Notebun: AI & Tech News

Nvidia and Baker Lab Release RFD4-Proteina Open Protein Design AI

hello@notebun.co (Notebun Newsroom) — Sun, 25 Oct 2026 09:45:00 GMT

Nvidia and the University of Washington’s Institute for Protein Design (IPD), led by Nobel laureate David Baker, have released RFD4-Proteina, an open AI model that designs new proteins by generating their amino acid sequences and 3D structures at the same time. The model was announced at Nvidia’s GTC Berlin conference on October 21, 2026, after its weights were posted on Hugging Face on October 19. Its developers pitch it as a faster, cheaper tool for designing proteins for drug discovery, diagnostics and industrial uses.

What is RFD4-Proteina?

RFD4-Proteina is a generative model for biomolecular design developed by the Baker Lab at the IPD and Nvidia’s generative AI research group. According to its Hugging Face model card, it is a flow-matching model paired with an autoencoder that produces protein sequences and all-atom structures jointly, rather than designing a backbone first and filling in the sequence later.

Scientists steer the output by supplying the molecular context they want to keep, such as a binding target, an enzyme motif, a small molecule or specific structural features. Nvidia describes it as part of its BioNeMo family and says it pairs programmable design with search, so researchers, and increasingly AI agents, can refine candidates against constraints and scoring goals.

The name nods to the IPD’s earlier RFdiffusion tools and to Nvidia’s Proteina line of protein models.

What can the model design?

The model card lists several supported tasks:

Protein binders that attach to a chosen target.
Enzyme motif scaffolding, which builds a protein around a catalytic site.
Antibody design.
Binders for small molecules and DNA.
Unconditional generation of single-chain proteins.

Nvidia says such designs could be used to interrupt disease processes or to detect disease in diagnostic tests. Analytics India Magazine reports that Baker also pointed to drug discovery and enzymes that break down pesticides as potential applications.

How RFD4-Proteina compares to RFdiffusion3

Nvidia says a research draft shows higher computational design success rates and faster generation than the baseline methods it was tested against, across four benchmark suites: protein binding, small-molecule binding, enzyme scaffolding and DNA binding. Analytics India Magazine reports that on a benchmark of 200-amino-acid proteins, Nvidia measured design generation roughly 13 times faster than RFdiffusion3.

Those numbers come with an important caveat. The model card says its benchmark metric counts designs that pass computational checks per GPU-hour, not proven activity in the lab, and that the published results come from the previous release rather than the current one. Every design still needs structural and experimental validation.

Specs, license and availability

The current release has about 435.6 million parameters, stored as roughly 1.74 GB of weights, and uses a unified vocabulary that covers proteins, nucleic acids and small molecules. It runs on Linux with Nvidia GPUs from the Ampere generation through Blackwell.

The model, weights and code are released under the permissive BSD 3-Clause license. Nvidia says RFD4-Proteina is accessible through a long list of platforms, including Amazon Bio Discovery, Anthropic’s Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI and Tamarind Bio. On Hugging Face, demo inputs are screened with SecureDNA, and flagged requests are refused.

Why it matters

Baker is a 2024 Nobel laureate in chemistry, and the IPD has a track record of releasing protein design software publicly. Making a jointly developed model open, with a permissive license, puts advanced design capability within reach of any lab with a suitable Nvidia GPU.

“We are moving from an era of discovering the proteins nature has made to one in which we can design entirely new ones,” Baker said in Nvidia’s announcement.

The release is also part of a broader Nvidia push into life sciences at GTC Berlin, where the company highlighted scientific AI agents built on BioNeMo and a partnership with Siemens Healthineers on AI-assisted radiation therapy.

What happens next

The real test for RFD4-Proteina will come in wet labs, where researchers will check whether its faster computational designs translate into proteins that actually bind, catalyze and fold as intended.

