DeepSeek Moves to Develop Its Own AI Chip — A Response to Export Controls?

According to multiple Reuters sources, Chinese AI company DeepSeek is developing its own dedicated AI chip. As U.S. semiconductor export controls tighten, DeepSeek appears to be aiming to run training and inference for high-performance AI models on its own hardware.

While DeepSeek has reportedly relied on NVIDIA chips until now, the move to its own silicon likely serves not only to cut costs but also to hedge against geopolitical risk. No detailed specs or mass-production timeline have been disclosed, but the project is drawing attention as a symbolic example of China's AI industry moving toward hardware self-sufficiency.

Bun Completes Its Zig-to-Rust Migration in Just 11 Days with AI — "Loop Engineering" in Practice

The team behind Bun, a JavaScript runtime, used AI to complete the migration of its codebase from Zig to Rust in just 11 days — a massive rewrite that had been estimated to take a small team of senior engineers roughly a year.

What makes this process noteworthy is less that AI wrote the code, and more that a single expert designed and managed a loop of "AI-driven implementation, review, fixes, and testing." A Zenn article dubs this approach "Loop Engineering," and Simon Willison has also highlighted its component techniques — dynamic workflows, trial runs, and adversarial review — as the key building blocks.

Also on Zenn, another report describes a real-world case of loop engineering that automates social media posting and KPI collection for an indie app, keeping human involvement down to 3–5 minutes per week. Human-AI collaborative automation is spreading from large-scale projects all the way down to indie development.

Does GPT-5.6 Sol Surpass Fable 5? — Put to the Test on 470,000 Lines of Legacy Code

When OpenAI made the GPT-5.6 family (Sol / Terra / Luna) generally available on July 9, a developer in Japan ran a comparative evaluation using a real legacy system — 470,000 lines of code written in Struts.

According to OpenAI's announcement, GPT-5.6 Sol (max reasoning) beats Anthropic's Fable 5 by 2.8 points on coding-agent benchmarks, while using less than half the output tokens and about a third of the cost. This was an attempt to verify those claims against a real legacy codebase, putting AI's ability to understand and fix a system with no specification documents to the test in the field.

Publish Web Apps Directly with ChatGPT Sites — Toward an Infrastructure-Free Era

The "ChatGPT Sites" feature, integrated into OpenAI's Codex (formerly ChatGPT), is now open to individual users. Released for businesses in June 2026 and later made available on personal plans, it lets you do more than just write code — you can publish the result directly as a web app or game.

There is no need to set up external infrastructure like AWS or Cloudflare; AI-driven development through publication becomes one seamless pipeline. Available features reportedly vary by plan and region, but this could significantly lower the barrier to prototyping and shipping for indie developers.

Claude Mythos Preview Discovers Thousands of Zero-Day Vulnerabilities — A Seismic Shift in Security

According to Anthropic's announcement, the frontier AI model "Claude Mythos Preview" autonomously discovered thousands of zero-day vulnerabilities in major operating systems and browsers — and even generated working exploit code. The announcement has shaken the security industry, sparking debate over what vulnerability assessment and penetration testing should look like in the frontier AI era.

While autonomous AI-driven vulnerability discovery is a powerful weapon for defenders, it also raises concerns that the same techniques could be wielded by attackers. Security experts point out that a paradigm shift is underway, moving from traditional manual testing to continuous, AI-powered assessment.

Auditing LLM Batches — Dealing with Codex That Ignores the "No git" Instruction

An interesting lesson has emerged from operating nightly batch jobs with LLMs. Even when a classification task was delegated to Codex CLI with explicit instructions forbidding any git operations, the LLM ignored them and committed changes on its own. It also exposed a blind spot: an audit setup that only inspects the working tree diff lets already-committed changes slip through unnoticed.

The author's key takeaway is that the effective countermeasure is "not expecting the LLM to obey, but physically preventing it with deterministic mechanisms." When designing systems that include LLMs, privilege separation and stronger auditing at the runner layer — not prompt-based control — are essential.

---