Month: March 2026

The Israeli video personalization company is launching Strata, a basic model that reportedly creates separate, editable layers for text, animation, footage and actors instead of a single flat file. It represents a direct challenge to the architectural limitations of diffusion-based video generators.

Every AI video model currently on the market produces the same thing: a flat file. You can view it, share it, and crop it around the edges, but if you want to change the font, adjust an animation, or swap out the background, you’re basically starting over again.

This limitation has kept AI video out of professional production workflows where video was always created in layers, with separate tracks for text, motion graphics, footage and audio that can be adjusted independently until final rendering.

Idomoo, the Israeli corporate video personalization company, today announces Strata, a generative AI foundation model that it claims is the first model designed specifically for multi-layered video output. Instead of generating pixels, as Idomoo co-founder and CTO Danny Kalish puts it, Strata generates structure: independent layers of typography, animation, motion paths, and synchronized audio, all assembled into what the company calls a “production-ready video draft.”

The 💜 of EU technology

The latest rumors from the EU tech scene, a story from our wise founder Boris and questionable AI art. It’s free in your inbox every week. Register now!

The difference is architectural. Standard diffusion models combine everything into a single tensor during generation; The spatial and temporal relationships between elements are baked into the pixels themselves.

According to Idomoo, Strata solves a different computational problem: it designs the entire composition, defining placement, contrast, movement, timing and pacing at all levels simultaneously, while enforcing brand guidelines. The resulting output can be edited at the layer level, just like a professional would work in Adobe After Effects.

Strata is part of Lucas, Idomoo’s AI video agent, which builds on the company’s existing next-generation video platform. One of his more technically specific skills is brand awareness: Lucas analyzes a company’s approved content to extract what Idomoo calls brand DNA, which includes design, narrative and assets.

Strata then applies this specification to every video generated through the platform, enforcing typography, motion cues, color values ​​and tone of voice throughout the output at scale. The intent is to eliminate the template workaround that most “AI wrapper” products currently rely on, where generated footage is limited to preset layouts.

Idomoo’s argument is that forcing content into templates results in a noticeable visual compromise; Instead, Strata designs individual blueprints for each video.

Personalization is the other axis. Because Strata’s output is layered rather than flat, individual data fields, names, account details, transaction histories and product images can be inserted into specific layers of the video composition in real time.

This is the core of Idomoo’s existing business model: the company’s platform already serves JPMorganChase, Verizon and American Airlines, among others, and generates personalized videos at scale for customer communications, onboarding and marketing.

Strata theoretically makes this personalization much more sophisticated because it occurs at the composition level rather than as an overlay on top of a pre-rendered clip.

The company is cautious about the scale of the launch. An early access version is currently being tested by several of its largest customers. It is available now via Lucas AI Video Agentbut Idomoo has not disclosed which customers are in the early access cohort, what benchmarks the model was tested against, or how it compares to standard diffusion models in terms of quality metrics.

The claim of “first base model specifically designed for layered video” is Idomoo’s own formulation and has not been independently evaluated. The Strata technology is patent pending.

It’s worth noting that Idomoo’s own platform documentation has so far explicitly stated that it uses standard AI base models rather than proprietary models. Strata represents a significant shift in this positioning: from a company applying AI to video to a company building foundational AI for video.

Whether the underlying architecture meets this framework will become clearer as enterprise customers move from early access to production deployments.

Founded in 2007 by Yaron Kalish, Danny Kalish and Assaf Fogel, the Ra’anana-based company has raised a total of $27 million, including a $9 million Series A in 2013 and an $18 million Series B in 2019. The company has been building its personalized video platform for enterprise customers for nearly two decades, giving it an unusual amount of structured video production data which it can train.

A new video (above) from South Korea shows the field testing and interaction capabilities of KAIST Humanoid v0.7, developed at the Korea Advanced Institute of Science and Technology (KAIST).

The impressive humanoid robot was developed at KAIST’s Dynamic Robot Control & Design Laboratory (DRCD) and uses actuators and other internally developed technologies.