Brain-Computer Interface Lets Unresponsive ICU Patients Control Sound

hello@notebun.co (Notebun Newsroom) — Sat, 24 Oct 2026 14:20:00 GMT

A noninvasive brain-computer interface allowed some brain-injured ICU patients who appeared completely unresponsive to control a sound using only their thoughts, according to a Columbia University study published October 22 in the journal Brain. About one in three of the 41 patients tested learned to raise or lower the pitch of a beeping tone within days of a severe brain injury. The researchers say the work could eventually give patients with hidden consciousness a way to communicate.

What did the Columbia study find?

The team, led by neurologist Jan Claassen of Columbia University Vagelos College of Physicians and Surgeons and NewYork-Presbyterian/Columbia University Irving Medical Center, tested the system on 41 clinically unresponsive patients with acute brain injury a few days after they were admitted to the hospital.

After a short training period, roughly a third of the patients were able to control the tone by thought alone. Every patient who had been identified as having cognitive motor dissociation, a form of hidden consciousness, through a separate exam was among those who succeeded. Some reached that point within a few days of their injury.

The paper, titled “Clinically unresponsive patients with acute brain injury activate a brain-computer interface,” lists Qi Shen as first author alongside a team of Columbia and NewYork-Presbyterian clinicians and researchers.

How does the brain-computer interface work?

Unlike implanted devices from companies working on brain chips, this system relies on standard EEG electrodes placed on the scalp, the same kind already used for routine brain monitoring in intensive care units.

The setup works in a few steps:

Patients hear a beeping tone through headphones.
A recorded voice asks them to change the tone’s frequency by imagining opening and closing their hands.
A bedside computer analyzes the EEG signals in real time and adjusts the tone accordingly.
The recorded voice offers encouragement and congratulates patients when they hit a target.

Because the hardware is already standard in ICUs, the researchers argue the approach could be far easier to deploy than alternatives. Functional MRI has previously been used to get yes-or-no answers from a single unresponsive patient, but it is not practical for routine bedside use.

What is cognitive motor dissociation?

Cognitive motor dissociation, or CMD, describes patients who appear to try to follow commands but cannot move their muscles to show it. Claassen’s team has linked the condition to a breakdown in communication between higher-order thinking regions of the brain and the motor regions that control movement.

In 2019, the same group used standard EEG and artificial intelligence to detect signs of hidden consciousness in 15% of unresponsive brain-injured patients, a finding that surprised the critical care field. Their research has also shown that patients with CMD are more likely to regain consciousness and recover significant day-to-day function. However, tools to identify CMD are still confined to research settings.

Why it matters

For families and clinicians, knowing whether a seemingly unresponsive patient is aware can shape decisions about care. A bedside tool that both detects awareness and offers a channel for basic responses could change how these patients are treated in the first critical days after injury.

Claassen is careful not to overstate the result. “Our brain-computer interface is only a first step,” he said in Columbia’s announcement, while adding that the work suggests such patients might one day be given a voice. He noted that EEG is already part of everyday ICU practice, which could help the technology spread if it matures.

Limitations and what happens next

The study shows control of a single tone, not open communication. Only about a third of patients could use the system, it required training, and public summaries do not specify how many of the 41 patients had CMD.

The researchers now plan to refine the interface so patients could signal needs such as pain or hunger. Participants and family members who tried the system called it helpful and suggested personal touches, including using a loved one’s voice, adding culturally relevant details, and building in ways to detect pain, hunger and awareness.

The work was funded by the National Institutes of Health, Paris Brain Institute America, the Canadian Institutes of Health Research, the Fondation du Centre hospitalier de l’Université de Montréal and the Royal College of Physicians and Surgeons of Canada.

Anthropic Launches Claude Duet, a Proactive AI Agent for Pro and Max

hello@notebun.co (Notebun Newsroom) — Fri, 23 Oct 2026 10:25:00 GMT

Anthropic has launched Claude Duet, a proactive AI agent that watches a user’s connected apps, flags what needs attention and carries out tasks in the background. Announced on October 21, 2026, and built on the company’s Opus 5.5 model, Duet is available as a research preview to Claude Pro and Max subscribers in the US through a waitlist. The launch is Anthropic’s answer to Meta’s Muse and OpenAI’s Dots, which arrived in September.

