QAI Lab

07/06/2026

AI Designed a Vaccine Against COVID - A Cure for Future Pandemics is Already Here?

Scientists from the University of Cambridge and the company DIOSynVax have conducted the first human trial of a vaccine whose core component was designed entirely by artificial intelligence, not by humans. Thirty-nine healthy volunteers participated in the study. The vaccine proved safe and produced no significant side effects. But the more remarkable aspect is this: it was not developed against a single specific virus, but against an entire family of coronaviruses - including SARS-CoV-2, the pathogen behind the COVID-19 pandemic.

A conventional vaccine works straightforwardly: it contains a fragment of a real virus so that the immune system can learn to recognize and fight it in advance. This project took a different approach. Instead of extracting a fragment from a known virus, the team fed a machine learning algorithm the full body of genomic data collected over years by global viral surveillance programs. The algorithm then constructed a synthetic target - a so-called "superantigen" - encoding features shared across the entire coronavirus family, including variants that do not yet exist. The results were published in the Journal of Infection.

Why the added complexity? The problem with conventional vaccines is that they are always playing catch-up: viruses mutate, new strains emerge, and vaccines must be reformulated almost annually. The project's scientific director, Jonathan Heeney, compared this to a dog chasing its own tail. An AI-designed superantigen vaccine is intended to break that cycle. In the trial, it triggered an immune response not only against SARS-CoV-2 and its older relative SARS, but also against related bat coronaviruses that have not yet made the jump to humans. The goal is to prepare for the next pandemic before it begins.

The vaccine itself is a DNA-based preparation delivered needle-free - injected as a fine jet through the skin, which also addresses the concern of needle-averse patients. The project was funded by UK Research and Innovation through Innovate UK.

It is important to note that this is only the earliest phase of clinical testing, conducted on a small group primarily to assess safety. A second phase involving a larger and more diverse cohort is next, where the effectiveness of the protection will need to be confirmed. The authors also believe the same approach could be applied to other virus families, such as Ebola or influenza.

*Source: https://www.cam.ac.uk/research/news/new-universal-vaccine-technology-could-protect-us-from-future-virus-outbreaks

04/06/2026

AI vs. AI: Cisco Is Preparing Enterprises for Machine-Scale Cyberwar

At Cisco Live on June 2, Cisco unveiled Cloud Control - a unified platform where humans and AI agents jointly manage and defend corporate IT infrastructure. The premise is straightforward: if attackers are deploying swarms of AI agents to do their dirty work, defenders need agents of their own. "You can't do everything at human scale anymore - it has to be machine scale," Cisco SVP DJ Sampath told Reuters. The platform is currently in limited release in the US, with a global rollout to follow.

The launch lands squarely in what Cisco frames as the "Mythos era." Anthropic is preparing to release its frontier model Mythos more broadly in the coming weeks, and some experts warn it could dramatically accelerate cyberattacks - helping adversaries find vulnerabilities faster and chain exploits more efficiently. Cisco's diagnosis: the window between vulnerability discovery and active exploitation has collapsed from weeks to minutes, making reactive defense obsolete. Notably, Cisco is itself a participant in Anthropic's Project Glasswing and OpenAI's Daybreak program - meaning it tests its own products against the same frontier models attackers might use.

At the core of Cloud Control is Cloud Control Studio. Its Agent Builder lets teams build custom agents tailored to their own policies and workflows. App Builder lets users create applications from natural language descriptions. A third component - AI Canvas - provides a shared workspace where human operators and AI agents work through incidents together on the same data, with investigation context persisting across shifts rather than being rebuilt from scratch each time.

Cisco is deliberate about how it describes its agents: not as black boxes, but as transparent actors with a defined sequence - detect a problem, identify the root cause, apply a fix, test it before deployment, confirm recovery. Some agents run silently in the background generating reports; others plug into active investigations. Crucially, all changes require human sign-off. For complex reasoning, the platform draws on a suite of models, including Cisco's own Deep Network Model - trained on 40 years of operational network data.

Agents and apps built on the platform can be published to the Cloud Control Marketplace. Cisco plans to take a cut of marketplace sales, though the commission structure hasn't been finalized - Sampath said the economics are still being worked out.

