Xiaomi today launched the XRING O3 mobile chip with 3nm process tech. It will show up with the Xiaomi 18 Fold smartphone next month. The company also revealed how the new self-developed chipset version differs from the past generation.
The tech giant hosted the Xiaomi XRING Chip Technology Communication Conference today, where it introduced its new-generation 3nm processor – XRING O3.
It is the first AI flagship mobile processor with major advancements. The tech maker has listed the respective chip on AnTuTu, where it scored 5228014 points.
During the event, Xiaomi revealed that XRING O3 comes with an upgraded chip architecture. It runs on TSMC’s 3nm process technology and has more than 24 billion transistors – a 26% increase compared to the XRING O2 chip.
XRING O3 is the world’s first mobile chipset to support LPDDR6 and delivers speeds up to 10,667Mbps with a memory bandwidth of 113.8GB/s – 48% more than XRING O1. It further has a die size of 133mm².
The new chip has undergone noteworthy changes – from AI to NPU and other chip components. The company has revamped it for its MiMo on-device AI models with up to 200 TPOS of tensor performance and 3.13 TFLOPS of vector performance.
(Image Credits: Weibo)
In addition, Xiaomi has improved power efficiency and responsiveness of the new chip with a revamped internal architecture. Thus, it offers 82ns latency.
AI performance for XRING O3 has been increased by 45%. Other changes for the chipset include two CPU AI acceleration units and eight neural-network GPU accelerators that expand the chip’s AI capabilities to a greater extent.
Besides the XRING O3, the Xiaomi chip briefing talked about XRINGO100 – an AI accelerator for large language models. It is the first to feature a 6nm 3D wafer-on-wafer advanced stacking design and provides 1.22TB/s bandwidth.
On the other hand is the XRING D100 – the first 3nm high-compute AI chip for smart driving in China.l
Lucy is a compact but capable 1.7B model focused on agentic web search and lightweight browsing. Built on Qwen3-1.7B, Lucy inherits deep research capabilities from larger models while being optimized to run efficiently on mobile devices, even with CPU-only configurations.
We achieved this through machine-generated task vectors that optimize thinking processes, smooth reward functions across multiple categories, and pure reinforcement learning without any supervised fine-tuning.
The main theoretical idea: reasoning as a “dynamic task vector machine”
The paper’s most important conceptual contribution is its treatment of the model’s <think> ... </think> reasoning process.
Rather than viewing the text inside <think> tags as merely an explanation or chain-of-thought transcript, the authors argue that the generation process itself can act as a dynamic task vector machine. In their framing, the model’s internal task representation is progressively:
constructed,
updated,
refined,
and aligned with the next tool action.
A useful paraphrase of the paper’s argument is:
The model should not merely generate a reasoning trace after deciding what to do. Its reasoning process should help construct and maintain the representation of the task that guides subsequent searches and answers.
This is intended to address a common problem in agentic systems: a model may know how to call a search tool but still produce unreliable, repetitive, or incoherent search trajectories.
The authors therefore focus less on teaching a fixed search procedure and more on stabilizing the reasoning loop that determines when and how the model searches.
Training method
Search-efficiency reward
The model is rewarded for visiting and examining retrieved documents rather than repeatedly issuing new searches. If it performs more searches than document visits, the paper applies a penalty.
This encourages Lucy to make better use of information it has already retrieved instead of continually broadening the search.
Efficient-thinking reward
The authors penalize excessively long internal reasoning. Their reward distribution is centered around approximately 35 reasoning tokens, with a strong penalty for unusually long reasoning spans.
The motivation is to combat “overthinking”—reasoning that is verbose, repetitive, or distracts the model from using tools effectively.
The paper explicitly says that these parameters were selected heuristically, based on observations from preliminary experiments.
Data and evaluation
The training dataset contains 10,325 question-answer pairs. It is described as an adaptation of the MuSiQue-Ans dataset, which contains multi-step questions requiring information gathering and reasoning.
For evaluation, the authors use SimpleQA and connect the evaluation system to an MCP server providing web-search and scraping tools.
The headline result is:
Lucy: 78.3% on SimpleQA with MCP
1.7B parameters
The paper compares this with several systems:
DeepSeek-671B with MCP: 78.2%
Jan-nano, 4B: 80.7%
Jan-nano-128k, 4B: 83.2%
The authors emphasize that Lucy performs approximately on par with the much larger DeepSeek model in this setup, while being far smaller.
However, the comparison must be interpreted carefully: the result is for tool-augmented evaluation with MCP, not necessarily for a pure, closed-book language-model test. Lucy’s web access is an essential part of its performance.
Emergent behavior: skipping unnecessary reasoning
One of the more interesting observations is that Lucy learned to suppress reasoning during predictable, low-decision operations.
The paper gives a simplified example:
When deciding what to search, Lucy generates a <think> segment.
When merely reading or processing a retrieved document, it may omit explicit thinking.
When deciding on a more complex follow-up search, it resumes thinking.
In paraphrase, Lucy appears to allocate reasoning selectively:
It spends computation on uncertain decisions, such as formulating a query, while skipping redundant deliberation during deterministic actions, such as reading a result.
The authors report a 17.8× latency improvement in the observed example. They also stress that this is only a single illustrative case and that its generality remains unproven.
This is one of the paper’s strongest practical ideas: reasoning need not be uniformly applied at every step. A useful agent may reason deeply only where the next action is ambiguous.
An important limitation: more thinking cannot replace missing knowledge
The paper is unusually candid about Lucy’s limitations.
The authors report cases where Lucy failed not because the needed information was absent from the web, but because the model lacked basic knowledge about the entities involved. As a result, it formed incorrect assumptions and generated poor search queries.
Their conclusion is that test-time reasoning has a ceiling:
Thinking longer cannot reliably compensate for missing conceptual grounding.
The paper contrasts Lucy with the larger Jan-nano model, which was better able to notice and correct its own mistaken assumptions during research. Lucy, by comparison, sometimes continued searching within an incorrect conceptual frame.
caveats are that:
the evaluation depends on external web tools;
the training dataset is relatively small and specialized;
several reward components are heuristic;
the correctness reward is described as a substring match, which may not perfectly measure semantic correctness;
the latency improvement comes from an illustrative scenario;
and the authors acknowledge that Lucy can fail when it lacks foundational knowledge needed to formulate sensible searches.
the book is not simply nostalgic about clubs and bowling leagues; it argues that community connection changes how institutions and everyday life function.
evidence draws from roughly 500,000 interviews over decades, alongside many other valid indicators
Bowling Alone: The Collapse and Revival of American Community is a 2000 nonfiction book by Robert D. Putnam. It was developed from his 1995 essay entitled " Bowling Alone: America's Declining Social Capital " in the Journal of Democracy. Putnam surveys the decline of social capital in the United States since 1950. He has described the reduction in all the forms of in-person social intercourse upon which Americans used to found, educate, and enrich the fabric of their social lives. He argues that this undermines the active civic engagement which a strong democracy requires from its citizens.
Contents
Putnam discussed ways in which Americans disengaged from community involvement, including decreased voter turnout, attendance at public meetings, service on committees, and work with political parties. Putnam also cited Americans' growing distrust in their government. Putnam accepted the possibility that this lack of trust could be attributed to "the long litany of political tragedies and scandals since the 1960s",1 but believed that this explanation was limited when viewing it alongside other "trends in civic engagement of a wider sort".1
A review in Kirkus Reviews praised the book for being understandable for non-academic readers, and said that overall it was an "exhaustive and carefully argued study." 6The Economist called it "a prodigious achievement." 7 C. S. Fischer, a sociology professor from the University of California, gave a positive review. Although he criticized a few of Putnam's interpretations of the data and felt that "social capital " was an awkward metaphor, he nevertheless called it "a 10-pin strike, a major contribution to study of social networks and social cohesion" with particular praise for its wide use of data.8
Everett Carll Ladd claimed that Putnam completely ignored existing field studies, most notably the landmark sociological Middletown studies,9 which during the 1920s raised the same concerns he does today, except the technology being attacked as promoting isolation was radio instead of television and video games.10
Other critics questioned Putnam's major finding—that civic participation has been declining. Journalist Nicholas Lemann proposed that rather than declining, civic activity in the US had assumed different forms. While bowling leagues and many other organizations had declined, others like youth soccer leagues had grown.11 He also points out that the thesis of Bowling Alone contradicts an implicit assumption of Putnam's previous book Making Democracy Work —that a tradition of civic engagement is incredibly durable over time.