What is Claude Duet?

Duet is designed to act before it is asked. Each morning it sends a short briefing on important items from a user’s connected accounts, then checks those accounts every few hours and sends an alert when something needs a decision or a task is finished. Anthropic is aiming it at professionals and small-business owners who have more to track than they can manage.

The agent runs on its own remote virtual machine with a built-in browser, so work can continue after the user closes a laptop or puts down a phone. It is offered as a standalone mobile app and on the web at claude.ai, and it reuses the memory, skills and app connections a user has already set up in Claude, so there is little to configure. All interactions take place in a single ongoing conversation thread.

What can Claude Duet do?

VentureBeat reports that Duet can tap more than 6,000 connectors and plugins through the Claude Marketplace. Examples Anthropic highlighted include:

Finance: Reviewing QuickBooks or Stripe for unpaid invoices and preparing reminders.
Sales: Using HubSpot data for outreach and assembling agreements for signature in Docusign.
People operations: Preparing and running Gusto payroll and drafting job listings for LinkedIn after approval.
Marketing and commerce: Building Canva campaigns and helping maintain Shopify stores.
Analysis and building: Preparing financial forecasts, building dashboards and creating web apps.

One early user, Amin Ebrahimi, who owns a business with about 110 employees, told VentureBeat the appeal is consolidation: “I don’t need to open up 100 tabs.”

How Anthropic is handling safety and privacy

Anthropic has put several guardrails around an agent that can act on a user’s accounts. Duet drafts actions but asks for approval before sending a message, making a payment, signing a document or deleting anything. According to VentureBeat, it is barred from trading securities or moving money between financial accounts, even if a user authorizes it.

MacRumors reports that Duet never sees passwords or logins when it is given account access, and that a dedicated classifier and safety system prompt check each message to prevent leaks of sensitive information. VentureBeat adds that the most restricted data, such as payment card and government ID numbers, is blocked outright, while some other sensitive information is used only on explicit instruction.

The company is also upfront about rough edges. The preview can misread a conversation and assign the wrong urgency, and users cannot yet watch Duet’s browser as it works.

How much does Claude Duet cost?

There is no separate charge for Duet during the preview. It is included with Claude Pro, which costs $20 a month or $200 a year, and Claude Max, which starts at $100 a month. Duet draws on the same usage allowance as regular Claude chats, and its background checks count against that limit, which is why Anthropic checks accounts every few hours rather than continuously.

Duet is not yet available on Team or Enterprise plans, and Anthropic has not announced a Duet API, post-preview pricing or a timeline for wider rollout.

How Claude Duet compares to Meta Muse and OpenAI Dots

Duet enters a fast-forming market. Meta launched its Muse agent on September 8, and on Meta’s October 21 earnings call Mark Zuckerberg said more than 10 million people in the US and Canada had set up a Muse agent in about six weeks, with roughly a quarter using it on any given day, Fortune reports. OpenAI launched Dots, agents powered by its GPT-6 Astra model, on September 29.

Muse has a free entry point, while Dots is aimed at OpenAI’s higher-priced Pro and business tiers. Fortune suggests Duet may partly be a defensive move, giving paying Claude subscribers a reason not to switch to a rival assistant. It also builds on earlier Anthropic products, including Claude Tag, which follows Slack discussions, and a September update that let Claude keep working on assignments after users close their laptops.

What happens next

Anthropic has not said when Duet will leave research preview or expand beyond its US waitlist. Fortune notes that compute capacity, which strained Anthropic earlier in the year, could affect how quickly access widens, though the company has not linked the waitlist to any shortage.

Meta Q3 Earnings: Revenue Up 26% as $9.9B Legal Charge Hits Profit

hello@notebun.co (Notebun Newsroom) — Thu, 22 Oct 2026 12:10:00 GMT

Meta Platforms reported third-quarter 2026 revenue of $64.6 billion on October 21, up 26% from a year earlier, as its advertising business kept growing quickly. But a $9.89 billion legal charge tied to teen mental health litigation pushed operating income down 37%, and investors focused on rising AI spending: shares fell about 3% the next morning. Meta also guided fourth-quarter revenue to $73 billion to $76 billion, above Wall Street’s expectations.