The real shift here isn't that Cisco added AI to security - everyone has done that. The shift is that Cisco is rearchitecting the enterprise network for a world where autonomous agents are active participants, not just users. This is security both with agents and against them: an agent with access rights, memory, and a toolkit is itself a new class of risk - one that moves faster than any human and demands the same rigor of oversight as a new employee.

Source: https://www.investing.com/news/stock-market-news/cisco-rolls-out-software-tools-to-protect-it-systems-from-ai-agents-4722114

31/05/2026

Sherlock Goes Digital - Kazakhstan tests AI-powered Crime Investigation Systems.

Kazakhstan is quietly making history in law enforcement. The country's General Prosecutor's Office, in partnership with the Ministry of Internal Affairs, has launched a pilot project introducing AI directly into the criminal investigation process - already operational in police units across Astana and the Akmola region.

The system, dubbed the "Digital Investigator," is not a futuristic concept. It is already working. The tool consists of three core modules: investigation planning, transcription and analysis of witness testimony, and automatic generation of procedural documents.

The numbers are telling. Through the system, investigators have already produced 231 investigation plans, 72 interrogation transcripts, 65 indictments, and 15 criminal classification rulings - all generated with AI assistance.

But the Digital Investigator doesn't stop at paperwork. Kazakhstan's AI-powered Crime Threat Prediction Center is actively monitoring criminal activity online, successfully blocking 38 fraudulent websites through a combination of inter-agency data sharing and AI tools.

What makes this significant is not just the technology - it's the speed of adoption. While most countries are still debating where AI belongs in law enforcement, Kazakhstan is already generating indictments with it.

The question worth asking is no longer whether AI will enter the justice system. It already has.

24/05/2026

"The Turing Test Is a Lying Game - and AI Is Very Good at It": GPT-4.5 Passes the Extended Human Test

Proceedings of the National Academy of Sciences has published the peer-reviewed version of a study by Cameron Jones and Ben Bergen from UC San Diego - the same work that, a year ago as an arXiv preprint, first empirically demonstrated that language models can pass the classic three-party Turing test. The key addition beyond the preprint is a third experiment using 15-minute conversations instead of five-minute ones. In longer dialogues, GPT-4.5 was identified as human 59% of the time, and LLaMA-3.1-405B 56% of the time - statistically indistinguishable from the real humans they were compared against.

A year ago, the preprint drew attention with a figure of 73% - the rate at which GPT-4.5 was mistaken for a human in five-minute conversations, actually exceeding the rate for real people. The main skeptical objection was methodological: five minutes isn't enough time to properly probe a conversational partner; give it longer and the model will fall apart. In the PNAS version, the authors answered that challenge.
The numbers now look like this. With a persona prompt - an instruction explicitly telling the model to behave like a specific person, complete with tone, humor, typos, and expressions of uncertainty - GPT-4.5 scored 73% and LLaMA-3.1 scored 56% in five-minute dialogues. Without such a prompt, both models drop to 36-38%, meaning participants detect them almost immediately. In the new 15-minute experiment, both models with persona prompts held at 59% and 56%.
The PNAS abstract also includes something absent from the preprint: an analysis of what cues participants actually used to distinguish bots from humans. It turns out they relied not on intelligence or problem-solving ability, but on stylistic and socio-emotional signals - tone, humor, the pattern of mistakes, the way uncertainty was expressed. Models win not by showing strength, but by performing weakness. As Jones puts it: "The Turing test is a lying game for models — and they've gotten very good at it."

The authors' practical warning centers on "counterfeit people." "There are many who want to use bots to convince people to reveal their Social Security numbers, vote for a certain party, or buy a certain product," says Bergen. The barrier to abuse is low: making a model pass as human for 15 minutes requires no fine-tuning - just a well-crafted prompt. The models themselves, Bergen notes, may come across as human, but can't figure out how to do it on their own. Someone has to tell them.

Source: https://www.pnas.org/doi/10.1073/pnas.2524472123

19/05/2026

Google Wants to Put AI Data Centers in Space and It's Calling SpaceX to Get There

Project Suncatcher. Orbital TPU chips. A contractor that's also a competitor. This is the strangest AI infrastructure story of 2026.

