to request market research, generate competitor analyses, prepare sales reports, draft content, compose presentations while demanding agents evaluate gaps and provide redirection of sub-agent outputs along the way. You are no longer writing fifty prompts to complete fifty tasks. You set the goal, and the orchestrating agent figures out the workflow.

Context engineering also means learning to provide just enough information at the right time. Overloading an agent with too much data slows it down and can lead it in the wrong direction. The best context engineers learn to prune data, manage memory, and decide what the agent needs next.

This changes how teams build and operate with AI. It turns AI into a teammate that can take on real projects, using your private data and tools to create outputs that are specific to your business. It also means you spend less time tweaking instructions and more time setting goals and evaluating final outputs.

WHY IT MATTERS

Context Makes Agents Useful: Agents can handle complex tasks across multiple steps without manual handholding, moving work forward while you oversee results. Duration of unsupervised task-execution time becomes a measure of agents’ ability and usefulness.

Your Data and Becomes Power: The same public model in the hands of others without proprietary data produces general results, but layered with your context, workflows and data that LLM-powered system becomes a competitive advantage.

Workflow Friction Drops: You do not need to build brittle scripts or chains of automations. You design goals and environments, and the agent handles the process.

Learning This Skill Pays Off: Context engineering is quickly becoming a core competency for teams that want to get practical value from advanced AI now rather than managing trials of models endlessly and waiting to find one that comes replete with reasoning and resources.

Just Jokes

Did you know?

Scientists at EvolutionaryScale have used AI to simulate 500 million years of evolution, creating a new fluorescent protein called esmGFP. This protein, designed by the AI model ESM3, does not exist in nature and was generated by mimicking evolutionary processes over half a billion simulated years.

The AI model was trained on a dataset of 770 billion protein sequences and functions similarly to a language model, predicting protein sequences and structures. The resulting protein shares 58% sequence similarity with its closest known counterpart from the bubble-tip sea anemone and exhibits fluorescence when synthesized and tested. This advancement demonstrates AI's potential in protein engineering, with applications ranging from medicine to environmental science.

Source

This Week’s Conundrum
A difficult problem or question that doesn't have a clear or easy solution.

The AI Sermon Authenticity Conundrum

A Finnish church recently let a language model write and deliver its midweek sermon. Worshippers listened. Some called it impressive. Others, cold. The words were right, the delivery smooth, but the weight behind both felt thin. Machines can gather centuries of scripture, weave compelling stories, and tailor messages to every fear and hope. But they cannot ache for the grieving or tremble with the guilty. They cannot weep with the brokenhearted or share the quiet terror of doubt.

Every sermon carries invisible weight. The preacher brings their own wounds, their own late-night prayers, their own fragile faith into the pulpit. Their words are not just doctrine. They are offering themselves. Even their failures carry grace. An AI sermon never flinches, never struggles, never costs the speaker anything.

The congregation may still find comfort. The message may still heal. But when every word costs nothing, how long before the sacred feels mechanical? When the preacher’s voice becomes an efficient simulation, does the community lose something essential, or simply adjust to a new kind of presence that no longer asks anyone to risk their soul?

The conundrum
If AI sermons soothe pain and strengthen faith, does comfort alone define sacredness? When the pulpit requires no vulnerability, no personal stake, no shared humanity, do we gain a purer message or lose the very thing that made the act holy?

Want to go deeper on this conundrum?
Listen/watch our AI hosted episode

Amazon Hits One Million Robots, Releases Warehouse AI Model
Amazon has now deployed over one million warehouse robots, each managed by a new custom-built AI model for logistics, inventory, and warehouse management. Notably, Amazon has released this model for others to use, not just internally.

Deeper Insight:
Amazon’s move signals the shift from proprietary robotics to open logistics AI, setting a new benchmark for automation at scale. By sharing its warehouse model, Amazon positions itself not only as a retailer but as a supplier of core infrastructure for global supply chains.

Apple Shifts AI Strategy, Courts OpenAI and Anthropic
Apple is moving away from in-house or acquisition-based model development, instead entering advanced talks to license models from OpenAI and Anthropic for a new, privacy-centric Siri. The goal is to run these models securely on Apple’s own hardware and data centers.

Deeper Insight:
Apple’s pivot reflects how even tech giants need to buy, not build, as the AI field accelerates. By doubling down on privacy and hardware integration, Apple aims to differentiate Siri in a market soon to be flooded with general-purpose AI assistants.

U.S. Senate Removes AI Moratorium From Federal Budget Bill
The U.S. Senate voted 99-1 to strip a ten-year federal moratorium on state-level AI regulation from the budget bill, allowing states like California, Colorado, Utah, and New York to pursue their own rules on deepfakes, data privacy, and automated decision-making.

Deeper Insight:
This keeps the “AI alignment laboratory” open for states’ use, letting them experiment with protections and oversight that may move faster than federal rules. Expect a patchwork of regulations and legal uncertainty for companies deploying AI across multiple states.

