Most enterprise RAG pipelines start the same way: a text parser converts web pages and documents into plain text so they can be chunked and indexed for retrieval. That conversion step destroys retrieval signals — and according to new research, it’s responsible for the majority of wrong answers.

A research team from UC Berkeley, Princeton University, EPFL and Databricks published a paper this week introducing PixelRAG, a system that skips that conversion entirely. Instead of parsing pages into text, PixelRAG renders them as screenshots, indexes those images and feeds retrieved tiles directly to a vision-language model reader. Tested across 30 million screenshot tiles covering all of Wikipedia, it outperforms text-based RAG across six benchmarks, improving accuracy by up to 18.1% over text-based baselines.

Parsers are the wrong place to look for fixes, according to the research team.

“Improving parsers is an endless process because every website requires special handling,” Yichuan Wang, lead author and UC Berkeley doctorate student, told VentureBeat.  “Our goal was to explore whether recent advances in VLMs make it possible to bypass that entire problem and build a retrieval system that works across websites without site-specific engineering.”

HTML parsers destroy the retrieval signals that enterprise RAG depends on

The goal of the researchers was to develop a clean end-to-end architecture.

“Modern web RAG pipelines often involve rendering, parsing, cleaning, chunking, and many other handcrafted stages,” Wang said. “Every stage introduces potential cascade errors and abstractions that move us further away from the original webpage. We were interested in whether we could eliminate most of that complexity and operate directly on the rendered page.”

Wang also noted that parsing inevitably loses information. Images, visual hierarchy, typography, emphasis (e.g., bold text), tables, and layout are either discarded or converted into imperfect textual approximations. 

“No matter how good a parser becomes, some information is fundamentally lost during the conversion,” he said.

The research identifies three ways text-based RAG loses the answer before it reaches the reader. All three were measured on SimpleQA, a standard benchmark of 1,000 factual Wikipedia questions:

• Parser loss (36.6% of failures). HTML-to-text conversion destroys structured content so completely that no text chunk in the corpus contains the answer.

• Rank loss (55.2% of failures). The answer exists in the corpus but gets outranked by keyword-dense infoboxes that land at rank 1 for 75.9% of queries, pushing answer-bearing paragraphs to rank 20 or lower.

• Reader loss (8.2% of failures). The correct content reaches the reader but flattened structure causes misattribution.

How PixelRAG works 

Unlike a standard LLM that reads only text, a vision-language model takes images as input alongside text, meaning it can read a rendered web page the way a human does, with layout and structure intact. “For many structured information extraction tasks, we believe modern VLMs have an inherent advantage because they can reason jointly over both content and layout rather than relying on a flattened text representation,” Wang said.

PixelRAG is built around that principle, replacing the text parsing pipeline with a four-stage system that operates entirely on rendered screenshots.

• Rendering. Pages are rendered using Playwright, a browser automation library, at a fixed 875-pixel viewport and sliced into 1024-pixel-tall tiles. Wikipedia’s 7 million articles produce roughly 30 million tiles. Assets are cached locally and rendered entirely offline.

• Indexing. Each tile is encoded as a single 2048-dimensional vector using Qwen3-VL-Embedding-2B and stored in a FAISS approximate nearest-neighbor index. The full index runs to approximately 120 GB in fp16 and supports incremental updates without full re-indexing.

• Training. The retrieval model is fine-tuned on synthetic contrastive data generated from the datastore, using dynamic hard-negative mining to filter false negatives. LoRA, a lightweight fine-tuning method that updates a small fraction of model weights, is applied to both the language model backbone and the visual encoder. Training on approximately 40,000 pairs completes in under three hours on a single H100.

• Storage. Raw screenshot tiles for Wikipedia require 5.6 TB, but a render-on-demand approach eliminates persistent storage: embed all tiles, delete the screenshots and re-render pages on demand at query time. The vector index requires approximately 120 GB.