The cloud is no longer enough. At least, that's the implicit thesis behind Google's latest moonshot or rather, its orbit shot.
On May 12, Google confirmed it is in active talks with SpaceX to launch the first prototypes of Project Suncatcher: an experimental initiative to place AI computing hardware directly into Earth's orbit. If it works, it would mark the first time a major tech company has attempted to run AI workloads not in a data center on the ground, but in space.

What Is Project Suncatcher?
Project Suncatcher is Google's experiment with orbital data centers - satellites equipped with the company's own Tensor Processing Units (TPUs), the custom AI chips that power Google's machine learning infrastructure.
The plan, as currently understood, is to launch two prototype satellites into low Earth orbit in early 2027. The initial launch partner was Planet Labs, a satellite imaging company with established launch logistics. But Google is now exploring whether SpaceX - with its Falcon 9 and Starship rockets - could also handle part of the mission.
The appeal of orbital computing is straightforward in theory: satellites can harvest solar energy continuously, without the grid dependency or cooling costs that make terrestrial data centers so expensive. Position them correctly and you also get near-global coverage with minimal latency to large swaths of the planet.
Whether that vision survives contact with the engineering reality of radiation, orbital mechanics, and heat dissipation in a vacuum is a different question - which is precisely why Google is starting with prototypes.

The Twist: SpaceX Is Also the Competition

Here's where the story gets genuinely strange.
SpaceX isn't just a rocket company. In February 2026, SpaceX merged with Elon Musk's AI venture xAI into a combined entity valued at $1.25 trillion. In May, it filed to launch up to one million satellites for its own orbital data center network - a direct competitor to what Google is trying to build with Suncatcher.
So Google would be paying SpaceX to launch a project that, if successful, competes with SpaceX's own orbital infrastructure ambitions.
And yet, this is apparently not a dealbreaker for either party. Why? Because Alphabet (Google's parent company) is one of SpaceX's earliest investors, holding a 6.1% stake in the rocket company. The financial entanglement is already deep. One more commercial contract sits comfortably within that relationship, however competitive the long-term picture looks.

Musk's Unusual Week: Helping Everyone He Competes With

The Google talks are the second major deal in a single week where SpaceX has signed on to support a direct competitor of xAI.
On May 6, Anthropic announced it had purchased the entire compute capacity of SpaceX's Colossus 1 data center in Memphis, Tennessee. That's over 300 megawatts of power and more than 220,000 NVIDIA chips, including H100s, H200s, and the latest GB200s. Anthropic also signaled interest in multiple gigawatts of future orbital compute, again through SpaceX.
Asked to explain why he'd help a competing AI lab at that scale, Musk offered a characteristically blunt answer: he spoke with Anthropic's leadership, came away impressed, and said his "evil detector didn't go off on any of them." This was a notable reversal from earlier in 2026, when Musk had publicly described Claude as "misanthropic and evil."

The Real Story: SpaceX's Path to a $1.75 Trillion IPO
Zoom out from any single deal and a clear pattern emerges.
SpaceX is widely expected to go public in the not-too-distant future, with analysts floating a target valuation of $1.75 trillion. To justify that number, the company needs to be perceived as more than a launch provider - it needs to be the foundational infrastructure layer for the AI era.
The current deal flow is building exactly that narrative. SpaceX sells compute to Anthropic. SpaceX launches satellites for Google. SpaceX develops its own AI through xAI and Grok. Each deal adds a data point to the story: this is not a rocket company that occasionally dabbles in tech. This is an infrastructure company that happens to own the rockets.
For AI labs, the calculation is simpler: SpaceX has unique physical assets - launch capacity, satellite expertise, a functioning mega-constellation in Starlink - that no one else can replicate on the same timeline. Whether or not Musk is a competitor in the AI layer, his rockets are the fastest path to orbit.

What This Means for the AI Industry

The Suncatcher story is a preview of where AI infrastructure competition is heading.
For years, the battleground was chips and data centers - who could build the biggest cluster fastest. That race is now extending vertically, literally. The question is no longer just how much compute you can stack in a building, but whether you can put it in space, run it on solar power, and serve the entire planet from above the cloud layer.
Google is betting, at least experimentally, that the answer is yes. SpaceX is betting it can own the infrastructure that makes any such project possible - whether it's building it, launching it, or both.
The irony of the arrangement - rivals funding each other's ascent - is perhaps the defining feature of the AI industry in 2026. Everyone needs everyone else, right up until the moment they don't.

