xAI Fires Engineer Who Raised Safety Warnings, Lawsuit Filed Over Grok
According to TechCrunch, a former xAI engineer has filed a lawsuit against the company and SpaceX. The former engineer claims they were fired just days after raising concerns about Grok's safety. Particularly noteworthy is the fact that the firing came right before SpaceX's historic IPO.
How AI companies handle internal safety warnings is an issue that affects the health of the entire industry. As a company led by Musk, xAI is constantly in the spotlight, and the question of how to balance safety against business speed is once again being put to the test. The outcome of this lawsuit could set an important precedent for internal governance and transparency at AI companies.
Astrophysicist Builds Black Hole Simulation with Codex
OpenAI highlighted the case of astrophysicist Chi-kwan Chan, who is using Codex to build black hole simulations. This work shows scientists leveraging AI coding tools to study physical phenomena in extreme conditions and help verify Einstein's general theory of relativity.
It's a fascinating example of AI coding tools expanding beyond software development into fundamental scientific research. The use of AI for generating and optimizing simulation code is expected to keep accelerating.
OpenAI Models and Codex Now Available on Oracle Cloud
OpenAI announced that its models and Codex are now available through Oracle Cloud. Enterprises can use their existing Oracle Cloud commitments (usage allocations) to adopt and deploy AI models while maintaining enterprise-grade security and governance.
This is part of a broader trend of OpenAI expanding its partnerships with major cloud providers. When enterprises adopt AI, being able to stay within existing cloud contracts lowers the barrier to adoption. Following Azure and AWS, availability on Oracle Cloud will be welcomed as an expansion of options.
Serious Quality Issues with Google's "Eloquent" On-Device Voice Input
Quality issues with "Eloquent," Google's newly released on-device voice input app, were reported in Reddit's LocalLLaMA community. When the reporter tried evaluating it with their own benchmark system, they found that in roughly half of the voice inputs, most of the words were missing from the output.
In the cases where accurate transcriptions came back (15 out of 50 tests), the WER (word error rate) was around 24%, on par with competing models. The problem isn't recognition accuracy — it's that the transcription itself fails. The reporter's hypothesis suggests that the transcriber is based on a chat-style AI model and may respond "about" the audio instead of transcribing it. In fact, when they ran Gemma 3n, an open model from the same family, directly, it returned responses like "I cannot transcribe this" in 11 out of 44 attempts.
Holster-scan: Detecting Package Hallucinations from AI Agents
A tool called "Holster-scan" was featured on Hacker News. It detects import statements for non-existent packages (hallucinations) in code generated by AI coding agents before the agent runs. Developed by nauta-ai, it is available on GitHub.
The problem of AI coding agents importing non-existent libraries not only causes runtime errors but can also pose a security risk. If a malicious attacker pre-registers a hallucinated package name, a supply chain attack becomes possible. The emergence of detection tools like this is important as a way to address the new challenges that come with the practical adoption of AI agents.
Pyrecall: Detecting "Catastrophic Forgetting" During LLM Fine-Tuning
An open-source tool called "Pyrecall" was introduced in Reddit's MachineLearning community. It is designed to detect "catastrophic forgetting" that occurs during LLM fine-tuning, and it also comes with a rollback feature for LoRA adapters.
Specifically, it takes snapshots of skill scores before and after fine-tuning, detects regressions, and can roll back on a per-adapter basis. While there is plenty of research on continual learning, practical tools have been lacking — this fills that gap. It is MIT-licensed and installable via pip install.