OpenAI and Anthropic Employees Ask US Government to Help Pace AI Development
According to a Bloomberg report on July 28, current employees at OpenAI and Anthropic have shared a letter asking the US government for **support in adjusting the pace of AI progress**.
The letter's details are laid out in the Bloomberg piece, but what stands out is that the very people working on frontier AI development are asking for mechanisms to control the pace of progress themselves. The fact that developers themselves are saying "we need outside help" suggests that AI capabilities may have reached a point where corporate self-regulation alone can no longer keep up.
The letter's specific demands and the number of signatories are unknown at this point, but as a voice from inside the AI industry, it's a significant signal.
---
OpenAI Publishes Field Report on Agentic AI in Scientific Computing
On July 28, OpenAI published a field report summarizing how agentic AI is actually being used in scientific computing.
According to the report, scientists are leveraging AI coding agents to modernize scientific computing software, achieving **accelerated software development and discovery** in fields such as genomics. It's a meaningful example of agentic AI moving beyond chatbots and being embedded in real research workflows.
The report also drew 23 points on Hacker News, with discussion around the theme that "much of the software for scientific computing is legacy and hard to maintain," and that AI agents are starting to fill that gap.
---
Semiconductor Stocks Plunge as AI-Related Selloff Deepens
According to the Financial Times, the selloff in AI-related stocks is deepening, with semiconductor shares falling sharply. MIT Technology Review also covered the OpenAI hack and the AI stock selloff side by side the same day.
With concerns over overheated AI investment mounting, market participants are moving to de-risk all at once. Whether this selling pressure is a temporary correction or a turning point for the AI bubble remains to be seen.
---
Microsoft Releases "Mage-VL" — A Codec-Native Streaming Multimodal Model
**Mage-VL**, released by Microsoft on Hugging Face, is a codec-native streaming multimodal foundation model.
Conventional multimodal models have mostly been designed to process images and audio through separate encoders, but Mage-VL handles them in an integrated fashion at the codec level, achieving efficiency and low latency in streaming responses. It's also attracting attention in Reddit's LocalLLaMA community.
Details on the architecture and benchmark results will presumably be revealed in an upcoming paper, but as an approach that calls into question the design paradigm of multimodal AI, it's worth watching.
---
LoRA Training DeepSeek-V4-Flash on 90GB VRAM — Training on GGUF Becomes a Reality
In Reddit's LocalLLaMA community, woct0rdho shared **a method to LoRA-train DeepSeek-V4-Flash (284B-A13B) on 90 GiB of VRAM**. It's an approach that trains directly on a GGUF-based model with no CPU offloading.
According to the poster, they achieved 19 s/it on a Strix Halo setup, implementing technical hurdles such as sliding attention, CSA, HCA, and mHC as vibe-coded Triton kernels that run faster than existing implementations.
That fine-tuning large-scale models is becoming feasible on consumer hardware is a big deal for the local LLM community.
---
The Threat of AI-Powered "Personalized Scams" Is Looming
A Washington Post opinion piece flagged the threat of AI-driven personalized cyberattacks, noting that AI is enabling mass individualization of phishing scams.
If attackers can use AI to mass-generate messages tailored to each individual target, defenses that rely on spotting "sketchy emails" may lose their effectiveness. The security industry is building AI-powered defenses in parallel, but there are concerns that the cat-and-mouse game between attack and defense will intensify.
---
Labwire — An Open Protocol for AI Agents to Control Lab Equipment
**Labwire**, released on GitHub, is an open protocol for AI agents to control laboratory equipment.
In automated scientific research, enabling AI agents to directly operate experimental equipment could dramatically improve reproducibility and efficiency. As MCP (Model Context Protocol) advances standardization between AI and software, this attempt to standardize the interface with physical hardware is drawing attention.
It appears to be an early-stage project, but it can be seen as one sign that the domain of AI agents is expanding from digital spaces into the physical world.