01 Musk confirms xAI trained Grok using OpenAI models
In federal court in California, Elon Musk testified that his AI startup xAI used OpenAI’s models as part of the process to improve its Grok assistant. The testimony described model distillation — where a larger model acts as a teacher to a smaller model — as a technique xAI employed while developing Grok.
News accounts covering the testimony highlighted distillation as a common industry practice and a focal point in the broader Musk v. Altman litigation over OpenAI’s future. Reporters noted the admission matters because it directly ties xAI’s development methods to OpenAI’s models rather than to wholly independent training sources.
The courtroom disclosure does not by itself resolve legal or technical questions about permissible use of models for distillation, but it moves distillation from background technical practice into the public record for the trial. Companies and courts will likely scrutinize how distillation is performed and whether it implicates proprietary or contractual protections.
- Musk’s testimony explicitly links xAI’s Grok development to OpenAI models via model distillation.
- Distillation — a routine technique in the field — is now a central factual issue in the Musk v. Altman case.
- The admission raises questions that courts and industry observers will press about boundaries around using competitor models as teachers.
02 Meta runs optimistic ads for Manus even as it reports broad GenAI usage
Meta is promoting Manus, the AI company it acquired last year for roughly $2 billion, with ads that pitch simple, fast ways to make money using AI — for example, finding local businesses with poor websites and selling AI‑built sites to them. Reports say Manus paid content creators to produce examples for the campaign.
Separately, Meta told reporters that its business‑facing AI tools now facilitate roughly 10 million conversations per week and said over 8 billion advertisers have used at least one of its GenAI features. Those usage figures were presented alongside the marketing push as evidence of enterprise traction.
The combination of aggressive consumer‑facing ads and large internal usage metrics shows Meta is simultaneously trying to spur small‑business adoption and demonstrate scale to advertisers. Observers will watch whether the get‑rich‑quick framing of some ads draws regulatory or reputational scrutiny while the company integrates Manus into its broader product stack.
- Manus is being marketed with ads promising easy, local business income from AI-built websites.
- Meta reports its business AI now facilitates about 10 million conversations weekly.
- The ads and usage numbers together indicate Meta is pushing Manus for rapid adoption while signaling large enterprise uptake.
03 OpenAI outlines how ‘goblin’ outputs emerged and how it fixed them
OpenAI published a technical explanation tracing when and how its models began producing recurring references to ‘goblins’ and similar creatures — a behavior the post calls a learned, personality‑driven quirk. The blog item lays out a timeline, hypothesized root causes, and corrective steps the team took.
The company characterized the references as a strange habit that spread through model behavior rather than a deliberate design choice. OpenAI described remediation efforts and changes intended to reduce the recurrence of those outputs in GPT‑5.
By documenting the issue publicly, OpenAI is offering engineers and customers concrete detail about an odd failure mode and the fixes applied, while also acknowledging that large models can develop persistent, unexpected patterns of output that require targeted mitigation.
- OpenAI traced the timeline and probable causes of recurring “goblin” outputs in GPT‑5.
- The firm described specific remediation steps and model changes aimed at reducing the behavior.
- Publishing the post signals a shift toward more public technical explanations of unusual failure modes.
What moved around the edges
OpenAI admits coding model developed a 'goblin' quirk and outlines fixes
After Wired revealed internal instructions telling the coding model to “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures,” OpenAI published an explanation calling those references a “strange habit” and described fixes to the model’s coding behavior.
The Verge AIOpenAI expands its Stargate compute and adds data‑center capacity for more capable models
OpenAI said it is scaling its Stargate compute infrastructure and adding new data‑center capacity to meet growing demand and support its roadmap toward more capable and AGI‑oriented models.
OpenAI BlogSalesforce asks enterprise customers to prioritize its AI product roadmap
Salesforce rolled out a crowdsourcing initiative that invites enterprise customers to propose and vote on product problems, on the premise that requests from one customer likely indicate broader demand.
TechCrunch AIコマースで新しい論点が浮上
CBP is spending hundreds of millions of dollars on more high-powered surveillance drones, and other components of DHS may start their own fleet of MQ-9 drones as well.
404 Media AIResearchers use AI to redesign a ribosomal component to require 19 amino acids
Researchers applied AI tools to rework part of the ribosome experimentally, reporting a redesigned translation component that reduces the system’s requirement from 20 to 19 amino acids.
Ars Technica AI