Hot AI summer isn’t coming.

It’s already sweating through the group chat.

With June officially underway, we’ve got fresh AI news, updates, model releases and more. 

And big drops hours away. 

Microsoft Build is teeing up a possible Copilot reset. NVIDIA just made local AI agents feel way less theoretical. CEOs are suddenly arguing over whether AI is really killing jobs, or just giving bad hiring decisions better PR. And next week, Apple’s WWDC gets its turn to prove whether it has a real AI answer or just another polished keynote.

On today’s episode of Everyday AI, we’re tracking the AI moves that matter and could shape your AI budget before summer even gets rude.

Let’s get you caught up. 

According to Fortune, Microsoft is working on an all-in-one Copilot super app that would bring Copilot chat, GitHub Copilot, Copilot Cowork, and a new agentic workflow feature called Autopilot into one experience.

Leaked screenshots reportedly show a proactive agent named Scout, plus separate tabs for GitHub Copilot and Cowork.

That matters because Copilot has become Copilot everywhere. The problem is not a lack of places to click, it is figuring out which Copilot owns which job.

The project is being led by Jacob Andreou, Microsoft’s executive vice president of Copilot. Reports say a Microsoft Build announcement is unclear, but a public launch could happen by the end of summer.

The larger move would be one interface for consumer and enterprise Copilot, instead of personal Copilot over here and work Copilot over there.

What it means: Microsoft’s AI strategy gets easier to buy when employees stop hunting through a Copilot junk drawer.

Welp, one clean command center can beat 10 scattered entry points. If this lands, the winner is the team that maps every Copilot to a real workflow before rollout chaos starts.

At Computex in Taiwan, NVIDIA announced Nemotron 3 Ultra, a 550-billion-parameter open-weights model built for long-running enterprise agents.

The model uses a hybrid Mamba-Transformer Mixture-of-Experts architecture with 55 billion active parameters. It is the largest Nemotron 3 release so far and the flagship engine of the family.

According to Artificial Analysis, Nemotron 3 Ultra scored 48 on its intelligence index, making it the strongest measured U.S. open-weights model.

The transcript also notes the global caveat: Chinese open models are still ahead overall.

Speed is the new wrinkle. A pre-release DeepInfra endpoint reportedly served more than 300 tokens per second, while similar large models from DeepSeek and Moonshot were closer to 50 to 100 tokens per second.

For agents, that throughput can shape how useful the system feels in real work.

What it means: Sheesh, the open-model race is getting practical fast.

A less famous model that runs much faster can still win the workday. For enterprise agents, waiting 10 minutes for the smartest answer feels dumb when fast enough moves now. Budget owners should care.

At Computex, NVIDIA moved into PC hardware with its first Arm-based CPU chips for consumer laptops and mini PCs under the RTX Spark name, with first products expected this fall from ASUS, Dell, HP, Lenovo, MSI, Microsoft Surface, and others.

The Microsoft Surface version is being framed unofficially as Microsoft’s MacBook Pro counterpunch in the high-end AI laptop market.

Jensen Huang positioned RTX Spark around local AI agents that can run tasks around the clock, not just a faster laptop chip. The first laptops will use an N1X processor made with MediaTek and built on TSMC’s three-nanometer process.

The memory setup could let users run models with up to 120 billion parameters locally. Pricing and independent benchmarks are still unknown.

What it means: Local AI is becoming procurement strategy.

If agents need files, privacy, speed, and long-running access, the laptop suddenly matters again. Nope. Not a chance on the victory lap yet, because price and benchmarks still decide whether RTX Spark is legit work gear or pricey sparkle.

In a Taiwan interview, NVIDIA CEO Jensen Huang pushed back on executives blaming layoffs on AI. He said the argument is lazy because some leaders are using AI as cover for broader cost cutting.

The timing is his main point. Generative AI became broadly useful at work recently, so it does not make sense to blame job cuts from two years ago on tools that were not yet widely productive.

The transcript framed May as a rough month for jobs, with cuts across Wix, Coinbase, Meta, Cisco, Cloudflare, ClickUp, Standard Chartered, and more.

The uncomfortable pattern is that public companies can announce layoffs, cite AI, and often watch the market respond favorably. Huang’s objection was that this creates fear while letting bad hiring plans hide behind new technology.

What it means: Welp, AI cannot be the scapegoat for every spreadsheet cleanup.

Leaders should show the workflow, the tool, the before-and-after metric, and the human plan. Otherwise, it is old-school cost cutting with a shiny AI sticker slapped on top. Everybody can see it.

Recent reports covered OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei softening earlier warnings about AI-driven white-collar job loss. Both companies are reportedly preparing IPO paths with estimated valuations above $1 trillion each.

Altman told Commonwealth Bank of Australia CEO Matt Comyn that he was wrong about the speed of AI job losses. He had expected more entry-level white-collar roles to disappear by now, while still saying the risk remains real.

His personal test changed the framing. After trying to delegate Slack and email messages to AI, Altman went back to handling them himself because human interaction mattered more than expected.

Amodei’s shift was sharper. He previously warned AI could eliminate 50% of white-collar jobs, but now says automation may let workers focus on more valuable parts of their jobs and produce more.

What it means: Yuuuuup, the panic headline and the business reality can both be loud.

Entry-level work can be in trouble without half the office disappearing. Smart leaders should redesign junior roles around AI supervision, judgment, and customer context before the labor market does it for them.

OpenAI is formally hiring for a new OpenAI Robotics effort, seeking hardware, operations, systems, and machine learning engineers to build robots useful in the physical world.

Sam Altman described a two-step vision. Near term, robots would help skilled workers build infrastructure. Long term, personal robots would help people with many physical tasks.

The effort grew from OpenAI’s world-simulation research, including DALL·E and Sora work. The idea is that video and world models can train robots in simulated environments before real-world testing, which could lower costs and make failures less risky.

OpenAI has previous robotics history with Dactyl, the robotic hand trained through reinforcement learning that learned Rubik’s Cube tasks before the team was shut down. It has also backed Figure AI, 1X Technologies, and Physical Intelligence.

What it means: Robotics is the sleeper enterprise story. Straight facts.

The first useful wave may be infrastructure work, not home chores. If world models make robot training cheaper, physical labor automation moves from sci-fi demo to budget conversation faster than many leaders expect.

Anthropic released Claude Opus 4.8, a point update to its flagship model, while holding back its more powerful Claude Mythos model for a potential release within a few weeks.

The company framed Opus 4.8 as a more reliable work partner for businesses worried about confident mistakes and risky actions.

The benchmark story is intentionally modest. Anthropic said Opus 4.8 brings small improvements over Opus 4.7, while putting more emphasis on honesty than raw test-score jumps.

The model is supposed to flag uncertainty, ask clarifying questions, catch its own mistakes, and avoid unsupported claims.

Pricing stays the same on the API side: five dollars per million input tokens and $25 per million output tokens. The model is available through Claude.ai, the Claude API, the console, and Claude desktop.

What it means: Honesty is great. Refusal cosplay is not.

Before switching defaults, test Opus 4.8 on retrieval, coding, writing, and messy analysis. The winning behavior is caution plus follow-through, because a model that politely stops too early still leaves the work on your plate.