On-prem

BYO power for AI bit barns may be the best way to ease the problem, says energy watchdog

Prices in the United States’ largest wholesale power market have nearly doubled in the past year thanks to demand from datacenters. And an independent watchdog predicts things will only get worse without some serious changes.

The PJM Interconnection serves all or parts of 13 states and the District of Columbia in the eastern US, including Northern Virginia, that’s got the densest cluster of datacenters in the world. The surge in wholesale power costs across PJM was outlined on Thursday by Monitoring Analytics, a firm that serves as the official market monitor for the Interconnection, in its Q1 2026 state of the market report. 

According to the report, the total cost per megawatt-hour (MWh) of wholesale power rose from $77.78 in the first three months of 2025 to $136.53 in the same period this year, an increase of 75.5 percent year over year. Monitoring Analytics didn’t mince words in its report, identifying datacenter load growth as the main driver of recent capacity market conditions and rising prices in PJM. 

“Data center load growth is the primary reason for recent and expected capacity market conditions, including total forecast load growth, the tight supply and demand balance, and high prices,” the report reads. “But for data center growth, both actual and forecast, the capacity market would not have seen the same tight supply demand conditions.” 

As for what might come next, the report doesn’t ignore the likely outcome of the current situation, either. 

“The price impacts on customers have been very large and are not reversible,” the report states, but the bad news doesn’t stop there. “The price impacts will be even larger in the near term unless the issues associated with data center load are addressed in a timely manner.” 

Based on the rest of the report, a timely resolution to the datacenter load issue shouldn’t be expected, at least not in a way that’ll benefit locals. 

For starters, Monitoring Analytics found that – like pretty much everywhere right now – power grids aren’t ready for the datacenter boom. PJM has taken steps to upgrade its power commitment and dispatch software to better operate its grid, but planned upgrades have been delayed multiple times with no planned implementation date on the calendar, per the report. 

“The current supply of capacity in PJM is not adequate to meet the demand from large data center loads and will not be adequate in the foreseeable future,” Monitoring Analytics asserted. 

Current plan: Shift the risk to everyone else

PJM has been planning a one-time backstop auction to procure new power generation for datacenter projects in the region at the request of the Trump administration and the governors of the states it serves, but Monitoring Analytics isn’t convinced the Interconnection is going about the process in the right way. 

The currently proposed auction structure, says the watchdog, would “generally shift significant risk to other PJM customers,” which is a temptation the group says “should be resisted.” 

“Other PJM customers, whether residential, commercial or industrial, should not be treated as a free source of insurance, or collateral, or financing for data centers,” the report continued. “Yet that is what most of the proposals related to a backstop auction actually do.” 

As for what PJM ought to be doing, you probably won’t need to rack your brain to figure that out: Monitoring Analytics says datacenters ought to be required to bring their own power. Such a rule, says the group, should include fast-track options for interconnection for BYOP datacenters, and otherwise a queue that would only connect datacenters when there is adequate capacity to serve them. 

“This broad bring-your-own new generation solution to the issues created by the addition of unprecedented amounts of large data center load does not require a continued massive wealth transfer through ongoing shortage pricing,” the analysts argue. 

When asked for its response to the problems raised by the Monitoring Analytics report, PJM told us that it was fully aware of the impact of electricity cost increases on its customers. 

“PJM is working with states and member companies to address these consumer impacts on multiple fronts, including extending market caps put in place since the 2025/2026 auction, authorizing multiple transmission expansion projects that are now in development, and reforming wholesale electricity market rules,” the Interconnection told us. Monitoring Analytics didn’t respond to questions.

Americans have become increasingly hostile to new datacenter projects driven by the AI boom, with 71 percent of respondents to a Gallup survey saying they opposed DC projects in their neighborhoods. Projects in multiple states have been abandoned recently due to pushback from locals, many of whom are concerned not only with electrical price increases, noise, and eyesores, but environmental harm as well. ®

Anthropic has launched a small business offering, signalling a new front in the AI platform rivalries.

Anthropic, the AI company best known for its Claude models, has unveiled a dedicated product for small businesses – a move that seems to mark a deliberate pivot beyond the large enterprise customers that have driven its success to date.

Claude for Small Business is a toggle-on feature within Claude Cowork, Anthropic’s task-automation platform. Once activated, it gives paying users access to 15 pre-built agentic workflows across finance, operations, sales, marketing, HR and customer service, and can be connected to software that many small businesses already use. Partner integrations include QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace and Microsoft 365.