In the video below, you can watch the bipedal bot walk, jog, and jump in an incredibly human-like manner. It also kicks a soccer ball at a goal (disappointingly, there’s no robot goalkeeper to challenge it) and performs a perfect moonwalk across astroturf. And it was the moonwalk that caused a stir in the comments on the video.

“Moonwalk was flawless,” one wrote, while another commented: “Okay, that was all impressive but you won me over with the moonwalk.”

In its robotics work, KAIST uses Physical AI, a form of AI technology that enables machines to understand and operate in the physical world, and helps explain why robots like the KAIST Humanoid v0.7 appear to move in such a human-like manner.

Instead of just “thinking in words” like typical AI, physical AI gives machines a sense of space and timing in real-world environments.

As part of KAIST’s broader collaborative intelligence initiative led by Young Jae Jang, the approach trains robots and systems to continuously learn through simulation and real-time feedback, rather than relying solely on huge historical data sets.

Essentially, physical AI connects the brain and body through the tight integration of software intelligence with hardware such as motors and sensors, so that the machines not only compute but also act, react and collaborate in complex environments, whether as part of fully automated factories or in humanoid robots that, for example, kick a ball.

Engineers are refining the KAIST Humanoid v0.7 with the goal of improving its mobile and dexterous capabilities, building on its existing walking and dynamic movement capabilities. Further integrating AI with mechanical hardware aims to make the robot perform more complex tasks such as carrying objects or operating machines, thus integrating physical AI into real-world humanoid robotic applications.

KAIST is one of South Korea’s top universities and is often compared to leading global technology schools such as MIT in the USA. KAIST was founded in the early 1970s to promote Korea’s scientific and technological growth. The focus is on research in areas such as AI, robotics, physics and engineering.

The indictment of Super Micro’s co-founder exposes not only a $2.5 billion scheme, but also a system that was never designed to stop such a scheme.

Somewhere in a rented warehouse in Southeast Asia, a man was using a hairdryer on a server box. Do not dry. To loosen the adhesive on a serial number sticker so that it could be carefully peeled off and pressed onto another machine, a machine that was never plugged in, never powered up, and never intended to reach its declared destination.

The real servers, which contain Nvidia’s most advanced AI accelerator chips, had already been repackaged in unmarked boxes and shipped to China. The doll, dressed in borrowed labels, waited for the examiners.

This scene, reconstructed from surveillance footage cited in a federal indictment unsealed on March 19, 2026, is the most accurate picture we have yet of how America’s semiconductor export controls actually work in practice. Not in theory, but in practice. It turns out the answer comes down to a hairdryer.

TNW City Coworking Space – Where your best work happens

A workspace designed for growth, collaboration and endless networking opportunities at the heart of technology.

Three men are charged in the indictment: Yih-Shyan “Wally” Liaw, 71, Co-Founder, Board Member and Senior Vice President of Business Development the super microcomputer; Ruei-Tsang “Steven” Chang, 53, general manager of the company’s Taiwan office; and Ting-Wei “Willy” Sun, 44, a contractor described by prosecutors as a “fixer.”

Between 2024 and 2025, they are said to have orchestrated the redirection of servers worth around $2.5 billion, many of which were assembled in the USA and integrate Nvidia GPUs, to customers in China via a front company in Southeast Asia.

At least $510 million in hardware was shipped during a single six-week window in spring 2025. Liaw and Sun were arrested. Chang, a Taiwanese citizen, remains a refugee.

The charges include conspiracy to violate the Export Controls Reform Act, conspiracy to smuggle goods out of the United States and conspiracy to defraud the government. These crimes are punishable by a total maximum prison sentence of 30 years.

Super Micro, the publicly traded San Jose company that makes the hardware at the center of the scheme, was not named as a defendant. It placed Liaw and Chang on administrative leave and ended their relationship with Sun. It said it had cooperated with investigators and maintained a “robust compliance program.”