Six benchmarks, 10x agent token savings and one unsolved problem

Researchers tested PixelRAG across six benchmarks spanning factual Wikipedia QA, table-based queries, multimodal QA and live news retrieval. They said it outperformed text-based RAG on all six, including on tasks where questions are answerable from text alone. On SimpleQA it reaches 78.8% accuracy versus 71.6% for the strongest text parser, widening to 48.8% versus 42.5% on structured table queries. Teams need Qwen3-VL-4B class models or above to see the benefit. Smaller models trail text retrieval by more than 12.5 percentage points.

The agent cost advantage is the strongest near-term case for PixelRAG. In benchmark testing, an AI agent using PixelRAG as its search backend ran on 3.6 million prompt tokens versus 37.5 million for text retrieval, at 2 to 4 times lower cost than alternatives including Google, while achieving higher accuracy. Image compression can cut that token budget by a further third.

Visual chunking is the main unsolved problem. Text-based RAG systems have spent years refining how to split documents into meaningful retrieval units based on topic, section or semantic content. PixelRAG currently has no equivalent: it slices pages by fixed pixel height, meaning a table or paragraph can get cut in half mid-tile with no awareness of content boundaries. 

“The text retrieval community has spent years studying chunking strategies, while visual retrieval has received much less attention,” Wang said. “We think this is an important area for future research.”

What this means for enterprises

The retrieval quality problem PixelRAG addresses reflects a broader market shift already underway. VB Pulse Q1 2026 data from qualified enterprise respondents found intent to adopt hybrid retrieval tripling from 10.3% in January to 33.3% in March, the fastest-growing strategic position in the dataset. PixelRAG’s own authors point to hybrid deployment as the most practical near-term path — layering visual retrieval on top of existing text systems rather than replacing them.

For teams already running RAG pipelines, the path to those savings is more straightforward than a ground-up rebuild.

“A practical path is to use PixelRAG as an enhancement layer alongside existing text retrieval systems,” Wang said. “Hybrid retrieval that combines both text and visual search is straightforward and is likely how many production deployments would evolve.”

Maine has taken its public data breach reporting portal offline after fraudulent breach disclosures were published on the state’s website, prompting a review of procedures to prevent abuse in the future.

Yesterday, BleepingComputer reported that fake data breach disclosures had been submitted to Maine’s official breach notification portal impersonating Discord and the multiplayer social virtual reality platform VRChat.

At the time, VRChat told BleepingComputer the filing was fraudulent and had been submitted using the name of a fictitious employee.

In a statement published Friday, the Maine Attorney General’s Office acknowledged that data breach “hoaxes” were submitted through the state’s reporting system.

“The Office of the Maine Attorney General has been made aware of an apparent abuse of our data breach reporting system,” the statement reads.

“After conversations with VRChat, one of two affected companies, it has become clear that the reported data breaches were hoaxes submitted by an unknown entity unrelated to either company. These false reports have been removed from the database. We have no knowledge of any recent legitimate data breach reports from either VRChat or Discord.”

The Attorney General’s Office says it has now temporarily disabled public access to the breach notification database while it reviews reporting procedures to reduce similar abuse in the future.

Prior to the shutdown, submitted breach notices were automatically published to the public database.

“We don’t have any independent knowledge of the breaches, the submitting entity fills out the information and it goes directly onto the site. We will review the one you’ve flagged, thank you,” Maine Attorney General’s Office told BleepingComputer.

The notice states that companies can continue to submit breach notifications through the reporting service, but members of the public seeking copies of disclosures must now contact the Attorney General’s Office directly.

Maine’s data breach portal is commonly used by journalists, researchers, and threat intelligence firms to monitor newly disclosed security incidents and determine whether organizations are reporting cyberattacks or data breaches affecting consumers.

The incident demonstrates how automatically published breach disclosures can be abused to spread misinformation and damage a company’s reputation.

The fraudulent VRChat filing claimed the company suffered a data breach impacting over 2.4 million people and included a fabricated employee contact name in the disclosure.

After BleepingComputer contacted VRChat about the filing, the company confirmed the disclosure was fake and stated it had not submitted the notice to Maine authorities.

