[AI DAILY NEWS RUNDOWN] Meta Fires 8,000, Musk Loses Megatrial, and Google’s Pivot (May 19 2026)

The Rundown: Cursor just released Composer 2.5, the company’s upgraded in-house coding model built on Moonshot’s Kimi K2.5, showing near-frontier benchmark performance at much lower token prices.

The details:

• Composer 2.5 nears Anthropic’s 4.7 and OAI’s GPT 5.5 across top development benchmarks, while showing a nearly 10% step up from its predecessor.

• An average CursorBench task costs under $1 with Composer 2.5, compared to up to $11 per task for Opus 4.7 and GPT-5.5 at similar score levels.

• Composer 2.5 was “partially trained on Colossus 2”, with Cursor also revealing it is currently training a larger SpaceXAI model with 10x more compute.

Why it matters: Months back, Composer 2 was swinging at the frontier at 1/10th the cost of GPT-5.4. Composer 2.5 picks up that thread, this time landing Opus 4.7-class scores while staying under $1 per task. With xAI’s Colossus compute muscle now fully behind Cursor, its next model could be the one that takes over the frontier.

Google’s AI glasses deserve the I/O spotlight

At Google I/O on Tuesday, we’ll learn where Gemini and Google’s AI devices go next, and I have to admit that one announcement is my most anticipated AI product of 2026.

Google’s biggest event of the year kicks off at 10:00 am PT at Shoreline Amphitheatre, right next to the Google campus in Mountain View. The Deep View’s Sabrina Ortiz and I will be on the ground at I/O to evaluate everything Google unveils and talk to the teams working on the products to unpack the details.

You can tune into the two-hour livestream on YouTube, and you can also follow my updates and commentary in real-time on X/Twitter at x.com/jasonhiner.

These are the most important developments to track:

• Gemini 4: We’ve all recently learned about Anthropic’s Mythos and OpenAI’s GPT-5.5, new models that claim to be so powerful that they aren’t yet ready to be released to the broader public. If Google is ready to match those capabilities, then I/O is the perfect time to unveil its latest model and share the company’s vision for the future of the technology, especially the balance between advanced capabilities and safety.

  
  
  

• Nano Banana: When Google released its Nano Banana image model in 2025, the realism of its images made every other AI imaging tool look pedestrian. However, OpenAI recently caught up with its new image model. In fact, generating images in ChatGPT is often faster and higher-quality for many queries now. Let’s see if Google has its next upgrade ready.

• Agents and coding tools: The past six months in the AI space have been dominated by the rapid advance of coding tools and AI agents, led by projects such as Anthropic’s Claude Code, OpenAI’s Codex, Perplexity’s Personal Computer, and, of course, OpenClaw and its spinoffs. Google has rarely been part of the conversation on agents or coding tools. Will it seek to change that at I/O?

• Project Aura: Done in partnership between Google and XREAL, think of Project Aura as spatial computing like Apple Vision Pro and Samsung Galaxy XR, but in a much smaller form factor that’s closer to a pair of very thick-framed glasses. Clearly, spatial computing has not been popular with the masses, and I don’t expect this product to change that. But it might offer a glimpse of the capabilities tomorrow’s smart glasses could have once the technologies shrink over time.

• Google AI glasses: Last year at I/O, Google wowed the tech world with its demo of AI glasses with a built-in display. That demo came across as far more functional and useful than the eventual competitor Meta launched last fall, the Meta Ray-Ban Display glasses. A lot of that is because Google also has Android and can do much deeper integrations with features like maps, language translation, notifications, and the camera app. And Google’s partnership with affordable glasses brand Warby Parker offers hope that it could deliver the product at a competitive price.

Musk loses OpenAI lawsuit after 3-week trial LINK

• A federal jury in Oakland rejected Elon Musk’s lawsuit against Sam Altman and OpenAI after less than two hours of deliberations, closing a three-week trial over claims the AI lab broke its nonprofit commitments.

• Judge Yvonne Gonzalez Rogers backed the advisory jury, dismissing Musk’s claims of breach of charitable trust and unjust enrichment as untimely, and told Musk’s lead counsel Steven Molo she could throw out any appeal “on the spot.”

• Musk had wanted the court to claw back up to $134 billion in “ill-gotten gains,” remove Altman and President Greg Brockman, and unwind OpenAI’s 2025 restructuring, while claims against co-defendant Microsoft were also dismissed.

Odyssey’s multimodal, multiplayer world models

Image source: Odyssey

The Rundown: Odyssey just marked two world model firsts back-to-back, dropping Starchild-1, which the company calls the first real-time, multimodal world model, and Agora-1, which lets multiple players interact inside the same AI-generated world.