In their 2017 book One Nation After Trump, Thomas E. Mann, Norm Ornstein and E. J. Dionne wrote that the decline of social and civic groups that Putnam documented was a factor in the election of Donald Trump as "many rallied to him out of a yearning for forms of community and solidarity that they sense have been lost." 12
"Bowling" (PDF). Spalding's Athletic Library. Vol. 1, no. 3. New York: American Sports Publishing Company. December 1892. Archived from the original (PDF) on March 27, 2020.↩
Bad bots account for 91% of the traffic reaching one retailer's DDR5 memory product pages, roughly 10 automated requests for every legitimate visit, according to a LinkedIn post this week from Jérôme Segura, VP of threat research at bot-mitigation firm DataDome. The figure updates research DataDome published in March, which measured the ratio at about 6:1 across several e-commerce sites.
Over the same period, the cheapest 128GB DDR5-6400 kit tracked by Tom's Hardware has reached $3,399, ten times its record low pricing. The Thales 2026 Bad Bot Report puts bad bots at 40% of all web traffic, less than half the share Segura reports on DDR5 listings.
DataDome's March report documented a single operation that generated more than 10 million blocked requests, with a one-hour sample showing 91 DDR5 listings each polled around 551 times, or once every 6.5 seconds. The requests carried cache-busting parameters to defeat cached stock data and targeted industrial module makers such as Micron and Apacer, as well as DIMM socket suppliers Amphenol and TE Connectivity, alongside consumer brands.
Latest Videos FromTom's Hardware
Segura wrote on LinkedIn that retailers carrying memory should assume their "traffic analytics are probably understating the problem significantly."
(Image credit: DataDome)
Keep in mind that the 91% figure comes from one unnamed retailer, and DataDome hasn't published a sample window or its methodology. DataDome also sells the product that blocks this traffic, and page requests ultimately aren't purchases, so the ratio measures monitoring pressure on listings rather than how much stock bots actually end up securing, which can’t easily be quantified.
Apacer CEO C.K. Chang said in July that DRAM supply to independent module makers could fall to 30% of 2026 levels next year, and TrendForce expects conventional DRAM contract prices to rise another 13% to 18% in Q3 after roughly 60% in Q2. Hyperscale buyers have reportedly placed deposits against most of 2027's DRAM output.
A 32GB DDR5-6000 kit that averaged $72 this time last year now averages $392 on PCPartPicker. While graphics card and console scalping in 2020 and 2021 ran against scheduled restocks that eventually outpaced resale demand, no equivalent restock date exists for consumer DDR5 before 2027 at the earliest.
Framework stopped selling standalone RAM in November last year, citing scalpers, and Micro Center sells its CPU, motherboard, and 32GB DDR5 bundles in-store only with a one-per-household limit. Newegg and Amazon have run repeated CPU-plus-memory bundles that price a 32GB kit well below its standalone cost, which ties each kit to a processor a reseller then has to move as well. On eBay, we found G.Skill's Trident Z5 Neo 32GB kit listed at $836.54 in December against a $429.99 retail price, roughly double, and about seven times the $117.99
As demand for artificial intelligence infrastructure surges, Alibaba Group Holding says it can deliver new data centres in a fraction of the usual time while cutting construction costs by 10 per cent through its proprietary modular architecture.
Using CUBE 5.0, Alibaba Cloud had slashed the delivery time for large-scale AI data centres to just 100 days, according to a report by state-backed newspaper China Securities Journal on Tuesday.
That compared with standard domestic delivery times of six to 12 months, and 12 to 18 months in the US, the report said.
Alibaba Cloud, the AI and cloud computing services unit of Alibaba, also expected to more than double its production efficiency this year, measured by the volume of modular data centres delivered within a given time frame, it added.
Unveiled in 2024, CUBE 5.0 increased the overall modularity rate across five key systems – power supply, cooling, security, intelligent management and fire protection – from 30 per cent to 90 per cent in its latest iteration, according to the report.
Unlike traditional construction, where power, cooling and IT systems are installed sequentially on-site, prefabricated modular data centres split infrastructure into separate components that are manufactured simultaneously in factories. These components are then shipped in container-like units and assembled on-site like building blocks.
HTML has been gobbling up swathes of what used to be JavaScript’s remit. This page lists a bunch of dynamic functionality that we can now achieve with just HTML.
prospect.org
Private Intelligence Firms Are Selling Dossiers on AI and Data Center Critics - The American Prospect
Daniel Boguslaw
7 - 9 minutes
Private intelligence firms are selling dossiers about critics of artificial intelligence and the data centers powering them, including trying to peddle them to federal regulators, according to documents obtained by the Prospect. The offerings show that even as voter hostility toward data centers has emerged as an election-altering, bipartisan issue, corporate spy shops are hawking reports that frame widespread dissent in the language of counterterrorism to both private- and public-sector clients.
The drivers of negative sentiment toward data centers and AI include environmental concerns, fears of job loss, and increased utility prices near the sprawling compounds housing server farms and processing facilities. More broadly, people across the ideological spectrum don’t appreciate Big Tech interests dictating their local economic development. Hundreds of grassroots organizations have emerged across the country, seeking to curb data center construction and limit the negative effects of AI.
More from Daniel Boguslaw
But the threat products and reports obtained by the Prospect are less concerned with altering the root cause of this dissent, and more focused on the threat posed by an enraged American populace to corporations’ bottom line.
In a product offering from the corporate intelligence company RANE Network sent to the Federal Energy Regulatory Commission, the corporate firm offered the federal agency—which regulates the interstate transmission and sale of electricity, gas, and oil—a tailored report on the threat posed by anti-tech sentiment.
“RANE can provide an assessment covering rising anti-technology sector sentiment among both the public and governments around the world, identifying specific implications for your organization,” the offering reads. Topics that RANE promises to cover include:
What has brought about an increase in anti-technology sector sentiment?
What role do misinformation and conspiracy theories play in shaping hostility toward the tech sector?
What risks does anti-technology sector sentiment pose to critical supply chains?
What are the potential implications of geopolitical tensions on the operations of technology companies?
How might increasing regulatory scrutiny impact the operations of major tech firms?
What actions can technology companies take to mitigate risks from anti-technology sentiment?
That email was sent in March 2025. This May, RANE offered a webinar on the coming anti-tech backlash. In the hour-long presentation reviewed by the Prospect, RANE cyber and intelligence analysts described the threat posed to technology companies by rising animosity toward Silicon Valley. RANE Network did not respond to multiple requests for comment.
“There’s backlash on a lot of fronts and there are a lot of actors converging on this. So we’ve seen protests from creatives who are worried about AI taking over their copyrighted works,” one analyst told the audience. “There’s also environmental protests. Labor unions are protesting. We’ve seen protests from employees who may have been laid off or are worried about potential layoffs.”
The analysts also warned about protesters concerned about civil liberties and human rights violations arising from AI, and “protests about the impacts on mental health and just general anti-tech or neo-Luddite sentiment that is really driving this activity.”
In 2020, RANE acquired the geopolitical intelligence firm Stratfor. Nine years prior, in 2011, Stratfor was breached by Anonymous, a hacker collective that published millions of internal emails on WikiLeaks. The internal communications detailed the company’s work monitoring the critics of major conglomerates, including Dow Chemical and Coca-Cola.
