At Google I/O, the company unveiled Managed Agents in its Gemini API — a service that promises to collapse weeks of agent deployment work into a single API call. It’s also a sign that Google believes its ecosystem, including the newly launched Antigravity CLI, is ready to own the execution layer end-to-end.

Before a single agent is written, teams are already spending days on the unglamorous work: standing up execution environments, managing sandboxes, wiring tool call infrastructure. Model providers like Anthropic have launched platforms to handle much of that work — but Google’s approach is different.

Google said in a blog post that Managed Agents in the Gemini API abstracts “away the complexity so that you can focus on your product experience and agent behavior.” The service is available in preview via new custom templates in Google AI Studio.

The growth has introduced a real architectural question: should agent management live at the execution layer — embedded in the model or its harness — or at the infrastructure layer, as a separate runtime?

Comparing Google’s approach

Until recently, agent orchestration relied on frameworks that sat above the model, directing agents and letting teams control routing and execution separately. That layer is now being absorbed by the platforms themselves.

Recent platforms like Claude Managed Agents embed orchestration at the model layer rather than on a separate runtime platform. The idea is that the model owns the reasoning and orchestration layers, and enterprises have control over execution. 

AWS, through new capabilities on Bedrock AgentCore, adds managed harnesses that stitch together the upfront tasks for deploying agents. Google’s approach goes further, optimizing the model, harness, and sandbox together and running everything in secure Google-managed environments.

René Sultan of Ramp, cited in Google’s announcement, said the shift is concrete: “The real shift with Gemini Managed Agents is that the agent runtime moves into the platform. With the sandbox, infrastructure and execution loop managed for you, developers can focus on productizing the agent’s domain-specific behavior and iterating at a completely different pace.”

The new orchestration reality 

Enterprises starting fresh with agents could find the platform offerings from Anthropic and Google strong, especially since they remove much of the difficulty of deploying agents while still maintaining some control. Google, however, is pushing for a more vertically integrated system, while Anthropic is betting on the model layer as an orchestration plane, and AWS focuses on authorization. 

But this also brings some risks, according to XYO founder and chief executive Arie Trouw.

“An additional risk is that developers will switch out what previously were deterministic services for what will now be probabilistic services, which can introduce unpredictable outcomes for the users at best, or data corruption at worst,” Trouw told VentureBeat in an email. “This is the classic example of having an amazing hammer and everything starting to look like nails. I’ve seen this pattern repeatedly as a developer and business founder myself in the past few decades.”

Scientific papers rely on readers trusting their information. That’s why it’s disturbing that a new study by researchers connected with Cornell and UCLA found 146,900 AI-generated fake citations in scientific papers hosted across four major research databases. 

A key limitation of large language models such as Gemini and ChatGPT is their tendency to produce plausible-sounding but incorrect information, a phenomenon known as hallucination. If a researcher relies on a chatbot to draft citations without verifying them, the model may generate references that are entirely fabricated.

While scientific papers are often hidden from the public eye, the research they report has a profound impact on our lives. Everything from the internet to lithium-ion batteries began as a research paper. 

But when scientists submit papers that cite AI hallucinations, it can erode faith in the quality of the research. 

Sloppy science

The research team analyzed 111 million references from 2.5 million scientific papers. They looked for citations with titles that the team could not match to any publication. While some of these instances were just spelling errors, the team also found hallucinations. 

Unscrupulous researchers had faked citations long before the rise of chatbots, so the team also examined the rates of unmatched citations in research published before 2023, when chatbots hadn’t yet become ubiquitous. 

“We find a sharp rise in non-existent references following widespread LLM adoption,” the authors write in the paper. 

The team also found that the bad citations were spread across many papers rather than concentrated in just a few. That suggests the problem is widespread, with many researchers relying on AI-generated references without fully verifying them.

Warning signs

Usha Haley, professor of management at Wichita State University, told CNET via email that she sees the proliferation of fake citations as a serious warning.

“Fake or AI-generated citations undermine trust in the scholarly record that provides the foundation on which peer review and cumulative knowledge rest,” Haley said. “Disturbingly, this skepticism is now coming from within academia itself and from early career scholars.”

The four databases where the researchers found the fake citations are arXiv, bioRxiv, SSRN and PubMed Central. These organizations, known as scientific repositories, play a major role in the research world. 

Before a paper is published in a scientific journal, the authors often upload it to a scientific repository, increasing its visibility and allowing the global scientific community to access it immediately. The new paper on AI hallucinating citations is currently hosted on arXiv.

Recently, arXiv has taken steps to stem the flow of false citations. The organization announced Tuesday that it will ban authors who submit work with hallucinated citations or with any sign of AI content that hasn’t been carefully checked. 