It is a strategy that makes sense, particularly in the massive US market. Small businesses account for 44pc of US GDP and employ nearly half the private-sector workforce, according to Anthropic, which added that their AI adoption has lagged behind larger enterprises.

“Small businesses make up nearly half the American economy, but they’ve never had the resources of bigger companies,” said Daniela Amodei, co-founder and president of Anthropic.

“AI is the first technology that can finally close that gap, which is why we’re launching Claude for Small Business, alongside training and partnerships to make sure AI shows up for the entrepreneurs and communities who need it most.”

The launch may signal that the battleground for AI user acquisition is shifting. While Anthropic has had significant success taking on its major rival OpenAI in the enterprise market, the latter is well ahead when it comes to small business, having released an Enterprise ChatGPT tier that included a small-team option back in 2023.

As part of the launch, PayPal and Anthropic have co-created ‘AI Fluency for Small Business’, a free online course teaching owners how to integrate AI into their operations.

“Together, we are equipping business owners and entrepreneurs with the tools, expertise, and trusted infrastructure they need to compete and thrive,” said Amy Bonitatibus, chief corporate affairs officer at PayPal. The course is available on demand; on completion, learners receive a shareable certificate.

In a bid to drive adoption of Claude for Small Business in the huge US market, Anthropic is taking the offering on the road with a 10-city US tour offering free, half-day AI training workshops for small businesses, kicking off in Chicago yesterday (14 May).

The new direction comes at a time when Anthropic is in discussions with investors to raise between $30bn and $50bn in new funding at a valuation of up to $950bn, a deal that would see the Claude maker surpass rival OpenAI as the world’s most valuable artificial intelligence start-up.

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

On Thursday, Microsoft shared mitigations for a high-severity Exchange Server vulnerability exploited in attacks that allow threat actors to execute arbitrary code via cross-site scripting (XSS) while targeting Outlook on the web users.

Microsoft describes this security flaw (CVE-2026-42897) as a spoofing vulnerability affecting up-to-date Exchange Server 2016, Exchange Server 2019, and Exchange Server Subscription Edition (SE) software.

While patches aren’t yet available to permanently fix the vulnerability, the company added that the Exchange Emergency Mitigation Service (EEMS) will provide automatic mitigation for Exchange Server 2016, 2019, and SE on-premises servers.

“An attacker could exploit this issue by sending a specially crafted email to a user. If the user opens the email in Outlook Web Access and certain interaction conditions are met, arbitrary JavaScript can be executed in the browser context,” the Exchange Team said.

“Using EM Service is the best way for your organization to mitigate this vulnerability right away. If you have EM Service currently disabled, we recommend you enable it right away. Please note that EM Service will not be able to check for new mitigations if your server is running Exchange Server version older than March 2023.”

EEMS was introduced in September 2021 to provide automated protection for on-premises Exchange servers, securing them against ongoing attacks by applying interim mitigations for high-risk (and likely actively exploited) vulnerabilities.

EEMS runs as a Windows service on Exchange Mailbox servers and is automatically enabled on servers with the Mailbox role. The security feature was added after many hacking groups exploited ProxyLogon and ProxyShell zero-days (which lacked patches or mitigation information) to breach Internet-exposed Exchange servers.

Admins with servers in air-gapped environments can also mitigate the flaw by downloading the latest Exchange on-premises Mitigation Tool (EOMT) version and applying the mitigation by running the script via an elevated Exchange Management Shell (EMS) with one of the following commands:

However, it’s important to note that applying the mitigation measures on vulnerable servers will cause issues, including:

• OWA Print Calendar functionality might not work. As a workaround, Microsoft suggested copying the data, taking a screenshot of the calendar you want to print, or using the Outlook Desktop client.
• Inline images might not display correctly in the recipients’ OWA reading pane. As a workaround, users are advised to send images as email attachments or use the Outlook Desktop client.
• OWA light (OWA URL ending in /?layout=light) does not work properly (this feature was deprecated several years ago and is not intended for regular production use).

Microsoft plans to release patches for Exchange SE RTM, Exchange 2016 CU23, and Exchange Server 2019 CU14 and CU15, but says that updates for Exchange 2016 and 2019 will only be available to customers enrolled in the Period 2 Exchange Server ESU program.