BleepingComputer also contacted Discord about the fraudulent notice submitted to the site but did not receive a response.

It is unclear how many additional fraudulent breach notices may have been submitted through the portal before the state suspended public access to the database.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Advertisement

Get the whitepaper

Just in time for America’s 250th anniversary, Disney Imagineers tapped Apple Vision Pro to help give one of their most iconic flight rides a patriotic makeover.

Disney has shared a brand-new behind-the-scenes video as part of the Disney Unscripted series on YouTube. This time, the company shows off what it takes to revamp one of its existing attractions.

The attraction in question is “Soarin’, at EPCOT, which has been rebranded to “Soarin’ Across America” for the 250th anniversary of the United States of America. Rather than focusing on wonders around the world, or in California for another version of the ride, Soarin’ Across America takes riders on an airborne adventure across the United States.

The reimagining requires a lot of work. From capturing all new aerial footage to crafting an all-new musical score, the project requires filmmakers, musicians, and Imagineers to work together.

In the video, we learn that Disney’s audio media designers donned the Apple Vision Pro to create a digital workspace during the music and sound effects mixing phase.

“So, usually for a Soarin’ attraction, we need to build scaffolding, but that was a ‘no-can-do’ for this project because we were on such an accelerated schedule,” Megan Duncan, one of Disney’s Senior Sound Editors, says in the video.

By using the Apple Vision Pro, a virtual digital workspace easily replaced that scaffolding and extra equipment. Most of the workflow only required an Apple Vision Pro, a custom desk attached to the flight simulator seats, and a small selection of audio mixing equipment.

SpaceX rented its Colossus 1 data centre to Anthropic not because it had surplus capacity, but because it could not make the facility work for its own AI models. Bloomberg reported on Friday that SpaceX encountered latency issues when trying to connect the Memphis site to two other data centre campuses located more than 10 miles away, compounded by aging network infrastructure.

The company had planned to train its most cutting-edge Grok models using a cluster of three facilities working together. Training large AI models requires ultra-fast connections between sites. If the links are older or lower bandwidth, they create delays that slow the entire cluster. SpaceX determined the facility would be more valuable generating revenue than sitting underutilised.

The hardware mismatch made things worse. Colossus 1 contains a mix of Nvidia chip generations, including Hopper and Blackwell systems alongside older accelerators. Colossus 2 and 3 were built more uniformly around Nvidia’s Blackwell chips. In a distributed training cluster, the workload is spread across machines that need to stay synchronised. Older chips create bottlenecks by forcing faster accelerators to wait. The cluster ends up performing closer to its slowest hardware, not its fastest.

The 💜 of EU tech

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

The result is that Anthropic is now paying $1.25 billion per month to use a facility that SpaceX’s own engineers could not fully utilise. Combined with the $920 million monthly Google deal, SpaceX is collecting approximately $2.17 billion per month in compute revenue from infrastructure it originally built for itself.

The revelation complicates the narrative SpaceX presented during its IPO roadshow. Musk’s company repeatedly stressed that Colossus 1 was built in just 122 days, exceeding industry averages. Speed of construction was a selling point. Bloomberg’s reporting suggests speed came at a cost: the facility was not built uniformly enough to serve as part of a larger training cluster.

SpaceX CFO Bret Johnsen said the company has not given up on internal AI services, including Grok. Musk has described the Anthropic arrangement as a 180-day lease with a 90-day mutual cancellation right, preserving the option to reclaim the capacity. “If compute gets super tight I said we might need it back at some point,” he said.

But Grok’s trajectory makes reclaiming the compute less urgent. Downloads fell from 20 million in January to 8.3 million in April. Paid conversion is a fifth of ChatGPT’s. Federal adoption has stalled. The product that was supposed to justify the data centre investment is underperforming, while the rental income from Anthropic and Google is now a $26 billion annualised revenue line. SpaceX built a data centre for AI training and accidentally became an AI landlord instead.