Meta Q3 2026 earnings at a glance

Here are the headline figures from Meta’s earnings release:

Revenue: $64.6 billion, up 26% year over year. Advertising accounted for $63.06 billion, according to Techloy.
Costs and expenses: $51.6 billion, up 68%, including the legal charge.
Operating income: $13.0 billion, down 37%, for an operating margin of 20% versus 40% a year earlier.
Net income: $9.86 billion, or $3.83 per diluted share, compared with $2.71 billion a year earlier.
Excluding the charge: Operating income would have been $22.88 billion and diluted earnings $7.40 per share. Techloy reports analysts had expected $5.46.

User growth was steady. Daily active people across Meta’s apps averaged 3.64 billion in September, up 3%. Ad impressions rose 10% and the average price per ad climbed 14%.

What is the $9.89 billion legal charge?

The charge stems from an agreement Meta reached in August 2026 with 52 state attorneys general to resolve litigation over teen mental health and social media. Meta says the agreement provides for payments of up to about $18.59 billion, of which it recorded $9.89 billion as a charge in the third quarter.

The charge explains most of the gap between Meta’s strong sales and its weaker profit line. Its Family of Apps segment, which includes Facebook, Instagram and WhatsApp, posted operating income of $17.1 billion, down from $25.0 billion a year earlier.

How much is Meta spending on AI?

Capital expenditures, including finance lease payments, reached $34.91 billion in the quarter. That heavy investment turned free cash flow negative, at minus $2.12 billion, compared with a positive $10.6 billion a year earlier. 24/7 Wall St. reports Meta ended the quarter with $82.4 billion in cash and marketable securities against $83.7 billion in debt.

For full-year 2026, Meta narrowed its capital spending range to $135 billion to $145 billion, raising the low end from $130 billion. It also lifted total expense guidance to $175 billion to $179 billion to reflect the legal charge. Looking to 2027, Meta declined to give a full-year revenue outlook but said it expects to sustain strong revenue growth while continuing to invest heavily in AI.

Muse, superintelligence and Reality Labs

CEO Mark Zuckerberg tied the quarter to Meta’s AI ambitions. “We hit major milestones this quarter toward delivering personal superintelligence for everyone,” he said in the release, pointing to the launch of Muse, Meta’s personal AI agent, and Muse Wonder, which he called the company’s first frontier model. Techloy reports that more than 10 million people in the US and Canada have signed up for Muse, with about a quarter of them using it daily.

Reality Labs, Meta’s metaverse and hardware unit, lost $4.1 billion in the quarter, slightly narrower than the $4.4 billion loss a year earlier. Meta expects Reality Labs losses of $18 billion to $19 billion in 2026, shrinking to around $15 billion in 2027.

Other notable details

Meta’s headcount stood at 66,775 at the end of September, down 15% from a year earlier. The company made no share repurchases in the quarter or in the first nine months of 2026, though it paid $1.36 billion in dividends and dividend equivalents.

The release also announced a new non-voting Class C common stock, effective in January 2027, to be distributed to shareholders as a one-time stock dividend.

Why the stock fell

Despite beating on revenue and guidance, Meta shares slipped about 3% to $677.20 in morning trading on October 22, according to 24/7 Wall St., which attributed the drop to the higher AI spending outlook overshadowing the beat. The reaction underlines a question hanging over all of Big Tech this earnings season: how long investors will reward rapid revenue growth while capital spending races ahead of cash flow.

What happens next

Meta’s report lands as investors across the sector weigh soaring AI spending against revenue growth. For Meta itself, the next markers are its fourth-quarter delivery against the $73 billion to $76 billion target, any 2027 capital spending guidance, and further payments under the attorneys general settlement.