BleepingComputer also reached out to Microsoft with questions about the attacks, but a response was not immediately available.

In October, weeks after Exchange 2016 and 2019 reached the end of support, the Cybersecurity and Infrastructure Security Agency (CISA) and the National Security Agency (NSA) released guidance to help IT admins harden Microsoft Exchange servers against attacks.

Over the last 5 years, CISA has added 19 Microsoft Exchange Server vulnerabilities to its list of actively exploited security flaws, 14 of which were also abused in ransomware attacks.

Automated pentesting tools deliver real value, but they were built to answer one question: can an attacker move through the network? They were not built to test whether your controls block threats, your detection rules fire, or your cloud configs hold.

This guide covers the 6 surfaces you actually need to validate.

Download Now

from the the-liability-tax dept

The recent “internet addiction” verdicts against Apple, Meta, and YouTube drew applause from those eager to see big tech take a hit. But look behind the headlines and the result is something else entirely. These cases won’t help children. They will fuel a litigation plague that raises costs, chills innovation and hits smaller companies the hardest. 

The legal theory behind these cases tries to work around Section 230 by shifting the focus from user content to product design. Plaintiffs argue that features like infinite scroll or “like” buttons create harm independent of users’ personal content. It is a creative argument. It is also a slippery slope with no clear limiting principle.  

Once product design becomes the hook for liability, any widely used product becomes a target. Newspapers, magazines and even packaged goods design headlines with catchy taglines to capture attention. Platforms do the same with feeds, to deliver value to their users. Labeling these as “addictive” design shouldn’t be seen as a viable path to sidestepping Section 230. 

This shift also has broader economic consequences. 

Trial lawyer lawsuits do not stay in the courtroom, they are priced into everything. Companies pay more for insurance, more for compliance, and more for legal defense. Those costs flow through to consumers in the form of higher prices and fewer options. At a moment when affordability dominates national conversations, this is a factor we cannot ignore. 

These cases are shaped by a litigation system that rewards scale and escalation. They are enormously expensive and often backed by third-party funders, which drives plaintiffs’ lawyers to seek the highest possible damages. In last month’s Los Angeles trial, plaintiffs asked for billions but secured just $6 million, about 0.5% of what was requested. Even that figure is diminished when measured against the cost of bringing the case. And when outcomes fall short, the incentive is to pursue more cases or larger awards to justify the investment. 

This burden is uniquely American. U.S. companies face a level of litigation exposure that most global competitors simply do not. That gap acts as an innovation tax on American firms, particularly small and early-stage companies that drive job creation and new ideas. We should be asking how to reduce that burden, not expand it. 

Roughly 80% of CTA’s members are small or early-stage companies. They do not have the budgets or legal teams to absorb years of litigation risk. For them, the threat of open-ended lawsuits is not theoretical. It shapes what they build, how they build it, and whether they can exist it at all.  

This is how an innovation economy slows without a single vote in Congress. Startups pull back, new features go unbuilt, and investment shifts away from risk. Over time, innovation slows, and momentum shifts from startups to incumbents. 

None of this means concerns about children’s online experiences should be dismissed. They should be taken seriously. But lawsuits are blunt instruments that do little to address the underlying issues. 

There are better and more effective paths. 

Platforms have already invested heavily in tools that give parents real control over how their children use technology. Supervised accounts, screen time limits, content filters, and transparency into usage patterns are improving quickly and becoming easier to use. Industry efforts like NetChoice’s Digital Safety Shield build on that progress by putting parents in charge rather than outsourcing decisions to courts. 

Congress also has a clear role. A national privacy law that protects personal data, including children’s information, would provide real safeguards while giving companies a consistent set of rules. What Congress should avoid is layering on vague obligations that invite more litigation. It’s delayed action for years. It should not delay further. 

And parents remain central. Technology has changed, but the need for engagement has not. Knowing what children are doing online, setting boundaries and staying involved matters more than any verdict. 

Social media is a powerful tool with real benefits and real risks. The right response is to manage those tradeoffs in a practical way that protects children without undermining innovation. 

Recent verdicts move us in the opposite direction. They reward litigation, raise costs and make it harder for the next generation of companies to succeed. 

We should focus on solutions that help children, not expand a system that is already very good at benefiting trial lawyers.

Michael Petricone is the Senior VP of Government Affairs at the Consumer Technology Association.