“The corpus of science is getting diluted. A lot of the AI stuff is either actively wrong or it’s meaningless. It’s just noise,” arXiv scientific director Steinn Sigurdsson told CNET’s Katelyn Chedraoui back in February. “It makes it harder to find what’s really happening, and it can misdirect people.”

Nvidia founder and CEO Jensen Huang is, perhaps, one of the greatest corporate hype men of all time when it comes to his company. He may even surpass Salesforce’s Marc Benioff when it comes to relentless optimism in his company’s future and revenues.

Even so, he delivers on the hype, quarter after quarter.

Instead of cautioning you to view the proclamation that he’s found a “brand new $200 billion TAM for Nvidia” with skepticism, I’d argue he’s earned a bit of trust.

Huang positioned this massive new market at the feet of Nvidia’s new CPU product, Vera, which was introduced in March. Speaking on Wednesday’s earnings call — after Nvidia posted another record-breaking quarter with $81.6 billion in revenue and forecast $91 billion for the next — Huang pitched Vera as a potentially transformative product. And one that already has promising sales figures.

But no matter how well Nvidia delivers, Wall Street harbors anxiety over what will knock Nvidia from its perch.

Lately, such fears have centered on the CPU. Nvidia is the king of the GPU, whereas historically the CPU markets were owned by companies like Intel and AMD. (Nvidia has made CPUs previously, of course, but that’s not its core business.)

For example, last month Amazon Web Services crowed about a giant contract it signed with Meta for millions of Amazon’s homegrown AI CPUs. Amazon CEO Andy Jassy has been clear that he thinks AWS can do AI chips, both GPUs and CPUs, at least as well, and possibly better than Nvidia.

But now, with the Vera CPU, which is sold alone and bundled with its Rubin GPU, Huang believes he’s unlocked “a major new growth driver” for his company because Vera is, he believes, “the world’s first CPU, purpose-built for agentic AI,” Huang said on the call.

“Vera opens a brand new $200 billion TAM for Nvidia, a market we have never addressed before, and every major hyperscaler and system maker is partnering with us to deploy it. The world is rebuilding computing for agentic AI and robotic physical AI. Nvidia sits at the center of these transitions,” hype man Huang said.

He explained that while the “thinking” part of an AI model uses GPUs, agents mostly run on CPUs. They use CPUs to do their assigned tasks and will, he predicts, run their own form of CPU-driven PCs.

Vera is for agents because it’s specifically designed to process tokens as fast as possible. This is opposed to classic cloud architecture CPUs designed with “cores,” or the ability to run multiple instances of apps as fast as possible.

That sounds logical, but with the major cloud providers as well as startups pursuing AI chip development, what makes him think that Nvidia will be the go-to source for agentic CPUs?

Because, Huang says, Nvidia has already sold $20 billion worth of standalone Vera CPUs this year and we’re only at the beginning.

“The world has a billion users, human users. My sense is that the world is going to have billions of agents, not today. I mean, we’re going to grow into it, but we’ll have billions of agents, and those billions of agents will all use tools. And those tools are going to be like PCs, just like us humans using using PCs today,” he said.

“We’re going to need a lot more CPUs,” he explained.

Airbnb started as a way to book someone else’s spare room. It is now trying to become the app you never leave during a trip. The company’s summer 2026 release adds boutique hotels, car rentals, luggage storage, and thousands of new experiences to the platform.

The hotel push is the most significant shift. Airbnb is partnering with boutique and independent hotels in 20 cities, including New York, Paris, London, Madrid, Rome, and Singapore. Each is selected for its neighbourhood location, design, and hospitality rather than chain affiliation.

The move gives Airbnb a foothold in markets where short-term rental regulations have limited its reach. New York City and Singapore both restrict short-term stays, effectively locking Airbnb out of some of the world’s most popular destinations. Hotels solve that problem.

Users will see hotel recommendations if they search for a one- or two-night city stay. A dedicated filter lets travellers view only hotel listings. Airbnb is also offering a price-match guarantee, promising to refund the difference in app credits if the same room is found cheaper elsewhere.

“There are a few examples of the types of trips for which a hotel is probably better suited, such as last-minute bookings, one-night stays, and business trips,” said Jud Coplan, VP of marketing at Airbnb.

Beyond hotels, the platform is layering on services designed to keep travellers inside the app for longer. Luggage storage is available through a partnership with Bounce across more than 15,000 locations in 175 cities. Car rentals launch this summer with 20 per cent credit back on first bookings. These join the grocery delivery and airport pickup services Airbnb rolled out earlier this year.