This sentence deserves to stay with you for a moment.

According to the indictment, the defendants and their co-conspirators communicated using encrypted messaging applications to coordinate what quantities of servers to order, which locations in China to ship them to and, crucially, how to hide the plan from the company’s own compliance team.

When an internal audit was planned, they deployed thousands of non-functioning server replicas in a warehouse rented by the front company. When a U.S. Department of Commerce inspector arrived to inspect the same facility, he used the same props and used heat guns to exchange labels and serial numbers before the visit.

The inspector, the indictment says, did not see the actual servers because they had already been sent to China. According to the public prosecutor’s office, an auditor from the company who should have been on site for a separate inspection was “outside the company and having fun at the expense of the fake company.”

The gap that was never a secret

The transshipment route through Southeast Asia is not a discovery. It is a well-known, documented and repeatedly highlighted feature of the export control architecture – one that U.S. trade analysts, think tanks and the Commerce Department itself have been warning about for years. Countries such as Malaysia, Singapore, Vietnam and Thailand have historically, as East Asia Forum analysts noted earlier this month, “lacked the enforcement infrastructure or political will to strictly monitor re-exports.”

According to an analysis published by The Diplomat, between April and July 2025, Vietnamese authorities intercepted more than 2,000 shipments that were falsely labeled as “Made in Vietnam” but were traced to Chinese factories. Malaysian tech hubs in Penang and Johor were reported for similar diversion practices.

DeepSeek, the Chinese AI lab that became a household name after releasing its model in January 2025, was accused in a report by Tom’s Hardware of setting up “ghost” data centers in Southeast Asia to pass exams and then passing on the GPUs.

According to an investigation by the Financial Times, China secured around $1 billion in advanced AI processors in the three months immediately following the last major tightening of U.S. export controls.

In other words, the pattern is not deviant. It’s structural. Controls are primarily enforced at the point of sale and at first shipment and are based almost entirely on the buyer’s declared end use and the downstream compliance of each intermediary in the chain. When the incentive to lie is hundreds of millions of dollars, there are limits to the honor system.

The company that survives itself

The appearance of Super Micro in this case is not a surprise, to say the least. The company has amassed a regulatory history that would be notable in isolation, but upon closer inspection suggests something more systemic.

In 2018, the company was temporarily delisted from Nasdaq for failing to file financial statements. In 2020, the company paid a $17.5 million fine to the Securities and Exchange Commission for what the agency called “widespread accounting violations”: more than $200 million in misrecorded revenue and understated expenses, which artificially inflated sales and profit margins.

The co-founder now facing federal charges, Wally Liaw, resigned from the company during this period. He returned as an advisor in 2021, was named senior vice president in 2022, and rejoined the board at the end of 2023.

In 2024, short seller Hindenburg Research released a report alleging new accounting irregularities, undisclosed related-party transactions, and, most notably, violations of U.S. export controls.

Ernst & Young, the company’s auditor, resigned shortly thereafter, saying it could no longer vouch for the accuracy of management’s financial representations. Super Micro has commissioned an independent special commission to review; No evidence of fraud was found.

Despite all this, Super Micro remained in the S&P 500. Sales in the last quarter were $12.7 billion.

There is a legitimate question in this number: At what point does the sample become a product? The compliance errors keep occurring. The managers involved always return. The stock continues to recover. The hardware keeps moving.

Whether Super Micro’s board and remaining leadership can provide a credible answer to this question will be of enormous importance not only to investors, but also to its credibilityThe possibility of the entire export control regime that they allegedly helped to circumvent.

Enforcement when the wind eases

The irony of this week’s indictment lies in its timing. The Trump administration in recent months has quietly relaxed export control stances that made shipping the hardware in question illegal.

In December 2025, the White House announced it would allow certain chips to be sold directly to approved customers in China.

In January 2026, the Bureau of Industry and Security released revised licensing rules that allow for case-by-case review rather than acceptance of denial for exports of previous generation AI hardware to mainland China.