Filed Under: innovation, internet addiction, liability, negligence, product liability, section 230, trial lawyers

Companies: meta

One of the key challenges of current multi-agent AI systems is that they communicate by generating and sharing text sequences, which introduces latency, drives up token costs, and makes it difficult to train the entire system as a cohesive unit. 

To overcome this challenge, researchers at University of Illinois Urbana-Champaign and Stanford University developed RecursiveMAS, a framework that enables agents to collaborate and transmit information through embedding space instead of text. This change results in both efficiency and performance gains. 

Experiments show that RecursiveMAS achieves accuracy improvement across complex domains like code generation, medical reasoning, and search, while also increasing inference speed and slashing token usage. 

RecursiveMAS is significantly cheaper to train than standard full fine-tuning or LoRA methods, making it a scalable and cost-effective blueprint for custom multi-agent systems.

The challenges of improving multi-agent systems

Multi-agent systems can help tackle complex tasks that single-agent systems struggle to handle. When scaling multi-agent systems for real-world applications, a big challenge is enabling the system to evolve, improve, and adapt to different scenarios over time. 

Prompt-based adaptation improves agent interactions by iteratively refining the shared context provided to the agents. By updating the prompts, the system acts as a director, guiding the agents to generate responses that are more aligned with the overarching goal. The fundamental limitation is that the capabilities of the models underlying each agent remain static. 

A more sophisticated approach is to train the agents by updating the weights of the underlying models. Training an entire system of agents is difficult because updating all the parameters across multiple models is computationally non-trivial.

Even if an engineering team commits to training their models, the standard method of agents communicating via text-based interactions creates major bottlenecks. Because agents rely on sequential text generation, it causes latency as each model must wait for the previous one to finish generating its text before it can begin its own processing. 

Forcing models to spell out their intermediate reasoning token-by-token just so the next model can read it is highly inefficient. It severely inflates token usage, drives up compute costs, and makes iterative learning across the whole system painfully slow to scale. 

How RecursiveMAS works

Instead of trying to improve each agent as an isolated, standalone component, RecursiveMAS is designed to co-evolve and scale the entire multi-agent system as a single integrated whole. 

The framework is inspired by recursive language models (RLMs). In a standard language model, data flows linearly through a stack of distinct layers. In contrast, a recursive language model reuses a set of shared layers that processes the data and feeds it back to itself. By looping the computation, the model can deepen its reasoning without adding parameters.

RecursiveMAS extends this scaling principle from a single model to a multi-agent architecture that acts as a unified recursive system. In this setup, each agent functions like a layer in a recursive language model. Rather than generating text, the agents iteratively pass their continuous latent representations to the next agent in the sequence, creating a looped hidden stream of information flowing through the system. 

This latent hand-off continues down the line through all the agents. When the final agent finishes its processing, its latent outputs are fed directly back to the very first agent, kicking off a new recursion round. 

This structure allows the entire multi-agent system to interact, reflect, and refine its collective reasoning over multiple rounds entirely in the latent space, with only the very last agent producing a textual output in the final round. It is like the agents are communicating telepathically as a unified whole and the last agent provides the final response as text.

The architecture of latent collaboration

To make continuous latent space collaboration possible, the authors introduce a specialized architectural component called the RecursiveLink. This is a lightweight, two-layer module designed to transmit and refine a model’s latent states rather than forcing it to decode text. 

A language model’s last-layer hidden states contain the rich, semantic representation of its reasoning process. The RecursiveLink is designed to preserve and transmit this high-dimensional information from one embedding space to another. 

To avoid the cost of updating every parameter across multiple large language models, the framework keeps the models’ parameters frozen. Instead, it optimizes the system by only training the parameters of the RecursiveLink modules.

To handle both internal reasoning and external communication, the system uses two variations of the module. The inner RecursiveLink operates inside an agent during its reasoning phase. It takes the model’s newly generated embeddings and maps them directly back into its own input embedding space. This allows the agent to continuously generate a stream of latent thoughts without generating discrete text tokens. 

The outer RecursiveLink serves as the bridge between agents. Because agents in a real-world system might use different model architectures and sizes, their internal embedding spaces have entirely different dimensions. The outer RecursiveLink includes an additional layer designed to match the embeddings from one agent’s hidden dimension with the next agent’s embedding space.