China has launched its first dedicated photonic computing laboratory in Shanghai, signalling that Beijing sees light-based chips as a strategic route around Washington’s tightening grip on conventional semiconductor exports. The Shanghai Key Laboratory of Integrated Photonic Computing Chips and Systems opened on 11 June at Shanghai Jiao Tong University, the state-backed Jiefang Daily reported.

The lab is a joint effort between the university and Shanghai-based Lightelligence, one of the country’s leading photonic computing startups. Lightelligence listed on the Hong Kong stock exchange in April, surging roughly 380% on its first day of trading, and claims to be the first company in the world to achieve large-scale deployment of hybrid optical-electronic computing, though that assertion has not been independently verified.

Why photons could matter for AI

Conventional AI chips push data through silicon circuits using electrons. Photonic chips swap electrons for photons, particles of light that travel faster and generate far less heat.

The theoretical payoff is significant. Photonic processors promise higher bandwidth, lower latency, and a fraction of the energy consumption, qualities that matter as training frontier AI models pushes data-centre power demands toward their limits.

Zou Weiwen, the lab’s director and a photonics professor at Shanghai Jiao Tong University, said optical computing was “an important pathway for achieving breakthroughs in computing power.” The facility will focus on photonic chip architectures, silicon-photonics integration, optical components, and the algorithms needed to make them commercially viable.

A hedge against export controls

The lab’s launch coincides with Beijing’s broader drive for technological self-reliance. Washington has restricted China’s access to advanced semiconductors since 2022 and has widened the rules repeatedly, forcing Chinese firms to hunt for alternatives.

That search has already shifted China’s AI chip strategy away from general-purpose GPUs and toward custom silicon. Photonics represents a more radical pivot, one that could let Chinese engineers sidestep lithography bottlenecks entirely by building on the country’s existing strengths in fibre optics and laser technology.

Chinese authorities have flagged photonics and photonic-electronic hybrid accelerator chips as strategic national priorities. Shanghai officials said they had mobilised coordinated funding across multiple science and technology programmes to back the effort.

Big ambitions, early days

Beijing is already pouring money into AI infrastructure through other channels. A reported $295 billion blueprint would build a nationwide network of data centres running largely on domestic chips by 2028.

Photonic computing, however, remains far from production-ready. Zou acknowledged that the field faces “fundamental scientific challenges,” citing the absence of a mature software and algorithm ecosystem capable of efficiently harnessing photonic hardware.

The gap between laboratory promise and commercial reality is wide. But with conventional chips increasingly hard to source and AI workloads growing exponentially, China is clearly willing to bet on the physics of light.

The 2026 World Cup is here, and if you’re still thinking about buying a new TV to watch the tournament in, we’d like it if you could take a beat and consider these five key features.

Big sports tournaments are usually when retailers bring out the big discounts, but before you snap up the cheapest deal you can find, we’ve laid out five features to give some thought to before you hit buy.

From size to HDR performance to motion processing, taking these five areas into consideration will help you in your search, and hopefully lead to you having the best AV experience to watch the tournament in.

Size

Bigger is, genuinely, better. Unless you’re not able to fit a bigger screen in your living room, we’d always recommend that you go for a bigger size than you currently have.

Advertisement

Advertisement

The scale is the obvious benefit. Jumping from 55- to 65-inches reaps positives in terms of immersion. And of course, if you have multiple people around for a watch party, then having a bigger screen means you aren’t all cramming for space on the sofa and craning your necks to see what’s happening.

The last few years have seen a rise in the number of affordable, large-screen TVs. TCL’s 98-inch C7K is available for £1999, but for something considerably less expensive but still plenty big, Sharp’s 70GK4245K could be yours for less than £450.

Sound

George Lucas once said that sound is 50% of the experience. He was talking about films of course, but we’d say the same applies to anything, especially if we’re talking about sports.

Hearing the roar of the crow, feeling the intensity when something happens on (or off) the pitch, or the hush of the silence before a penalty is taken – sound matters and brings immersion to the experience of sports. So don’t buy a TV with tinny sound.