Denmark Proposes Copyright Protection for Your Likeness
Denmark has proposed a law granting individuals copyright over their face, body, and biometric data. This would let people legally challenge deepfakes or unauthorized uses of their likeness as a property right, not just a privacy concern.

Deeper Insight:
If adopted, this would set a new global benchmark for digital rights. It treats your digital identity as something you own. Similar protections could spread to other countries, especially as AI-generated media blurs the line between real and fake.

Meta AI Wants Access to All Your Photos
Meta recently began prompting users for access to the entire camera roll on their devices, with unclear terms about how often and for how long Meta can access and use those images. The move has sparked privacy concerns.

Deeper Insight:
Data access creep remains a major issue in consumer AI. As companies seek ever more training data, transparency and control over personal content are set to become major battlegrounds for regulators and tech giants.

Cloudflare Launches Pay-Per-Crawl to Monetize Data Scraping
Cloudflare is rolling out a pay-per-crawl system that allows website owners to charge AI agents and bots for access to their content. News sites and stock photo libraries are early adopters, with Cloudflare acting as a gatekeeper.

Deeper Insight:
This is a direct response to AI’s voracious appetite for web data. If widely adopted, pay-per-crawl could force AI companies to strike licensing deals for high-value content, and shift the economics of how the web is indexed, searched, and used for model training.

Spotify’s Velvet Sundown Mystery: AI Band or Marketing Stunt?
A psychedelic rock band called Velvet Sundown racked up over 550,000 Spotify listeners in just two weeks, despite lacking real members, social media, or live shows. Their photos appear AI-generated, and speculation is growing that this is a test of how AI music can exploit streaming algorithms and generate passive income.

Deeper Insight:
As AI-generated “slop” content floods platforms, music and video services will need new ways to identify, manage, or even ban inauthentic creators. The line between creative innovation and manipulation is getting blurrier.

OpenAI’s $10 Million-Plus Enterprise Consulting Arm
OpenAI has launched a consulting service for large enterprises, charging at least $10 million per engagement to customize AI models and develop AI applications on site. The company has recruited former Palantir engineers to deliver these hands-on digital transformations.

Deeper Insight:
OpenAI is expanding beyond API sales into high-margin enterprise consulting, directly competing with the likes of Accenture and Palantir. This signals the rise of “AI transformation” as a service and raises the bar for what big organizations expect from AI partners.

SongScription: Shazam for Sheet Music
SongScription is a new tool that transcribes any audio, from a YouTube link to a hummed melody, into piano sheet music. It offers a “whisper for sheet music” experience and could become an essential resource for students, educators, and musicians.

Deeper Insight:
Tools like SongScription break down barriers for music learners and creators, making musical knowledge more accessible. The technology also points toward a future of AI “music agents” that can analyze, transcribe, and even create scores in seconds.

Cursor Web App Adds Background Coding Agents
Cursor has launched a web app to let users orchestrate and manage multiple background AI coding agents. This feature is part of a trend toward more multi-agent workflows in coding tools.

Deeper Insight:
As AI-powered development environments become more sophisticated, the ability to coordinate multiple agents working on different tasks in parallel is turning into a must-have feature for serious coders and teams.

Grammarly Acquires Superhuman to Build AI Productivity Suite
Grammarly acquired the AI-powered email platform Superhuman, aiming to create a multi-agent productivity platform focused on the email Inbox.

Deeper Insight:
Productivity apps are rapidly evolving from single-use tools to AI-powered platforms that handle writing, scheduling, and communication. As Google and Microsoft ramp up competition, user loyalty may hinge on which app offers the best personalized workflow assistants.

Sakana AI Releases Adaptive Branching Monte Carlo Tree Search (ABMCTS)
Sakana AI introduced ABMCTS, a new inference-time scaling algorithm that combines the outputs of multiple models to create better ensemble results. The code and paper are open source.

Deeper Insight:
By orchestrating different models like a foreman assigning tasks, ABMCTS could become a building block for next-generation “collective intelligence” systems, especially as organizations look for ways to mix and match the best models for different challenges.

Google Adds Notebook LM and AI Tools to Education Suite
Google is integrating Notebook LM and other AI-powered tools into Google Classroom and its education suite, aiming to help teachers and students build lesson plans, generate content, and support learning.

Deeper Insight:
AI is becoming an essential part of digital classrooms, not just for students but for educators designing personalized content and support. Google’s move will accelerate adoption and reshape expectations for tech-enabled teaching.

Quantum Computers Beat Classical—Now Unconditionally
Researchers at USC developed methods that allow quantum computers to outperform classical computers exponentially, regardless of conditions. This represents a major leap in closing the “noise gap” that had previously limited real-world applications.

Deeper Insight:
Unconditional quantum advantage could soon move from theory to practical breakthroughs, pushing quantum computing closer to the mainstream. The timeline for widespread impact is still a few years out, but the gap between science fiction and reality just shrank.