During training, first, the inner links are trained independently to warm up each agent’s ability to think in continuous latent embeddings. Then, the system enters outer-loop training, where the diverse, frozen models are chained together in a loop, and the system is evaluated based on the final textual output of the last agent. 

The only thing that gets updated in the training process is the RecursiveLink parameters and the original model weights remain unchanged, similar to low-rank adaptation (LoRA). Another advantage of this system comes into effect when you have multiple agents on top of the same backbone model. 

If you have a multi-agent system where two agents are built on the exact same foundation model acting in different roles, you do not need to load two copies of the model into your GPU memory, nor do you train them separately. The agents will share the same backbone as the brain and use the RecursiveLink as the connective tissue.

RecursiveMAS in action

The researchers evaluated RecursiveMAS across nine benchmarks spanning mathematics, science and medicine, code generation, and search-based question answering. They created a multi-agent system using open-weights models including Qwen, Llama-3, Gemma3, and Mistral. These models were assigned roles to form different agent collaboration patterns such as sequential reasoning and mixture-of-experts collaboration. 

RecursiveMAS was compared to baselines under identical training budgets, including standalone models enhanced with LoRA or full supervised fine-tuning, alternative multi-agent frameworks like Mixture-of-Agents and TextGrad, and recursive baselines like LoopLM. It was also compared to Recursive-TextMAS, which uses the same recursive loop structure as RecursiveMAS but forces the agents to explicitly communicate via text.

RecursiveMAS achieved an average accuracy improvement of 8.3% compared to the strongest baselines across the benchmarks. It excelled particularly on reasoning-heavy tasks, outperforming text-based optimization methods like TextGrad by 18.1% on AIME2025 and 13% on AIME2026. 

Because it avoids generating text at every step, RecursiveMAS achieved 1.2x to 2.4x end-to-end inference speedup. RecursiveMAS is also much more token efficient than the alternative. Compared to the text-based Recursive-TextMAS, it reduces token usage by 34.6% in the first round of the recursion, and by round three, it achieves 75.6% token reduction. RecursiveMAS also proved remarkably cheap to train. Because it only updates the lightweight RecursiveLink modules, which consist of roughly 13 million parameters or about 0.31% of the trainable parameters of the frozen models, it requires the lowest peak GPU memory and cuts training costs by more than half compared to full fine-tuning.

Enterprise adoption

The efficiency gains — lower token consumption, reduced GPU memory requirements, and faster inference — are intended to make complex multi-step agent workflows viable in production environments without the compute overhead that limits enterprise agentic deployments. The researchers have released the code and trained model weights under the Apache 2.0 license.

Looking for the most recent Mini Crossword answer? Click here for today’s Mini Crossword hints, as well as our daily answers and hints for The New York Times Wordle, Strands, Connections and Connections: Sports Edition puzzles.

Need some help with today’s Mini Crossword? I’ll tell you, 6-Across was a completely new fact to me, and I even had those particular pets when I was a kid. Read on for all the answers. And if you could use some hints and guidance for daily solving, check out our Mini Crossword tips.

If you’re looking for today’s Wordle, Connections, Connections: Sports Edition and Strands answers, you can visit CNET’s NYT puzzle hints page.

Read more: Tips and Tricks for Solving The New York Times Mini Crossword

Let’s get to those Mini Crossword clues and answers.

Mini across clues and answers

1A clue: De Armas of “Knives Out”
Answer: ANA

4A clue: Jack ___, five-time “S.N.L.” host
Answer: WHITE

6A clue: Class pets that thump their feet to warn of approaching kindergarteners
Answer: GERBILS

8A clue: Boxer Muhammad or Laila
Answer: ALI

9A clue: Reach 0% battery
Answer: DIE

10A clue: Big name in hair loss prevention
Answer: ROGAINE

12A clue: Jack ___, six-time “S.N.L.” musical guest
Answer: WHITE

13A clue: Attempt
Answer: TRY

Mini down clues and answers

1D clue: “Okay then …”
Answer: ALRIGHT

2D clue: Catch, as a criminal
Answer: NAB

3D clue: Notable feature of lemon and grapefruit juice
Answer: ACIDITY

4D clue: Underneath
Answer: BELOW

5D clue: Actor Kevin of “Dave”
Answer: KLINE

6D clue: Long-nosed fish
Answer: GAR

7D clue: Understand
Answer: SEE

11D clue: Make public, as a grievance
Answer: AIR