That’s easier said than done when even TVs that rack £3000 asking price have a sound that’s average. And a TV that has good sound might not have as good picture. As always, if you know (from reading our reviews, of course) that TV sound is on the weaker side, give it a boost with an external sound system.

Advertisement

Advertisement

We’d also avoid most of the built-in audio modes on TV, such as sports. Very rarely do they provide the kind of all-encompassing, immersive experiences they suggest they can.

HDR

While not every sports tournament is produced and broadcast in 4K, the last few football tournaments have been in available in HDR. For the 2026 World Cup, you can view the tournament in 4K HLG HDR on the BBC iPlayer.

More expensive TVs offer a better HDR experience because they can hit higher levels of brightness and produce a better colour experience. If you want to watch the World Cup in the best way possible, we’d suggest having a look at 4K TVs priced within the £1000 to £2000 price range for a better HDR experience. We have you covered with out best 4K TV list.

Picture mode

Advertisement

Leading on from the previous point is picture mode. Vivid (or Dynamic) is an option for some, but we find that too garish in terms of brightness and colours; and also brings in issues with the motion processing negatively affecting picture quality.

Advertisement

Film (or Movie) may offer the best, most accurate colours; but this mode is often for watching in the dark or when the curtains are drawn (considering some of the match times, this might be more useful).

The picture mode we’d suggest you watch the World Cup in, is Standard mode. Standard mode gives blues and greens a boost – helpful for bringing that rich green tone of the grass – and while it adds some processing to the mix, it’s less heavy on the picture than it would be with Vivid.

It’s also brighter than Film modes and will have more of impact if you’re watching during the day, but a lot of the matches at the World Cup will be on evening/night-time in the UK.

Motion Processing

If you’re going to use Standard picture modes (or any mode other than Film/Cinema), your TV is going to automatically add some motion processing unless you dive into the settings and disable.

Advertisement

Advertisement

If you prefer motion processing for your sports, there are some TVs that do it better than others. Sony, Panasonic, LG are towards the top of the list; Samsung not far behind without tweaking the settings a little bit, with the likes of TCL and Hisense behind and a little less consistent.

Motion processing performance can vary depending on the price. Some cheaper TVs do away with it completely (Roku models tend not to have it), but sometimes it’s better to have an affordable TV that doesn’t do it, than one that does it poorly.

Congress failed to extend a key surveillance law on Thursday night, according to a report by Politico. This effectively means that Section 702 of the Foreign Intelligence Surveillance Act (FISA) will expire for the first time since 2008, as the House isn’t expected to vote again until June 23.

The House rejected a proposal that would’ve extended the law until July 2, on a 218-198 vote. The extension actually required a two-thirds majority, but didn’t even get a simple majority. Nearly 20 Republicans joined with Democrats to block the motion. A few hours later, Oregon Senator Ron Wyden blocked a couple of proposed extensions for the law in the Senate.

This law has been around nearly 20 years through multiple presidencies from both parties. So what’s the issue right now? There are some who don’t like it when the government engages in massive warrantless surveillance programs, sure, but that never stopped the law from being renewed before. Reporting indicates that Congress was close to a three-year extension, until President Trump announced he planned to install political ally Bill Pulte as director of national intelligence.

Democrats have raised concerns over Pulte’s appointment on the grounds that he has no intelligence experience and fears that he could use sensitive information gathered via Section 702 for political or personal purposes. Pulte regularly insinuated Fed board member Lisa Cook fired engaged in mortgage fraud, an allegation that has since been debunked; Cook was removed from her post by President Trump last August.

Trump has since nominated Jay Clayton, the top federal prosecutor in New York City, for the intelligence job. However, he has suggested that Pulte could take the job on an acting basis. “There needs to be a clear guarantee that Mr. Pulte will not serve as acting DNI,” Senator Mark Warner wrote in a statement.

As for Section 702, it lets the government conduct warrantless surveillance of foreign targets located outside of the United States. It also allows agencies like the NSA and the FBI to spy on Americans if the action is “reasonably likely” to collect information about foreign intelligence.