On the experiences side, Airbnb is adding guided visits to 3,000 landmarks and more than 2,500 food experiences, putting it in direct competition with Viator and GetYourGuide. The app is being redesigned with a new home screen that surfaces stays, experiences, and services in one view.

The company is not launching a formal loyalty programme. But it is offering credits for first car rental bookings and up to 15 per cent back on hotel stays, a strategy that looks like it is testing the economics of retention without committing to the overhead of a full points system.

On the AI front, Airbnb is taking a notably different path from competitors. While Google, Expedia, and others have built AI-powered itinerary planners, CEO Brian Chesky said during the Q1 2026 earnings call that a chatbot is not the right interface for travel.

Instead, AI is being embedded in quieter ways. Hosts can now enter an address and let AI auto-fill listing details. Guests get AI-generated review summaries with category tags for location, amenities, and family-friendliness. A new comparison tool shows AI-generated summaries of properties saved to a wishlist.

The biggest AI investment is in customer service. Airbnb’s AI support bot, which launched in the US last year, now handles 40 per cent of all customer queries. The company is expanding it globally with support for 11 languages and adding interactive cards that let users modify bookings or resolve issues directly in the chat. A voice-based AI assistant is planned for later this year.

Chesky also disclosed that AI now writes 60 per cent of Airbnb’s new code, a figure that speaks to how deeply the company is integrating the technology into its own operations, not just its products.

The summer release is Airbnb’s clearest signal yet that it sees its future as a full-stack travel platform, not just a place to find a quirky flat. Whether travellers want one app for everything, or prefer specialist tools for hotels, cars, and experiences, is the bet the company is making.

After exiting Opening.io to iCIMS and spending two years on the investor side as a partner at Delta, Andreea Wade is back in the founder seat, and this time the moonshot is hers.

“You always hear there are no operators in VC. So I really thought, ‘OK, this is what I want to do’,” says Andreea Wade about her decision to join VC firm Delta Partners back in 2024. That was the plan, until November of last year, when her former co-founder Adrian Mihai sent her a 3am message. He had just beaten the numbers on a research paper that, in theory, solved one of AI’s most invisible infrastructure problems, explains Wade.

“I leaned in a bit more, and I was like, ‘Can I help you?’ Because that’s my thing. How can I help?” What started with a spark of interest finished with the decision to drop her new VC career with Delta Partners and return to the start-up world to co-found Univec.ai with Mihai.

It wasn’t supposed to go like this. When Wade joined Delta Partners as a general partner (a rarity at the partner-led firm) after exiting Opening.io, the AI talent intelligence company she’d built with Mihai, to iCIMS, the message was clear.

“The guys were like, ‘Oh, this is a job for life. You come in, that’s it’. Because that’s how VC works. You raise a fund, you have to be there, especially as a partner.”

Wade got it. She’d done the building thing. Now she’d do the helping-others-build thing, and she was good at it. As a self-described “founder whisperer”, she threw herself into the role and found she had a special insight born from sitting at the other side of the table.

“Regardless of what people were saying to me, how they were saying it, I knew exactly where they were, even if the words were not necessarily pointing at the thing.”

She liked being useful. “Where I come alive is when founders need help. I’m like, ‘OK, sleeves up, let me help you with the raise, with the rebrand, with whatever it is’.” But there are only so many companies a partner can get behind in any given year. So when Mihai (her co-founder of two decades, the cool kid from her hometown who once won the Romanian national programming olympiad) landed on something that might genuinely change a market, founder-whispering wasn’t going to cut it.

The problem he’d cracked sits “deep, deep in AI infrastructure”, says Wade, invisible to most, but foundational. AI models speak in vector embeddings, a layer of numbers that turns text and other content into something machines can reason about. Every vendor (OpenAI, Google, Anthropic) has its own embedding models, each effectively speaking a different language. Worse, they get deprecated. “Every single time a model is killed, you have to redo it again. Imagine that. You’ve trained, you paid all this money to train on all the poetry in the world, and a year from now, six months from now, a month from now, it’s equal to zero.”

Research to real world

Until recently, the only published solutions lived in academic papers. Mihai beat them. Univec.ai now has 87-plus bridging models that translate between embedding spaces without re-embedding from scratch. They’ve open-sourced a chunk, and are publishing benchmarks and model cards for every release – partly because the market doesn’t yet know it has this problem, says Wade.

That last bit is critical. When Wade showed the work to a hugely experienced AI lead in her investor network, his reaction was immediate. He’d only ever seen the underlying research paper. He told Wade: “Andreea, 75pc of the companies in our portfolio will not know that they have this problem, but every single one of them will.”

