Just Google DeepMind has released a up-to-date version of its AI-powered Gemini model that can control a range of different robots – including humanoids capable of performing dexterous tasks such as screwing in lightweight bulbs and tying garbage bags.

Gemini Robotics 2 combines several different artificial intelligence models into one system. Together, they allow the robot to understand its environment and learn how to behave in it. A vision language model (VLM) that understands images and video can communicate with humans and draw conclusions on how to perform various tasks. Two vision language-based action models (VLA), trained to understand movement in physical space, control the movement of the robot’s entire body, as well as the movements of its grippers or hands.

In pre-launch video demonstrations, the company showed several different robots performing elaborate tasks on their own using a connected model. In one demonstration, Apptronik’s Apollo 2 robot used Edged’s hands to organize shelves. Google DeepMind trained a model to perform these tasks using a combination of human teleoperation, video examples, and simulations – AI models are not yet capable of performing a wide range of elaborate tasks without special training.

While Anthropic and OpenAI have taken the lead in chatbots and AI coding tools, Google has more experience in robotics research and published critical Work on the apply of artificial intelligence to train robots to perform useful activities. This publication is another sign that the search giant is betting that artificial intelligence will need to break free from the digital sphere to realize its full potential. (It previously worked with Boston Dynamics, a leader in legged robots, to supply the brains for these machines.)

“This is another milestone on our path to achieving what we call physical AGI, which means we will have a robot that can do everything a human can do,” Carolina Parada, director of robotics at Google DeepMind, tells WIRED.

But giving pioneering AI models access to robots to roam workplaces or homes and manipulate objects comes with risks. Previous research has shown that using pioneering artificial intelligence to control robots can cause unexpected and sometimes unsafe behavior. The idea that these models can take sudden or unwanted actions in the digital realm became clear recently when an unreleased AI agent developed by OpenAI hacked several systems.

“The safety issue is even more pressing because they are exposed to so many other situations,” Parada says. “There will be a lot of uncertainty, so it’s important to better understand security.”

Parada says Google takes a multi-layered approach to security, using guardrails at every layer of the model. It also introduces ASIMOV-Agentic, a up-to-date benchmark for measuring the safety of various artificial intelligence systems working together to control a robot. The benchmark detects whether a command will result in a harmful or insecure result.

The company’s CEO, Demis Hassabis, previously told WIRED that he hopes to develop an AI operating system for a wide variety of robots, similar to the Android operating system for smartphones.

Unlike many others enormous technology platforms LinkedIn has decided not to spend aggressively on expanding its AI data centers this fiscal year. Executives at the professional social network tell WIRED they plan to keep GPU investment steady, while demand for processing power and storage also remains unchanged.

The expense calculations are for LinkedIn’s fiscal year, which began last month and ends next June. The company says it has managed to avoid substantial spending on AI hardware because it has found ways to employ existing GPUs twice as efficiently over the past six months. LinkedIn’s plan may not come to fruition yet as AI hardware requirements change rapidly, but executives say the company has already taken into account rising memory chip prices.

“One of the goals we’ve set for ourselves is to basically keep the compute area flat, or as flat as possible, when sending higher compute-demanding components to production,” says Erran Berger, LinkedIn’s chief technology officer for engineering. “That’s quite a bold statement in today’s world.”

Berger and Raghu Hiremagalur, LinkedIn’s chief technology officer for infrastructure, say they want to be conservative with spending and that the fresh restrictions will motivate engineering teams to be more imaginative when developing the many fresh generative AI features LinkedIn plans to launch. Berger says he believes the efficiency gains could accelerate over time, allowing LinkedIn to better take advantage of its data center expansion when it eventually increases its budgets again.

“I really want to reiterate that for a company of our scale, let’s say an entire year of doing this without incremental storage and compute is no mean feat, but it took a lot of work to get there,” says Hiremagalur.

Companies like OpenAI, Meta, and Google are raking in all the money they can and forming unlikely partnerships to build, equip, and operate massive data centers filled with the latest computer chips. Labor and parts shortages have held up many projects, and many companies have had to limit customer employ of some AI tools. However, there are growing questions about whether continued investment in artificial intelligence is sustainable. LinkedIn, with more than 1.3 billion users, is perhaps the biggest company yet to publicly address spending concerns by bucking the construction boom.

“This is encouraging for the industry,” says Songyee Yoon, managing partner at Lead Venture Partners and a board member of server maker HP. “This suggests that AI is starting to move from experimentation to production discipline. The companies that win will not simply be the ones that spend the most on infrastructure.”

Owning it

A few years after Microsoft acquired LinkedIn in 2016, the company tried to move to its parent company’s Azure cloud service, but it didn’t make economic sense to squeeze the giant social network into general-purpose data centers. “Microsoft Azure was growing like crazy, the level of customer demand was enormous, and at the same time we saw explosive growth on the LinkedIn side,” says Hiremagalur.

