1. [AINews] Jeff, Sanjay, Oriol, and Quoc depart DeepMind; Demis to Chair; Koray to SVP — what is going on at GDM??? — AINews

  • Why read: Four top researchers leaving Google DeepMind marks a major shift and the start of a new "autoresearch" venture.
  • Summary: Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le left DeepMind on good terms to start Discovery Loop, an autoresearch company. Meanwhile, Demis Hassabis becomes Chair and Chief Scientist to focus on long-term strategy and Isomorphic Labs. Koray Kavukcuoglu takes over as SVP. Like past exits to Anthropic and OpenAI, this move makes people wonder if Google's structure limits basic AI research. But because Google is backing Discovery Loop, the split looks like a friendly spin-off, not a messy breakup.
  • Read more

2. Outlook where frontier AI is headed next 18 months: — Mike Knoop

  • Why read: Explains how top AI labs plan to build better models using automated training loops and verifiers.
  • Summary: Today's AI reasons well in math and coding because we have plenty of data and reasoning traces. Expanding this to other fields is hard because human-driven training loops are slow and costly. To fix this, labs are moving to recursive self-improvement. They use coding agents to create symbolic world models that verify work automatically. This lets AI test ideas safely in offline sandboxes, like git repos, instead of live production. As these systems produce massive amounts of training data, future models will absorb the reasoning directly and drop the need for the external setup.
  • Read more

3. The next generation of MCP — The Cloudflare Blog

  • Why read: The Model Context Protocol (MCP) is now stateless, making it much easier to build and scale AI agents.
  • Summary: The new MCP 2026-07-28 spec drops stateful connections. Before, running an MCP server meant dealing with sticky sessions, open streams, and message replays, which drove up costs and maintenance. Now, MCP servers run cleanly on serverless platforms like Cloudflare Workers without persistent state. The update includes fresh SDKs for TypeScript, Python, Go, and C#. For developers, connecting agents to tools is now simpler. For operators, it means easier deployments, cleaner autoscaling, and less overhead.
  • Read more

4. Knowing When to Stop: The Art of Making a Loop Converge — Yoko

  • Why read: Explains why AI agents get stuck or fail, pointing to verifiers as the missing piece in loop engineering.
  • Summary: Autonomous agents don't know when they are finished. The surrounding system has to tell them they are done. As developers shift from prompting to automated loops, an agent's success relies entirely on its verifier. If a verifier just checks if tests pass, the agent might memorize the test inputs instead of solving the real problem. Good loop engineering ensures every cycle gets the agent closer to the actual goal. Without a strict stopping condition, agents chase the wrong signals and waste compute.
  • Read more

5. Why Agents Started Dreaming (And What is Dreaming)? — mem0

  • Why read: Introduces "dreaming," a background process that cleans up agent memory to stop performance from degrading over time.
  • Summary: Agents that run continuously extract and store facts, eventually creating a messy, contradictory recall system. Because they write memories quickly during chats with a small context window, they often duplicate facts instead of updating them. Developers are fixing this with "dreaming": a background job that cleans the memory store while the agent idles. It merges duplicates, updates old information, and combines related facts into better summaries. This treats agent memory like standard database compaction, keeping the system reliable without retraining the model.
  • Read more

6. I’ve Seen This Movie Before: What AI Builders Can Learn From Data Engineering — Josh Rosen

  • Why read: A reminder that data engineers solved many of the workflow problems AI builders are struggling with today.
  • Summary: LLMs offer new capabilities, but moving information between systems and coordinating work are old problems. Data engineering uses directed acyclic graphs (DAGs) to map and execute dependencies. AI builders should adopt DAGs instead of trusting unpredictable agent loops. AI workflows also look exactly like ETL pipelines: tools extract context, the model transforms it, and the system loads the result. Separating these steps and keeping orchestration predictable lets developers use flexible LLMs while keeping production systems reliable.
  • Read more

7. Scoping and curating eval datasets — Annabell

  • Why read: A practical guide for replacing vibe checks with real evaluation datasets for production AI.
  • Summary: When AI apps leave the prototype phase, developers need actual datasets to measure quality and catch bugs. Start with 15 to 30 rows covering the most common inputs and define a clear grading rubric. As the app grows, split datasets by task: one for end-to-end performance, another for single steps, and a third for edge cases. Pull these examples from production logs, existing FAQs, and expert-written synthetic data. This keeps evaluations organized and ensures the metrics actually tell you what to fix.
  • Read more

