From international AI news on August 2, we cover a new approach to agent design, a very real side effect of AI slop, and a sharp technical experiment from the community.
Meta AI Prototypes a Dedicated Memory AI to Prevent Agent "Forgetting"
On long-running, complex tasks, AI agents can forget errors they have already diagnosed and end up repeating the same failed steps. To tackle this problem, Meta AI is experimenting with an approach that **places a separate "memory agent" alongside the main agent**.
The memory agent manages a structured memory bank and decides for itself when to alert the main agent and when to stay silent. Its defining feature is that it doesn't just record everything — it controls what should be delivered, and when. Improvements of up to 8.3 points have been reported on two benchmarks.
How to design long-term memory for agents is a topic that directly affects the reliability of autonomous tasks. The idea of controlling not only memory storage but also the timing of retrieval is likely to influence future agent design.
Apple's Bug Bounty Crippled by AI-Generated Fake Reports
Apple's bug bounty program has been struggling with a flood of AI-fabricated bug reports. With the review pipeline clogged, the company has responded by capping the number of submissions per researcher.
That side effect has now taken a real-world form. Bynario, an Italian startup, reportedly discovered a severe macOS vulnerability that could have fetched up to $200,000 on the black market, but was initially unable to report it because of the cap.
Alongside advances in AI-driven vulnerability discovery, the mass production of fake reports with AI has started to strain triage operations. It's a case where the design of bounty programs themselves is being called into question in the generative AI era.
A "16.5-Trillion-Parameter Empty Model" Appears on Hugging Face
[`tsfrm/vacuum-16t`](https://huggingface.co/tsfrm/vacuum-16t), published on Hugging Face, declares one of the largest parameter counts on the hub while being **a model with essentially no capabilities whatsoever**. Its creator explains that it was released as a jab at labs and companies that brag about having "the world's largest model."
The trick lies in the safetensors specification. Hugging Face computes parameter counts from the header alone, without reading the actual tensor data. So by lining up 3,841 tensors of shape `[65536, 65536]` and filling all the actual data with `0x00`, the model was listed at the top of the rankings as an apparently 8.25TB, 16.5-trillion-parameter model. The data actually transferred was reportedly only about 692KB.
The experiment highlights two technical facts. One is deduplication through Xet's content-defined chunking: byte-identical blocks are transferred only once. The other is that **storage allocation quotas are not deduplicated** — you are billed for the full apparent size. The context window is also declared as `2^32` (about 4.3 billion tokens), the maximum single tensor dimension the parser accepts.
It's a sharp, community-driven proof of concept that visualizes just how easily the parameter-count metric can be gamed.
Side Note: A Performance Hit from Logging Claude Code's History
In Japan's developer community, a case has been reported where someone built a CLI that automatically collects Claude Code and Codex work history into a local SQLite database and ran it via hooks — and Claude Code became noticeably slower as a result. They eventually identified the bottleneck and cut processing time from 268ms to 17ms. It's a valuable, measured piece of insight into how hook design affects the editing experience, and it should be useful reference material for developers extending Claude Code.