Agile Robots Unveils Agile ONE SN Humanoid Built on Nvidia GR00T

hello@notebun.co (Notebun Newsroom) — Wed, 21 Oct 2026 15:40:00 GMT

Munich-based Agile Robots has unveiled Agile ONE SN, a compact humanoid robot designed as a ready-made research platform, at Nvidia’s GTC Berlin conference on October 21, 2026. Built on Nvidia’s Isaac GR00T robot software and powered by a Jetson Thor onboard computer, the robot is meant to give labs a standard humanoid they can start experimenting with in days rather than weeks. Nvidia says it is scheduled to be available in early 2027.

What is Agile ONE SN?

Agile ONE SN is a “reference” humanoid: a complete package of body, sensors, computer and software intended for robotics researchers and developers rather than consumers or factories. Instead of each lab assembling its own hardware and software stack, the idea is to offer one validated platform that works with Nvidia’s simulation and training tools out of the box.

Nvidia says the approach can cut setup time for researchers from weeks to days by tying together simulation, training and deployment. The company named Stanford University and ETH Zurich as leading institutions that have shown interest in the robot.

Agile ONE SN specs

According to Nvidia’s announcement, the robot is built on Agile Robots’ Agile ONE S body and includes:

Size: 1.40 meters (4.59 feet) tall and 40 kilograms (88 pounds).
Mobility: 31 degrees of freedom across the body, with a 3-kilogram payload per arm.
Hands: Two five-fingered WUJI Hand 2 units, each with 20 degrees of freedom, tactile sensing and 1 kHz force control.
Compute: Nvidia Jetson Thor, rated at up to 2,070 FP4 teraflops with 128 GB of unified memory.
Sensors: Cameras in the head and wrists, lidar and inertial measurement units.
Safety: Planned integration with Nvidia Halos, the company’s robotics safety platform.

No price has been announced, and GadgetScout notes that no precise launch date or UK sales channel has been confirmed.

Why Nvidia is backing a reference humanoid

The launch was one of several robotics updates Nvidia made at GTC Berlin, which ran from October 20 to 22. Alongside the robot, Nvidia made Isaac Lab 3.0, its open-source robot learning framework, generally available and introduced Isaac OS, an open-source runtime built on its Holoscan platform that links AI models, sensors and robot controls in real time.

The company also published a library of more than 2,000 simulation-ready robot and factory assets on Hugging Face. A standard humanoid running on Nvidia’s chips and models gives researchers a common target for all of these tools, and gives Nvidia another way to anchor the emerging humanoid ecosystem to its hardware.

Nvidia CEO Jensen Huang framed the moment in sweeping terms, saying “the ChatGPT moment for robotics is fast approaching.” Agile Robots founder and CEO Zhaopeng Chen, for his part, argued that usefulness at scale is the real breakthrough humanoids need.

Who is Agile Robots?

Agile Robots was founded in 2018 as a spin-off from the German Aerospace Center (DLR) and is headquartered in Munich, according to Crypto Briefing. A 2021 funding round led by SoftBank valued the company at around $1 billion, and its investors include SoftBank and Xiaomi.

The company showed a full-size humanoid, the Agile ONE, in November 2025. Crypto Briefing reports that production of the standard Agile ONE S and the new SN version is planned to start by the end of 2026, with output projected at 1,000 to 2,000 humanoid robots in 2027. The outlet also lists NEURA Robotics as a competitor in Europe’s humanoid market.

How it fits into the humanoid robot race

Humanoid robots were a major theme of GTC Berlin, where Nvidia listed European robotics companies including ABB Robotics, KUKA, NEURA Robotics, Humanoid and Universal Robots among those building on its platforms. Most of these efforts target industrial work, while Agile ONE SN is pitched squarely at the research labs that develop the next generation of robot skills.

That positioning matters because research groups often spend significant time on integration before any learning experiments begin. A shared platform could also make results easier to compare across institutions.

What happens next

The key milestones to watch are the early 2027 availability window, final pricing, and whether Agile Robots hits its production targets. Integration with Nvidia Halos safety tools is also still listed as planned, with no timeline given. If the platform ships as described, it would give research labs an off-the-shelf humanoid designed around Nvidia’s robot learning stack.