The specter of hacktivism continues to haunt RANE analysts. “Anti-establishment” and “anti-capitalist groups” alongside Anonymous were referenced during the presentation before a warning that “anti-tech and environmental protest … groups also have a history of some of them going beyond just protest movements to also include, you know, violent extremism,” an analyst said.
In addition to describing the general patterns of concerning groups, presenters also discussed surveilling “online chatter.” The analysts discussed how movements emerge from online spaces, and warned that things can spiral out of control quickly, which is all the more reason to surveil those spaces.
The push to surveil and monitor techno-skeptics comes at the same time that federal agencies have begun circulating their own surveillance memorandums of a new domestic extremist category: “anti-tech extremism.” As Wired reported in May, this new threat category is being used by the FBI and DHS to target constitutionally protected speech and assembly, regardless of whether or not participants have committed, or intend to commit, a crime. In leaked State Department documents obtained by journalist Ken Klippenstein, Secretary of State Marco Rubio informed employees that the agency was monitoring a new alliance of “militant anti-tech and eco-terrorist movements.”
Kroll, the world’s largest investigation and corporate intelligence firm, has also begun selling “risk intelligence” on “public controversy, labor tensions, executive visibility, activism, geopolitical pressure” and “reputational events” to data center operators, according to a June article posted on the firm’s website. Those interested in Kroll’s investigative and monitoring services are encouraged to reach out to its enterprise security risk management team.
Meanwhile, the online surveillance company Liferaft—a subsidiary of the world’s second-largest security firm, Securitas—published a report on July 31 summarizing months of online surveillance of alleged “Threats to AI Infrastructure and Executives.” In the report, Liferaft scanned and analyzed thousands of social media and online forum posts referencing “anger at AI executives,” “opposition to new data centers,” and “grievance over the resources these facilities consume.”
Liferaft’s monitoring tracks the ebb and flow of hostility toward AI companies and executives over the course of several months, but also explicitly warns of a shift in so-called organizational intent, “from personal frustration to coordinated rhetoric … [that] is something threat intelligence professionals pay close attention to, because it tracks with how movements organize.”
The same month Liferaft published its report, hundreds gathered in San Francisco’s downtown to protest AI companies and the existential risks they pose. Anthropic, Google DeepMind, and OpenAI offices were all visited by protesters, including the AI theorist and spiritual leader of the AI-skeptic movement Eliezer Yudkowsky, whose book title “If Anyone Builds It, Everyone Dies” has become a rallying cry for critics of sentient AI. The constitutionally protected protest fits neatly in all three firms’ criteria as a surveillance target.
GRAND LAKE — The future of the Colorado River, which provides water to communities across Colorado, is more clear — at least for the next two years.
The federal government released Friday its near-term management plan for the river system, which provides vital water supplies to nearly 40 million people. The river, which begins in the hills above Grand Lake in western Colorado, is the workhorse of the West and the reason communities and economies have been able to grow to what they are today. But it has also been overstressed by rising temperatures, ongoing drought and persistent demands while water regulations have been slow to address the looming need for water restrictions.
Now, downstream states will bear the brunt of those mandatory water cuts, totaling 1.25 million acre-feet, under the new river plan for 2027 and 2028. Upstream states, like Colorado, are being called on to conserve water and help avoid crisis lows at the river basin’s key reservoirs, but do not have to cut their water use.
“These decisions provide a water management strategy for Basin stakeholders to respond to the prolonged drought by incorporating flexible tools and voluntary actions while leaving room for consensus agreements,” Andrea Travnicek, the Bureau of Reclamation’s assistant secretary of water and science, said in a statement Friday.
The basin’s communities have endured three years of heated debates over how to manage the river’s key reservoirs, Lake Mead and Lake Powell, after 2026, when the current rules expire.
In the past, the river’s management has frequently been determined by agreements between the basin states. But the states failed to reach an agreement during tense negotiations over everything from the level of mandatory cuts to the time period for the next set of rules. Those negotiations are not over: The federal government has said it will incorporate a seven-state agreement into its management plan as soon as the states can reach a consensus.
Washington officials moved on with the federal planning process when states missed several deadlines to contribute a united proposal. It announced this summer that the next set of rules will last for 10 years, and states and tribes will weigh in on reservoir management with new operational plans every two years.
The Department of the Interior and its subagency, the Bureau of Reclamation, said the 2027-2028 plan is based on feedback from the states and tribal nations in the basin.
The Secretary of the Interior, who has the authority to decide shortages for the downstream states, plans to cut releases from Hoover Dam by 1.25 million acre-feet each year for the next two years.
If the Lower Basin states stick to their proposed plan for distributing the cuts, Arizona would reduce its water use by 760,000 acre-feet, California would cut its use by 440,000 acre-feet and Nevada would have to cut its use by 50,000 acre-feet.
California applauded Lower Basin states for stepping up in a statement Friday.
“California, Arizona, and Nevada have shown that states can compromise, make difficult decisions, and reduce water use when the river demands it,” JB Hamby, Colorado River negotiator for California, wrote. “But three states cannot carry the responsibility of all seven.”
Colorado’s negotiating team did not respond to requests for comment Friday morning.
(Who takes cuts has been a tension point in negotiations. While Lower Basin states have argued everyone needs to share water cuts, Upper Basin officials say they already have water cuts imposed by nature and do not use their full share of Colorado River water. Upstream states also allege that their downstream counterparts have repeatedly gone over their legal water allowances.)
“We remain committed to developing a solution that works for all the Colorado River states, tribes, and interests,” wrote Becky Mitchell, Colorado’s top negotiator and Commissioner to the Upper Colorado River Commission in a prepared statement Friday. “We also acknowledge that the Upper Division States face mandatory cuts already, including water rights that predate the 1922 Colorado River Compact by decades.”
The plan for the next two years also includes voluntary conservation targets for the downstream states totaling 700,000 acre-feet of water savings over the two-year period. That’s in addition to the required cuts.
One acre-foot roughly equals the annual water use of two to four households. Colorado uses about 5.3 million acre-feet of water each year from multiple major river systems. It used less than 2 million acre-feet of Colorado River water on average between 2021 and 2025, according to federal records.
While the Interior Department can decide shortages for the downstream states, it does not have the same authority in the four upstream states, including Colorado. Its authority mostly involves managing federal water projects, like Blue Mesa Dam and its sister dams, Morrow Point and Crystal, in western Colorado.
Colorado and the other upstream states, New Mexico, Wyoming and Utah, will likely enter into an agreement to use those federal dams, including Blue Mesa, to help avoid critically low water levels at Lake Powell, on the Utah-Arizona border.
The bureau is aiming to keep the reservoir’s water elevation above critical levels over the next two years. By setting the minimum elevation at 3,510 feet above sea level, officials said they will have a buffer before the water drops to more problematic lows.
“That will give us a little bit more time on decision making if we need to make some changes somewhere else with the system,” Travnicek said in a press conference Friday.
If the water level falls below 3,490 feet, Glen Canyon Dam can no longer generate hydropower, cutting a source of renewable, cheap and reliable energy for communities across the West.
Powell’s elevation was 3,519 feet above sea level Thursday. In the worst case scenario, it’ll fall below 3,510 feet by December, but the bureau anticipates it can keep the water level above the target elevation, according to federal projections.
The combined contents of Lake Powell and Lake Mead have not been this low since before Lake Powell began filing. Both Lake Powell and Lake Mead have hit record lows the last few weeks.
The new management plan alone cannot reduce the risk to the entire Colorado River Basin, said Jennifer Pitt, Colorado River program director for The Audubon Society. She’s pushing for substantial investment in conserving water and improving the watersheds that feed the river system.
“What’s clear after seeing the projections for reservoir elevations over the next couple of years is that the buffer the Colorado River reservoirs provided over the last many decades is gone,” Pitt wrote in an email to The Colorado Sun. “There is real risk for the cities, farmers and ranchers, Tribes, and the entire natural world that depends on the Colorado River.”
