NVIDIA releases Nemotron 3 Ultra as Apple rebuilds Siri

NVIDIA introduced Nemotron 3 Ultra on June 4, 2026. The open model comes with weights, training data, and training recipes so developers can adapt and deploy it themselves.¹² It is the most capable model in the Nemotron 3 family, with 550 billion total parameters and 55 billion active during inference. NVIDIA says it supports a context window of up to one million tokens.¹²

Its hybrid mixture-of-experts architecture combines Mamba-Attention and LatentMoE.¹ NVIDIA positions the model for long-running AI agents and reports higher throughput than selected competing models on long outputs. Those comparisons come from NVIDIA itself.¹² For companies, the release provides another way to adapt a model to their own tasks rather than relying solely on an external service. Before deployment, they should check the applicable license terms for the weights and training materials, as well as compute requirements.

OpenAI announced an updated memory system for ChatGPT on June 4, 2026. It is designed to retain preferences and other important details more reliably across conversations by synthesizing memory in the background. Users can review and edit a summary of what ChatGPT remembers. The rollout starts with Plus and Pro users in the US, with more countries planned.¹²

The update addresses a challenge of long-term use: Stored details can become stale or lose relevance as circumstances change. OpenAI also provides a setting to return to the previous saved-memories experience. For businesses, the visible summary offers a way to check retained information, but it does not replace internal rules on what confidential material employees may enter into ChatGPT.²³

At WWDC on June 8, 2026, Apple outlined a new architecture for Siri and Apple Intelligence. On-device models are intended to handle simple requests. The most demanding tasks will go to AFM Cloud Pro, a cloud model running on Nvidia GPUs. A developer beta for the new Siri was described for June, with a public release planned for later in the year.¹²³

The tiered approach is meant to provide more capable answers without sending every request to the most powerful cloud model. Apple compared AFM Cloud Pro with Google's frontier Gemini models, while reporting also describes a role for Google Cloud and Gemini technology. Whether Gemini itself processes requests or primarily helped develop Apple's model remains unclear. The announcement does not mean Siri is being replaced by Google's consumer Gemini app.²³⁴

NVIDIA introduced Cosmos 3 on June 1, 2026, at GTC Taipei. The company describes the open foundation model for physical AI as a system combining visual reasoning, world generation, and action prediction. Cosmos 3 Super and Cosmos 3 Nano were available at launch, while Cosmos 3 Edge was announced for real-time inference. Developers can obtain the models through Hugging Face and try Cosmos 3 on build.nvidia.com.¹²

Cosmos 3 is intended to support development of robots and other systems that must perceive their surroundings and plan actions. The model family handles text, images, video, ambient sound, and actions. Open availability gives teams a starting point, but does not settle whether a model fits every use case. Before integrating it, teams should review the model cards, license terms, and deployment requirements.¹³

Anthropic says Claude authored more than 80% of the code merged into its production systems as of May 2026.¹² The company published the figure in early June as a measure of its internal software development workflow.¹ It refers to Claude-authored code that was merged, not every suggested change or all the work involved in delivering a product.¹²

The figure shows how extensively Anthropic uses AI tools in its own development process.¹² It does not mean Claude ships changes independently or improves itself without human involvement.¹² Engineers still review and integrate the code.² The measurement also comes from Anthropic’s own workflow: it cannot, by itself, establish what share of code AI writes at other companies or how much their engineering productivity would change.¹²

Anthropic introduced the Anthropic Institute on June 4 as a research effort. It will examine what increasingly capable AI systems could mean for further AI development and for decisions society may face. The company pointed to internal data showing that Claude is already accelerating its software development. The effort also aims to work with external stakeholders on how to respond if AI systems eventually become capable of building more powerful successors.¹²

Anthropic describes recursive self-improvement as a possibility, not an achieved or inevitable capability. It also says it is unclear whether Claude can exercise the research judgment needed to choose the right problems to pursue.¹ That limits what can be concluded from the company’s productivity figures. As risks grow, Anthropic says AI developers may eventually need a coordinated, verifiable pause in development.³

Meta has removed face-recognition code from the Meta AI companion app for its smart glasses. The Electronic Frontier Foundation reported that a June 5 app update removed the code.¹ The dormant feature had been found in an app reportedly installed on more than 50 million phones.²³ Meta said face recognition had not been enabled for consumers and that it had made no final decision about launching it.²³

The code was designed to detect faces and compare them with a local database on the user’s phone.²³ The case shows how code for an inactive feature can still be distributed through a widely used app. The reporting does not establish that users could identify people with the glasses or that biometric data was sent to Meta.²³

OpenAI launched Codex for Open Source in early June for maintainers of active, widely used public open-source projects.¹² Selected applicants receive six months of ChatGPT Pro with Codex access, plus API credits for project work.¹² Applications are available through a form.¹ The offer targets core maintainers, not everyone who contributes code to a public project.³¹

The program puts OpenAI’s coding tools into ongoing open-source maintenance: The credits are intended to support work such as code review, automation, and release workflows.³¹ The available information does not establish how much API credit each project receives.³¹ Codex Security access is also assessed case by case rather than guaranteed to every selected applicant.³¹

Netflix is testing a voice interface to help viewers find movies and shows. Chief product and technology officer Elizabeth Stone said at the Bloomberg Tech conference that the company is using generative AI and natural language processing for content discovery.¹ The aim is to tailor suggestions more closely to individual viewers and make it easier to decide what to watch next. The feature remains an experiment rather than a generally available service.²¹

Netflix already uses algorithms to personalize recommendations; a voice interface would add a conversational way to find titles.²³ That could be useful when viewers can describe what they want but do not have a specific title in mind. Whether the tests will become a permanent feature, and how well the suggestions work in everyday use, remains unclear.¹