It’s the kind of “beautiful problem with beautiful solutions for the geeks within infrastructure” that Wade and Mihai have tackled before. At iCIMS, they were sitting on 600 million CVs, which Mihai corrected to trillions of data points.

“I remember building all this kind of marketing speak, and Adrian going, ‘It’s actually trillions, but don’t say trillions because it sounds like gazillions, so just say billions’.”

They also built one of the early vector databases, before that was a category, and didn’t spin it out. “We still had a little bit of regret on not turning that into a company.” This time, they’re not making that mistake.

What followed for Wade was a few weeks of long walks at the end of last year, and an honest reckoning. “I was already solutioning in my head. I was already working. I was already there before I was there. I just felt alive in a way that I haven’t felt in a long time.”

So she told Mihai she was in. Then came the hard part – breaking the news to her partners at Delta.

“I was having 50 heart attacks at the same time,” she says of the Monday morning she told them. When she finally got the words out, she was taken aback by the level of understanding. “They were like, ‘You need to do what you need to do, and don’t worry about anything else’.”

The resilience of immigrant founders

It’s far from Wade’s first reinvention. She arrived in Ireland 24 years ago, at 23, on an inter-company transfer through Chubb, the only way she could get here before Romania joined the EU – Wade was born in Romania to Hungarian parents. Back home, she had been a senior editor on an advertising magazine. Her first Irish gig was patrolling Coca-Cola warehouses in Drogheda on night shift.

“It’s raining, it’s cold, things are creaking, and I’m patrolling. I remember thinking if my friends, my parents at home, could see me, they’d be like, ‘What are you doing?’” Several decades later, Wade will get her Irish citizenship on 22 June.

She has leaned on that arrival story before, mostly privately. But the resilience of immigrant founders is something she finds herself returning to. “You genuinely can only depend on yourself. Something goes wrong? You can’t move into your parents’ house. It’s sink or swim.”

Between security work, journalism, a stint running an underground metal festival in Romania (Dark Bombastic Evening, or DBE, which still runs), the product curriculum at the Digital Skills Academy that start-up scene regular Gene Murphy handed her two weeks before launch, time as head of product at Independent News & Media and her own start-up branding consultancy called Brandalism, Wade has picked up a broad set of skills.

The common thread, she says, is being able to explain complicated things plainly, which is useful when your second company is building a new category of AI model that most of the market hasn’t realised it needs.

And what’s next? Well, Univec.ai will start a fundraising process in the next couple of months. The inbound interest is already there, says Wade, from European and US funds, generalists, infrastructure specialists and female-focused funds.

Wade is particularly interested in the infrastructure specialists, given the brand new start-up’s mission. “We want to contribute,” she says. “OpenAI and the others are building the foundational models. There are slices within infrastructure where we want to make our own contribution to AI.

“We want to build a new category, and be the leaders in it.” Given her and Mihai’s track record as founders, you would not bet against them. Job for life, indeed.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

You’ve probably come across LinkedIn posts that sound way too polished. These feel inauthentic while trying to sound motivational and strangely empty. The kind that turns a basic workplace thought into five neat paragraphs that push a fake lesson, and a comment section full of robotic applause.

Well, LinkedIn is now calling this a problem. The platform says it is taking new steps to reduce the reach of what it calls “AI slop,” referring to low-effort, AI-generated content that may sound clean on the surface while offering little original thought, expertise, or lived perspective.

How LinkedIn is cracking down on AI noise

LinkedIn’s Laura Lorenzetti says AI can be useful for refining language, although posts and comments still need to reflect the person behind them. So the company is building technology systems with its editorial team to identify signals of generic AI content. These systems are being trained to distinguish between posts that add perspective, context, or expertise and posts that feel repetitive, polished, and empty.

It doesn’t apply to full posts as the new system will recognize and act on comments created at scale using automation tools. These will include comments that have little to no human involvement. LinkedIn is also targeting replies that merely restate the original post without adding anything of substance.

AI slop is getting buried

LinkedIn isn’t saying every AI-assisted post will be punished. The focus is to make AI-generated content less present. When the platform detects such posts, it will be less likely to distribute it beyon the poster’s immediate network.

LinkedIn mentioned that early testing has been encouraging, with its systems correctly identifying generic content 94% of the time. The company also says members are already seeing fewer of these posts from outside their networks.

Alongside this, verification is playing a big role in fighting bots and fake AI profiles as well. With more than 100 million verified members, this could reduce the exhausting AI noise from dominating users’ feeds. It’s about time LinkedIn has started its fight against AI, with other companies like Meta and YouTube also readying tools against the avalanche of AI-generated content.