The AI industry’s boldest promise right now is that AI will soon improve itself, with almost no need for human oversight. LLMs can already write code, generate synthetic data for training, and optimize the computer chips they run on. Forecasts of explosive AI progress predict that what researchers call recursive self-improvement is on the horizon.
But a new study suggests that it might take a while for us to get there. The researchers behind it found that AI agents are not yet capable of conducting open-ended AI research—free-form investigations that have no clear-cut answers and require judgment and taste, which may be integral to building self-improving AI.
A multi-institution group of researchers, led by Peter Kirgis and Sayash Kapoor at Princeton University, found that AI agents could solve the engineering problems necessary to do AI research but lacked the judgment and creativity to produce original research at the caliber of papers accepted by a top machine-learning conference. The gap suggests that some of the hyped-up timelines for automating AI research may be running ahead of the evidence.
Most existing research on how agents can automate AI research evaluates their ability to complete narrow tasks with checkable answers, such as solving engineering problems or post-training small language models against a benchmark. But making progress in AI research also requires open-ended thinking—choosing a set of hypotheses, deciding what evidence would settle a question, or knowing when to start over.
To test agents on those kinds of skills, the researchers in the study proposed a new method of evaluation called “shadow evaluation,” which requires the AI to answer a research question from a high-quality unpublished paper.
The researchers asked Anthropic’s Claude Opus 4.8, running on open-source software called OpenClaw, to tackle such questions, in this case from two papers submitted to the prestigious machine-learning conference NeurIPS 2026.
The first question was whether a large language model’s “personas,” which determine its behavior, can be controlled by editing the model’s weights (the billions of numbers that store everything it learns during training). The other asked how to design a detector that points out when a model that makes predictions based on spreadsheet data has become unreliable. Because the papers had not been made public, the agents could not memorize the answers from their training data or find them online.
The agents were given six days, $3,000 in Anthropic API credits, a GPU budget to run the experiments, their own virtual computers, and access to the open web to produce a research paper worthy of publication at a top-tier AI conference. The papers’ original authors graded the agents’ papers as they would evaluate one submitted to a conference.
Those authors rejected both papers.
The agents were capable of all the engineering required to conduct the research, the human scientists found. The agents reviewed the literature, ran hundreds of experiments, and compiled the results.
“On the other hand, the agents were unambiguously bad at carrying out the research itself,” says Kapoor. They ran bizarre experiments (in some cases testing their hypotheses on tiny synthetic datasets), struggled to write intelligibly about their work, and made no novel contribution to their fields. “The papers were nowhere close to the mark when it came to being at the quality of a top AI conference,” he says.
That’s because the agents struggled to muster the creativity and judgment necessary for conducting research. They didn’t do enough to explore different ideas, and they committed to unpromising approaches too quickly. Though the agents developed novel and ambitious hypotheses resembling those that the original authors themselves started with, they rejected them on the basis of very limited data. And they couldn’t backtrack from failing approaches. They could make small pivots but could not fundamentally rethink their approach or try new ones from scratch.
The agents also failed to incorporate feedback from subagents or external AI reviewing tools. Instead of revising their methodology, the agents narrowed their claims and added caveats. They also couldn’t effectively use resources, such as tokens, compute, and time. And they couldn’t follow instructions about things like how much time to spend on different phases of the research or how long their paper could be.
For all their failures, the agents didn’t engage in the misbehavior that researchers call “reward hacking,” hiding or misrepresenting experiments or data. Although subagents, or helper AIs that the main agent spawns to handle pieces of the work, occasionally hallucinated or misrepresented the results, these were caught by the orchestrator agent, the lead AI supervising the project.
The reason AI models are good at research engineering but not at open-ended research may come down to how they’re trained, says Kapoor. Models get good at whatever they can be drilled on in a training regime called reinforcement learning, which is easier to apply to tasks whose success can be checked automatically. “But it’s harder to create environments to train these models when the task itself is open-ended,” he says.
Kapoor says the team is now conducting the experiment with Mythos, Anthropic’s most advanced model, which launched in April. It was subsequently required by the Trump administration to meet various safety restrictions and is now available only to approved organizations. Anthropic did not respond to a request for comment.
There are some limitations to the study. It covered just two research papers, and the original authors knew the papers they were grading were generated by AI agents, which could have colored their evaluations. And the researchers had substantial discretion in designing and executing the study, meaning that their preexisting beliefs and biases could have slipped into the results. Evaluations of open-ended research trade some objectivity for a much richer test than any benchmarks can offer.
Still, the results may temper the claims that recursive self-improvement is on the horizon. In June, Anthropic published a blog post titled “When AI Builds Itself,” charting its progress toward models that speed up their own development. In July, OpenAI advertised the fact that its new model GPT-5.6 Sol had helped post-train a smaller model, saving researchers weeks of work.
The new finding may echo what AI companies are finding internally, regardless of their most optimistic public statements. Anthropic cofounder Jack Clark wrote in his newsletter Import AI that it rhymes with what the company found when it tried to automate some aspects of AI safety research.
“There’s a certain absence of valuable, intuitive creativity in today’s AI systems, and though they’re extraordinarily capable engineers they seem to have a certain property of rote, formulaic thinking that might prevent them [from] being good researchers,” he wrote. He called AI systems’ lack of creativity a “bearish signal on short recursive self-improvement timelines.”
AI companies do have every incentive to develop AI systems that can rapidly accelerate their own progress, just as they did to make the models better at coding. OpenAI has made building an automated AI researcher an explicit goal, and Anthropic identifies self-improving AI as the industry’s next milestone.
“If there is investment and then conscious effort toward this direction, I feel like there would be interesting progress, even if it’s failing currently,” says Najoung Kim, a professor of linguistics and computer science at Boston University who researches how AI agents can automate AI research but did not work on the study. On the other hand, it’s possible that AI progress may be bifurcated. AI systems might race ahead on narrow tasks—the kind that can be scored—while advancing slowly on open-ended research.
The big open question, then, is how crucial open-ended research is to recursive self-improvement—whether AI systems can grind their way there without it, simply by improving on the narrower tasks. “If we look back to the biggest advances in the field, the invention of transformers or the invention of big new architectures that allowed us to make a lot of AI progress—all of those did require creative leaps,” says Kapoor.
“That said, others have this hypothesis that all of what we need for transformative AI, in particular for recursive self-improvement, is already there.” That would include making a model train faster and boosting its benchmark scores.
“That’s frankly the trillion-dollar question right now,” he says.
Republican lawmakers introduced multiple bills targeting utility-scale solar
The Corporation Commission is increasing fees and decreasing reimbursements for rooftop solar customers
Solar and battery projects continue to be built, but natural gas capacity is expanding faster
It has been a rough year for solar advocates in the nation’s sunniest state.
Republican lawmakers, state regulators and some utility companies have historically been unfriendly to renewable energy resources. But this year they’ve attempted to block the construction of utility-scale solar on state land, repeal tax incentives for rooftop solar systems and impose higher fees on ratepayers who choose to adopt solar.
While these attempts were ultimately unsuccessful, there’s no indication that opposition to Arizona’s solar industry is abating anytime soon.
This comes amid what an industry advocate called “unprecedented” federal hostility toward renewable energy, driven by President Donald Trump’s plans to reinvigorate America’s coal industry and accelerate nuclear energy development. Major cuts to federal funding and tax incentives have further compounded state-level attacks.
“It’s death by a thousand cuts,” said Autumn Johnson, the executive director of the Arizona Solar Energy Industries Association.
Arizona is consistently ranked as the sunniest state in the U.S., but solar only makes up 13% of the energy powering its electric grid. Natural gas is by far the largest source of energy used in the state, while utility companies are moving toward thermal energy sources and away from solar and wind projects.