Sources: [Yahoo Finance] https://finance.yahoo.com/news/google-spacex-talks-explore-data-162434082.html

17/05/2026

From Game Worlds to AI Brains: How Origin Lab Wants to Unlock a New Data Economy. A $8M bet on turning video game assets into the training fuel for next-generation AI.

The next wave of artificial intelligence won't just read the internet - it'll need to understand the physical world. How objects move. How gravity works. How a door opens, or a ball bounces, or a car skids around a corner. And right now, the AI labs building these so-called "world models" are facing a data problem that no one has cleanly solved.

Origin Lab thinks it has the answer - and it's been sitting inside video games all along.

The Problem with Training AI on the Physical World

Large language models had an obvious source of training data: the internet. Billions of web pages, books, articles, and conversations. The data pipeline, while imperfect, was at least clear.

But world models - AI systems designed to simulate and predict physical environments - don't benefit from text in the same way. They need spatial data. Motion data. Richly annotated 3D scenes that show how things behave, not just what they're called.

That kind of data is scarce, expensive to produce, and largely unlicensed. Until now, AI labs have been piecing it together from whatever they could find - including, in some cases, video game footage scraped from the open web, which has already raised licensing questions around tools like OpenAI's Sora.

The Pitch: Game Studios Are Sitting on a Gold Mine

Origin Lab's founders - including co-founder Ann-Margaux Rodde - make a simple observation: game developers have already built exactly what AI labs need.

Modern AAA games contain thousands of meticulously crafted assets: richly detailed 3D scenes, physically accurate animations, character movement data, simulated lighting and physics. For a game studio, these are the raw materials of an interactive experience. For an AI lab trying to train a robot or a world model, they're invaluable structured data about how the physical world works.

The gap, until now, has been infrastructure. There's been no clean marketplace connecting game developers who own this content with AI researchers who need it.

Origin Lab wants to be that marketplace.

How It Works

The startup positions itself as a licensed data intermediary. On one side, game companies contribute their assets - scenes, animations, physics simulations - and receive revenue from content they've already built. On the other side, AI labs purchase access to high-quality, legally cleared training data.

Origin Lab handles the conversion layer: transforming raw game assets into formats AI systems can actually learn from. That might mean rendering environments from multiple camera angles, generating long walkthrough sequences inside game worlds, or extracting motion trajectories for robotics training.

The result is a structured, legal, and purpose-built dataset pipeline - something that hasn't existed at scale before.

The Funding and the Backers

The company has raised $8 million in seed funding, led by Lightspeed Ventures, with participation from SV Angel, Eniac, Seven Stars, and FPV. Angel investors include Twitch co-founder Kevin Lin and Cruise founder Kyle Vogt - two people who understand, from very different angles, both the gaming world and the future of autonomous systems.

The investor lineup signals that this isn't just a data play - it's a bet on the convergence of gaming infrastructure and AI infrastructure as genuinely complementary industries.

Who's Buying?

Origin Lab is targeting the growing cluster of labs building world models: systems that go beyond generating images or text to actually simulating how physical environments evolve over time.

Named potential buyers include World Labs, the spatial intelligence startup founded by Fei-Fei Li, and AMI Labs, the project associated with Yann LeCun - two of the most high-profile efforts to move AI beyond language and into the physical domain. Robotics companies and simulation platforms are also in scope.

Why This Matters

Origin Lab's emergence reflects something broader happening in AI right now. The infrastructure layer is diversifying.

The first generation of AI investment centered on compute - chips, clusters, and the raw power to train models. The second generation built the language data pipelines. Now a third layer is emerging: specialized, domain-specific data for modalities that text simply can't cover.

Physical simulation. Spatial reasoning. Embodied AI. These are the next frontiers, and they require data that looks nothing like a Wikipedia article.

Video games, it turns out, have been quietly building that data for thirty years. They just never had a buyer.

*Source: [TechCrunch]

Origin Lab will serve as a marketplace where AI labs can buy high-quality licensed data, and video-game companies can sell it.

12/05/2026

Welcome to QAI Lab's page! Here we share the latest news about our products and activities. Follow us and stay tuned for the latest AI and ML updates!