The details:

• Starchild-1 can generate synchronized audio and video on the fly while taking in and adjusting to user inputs, with no fixed generation length.

• Agora-1 can host up to 4 players in one AI-generated world stream, demoed via a GoldenEye video game simulation where every pixel is produced live.

• Agora keeps a shared game state across participants, tracking agent details like position and health as actions change the world.

• The company frames Agora as an early preview for multiplayer games, robotics, and agents training together inside simulations.

Why it matters: Many of the sharpest minds in tech believe world models are the future of the industry, and these previews look like a major move up the capability ladder. Going from rendered clips to live, adjustable shared streams opens whole new avenues for both creative (gaming, storytelling) and simulation (robotics, AI training).

Apple lets Vision Pro users control a wheelchair with their eyes LINK

• Apple is rolling out a new Vision Pro feature that lets people steer motorized wheelchairs using the headset’s eye-tracking system, arriving later this year alongside other accessibility updates previewed before Global Accessibility Awareness Day on May 21.

• The power wheelchair control function works with TOLT and LUCI alternative drive systems through Bluetooth or wired connections, skips frequent recalibration, and holds up across different lighting conditions, according to Apple.

• Pat Dolan, who has lived with ALS for 10 years and founded GeoALS, said the option to drive his power wheelchair on his own is “gold” to him and called the feature life-enhancing.

Meta reassigns 7,000 staff to AI LINK

• Meta is moving 7,000 of its staff into four new groups built around artificial intelligence, an internal memo from HR head Janelle Gale told employees on Monday, two days ahead of planned layoffs.

• The four new organizations will build A.I. tools and apps, run on what Gale called “A.I. native design structures,” and have fewer managers per employee than other parts of Meta, with role details coming Wednesday.

• The reassignments come just before Meta cuts roughly 8,000 workers, or 10 percent of its staff, on Wednesday, with U.S. employees getting 16 weeks of severance plus two extra weeks per year worked.

OpenAI cofounder Andrej Karpathy joins Anthropic LINK

• Andrej Karpathy, a founding member of OpenAI and a well-known AI researcher, is heading to Anthropic, saying in a post on X that he’s eager to get back into research and development on large language models.

• Karpathy called the next few years at the frontier of large language models “especially formative,” and recently said he was blown away by the progress of agentic AI for coding after dismissing agentic abilities months earlier.

• He had been running Eureka Labs, his AI education startup, and said that work still matters deeply to him and he plans to pick it back up later; he previously helped build Tesla’s Autopilot before leaving OpenAI in 2024.

Google and Blackstone to create new AI cloud company LINK

• Google and Blackstone are teaming up to launch a new U.S.-based AI infrastructure company, with the asset management firm putting $5 billion in equity capital into the venture, which Blackstone announced on Monday.

• Google will supply its tensor processing units to the company, bringing the first 500 megawatts of compute capacity online by 2027, with plans to scale up further over time, according to Blackstone’s statement.

• The unnamed company will be led by Benjamin Treynor Sloss, Google’s former chief programs officer, and the Wall Street Journal reports that Blackstone will hold a majority stake and has already picked likely data center sites.

What Else Happened in AI on May 19th 2026?

Anthropic acquired Stainless, the startup behind its official SDKs and MCP server tooling, adding the team previously responsible for Claude’s developer libraries.

OpenAI partnered with Malta to offer free ChatGPT Plus to every citizen who completes a national AI literacy course, the first country-wide deal of its kind.

Amazon rolled out Alexa Podcasts, a new NotebookLM-style custom podcast creator in Alexa+ that creates a conversation between two AI hosts on any topic.

Meta is laying off as many as 8,000 employees this week and is no longer planning to hire for another 6,000 open roles, coming as part of the company’s AI efficiency push.

OpenAI announced a new partnership with Dell to run Codex inside corporate data centers, connecting the coding agent to enterprise internal systems.

Andrej Karpathy has joined Anthropic

NYT: Before Mass Layoffs, Meta Reassigns 7,000 Workers to Focus on A.I.

NY Post: Meta employees told to work remotely Wednesday as company prepares to slash 10% of workforce: report

WSJ: Google and Blackstone to Create New AI Cloud Company

FT: Google DeepMind’s Demis Hassabis emerges as early Anthropic investor

WSJ: The Art of War, Elon Musk Edition: How to Lose a Lawsuit and Still Claim Victory

WSJ: The American Rebellion Against AI Is Gaining Steam

FT: Tech groups score win on clean energy rules for gas-powered data centres