Johnson represents solar energy businesses as a lobbyist at the Legislature and Corporation Commission, where she has already noticed a drop-off in participation from renewable energy businesses.
“What I hear about is, ‘this company can’t renew their membership because they are cutting back on this state,’ or ‘they’re deprioritizing Arizona,’ or ‘they’ve gone belly up entirely,’” Johnson said. “Our membership numbers are down probably 25 or 30% over the last two years.”
Lawmakers take aim at solar
Republican lawmakers pushed three bills this year to discourage utility-scale solar development. All came from Reps. Ralph Heap and David Marshall, who launched a bid to unseat incumbent Republican Commissioners Kevin Thompson and Nick Myers.
The first bill would have classified utility-scale solar projects as public nuisances if they were located within four miles of a residence. The second would have required utility companies to get 85% of their electric generation from “reliable sources,” excluding wind and solar from the definition of “reliable.”
The third bill would have prohibited the State Land Department from using solar score maps to evaluate state trust land for potential solar development. Additionally, drafts of ASLD’s continuation bill, which was meant to extend the department’s statutory life, included a moratorium on solar and wind projects on state land until 2029.
While none of those proposals ultimately came to fruition, it made for a dizzying legislative session for solar advocates like Johnson. She said this year marked the first time she has had to weigh in on an agency continuation bill or state budget negotiations.
“[This year] was a lot worse, but more from an optics standpoint than from an actual outcome standpoint,” Johnson said.
Heap, who unseated Myers in the July 21 primary, told the Arizona Capitol Times in May that his legislation was driven by Trump’s characterization of renewable energy sources as unreliable and uneconomic.
“The utilities, [Arizona Public Service] in particular, gave the Corporation Commission their 15-year plan of how they’re going to generate energy between 2023 and 2038, and that plan was net zero, green energy, run the state on wind, solar and batteries,” Heap said. “And since then, almost all the projects that have gone in the state, up to 90% now, are green energy projects.”
Natural gas versus solar
By all accounts, utility companies are building more solar projects. Between 2025 and the first eight months of 2026, the Commission approved line siting proposals for 12 solar and battery storage projects.
However, Johnson noted that new natural gas generation is still outpacing new solar generation as utility companies look to add more resources to meet skyrocketing energy demand.
“They’re still adding renewables … but they’re building a ton of new gas over the next decade, so the percentage in Arizona that is actually renewables is flat or slightly decreasing,” Johnson said.
Myers, the Commission’s chair, said in a statement that the line siting approvals demonstrate the Commission’s commitment to an all-of-the-above energy approach.
“Arizona is building an increasingly flexible energy portfolio to meet record-setting demand,” Myers said. “The Commission’s siting decisions are helping bring new resources online while supporting the infrastructure needed to deliver electricity to Arizona customers.”
But the Commission has also moved to disincentivize residential rooftop solar this year by considering a proposed increase in APS’s “grid access charge” and adjustments to solar export rates.
The Commission adopted the GAC in the 2022 APS rate case after an administrative law judge recommended setting rates for rooftop solar customers 15% higher than non-solar customers. APS did not specifically propose the GAC, but suggested in rate case proceedings that solar customers are underpaying the cost incurred to serve them.
The Court of Appeals ruled in June that the Commission must either conduct a rehearing on the GAC or vacate it entirely, citing due process concerns. Meanwhile, APS is seeking to increase the GAC in its ongoing rate case.
And on Aug. 12, the Commission approved requests from APS and UNS Electric to lower export rates that reimburse residential rooftop solar customers for energy they provide to the grid by 10%.
“Rooftop solar is one of the few actions Arizona ratepayers can take to control their energy needs and meaningfully reduce monthly bills,” the solar groups wrote in a letter to the Commission. “Unfortunately, this Commission’s continued approval of the maximum reductions to the [export rates] have made rooftop solar harder to pencil out for families across the state.”
The upcoming Commission election could offer some hope for clean energy advocates if the two Democrats running can manage to upset the all-Republican balance on the body. But Johnson said she’s yet to see an entirely solar-friendly Commission.
“Each Commission that I’ve observed has become more and more difficult for solar in the seven years that I’ve been doing this,” Johnson said.
Two employees of a navigation center serving San Francisco's homeless population have been arrested for allegedly harboring a man charged in the shooting of a police officer, the department announced Friday.
According to officers, 36-year-old Oliver Barcenas and 45-year-old Erika Porter were booked into San Francisco County Jail on Thursday. Barcenas and Porter worked at the Bayshore Navigation Center at 125 Bayshore Boulevard, near where an officer was shot on the night of May 31.
"Attempting to assist an armed gunman evade capture after shooting a police officer is unconscionable," Chief Derrick Lew said in a statement. "Thankfully, our officers captured this violent suspect despite the efforts by two individuals to help him."
The incident began around 10:30 p.m., when police were alerted about a suspect vehicle entering the city on the Bay Bridge with the help of an automated license plate reader.
Court documents said the vehicle, a 2021 gray Toyota, was believed to have been used in an armed robbery in Hayward. Officers found the Toyota near Fifth and Folsom streets and attempted to stop the vehicle.
The suspect allegedly refused commands from officers and fled northbound on Fremont Street. The pursuit, which reached speeds of up to 60 mph, ended when the car struck a concrete barrier on the Bayshore Boulevard on-ramp to Highway 101.
Bodycam footage from a shootout between a robbery suspect and police in San Francisco on May 31, 2026 that critically wounded an officer. San Francisco Police Department
Following the crash, officers ordered the suspect and a passenger to exit the vehicle. Bodycam footage released by police showed the driver emerging from the vehicle and opening fire on two police officers, striking Officer Brittany Taylor in the leg. Several officers returned fire, striking a passenger in the vehicle.
The passenger, later identified as Ariunsanaaa Dolgorsuren, received medical treatment for his injuries and was arrested.
The suspect, later identified as 36-year-old Norris Reed, left the scene on foot. Officers were called to the navigation center around 12:15 a.m. on June 1 and Reed was arrested.
Police said two firearms were recovered from Reed's waistband at the time of his arrest. Reed has since been charged with attempted murder and other felonies in connection with the shooting.
Officers said Friday that a review of surveillance footage showed Barcenas providing Reed with clothing to conceal his identity despite knowing he was possessing a firearm when entering the center.
Meanwhile, Porter is accused of making false statements to police when officers made an earlier visit to the navigation center while they were searching for the suspect.
Barcenas and Porter are accused of obstruction and harboring an attempted murder suspect. According to officers, Barcenas is on federal probation for being a felon in possession of a firearm and ammunition.
Police said Friday that Taylor has since been released from the hospital and is continuing to recover from her injuries.
A new U.S. regulatory framework for fusion machines is set to establish a clear, predictable regulatory pathway that will likely accelerate licensing.
July 22 - The Nuclear Regulatory Commission (NRC) in February proposed a new framework to regulate fusion machines under the Byproduct Material Framework, underscoring the lower safety risks of fusion devices.
The new rules come after fusion developers complained that stiff regulations were holding up the industry because their devices have long been regulated as if they were fission reactors, even though they do not pose meltdown risks and do not generate long-lived nuclear waste.
Instead of requiring exhaustive safety analysis to assess the risks of fusion devices, the proposed regulatory framework focuses on the radioactive materials used by these machines, like tritium, as well as waste management, decommissioning, and radiation protection.
Under the new rules, “licensing happens on the scale of months to a year rather than the years or even a decade a fission reactor can face, because the focus is on safely controlling the material, not reviewing a reactor design,” said Andrew Proffitt, Senior Director of Regulatory Policy at fusion developer Helion Energy.
"The industry supports this rule making. It reflects years of work between regulators, developers, and the public, and it provides the certainty companies and investors need to bring commercial fusion online,” Proffitt told Reuters Events.
Unlike existing nuclear facilities that rely on fission—the splitting of heavy atoms like uranium—fusion companies are developing first-of-a-kind devices to fuse light atoms together. This is the same reaction that powers the sun, and while it promises virtually limitless, carbon-free energy, it has never been achieved on a commercial scale.
