Two Teams Simultaneously Crack a Quantum Cryptography Problem with GPT-5.6 — What Does "Independent Discovery" Mean Now?
The Decoder reports that two research teams, working independently of each other, solved the same open problem in quantum cryptography using OpenAI's "GPT-5.6 Sol Ultra," submitting their papers just three hours apart. The article also quotes a researcher: "Whenever an open problem comes up, the first thing we do now is try whether GPT can solve it."
Purely on the technical side, this shows that frontier models have reached the point where they can solve open human problems. But what stands out is the debate it has sparked. When everyone relies on the same handful of models, the same problem tends to yield the same solution at the same time. That is a situation that shakes the very concept of "independent discovery," and it raises a new question for the research community about how priority and originality should be handled going forward.
Hints of GLM 5.3 — A Branch Appears in Z.AI's SDK Repository
A branch named "glm-5.3" has been spotted in the repository of the Java SDK that Z.AI (Zhipu) publishes, and it became a talking point on LocalLLaMA as a sign that the next model, "GLM 5.3," is in development. At this point, all that can be confirmed is that the branch name exists — nothing is known about its capabilities, release timing, or other details, so we will have to wait for an official announcement. It is a minor tidbit, but one that also reflects the community's habit of hunting for leaks at the earliest opportunity.
A Rush of New Open-Weight Models — 39B MoE, Fast Coding, and a 1B Model Built from Scratch
The open-weight camp has seen a steady stream of new model releases.
AI9Stars has released "G9v3-39A5B," a MoE model with 39B total parameters and 5 active experts. It is available under the Apache 2.0 license, which permits commercial use, and it targets reasoning, coding, tool use, and everyday assistant tasks, aiming for stronger reasoning than previous 3B models. It is said to support both Think / No Think modes.
Among coding models, "KAT Coder 2.5 dev" has been drawing attention on LocalLLaMA. One user evaluated it as "fewer tokens, faster, and more accurate than Qwen 3.6 35B A3B. In my environment it is 27B-class but 5x faster, completely outclassing the Gemma 4 models." That said, this is one individual's subjective benchmark, and results vary considerably depending on hardware and usage, so treat it as a rough reference at best.
In addition, Hacker News featured "AQ," a 1B academic LLM built from scratch by a two-person team in India. It is small in scale, but it is attracting interest as a real-world example of a model trained entirely from zero.
Korean AI Chip Designer DeepX Valued at $2.2 Billion as It Kicks Off Series D
Bloomberg reports that DeepX, a South Korean AI chip designer, is raising funds at a valuation of about 3.14 trillion won (roughly $2.2 billion), around four times its previous valuation. Headquartered in Pangyo, the company has secured 42 billion won as the first tranche of its Series D from existing investors BNW Investment and DS Asset Management. Against the backdrop of the AI boom, it shows that investor appetite is reaching strongly upstream into chip design.
A Hands-On Review of Rokid's AI Glasses Arrives — What They Can and Can't Do
ITmedia has published a hands-on review of the "Rokid Smart AI Glasses," which come equipped with both a camera and a display. The product went on general sale on July 10, and the review takes stock of what the glasses "can and can't do" at this point, exploring where AI glasses stand today and where they might be headed. For a device category where the walls of technology and usability tend to show up just past the hype, it makes for a useful look at where the line of practicality currently sits.
Around the Developer Ecosystem — Claude Code Weekly Digest, and Refactoring That Cut Tokens by 83%
A Qiita article that rounds up Claude Code changes on a weekly basis has published its edition for the week of 2026/08/02. It covers v2.1.213 through v2.1.220, with a caveat that the content is an AI-generated summary of the CHANGELOG and may contain errors.
Also on Qiita, an experiment was reported in which refactoring reduced AI input tokens by 83%. The argument: in an era when we have AI agents modify our code, the value of refactoring can be measured not only in "readability and maintainability" but also in economic terms — "fewer tokens consumed, and thus lower cost, the next time you have AI make changes." As a new yardstick for cost awareness, it makes for an interesting read.