A rule known as the “Affiliates Rule,” intended to close loopholes surrounding Chinese subsidiaries, was suspended for a year almost immediately after it was published.

This creates a strange political geometry. The Justice Department is prosecuting men for sending chips, which US policy is beginning to allow at the same time.

There is a version of the story in which this tension resolves smoothly: the government enforces the current rules while adapting them to the future, and the two paths do not contradict each other.

There is another version in which enforcement becomes selective, a tool to signal toughness while the underlying architecture quietly weakens. Which version actually unfolds is a question worth watching closely.

Congress was watching, and not quietly. The BIS received a 23% budget increase for fiscal year 2026, with bipartisan support and explicit funding for semiconductor law enforcement. Several lawmakers have sought congressional oversight of export permits, frustrated by what they say is inconsistent executive authority.

What none of this solves is the fundamental architecture of the problem. Export controls enforced at the point of sale, based on declared end use and monitored by corporate compliance teams who can be fooled with a hairdryer and a rented warehouse, are ultimately not a system designed for the scale of economic incentives currently in play. The chip war has raised the stakes far beyond the level the honor system was intended for.

The servers have already arrived. The stickers have been carefully reapplied. The dummy machines were ready for inspection. And somewhere in a data center in China, the real hardware is running, training models, refining weights, closing the gap.

The inspectors are still on the way.

Google is once again testing how artificial intelligence can transform the way users interact with information online – this time by changing one of the most fundamental elements of search: headlines. The company has begun experimenting with replacing original news headlines in search with AI-generated alternatives, raising concerns among publishers and questions about accuracy and editorial control.

AI is starting to rewrite the “10 blue links” of the web

In a recent experiment, Google started showing AI-generated headlines in its traditional search results instead of publishers’ original titles. This represents a significant departure from previous practices where Google shortened or slightly modified headlines for clarity.

Reportedly, these rewritten headlines can sometimes change the meaning or tone of the original article. In some cases, nuanced or critical headlines were shortened to generic or misleading wording, potentially giving users the wrong impression of the content before they even clicked on it.

Google describes the feature as a “small and narrow experiment” aimed at improving search results matching user queries and increasing engagement. However, the company has not revealed how extensively it is testing the feature.

Why this is a concern for publishers

Headlines are a crucial part of journalism – they are carefully crafted to accurately convey a story while attracting readers. By replacing them with AI-generated versions, Google is effectively taking control of how articles are presented to users.

News Unsplash

Publishers fear this could undermine their editorial voice and credibility. If an AI-generated headline misrepresents a story, readers may blame the publication rather than Google, potentially damaging trust.

There are also concerns about transparency

Currently, there is little to no evidence as to whether a headline has been altered by AI, making it difficult for users to distinguish between original content and machine-generated content.

This development follows similar experiments at Google Discover, where AI-generated summaries and headlines have already been criticized for inaccuracies and misleading wording. For everyday users, the change could subtly alter the way information is consumed. Headlines play an important role in shaping first impressions, and even small changes in wording can influence how a story is interpreted.

Google News Unsplash

When AI-generated headlines prioritize engagement over accuracy, users are more likely to click on content that isn’t quite what they expected. In some cases, they may also miss important context or nuance that was present in the original headline.

At the same time, Google argues that AI can help make headlines more relevant to searches, potentially improving discoverability and making it easier for users to find what they’re looking for.

What’s next?

While the current rollout is limited, the experiment could signal a broader shift in the way Google integrates AI into search. The company has historically introduced features as small tests before expanding them into full-fledged products.

Publishers and industry observers will likely push for clearer labeling, more transparency and possibly opt-out mechanisms as the feature expands. Meanwhile, regulators may also pay closer attention to how AI-generated content is presented, particularly if it impacts the distribution of news and public information.

For now, the experiment remains a “canary in the coal mine” — a sign of how AI could increasingly influence not only what users see online, but also how that information is compiled and understood.