The regulatory update comes as fusion developers are enjoying a boom in investments that have paved the way for several technological breakthroughs in recent months. Established players including Helion Energy, TAE Technologies, Commonwealth Fusion Systems (CFS), and Type One Energy have already started building advanced prototypes.
The proposed framework “acknowledges that fission and fusion are different beasts,” said Robb Hughes, Head of External Affairs at Realta Fusion.
“Fission reactors pose a risk of meltdown and create long-lived radioactive materials, so fission simply needs more stringent regulation than fusion, for which those same hazards do not exist,” Hughes told Reuters Events.
The new rules come amid a flurry of activity under President Trump aimed at fostering a U.S. “nuclear renaissance.” This includes new NRC policies to expedite the approval of small modular reactor (SMR) designs, streamline environmental reviews, and modernize radiation safety standards.
“Fusion has the benefit of an advantageous safety profile compared with fission, creating an opportunity for a lighter regulatory burden and an addressable market for fusion that we believe could eclipse the SMR market,” a spokesperson for General Fusion told Reuters Events.
For exclusive nuclear insights,sign upto our newsletter.
The rules align with the ADVANCE Act of 2024, which establishes that fusion machines should be regulated as particle accelerators rather than traditional nuclear reactors.
The rules do not guarantee faster deployment, but they give “developers something they did not have before: a clear licensing framework,” said Spencer Toohill, Chief of Staff for Nuclear Energy Innovation at The Breakthrough Institute
The new fusion framework is designed to be flexible, allowing developers to “manage design-specific hazards without repeatedly asking the regulator for permission to depart from prescriptive requirements that may not fit their technology,” Toohill told Reuters Events.
Whereas fission developers are often required to seek amendments during the lifecycle of their assets, the “more permissive, materials-oriented framework for fusion reduces the need to ask for so many license amendments,” she said.
While the NRC is legally mandated to finalize commercial fusion regulations by December 31, 2027, the agency and the Fusion Industry Association (FIA) aim to finalize and adopt the rules by October 2026.
Download ourexclusive reporton the latest funding leaps in fusion energy.
State licensing
Under the new framework, state regulators will hold primary responsibility for licensing and overseeing most commercial fusion facilities.
The Organization of Agreement States (OAS)—a longstanding coalition of 40 U.S. states that have signed formal agreements with the NRC—will coordinate the implementation of the new rules.
Under the umbrella of OAS, member states will “engage, communicate, exchange best practices” to ensure that the new regulatory framework is applied in a consistent way across state lines, Jeff Merrifield, the leader of Pillsbury Law’s Nuclear Energy Team, told Reuters Events earlier this year.
Nonetheless, Toohill said she expects some degree of “variation in implementation from state to state,” especially for early demonstrations.
Helion has secured regulatory approvals from the Washington State Department of Health, including a Radioactive Materials License (RML) and a Radioactive Air Emissions License (RAEL) for its Orion facility in Malaga, Washington, the company said on June 16.
“After years of experience working with Washington state regulators on earlier prototypes, Helion is the first company in the world to have the regulatory permits and licenses in place for a fusion power plant,” Proffitt said.
Other fusion developers, including CFS in Massachusetts, TAE Technologies in California, and Type One Energy in Tennessee are similarly working with their state regulators to secure licenses.
“These are the first of their kind applications that we're going to see over the next few months or year, and they are going to pave the way and set the pace for other states to follow suit,” Merrifield, who also works as outside counsel for FIA, said.
--Editing by Eduardo Garcia
Opinions expressed are those of the author. They do not reflect the views of Reuters News, which, under the Trust Principles, is committed to integrity, independence, and freedom from bias. Reuters Events, a part of Reuters Professional, is owned by Thomson Reuters and operates independently of Reuters News.
Mark is an energy reporter for Reuters Events, part of Reuters Professional. Based in New York, Mark has two decades of experience in commodities, writing about energy policy, markets, history and consumer impact. His articles have featured in the Washington Post, Boston Globe, Houston Chronicle and Bloomberg Markets, among other publications.
The Hobbs administration says a federal judge went for the nuclear option too soon when she turned over control of the state’s prison healthcare system to a receiver without exhausting other remedies in a lawsuit that has dragged on for 14 years, and is asking the judge to pause doing so.
“Jumping from the current tools that are yielding measurable results to the Receivership Order’s ‘remedy of last resort,’ before employing any intermediate measures, is an abuse of discretion,” Arizona Department of Corrections, Rehabilitation and Reentry attorney Mary O’Grady wrote in a brief aiming to persuade Judge Roslyn Silver to issue a stay on her July order appointing Annette Chambers-Smith to oversee the system.
Since ruling in 2022 that the healthcare system in Arizona prisons is so substandard that it amounts to a violation of the Eighth Amendment rights of inmates, the state has only had about three years to comply with the terms of an injunction that the Corrections Department and the plaintiffs in the class-action lawsuit agreed to.
“Since it made its constitutional findings, the Court has not attempted to improve compliance through any intermediate measures, such as remedial plans or sanctions,” O’Grady argued.
In July, Silver appointed Chambers-Smith and granted her the sole power to determine what the state needs to do to stop violating the constitutional rights of inmates.
The plaintiffs, represented by the ACLU Foundation of Arizona and Prison Law Office, said it is “outrageous” that the Department of Corrections would argue that the court skipped steps and that its legal position effectively attempts to hand-wave away the myriad failures of the system that were documented prior to the 2022 injunction.
Prior to that injunction, Prison Law Office attorney Sophie Hart wrote, the court did escalate its attempts to bring the state into compliance through fines, contempt orders, expert evaluations and a trial. Ultimately, she argued, the injunction “did not fundamentally alter what the Department is required to do.”
And Silver, Hart noted, has previously rejected the state’s attempt to reframe evaluation of its compliance by ignoring what happened before 2022 as a “profound misunderstanding of the law and facts.”
The class-action lawsuit stretches back to 2012, when prisoners sued the department and alleged that its shoddy health care violated their Eighth Amendment right against “cruel and unusual punishment.” The 14 years since have been marked by escalating judicial intervention, culminating in Silver ordering the prison health care system into receivership.
In 2014, the prisoners and the state reached a 103-point agreement on health care performance. But when the Corrections Department repeatedly failed to comply with those standards, the court twice held the agency in contempt and levied fines totaling more than $2.5 million.
Following a 15-day bench trial in 2022, the court issued a 200-page order declaring the health care system “pervasively and systemically unconstitutional.” The next year, a permanent injunction was issued with specific “quality indicators” to assess how the prison system was complying with the court’s directives.
In 2024 and 2025, Silver wrote, the state repeatedly and flagrantly violated those orders, and a pilot program aimed at testing a new model for prison health care failed disastrously — it was only partially implemented at one of the two test sites, and was only in place for eight days before it was scrapped.
“The Court’s patience has run out. Too many individuals are needlessly suffering while Defendants have deployed many delay tactics,” Silver wrote in a February order placing the healthcare system into receivership. “The Court has exercised restraint for much of this litigation, to the point that anymore tolerance of unconstitutional healthcare becomes judicial indulgence.”
But the Corrections Department says that Silver asked the wrong question. Instead of focusing on whether the department has complied with the injunction, she should have asked whether a receiver is the only way to achieve compliance.
“Will a receivership achieve compliance more quickly and efficiently than the remedies already in place? Plaintiffs’ opposition elides that question,” O’Grady argued.
The plaintiffs, who asked Silver to appoint a receiver, said that, after nearly 15 years of litigation and ongoing failures of the healthcare system — including documented deaths as a result — the answer to either question leads to the same result: The state has consistently refused to do what it has been ordered to do and has proven it won’t take the steps needed to comply.