8. The MVP is dead. Long live the MVD! — Sajith Pai

  • Why read: Building a product is so easy now that distribution, not software, is the bottleneck for new startups.
  • Summary: AI coding tools make shipping a Minimum Viable Product (MVP) fast and cheap. Because of this, MVPs no longer impress investors. The real hurdle is customer acquisition. Founders now need a "Minimum Visible Distribution" (MVD)—like an active community, a personal audience, or strong partnerships—before raising money. Without past startup exits or deep industry ties, an MVD proves a founder can actually reach users. Building an audience now comes before building the app.
  • Read more

9. The End of No Code — Philip Zeyliger

  • Why read: Argues that LLMs are killing no-code platforms because AI makes standard code accessible to everyone.
  • Summary: Bending Spoons buying Airtable for $1.28 billion signals the end of the no-code boom. No-code tools won originally by letting operators bypass IT and build logic on easy databases. Now, AI agents can write and maintain real code. Instead of getting locked into closed, proprietary platforms, businesses can use AI to build on open standards like Linux. The future of software is standard, boring infrastructure managed by AI, not niche no-code tools.
  • Read more

10. The 10 things I’m telling CEOs about AI — Greg & Taylor

  • Why read: Direct advice for executives on how to fund, frame, and lead AI adoption.
  • Summary: CEOs need to drive AI adoption themselves, not just wait for their teams to prove the ROI. Companies should lower their short-term expectations but raise their long-term ones, as the investments take time to compound. Leaders also need to build AI agents themselves to understand what the tech can and cannot do. Finally, companies are ignoring the human element: employees are afraid AI will take their jobs and are resisting the tools. Getting AI to work requires managing that fear and changing how people work.
  • Read more

11. They Buy Dead SaaS and Make It Profitable (Hint: AI is not a real business) — Sangram

  • Why read: Argues that the market is currently rewarding tight operations and cash flow over AI innovation.
  • Summary: Bending Spoons just went public at a $25.7 billion valuation. Their playbook: buy struggling SaaS companies like Evernote, cut costs, raise prices, and fix the economics. Meanwhile, many innovative AI startups are burning cash with no real revenue. The market is paying for operational discipline, not cool tech. For founders, the takeaway is simple: the business model matters more than the features. AI won't save you if you can't make money.
  • Read more

12. Meta Needs AI but maybe not AGI — Rob Tanner

  • Why read: Questions why Meta is spending billions on Artificial General Intelligence when it doesn't fit their core business.
  • Summary: Meta uses machine learning to run ads and recommend content, but its push for AGI lacks a clear business case. The company is spending $130–145 billion on infrastructure and building enterprise coding tools—a market where it has zero sales experience. The only real logic is defense: Meta wants to own the next computing platform so it doesn't have to rely on Apple or Google. But like its Metaverse gamble, Meta is burning cash on AGI without an obvious way to turn it into advertising revenue.
  • Read more

13. Ownership — Thorsten Ball

  • Why read: Explains what it actually means for an engineer to own a problem end-to-end.
  • Summary: "Owning" a problem means taking it from discovery to the point where no one else has to think about it. It means questioning the root cause, finding edge cases, and mapping the data flow. Writing code and passing tests isn't enough. Owners deploy the fix, manually verify it in production, and communicate the changes to the rest of the team so nothing breaks downstream.
  • Read more

14. The Big Management Lie: Overpromising — staysaasy

  • Why read: Warns managers not to overpromise promotions or bonuses to keep top performers happy.
  • Summary: Managers often promise promotions or raises they can't guarantee just to keep anxious employees from quitting. A promise gives the employee a short burst of happiness. But if budgets freeze or reorgs happen and the reward falls through, the employee's motivation tanks permanently. Staff see broken promises as lies. Being honest about uncertain rewards is much safer than making promises you might not be able to keep.
  • Read more

15. Autonomous Company Infrastructure — naïve

  • Why read: Looks at the new infrastructure being built to let AI agents run entire companies on their own.
  • Summary: Naïve just raised $28.5M to build an operating system for AI agents. Right now, developers spend months connecting payments, hosting, and databases because those tools were built for humans. Naïve uses a single config file and API to spin up a full stack, complete with serverless agents and a shared memory layer. By cutting compute overhead and routing tasks to the cheapest capable models, they want to make autonomous companies financially viable.
  • Read more

Themes from yesterday

  • Agent Infrastructure: The focus is shifting to plumbing. We're seeing stateless MCP, memory cleanup processes, and AI builders adopting standard data engineering tools like DAGs to manage loops.
  • Distribution over Features: The market is rewarding business fundamentals. Startups need an audience before a product, and public markets are paying for cash flow and price hikes over pure tech innovation.
  • Frontier AI Shifts: Big moves at the top. DeepMind lost senior researchers to a new startup, and people are starting to question if Meta's massive spending on AGI actually makes sense.