The Era of Microsoft Openly Competing With OpenAI and Anthropic

On July 29, at its earnings briefing for Wall Street, Microsoft announced a string of its own AI models, harnesses, and competitors to Mythos. While the company continues to call its relationship with OpenAI and Anthropic a "partnership," it made clear that it intends to pursue sustainable growth on its own AI foundation.

Microsoft has long depended on OpenAI's reasoning models, but this announcement put its in-house models front and center. This looks like more than a simple model swap: Microsoft, which controls Azure's massive delivery infrastructure, appears to be moving to cement its own position in the "platform layer" of AI as well. For its part, OpenAI has been pushing toward infrastructure independence with initiatives like Stargate, so the relationship between the two is arguably shifting, step by step, from a "strategic partnership" to a mix of cooperation and competition.

MCP Achieves "Statelessness" in Its Fifth Specification Revision

The Model Context Protocol (MCP) specification was published on July 28 as its fifth revision (MCP 2026-07-28). The heart of this revision is "statelessness."

Historically, MCP was a "connection spec" in which the server and client held session state between them. With this revision, a stateless request/response model has been added, allowing MCP servers to be treated as stateless API endpoints. In other words, MCP has moved up a layer in the architecture — from "connections you open each time" to "plumbing you can leave in place."

The practical impact is significant. Running MCP servers in containerized and serverless environments becomes standardized, making scaling easier. The specification around authentication has also been fleshed out, with production use in mind. As a foundational technology underpinning the AI agent ecosystem, MCP has taken a major leap in maturity with this revision.

AI Agent Approval Gates — What Happens After Rejection Is the Real Challenge

When you hand production operations (deploys, data deletion, refund processing, and so on) to AI agents, the standard approach is a human approval gate — human-in-the-loop. Before a dangerous operation, the agent sends an approval request and waits; once a human approves, it proceeds.

But once you actually implement it, you discover that production reality is not "building the approval UI." What matters is "**designing everything that happens when the request is not approved**."

Concretely, the problem is the agent's behavior after an approval is rejected. Simply stopping stalls the entire task. Does it propose an alternative, skip the step, or defer the decision to a human? The right response also changes depending on whether the rejection was about "timing" or "substance." The realization that an approval gate is not "a technique for stopping things" but "a technique for designing the state transitions after the stop" is a lesson worth holding onto for the many teams pushing agent operations forward.

Banned From Wikidata Overnight — 109 Items Deleted in Two and a Half Minutes

A developer tried to register their own work on Wikidata as a source for LLM citations, only to find one morning that their account had been blocked indefinitely. Citing a "promotion-only account," the platform bulk-deleted the 109 items they had spent six weeks creating — in about two and a half minutes.

They believed each individual edit complied with the rules, but the pattern — "creating a large volume of items, all related to oneself" — apparently matched the platform's spam-detection criteria.

This case carries an important lesson for content strategy in the AI era. The motivation itself — "information listed on Wikipedia or Wikidata is more likely to be cited by LLMs" — is understandable, but mass editing without a deep understanding of each platform's rules and how moderation actually works can lead to irreversible consequences.

DeepSeek v4 flash q2 on Consumer Hardware — Hitting 10 tok/s

Benchmark results have been published for running the q2-quantized build of DeepSeek v4 flash (about 90GB) on a Threadripper 1950x + 128GB RAM + RTX 5060Ti 16GB system.

Without a GPU (CPU only), both prefill and decode ran at around 3 tok/s, but enabling GPU offload improved that to 10–15 tok/s for prefill and 10 tok/s for decode. The finding that a 90GB-class model can run at practical speeds on a high-end but consumer-grade configuration with 128GB of system memory and 16GB of VRAM is encouraging news for the local LLM community.

Meanwhile, in that same LocalLLaMA community, someone shared the experience: "I bought an RTX 5090 to break free from API dependence and ended up building a mini data center." With 100B-class models arriving one after another, even personal local setups will constantly face the trade-off between capacity and performance.

Do "AI Self-Evaluation Loops" Actually Work?

In 2026, you see the phrase "PROMPTS ARE DEAD. LOOPS ARE THE NEW MOAT." everywhere. It refers to a workflow in which one AI generates output, another AI evaluates it, and if it fails, the loop automatically retries and improves.

Intuitively, this approach seems like it should help, but those who have actually put it to the test report that the results are anything but straightforward. The evaluator AI has its own biases and limitations, so quality does not simply improve the more you loop. Real-world operation reportedly requires a great deal of tuning: designing evaluation criteria, setting loop termination conditions, balancing cost against quality, and more.

"Loops are powerful, but they are not magic" — that seems to be the lesson to draw from this experiment.