The crux of the Corrections Department’s argument is that allowing the receiver to take the reins now, while the state appeals the receivership order, results in irreparable harm and violates the principles of federalism by stripping the state of its inherent authority.
While the plaintiffs have argued that $2.2 million in contempt fines can be used to pay for the receiver during the appeal, the Department of Corrections noted that the costs are likely to far exceed that. O’Grady pointed to California, where a judge put its prison healthcare system into receivership more than 20 years ago, which spent more than $180 million on its receiver in the first five years, ranging from $11 million to $93 million in a single year.
And on top of that, the state argued, using the contempt fines to pay for the receiver would actually create irreparable harm because those fines were designed to be coercive and the Corrections Department can still recover them through compliance. But if that money gets spent on a receiver, that ability goes away if the appeals court reverses Silver’s order.
O’Grady also argued that Silver also irreparably harms Arizona’s authority because “rather than temporarily enjoining state officials from enforcing their enacted policies, the receivership displaces the officials’ authority altogether.”
“Any time a State is enjoined by a court from effectuating statutes… it suffers a form of irreparable injury,” she wrote.
There’s also the broader issue of the money needed to comply with the litigation. Although Gov. Katie Hobbs initially proposed more than $118 million in new funding for compliance, but the final budget she negotiated with the Republicans who control the state legislature provides only about a third of that — and $45 million of that money is contingent on the Corrections Department and its private healthcare vendor, NaphCare, signing a contract amendment with automatic penalties for persistent job vacancies.
That amendment is not yet in place, but an ADCRR spokesman told the Arizona Mirror that the agency is “currently working through the contract amendment process” with NaphCare.
The Department of Corrections also said that Silver relied too heavily on inmate deaths due to poor medical care.
“Evaluating a system serving 25,000 patients by reference to its worst outcomes will inevitably overstate systemwide risk,” O’Grady argued. “If that method were sound, any time inadequate medical care is provided in a correctional setting, the entire corrections system would be exposed to court-managed institutional reform.”
But Hart countered that the court’s methodology went far beyond merely the documented deaths, and it was based on the state’s own self-reported compliance data.
“Defendants do not dispute this evidence. Instead, they take aim at this Court’s analysis,” she wrote.
ETS, TED and the Khan Academy have jointly launched a new AI degree model at a fraction of traditional college costs, vowing to create a globally accessible education pathway backed by global corporate giants.
Sir Nick praised Efekta's AI platform, built in Europe and developed in emerging markets. Photo: Unsplash.
Launched at the TED conference in Vancouver last month, the new Khan TED Institute was developed in partnership with the likes of Google, Microsoft and Accenture to ensure “deep relevance with the rapidly changing nature of work” said ETS.
The model leverages AI to deliver remote undergraduate degrees at $10,000, with founding partners vowing to expand access to learning pathways and AI education, without losing site of the human skills required in employment.
“The partners see the Khan TED Institute as a complement and alternative, designed for learners who want a different pathway – one that is more affordable, globally accessible, and closely aligned with the modern workforce,” ETS CEO Amit Sevak told The PIE News.
He emphasised the institute would be a “new addition to the ecosystem of higher education, not a replacement”.
Alongside those who can’t access or afford traditional university education, the new program is aimed at those with existing qualifications who want to re-skill in the AI age and supplement their education.
With applications set to open in 12 to 18 months, prospective students are currently being invited to express their interest to learn more, with educators, corporate firms and philanthropists also encouraged to reach out.
“For nearly 40 years, TED has been connecting people with powerful ideas – and in the process, quietly educating millions,” said TED CEO Logan McClure Davda, emphasising the need for learning to keep pace with the AI era.
This is an innovative new addition to the ecosystem of higher education, not a replacement
Amit Sevak, ETS
The inaugural program will be a bachelor’s degree in applied AI, designed to prepare students for careers across technology, consulting, finance, product management, and other knowledge‑intensive fields.
Unveiling the new model, Khan Academy CEO Sal Khan said each founding firm brought something different to the partnership, with his company specialising in “competency and mastery-based learning”.
TED, meanwhile, will bring the “’durable skills’ of communication, collaboration and creativity” and testing giant ETS will “pioneer new forms of assessment” for the AI world.
The partners are seeking American degree accreditation, with plans to quickly scale beyond the US.
Rather than learning from a traditional university faculty: “Students will participate in structured group work, live sessions with TED speakers and thought leaders, dialogue sessions, and collaborative projects”, Sevak explained.
He emphasised there would be “no wasted seat time” and students would engage in peer-based tutoring and competency based assessments, with access to AI-enabled learning support.
Europe looks beyond the region in global student recruitment push
European countries are increasingly looking beyond their borders to recruit international students, signalling a shift in higher education strategy towards global competition, a report by the European Association for International Education (EAIE) and the Academic Cooperation Association (ACA) has found.
Youth mobility talks purportedly broke down over a cap on the number of people coming into the UK and tuition fees for EU students. Photo: iStock.
“An increasing number of countries show interest in venturing beyond Europe,” the report notes, pointing to growing efforts to recruit students from regions such as Asia, Africa and the Americas.
China has long been a major source of international students, with thousands studying abroad each year. However, South Asian markets, particularly India, are also emerging as key recruitment targets.
An increasing number of countries show interest in venturing beyond Europe Mapping the internationalisation strategy landscape across the EHEA
The shift is reflected in how individual countries are positioning themselves internationally.
In France, the government’s “Choose France” strategy aims to attract 500,000 international students by 2027, supported by measures such as simplified visa processes and expanded English-taught programs.
However, the country has recently announced that it will up fees for non-EU students as part of the strategy, as it moves to shape international recruitment. Critics have blasted the move as a “blow to university autonomy”.
Germany’s latest internationalisation strategy emphasises diversifying global partnerships and retaining international talent, particularly in a changing geopolitical climate.
Meanwhile, Ireland’s Global Citizens 2030 strategy focuses on building its reputation as a “ first-choice destination ” for international learners and researchers, with plans to expand its global presence through talent attachés and new mobility schemes. Ireland has also enjoyed a surge in interest from US and Indian students recently, with international enrolments reaching their highest point to date.
Norway is also targeting key markets outside Europe – including the US, India and China – through its Panorama strategy, which links higher education, research and innovation partnerships.
The report, which analyses policies across 47 countries in the EHEA, shows that internationalisation strategies have matured significantly. Just 13 countries – including Austria, the Czech Republic, Italy, Switzerland and the United Kingdom – currently have a dedicated international education strategy in place.
Others pursue internationalisation through broader education, research or economic policies, with the report stressing that the absence of a formal strategy does not necessarily indicate weaker performance.
A notable trend is the shift towards “internationalisation for all”, with countries placing greater emphasis on widening access and ensuring participation from underrepresented groups.
However, the report highlights persistent challenges, including visa restrictions and difficulties in recognising international qualifications – issues that continue to hinder global mobility. These challenges are compounded by a more volatile global environment, with geopolitical tensions and shifting political attitudes influencing national strategies.
As a result, future policies are expected to become more adaptive, with greater emphasis on flexibility, crisis management and long-term resilience.
Looking ahead, the report suggests that national strategies will become more evidence-driven and outward-facing, with a stronger focus on communicating the value of international education to governments and the public.
OpenAI, Anthropic, Amazon and Microsoft are bankrolling a new nonprofit to retrain American workers for the AI economy. RAISE US, led by Gina Raimondo, has already raised more than $500m.
The biggest names in AI agree on at least one thing. The technology they are building could upend the American job market, and almost nobody is ready. So they are writing cheques. Gina Raimondo, the former US commerce secretary, and Eric Holcomb, the former governor of Indiana, have launched RAISE US. The nonpartisan nonprofit will work with governors and employers to help workers through the disruption.
The group has secured more than $500m so far. The target is $1bn in multi-year commitments. It launches with more than two dozen of America’s largest companies and philanthropies behind it. Raimondo will serve as chief executive.