Google Willow Quantum Chip Cuts Magic State Errors 40-Fold

hello@notebun.co (Notebun Newsroom) — Mon, 19 Oct 2026 13:05:00 GMT

Google Quantum AI says it has shown on real hardware that a technique called magic state cultivation works on its Willow superconducting processor. The result, announced on October 15 and published in the journal Science, cut errors in so-called magic states by roughly 40 times and produced T-states with a record 99.99% logical fidelity. The work targets one of the most expensive steps on the road to a useful, error-corrected quantum computer.

What is a magic state in quantum computing?

Error-corrected quantum computers can run some operations cheaply, but a fully “universal” machine also needs operations that go beyond what classical computers can simulate. Magic states are specially prepared quantum states that make those harder operations possible without breaking the machine’s error protection.

The catch is supply. Each magic state is consumed when it is used, so a large computation needs a constant stream of them. Making them with low error rates is costly: Google cites estimates that up to 95% of a future quantum computer’s capacity could be spent just producing these states.

How magic state cultivation works

The cultivation approach was proposed in 2024 by Google researchers Craig Gidney, Noah Shutty and Cody Jones as a leaner alternative to magic state distillation, the traditional and resource-hungry method. Instead of combining many noisy states into one good one, cultivation starts with a single noisy state injected into a small region of the chip.

A series of coordinated checks then looks for errors. If a check fails, the attempt is thrown out and restarted; if it passes, the state is gradually grown into a purer, lower-error version. In a blog post, Google researchers Emma Rosenfeld and Craig Gidney compare the process to a filter that only lets clean states through.

What Google achieved on Willow

According to Google and coverage of the Science paper, the team reported:

About a 40-fold reduction in errors compared with states that were not cultivated.
A logical fidelity of 99.99% for T-states, which Google describes as a record.
A successful transfer of a cultivated state from a distance-3 color code into a larger distance-5 surface code, the kind of code used for general error correction.

That reliability has a price. ScienceAlert reports that in the strictest runs only about 8% of attempts were kept, with the rest discarded. Moving the state into the larger surface code also added new errors, which the researchers flag as the next problem to solve.

The project also hit an unexpected obstacle. Early measurements appeared to contradict physics, and the team traced the issue to a faulty assumption in its standard fidelity-measurement method, which was skewed by small lab fluctuations such as temperature changes. The researchers built a new technique, called kickback tomography, to measure the states accurately. In the paper, they write that the findings move cultivation “from a theoretical study into an experimental reality.”

Why it matters for fault-tolerant quantum computers

Because magic states are expected to dominate the hardware budget of large quantum machines, any method that produces them more cheaply could shrink the size of the computers needed for practical tasks. Google frames the result as bringing it one step closer to a useful, error-corrected quantum computer.

The company is careful about the limits. Its researchers say cultivation still has to be shown at a larger scale, and that many technical challenges remain. Willow is not a fault-tolerant computer today, and this result does not change that on its own.

What it means for encryption and “Q-day”

The announcement quickly drew attention from the cryptocurrency world, where a sufficiently powerful quantum computer could one day break the digital signatures that secure transactions. The Block noted that Coinbase had published technical priorities for a post-quantum Bitcoin transition the same week, and that Google has set an internal goal of moving to post-quantum cryptography by 2029.

Experts quoted by The Block urged calm. Alex Pruden, CEO of Project Eleven, said the technique was already built into earlier estimates of the resources needed to break encryption, so it does not shift timelines so much as make those estimates more credible. Castle Island Ventures co-founder Nic Carter said quantum timelines are generally compressing but that there is no imminent threat.

What happens next

Google’s next steps are to scale cultivation up and to reduce the errors introduced when cultivated states are handed off to larger error-correcting codes. Those milestones, rather than a single record fidelity number, will determine whether cultivation becomes a standard building block for the fault-tolerant machines Google and its rivals are racing to build.