As one would expect, authorities have played fast and loose with that whole “reasonably likely” thing. Law enforcement agencies have been caught with their hands in the data cookie jar a lot since 2008. The surveillance-based FISA court found tens of thousands of improper database searches in 2017 and 2018 alone. A judge also ruled in 2019 that the FBI and NSA committed multiple violations of either the law or privacy-minded court orders when collecting data from phone and tech companies.

House Democrats are pushing for “meaningful reforms” of the law. “Section 702 is a critical foreign intelligence authority, but we cannot in good conscience vote for reauthorization without significant reforms to protect both national security and the constitutional privacy rights of Americans,” House Minority Leader Hakeem Jeffries and other leaders said in a joint statement.

Entrepreneur and former presidential candidate Andrew Yang has a theory about where the next wave of startup opportunity lies, and it starts with a question most founders aren’t asking: what if the business model was giving money back instead of extracting it?

Yang was inspired by Mark Cuban. Not by his wealth, or his celebrity, but by Cost Plus Drugs — Cuban’s startup that sells pharmaceuticals at cost. Yang made a list.

“Housing, education, food, fuel, transportation, media, and wireless,” Yang told TechCrunch on a recent episode of Equity. “The things we all spend money on.”

He picked wireless and last September launched Nobile Mobile, a new mobile virtual network operator that provides cell service for a fraction of what traditional carriers charge and gives customers money back if they use less data. 

As AI threatens to compress wages and displace workers, Yang sees a business opportunity in bringing down the cost of living. Cost Plus Drugs, Noble Mobile, dumb phone makers like Light Phone, and even online grocery store Misfits Markets are early examples of an emerging business category where the startup’s value proposition is the margin it gives back to the customer.

“AI is going to suck up a lot of the value and the jobs, and then Americans are going to look up and say, ‘How do I meet basic needs?’” Yang said. He believes meeting people’s needs “less expensively” is “a very rich vein of opportunity.” 

That instinct didn’t emerge from nowhere. Yang first launched himself into the public eye during his 2020 presidential campaign, during which he advocated for Universal Basic Income as a means of combating AI-related workforce displacement and wealth concentration. The campaign didn’t succeed but the thesis has only grown more relevant.

Yang is still an advocate for UBI, arguing that the value generated by AI companies needs to be redistributed into the hands of the average American. But whether the government will be the vehicle for that redistribution, or whether it will just use any collected wealth to “plug a hole and do something not terribly productive,” Yang is less certain. 

“There is room for a direct connection between the money and the people,” he said. 

That’s where the market comes in. Where policy fails, Yang argues, market incentives can step in. Noble Mobile is his attempt to prove the point. Since its launch last September, the company has grown to “thousands and thousands” of customers and is bringing in “millions in revenue.”

“We’re unit profitable per customer, but we just share the profits with our subscribers with the idea that it’ll make you happy, you’ll stay around, and maybe you’ll tell your friends and family,” Yang said. 

The pitch is simple. Yang noted that the average monthly savings of $50, invested and compounded over 40 years, could amount to $24,000 — enough for a retirement down payment. And in this economy, who isn’t thinking about little ways they can upgrade their personal finance?

Whether investors will share that enthusiasm is another question entirely. Even if the opportunity is real, capital is concentrated heavily in AI right now, while consumer-facing businesses with thin margins and a social mission are a hard sell.

“I had at least one investor say to me around Noble Mobile, ‘Love you, Andrew, want to work with you — if you could just make this an AI company, we’ll invest,’” Yang said. 

The tide might be changing, though, simply because even the most wealthy, extractive companies need an economy in which consumers have enough buying power to purchase their products. 

“The value being concentrated in the hands of a handful of folks and firms is just bad for everybody,” he said. “There are some folks I know in Silicon Valley who are open to that for a variety of reasons…[like] they just don’t want to have to hire private security.”

Yang encouraged founders and investors to take on problems they’re passionate about and find a way to build a valuable enterprise on top of it.

“Think bigger and more broadly about trying to tackle problems and don’t subscribe so much to groupthink, because there are some valuable opportunities out there,” he said.