Its anchor backers are striking. OpenAI, Anthropic, Microsoft and Amazon are fierce rivals, racing to build the most powerful models. All four have signed on. So have Bank of America, IBM, Mastercard, AMD, Eli Lilly and the Rockefeller Foundation, among others.
“America has a technology strategy for leading the global AI competition. It does not yet have a people strategy, and we cannot lead without one,” Raimondo said. “If we build the best AI systems in the world and leave millions of Americans behind, we won’t have won anything. We’ll have automated our own decline.”
Why the money is moving now
The launch lands at an anxious moment. Estimates of how many jobs AI will displace range wildly. They run from half of all entry-level white-collar roles to a few thousand here and there. Layoffs remain low for now. But worker sentiment toward AI has been worsening, and some firms have already blamed the technology for deep cuts.
That nervousness is shared by the very companies funding RAISE US. The same labs preparing to go public at trillion-dollar valuations are now helping pay to soften the blow. Two of the new backers, IBM and Workday, have themselves cited AI in recent job cuts.
Public mood is part of the story too. Anger over AI has grown louder. It runs from data-centre power bills to the steady drumbeat of layoff headlines. Washington has done little. Congress has not addressed the disruption, and the White House has played down the risk of widespread job losses.
Into that vacuum steps a privately funded coalition. “This is an independent effort,” Raimondo said. “It’s the first one I know of where competitors in the tech industry have put aside their competition to say, ‘We’re going to write big checks and, in the service of our country, do what we can to figure out this transition.’”
How RAISE US plans to work
The strategy runs through state capitals. RAISE US will start with governors in Utah, Arkansas, Maryland and Connecticut, a deliberately bipartisan mix. States control community colleges, credentialing and business incentives. Those levers decide whether employers retrain workers or let them go. That makes states the natural place to test ideas.
The bulk of the budget will fund pilot programmes. A lean team of about 15 staff and consultants will oversee them. In Arkansas, the group is backing an AI-powered career navigation platform called Arkansas LAUNCH. In Maryland, it will expand a “service year” for recent high school graduates into fields with shortages, such as health care. It will also stand up an accelerator to help displaced workers start businesses.
Other ideas borrow from the social safety net. RAISE US wants to pilot “wage insurance” for workers who take a lower-paying job rather than dropping out of the labour force. It also wants to test short-time compensation to keep people employed through a transition.
The group wants to keep workers in the jobs they already have, too. It plans to offer technical help to companies that would rather retrain staff than replace them. Microsoft says it has found a model that works. It cross-trains its junior lawyers across the business and equips them with AI skills, so they can be moved as roles change.
“You can think of doing that with almost any job we have,” said Brad Smith, vice chair and president at Microsoft. “It creates an opportunity to transfer people from jobs that are being eliminated to jobs that are being created.”
A crowded and skeptical field
RAISE US enters a busy lane. Governments and companies have floated a wave of fixes. The corporate giants now in workforce programmes have learned that retraining is hard, and historically not very successful. Raimondo herself called past efforts “ineffective.”
That history invites doubt about a corporate-funded group steering policy. The structure tries to answer it. RAISE US also runs a policy lab, funded by philanthropies rather than companies. The aim is to keep its recommendations at arm’s length from its sponsors.
The wider debate is sharper still. Senator Bernie Sanders has suggested seizing half the stock value of top AI firms for a public fund. OpenAI itself has floated robot taxes and a national wealth fund. RAISE US is more modest. Its board includes Liz Shuler, president of the AFL-CIO, a signal it wants labour at the table.
Notably, the group is not sold on universal basic income, the Silicon Valley favourite. Its leaders argue that work offers more than a wage. They would rather build pathways into new jobs than simply send out cheques.
The unanswered question
The deeper uncertainty is whether AI will create jobs as fast as it destroys them. Even backers concede they cannot be sure. The companies bankrolling RAISE US are themselves split on the threat. Some bosses warn of upheaval. Others, including Sam Altman, argue an AI jobs apocalypse is unlikely.
Raimondo frames the effort as insurance against the worst version of the transition. She points to the Committee for Economic Development, formed by big businesses in 1942 to ease soldiers back into the economy after the war. The parallel is pointed. It is a moment of upheaval met by employers acting before the government did.
“I think this technology will lead to more productive people, new jobs and new industries, and I want to get there,” Raimondo said. “But I also worry about the transition, and a window where people could get hurt.”
The ambition is real, and so is the money. The funders themselves struggle to predict this shift. Whether $500m and a handful of state pilots can blunt it is the question RAISE US now has to answer.
Investing.com -- Reddit shares rose more than 11% in pre-open trading Friday following S&P Dow Jones Indices’ announcement that the company will be added to the S&P 500 index.
“We’re proud to be added to the S&P 500, recognizing the growth, momentum, and consistency we’ve established as a public company,” said Drew Vollero, Reddit’s Chief Financial Officer. “Reddit’s raw materials are special, and there is so much potential to continue building great products and scaling profitably. Our focus remains on executing against the many opportunities ahead.”
The social media platform will enter the benchmark index before the market opens on Aug. 18. Reddit will take the place of AvalonBay Communities, which Equity Residential is acquiring.
Google has announced three new subsea cables in the Americas.
The search giant this week announced three new subsea cable systems - Alisios, Canoa, and OlaLuz – linking south, central, and north America.
The three cables form part of the Americas Connect initiative alongside the previously announced Firmina, Curie, Nuvem, and Sol cables.
The Alisios subsea cable system will connect the Dominican Republic, Panama, and Chile.
Canoa will connect the Dominican Republic to Bermuda, where the Sol and Nuvem cables land.
The OlaLuz cable will connect the Dominican Republic directly to Florida.
Google is also introducing a new branch of the Firmina subsea cable that will land directly in the Dominican Republic.
Capacities, landing points, and project timelines weren’t shared.
“Named after regional maritime traditions, these cables will create highly resilient, diverse routes connecting Chile, Panama, the Dominican Republic, Bermuda, Florida, and beyond,” Bikash Koley, Google VP of global infrastructure, said on LinkedIn. “By creating geographically diverse rings across the Pacific Coast, Caribbean Sea, and Atlantic Ocean, we’re helping bridge the digital divide, connect communities, and drive economic growth in partnership with local governments.”
The new cables will link with Google’s existing subsea infrastructure in the region. Sol and Nuvem link North America and Bermuda to Europe. Firmina and Curie link Brazil and Chile to the US.
The Alisios cable is named after the vientos alisios - the Spanish term for the trade winds that blow across the tropics and the Caribbean, historically used by sailors to navigate between continents. Canoa comes from the Taíno word for canoe, harkening to the longstanding maritime culture in the Caribbean and North Atlantic. The name OlaLuz combines the Spanish words for "wave" (ola) and "light" (luz), referencing the waves of light that carry data along the cable on the ocean floor.
“As we welcome the announcement of the Alisios, OlaLuz, and Canoa subsea cable systems, the Dominican Republic is pleased to continue partnering with Google to bolster digital connectivity and innovation as part of Americas Connect,” said Luis Rodolfo Abinader Corona, president of the Dominican Republic. “Our participation in the Caribbean and Latin America telecommunications network allows us to join our partners in strengthening connectivity across the Americas, serving as an integral node in this expanding digital ecosystem. Our shared vision supports a leap forward in the DR government’s quest to bridge the digital divide, foster local talent, and drive tech-based economic opportunity for our people. This milestone is yet another piece of our collaboration with Google propelling the Dominican Republic into a new era of global digital integration.”
The Dominican Republic currently has three data centers, according to Data Center Map, all located in Santo Domingo and operated by Kio or NAPCaribe (x2). Six cables land there across five landing points.
Google has invested in more than 30 subsea cables in the past 15 years, more than any other hyperscaler, both privately and part of multi-party consortia. The search giant is developing multiple cable landing stations and “connectivity hubs” across a number of locations to support its network build-out.