In 2022, LinkedIn took full advantage of its own data centers in Oregon, Texas and Virginia. This owner has given LinkedIn significant control over every detail of its technology, well-preparing it to meet the realities of the fresh era. Around the same time, LinkedIn began developing AI-powered assistants that could assist users write messages, find jobs, and recruit candidates. The undertaking was not low-cost. “The cost of each query coming to our site increases over time,” says Hiremagalur, adding that the amount of data LinkedIn stores is doubling annually. “It’s not a sustainable place.”

# Entry

Practical decoding of constraintsalso known as structured generation or guided decoding, involves engineering strategies to force a enormous language model (LLM) to produce text output that strictly follows a specific data schema, grammar, or regular expression() at the level token selection scene.

With this article’s introductory guide to practical constraint decoding, you’ll no longer have to beg your model to “output valid JSON without including any reductions,” just provide an example. Constraint decoding makes it mathematically impossible for LLM to deliver anything beyond the defined constraints.

# How does practical constraint decoding work?

While the typical LLM generation process acts as a “leap of faith” in which you pass a hint to the model and it may get exactly what you’re looking for (or not), practical constraint decoding requires a subtly different approach. It treats prompting and text generation as a unique, interwoven program. This allows you to lock out certain characters that are crucial to maintaining the specific required syntax, allowing the model to “fill in the blanks” between them.

How about we go into more detail? When LLM sends another token of its response, it initially creates a vector of raw results, or logits — one for each possible tag in the available vocabulary. This usually means thousands of possible options to choose from.

But when using practical constraint decoding, something happens before the inference process starts: a finite-state machine is builtduring which the target constraint is compiled – for example using the Pydantic model in Python. At a given inference stage, the finite state machine evaluates the current state and provides: list of allowed subsequent tokens. This “white list” exists used as a mask on the raw LLM logit vectorso that for any token outside this list its logit is set to negative infinity, i.e -inf in Python.

After the masking process, the model normally performs a softmax normalization and sampling process (based on parameters such as temperature, top-p or top-k) on the “surviving tokens” to finally select the most likely one and generate it.

It may seem that applying this process to an entire vocabulary of many thousands of words would significantly snail-paced down model inference. Good news: this is not the case. Contemporary Python libraries take the unchanging LLM vocabulary and pre-compile it before the user starts typing the prompt. The state machine will not have to search through the entire vocabulary and the latency overhead will be drastically reduced.

So what is the current gold standard for implementing practical constraint decoding? Probably, guidelines the library deserved this distinction. It allows us to define and pass Pydantic models, JSON schemas, or directly to a wrapped version of a pre-trained model, thus limiting its freedom in generating results.

# Example

Let’s go through an example. First install the outlines:

pip install outlines[transformers]

Now let’s move on to the code:

from pydantic import BaseModel
import outlines
from transformers import AutoTokenizer, AutoModelForCausalLM

class UserProfile(BaseModel):
    name: str
    age: int
    is_active: bool

model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"

llm = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

model = outlines.from_transformers(llm, tokenizer)
result = model("Extract the user: John is a 34 year old pilot.", UserProfile)

print(result)

Exit:

{"name": "John", "age": 34, "is_active": true}

This example showed how to exploit the outline library to wrap a pre-trained model along with its tokenizer and restrict it to output JSON objects defined by a custom class we defined – called UserProfile and Pydantic inheritance BaseModel.

# Summary

Strengths:

• If used correctly, it provides a 100% guarantee of correct syntax, eliminating the need to parse blocks in the code.
• This helps dramatically save tokens in tooltips, no longer requiring several token-consuming examples to, for example, show the model what a valid JSON object should look like.
• It helps democratize diminutive models by transforming a “small” 1B parameter model, which would otherwise compromise JSON generation exploit cases, into a reliable data builder.

Limitations:

• If LLM needs to say it can’t answer something, but the schema forces it to generate an integer, for example, it will do so and will no longer be sincere in edge cases.
• When first running the Pydantic vs. LLM schema, you may experience a freeze for a few seconds while building the finite state machine, making the first run much slower, although subsequent runs will be smoother.

This article presents practical constraints decoding, exploring why it is necessary in certain LLM-based situations, how it works, and what is the most commonly used solution in the current landscape: the contour library. An example of its exploit is also provided.

Ivan Palomares Carrascosa is a thought leader, writer, speaker and advisor in the fields of Artificial Intelligence, Machine Learning, Deep Learning and LLM. Trains and advises others on the exploit of artificial intelligence in the real world.

I recently got one to see what happens when you jailbreak some of the most powerful AI models in the world.

Don’t worry – this AI manipulation wasn’t used to hack anyone or build a nuclear bomb. I’ve simply seen firsthand how susceptible some borderline models are to letting go of their guardrails.

FAR.AI, a California-based nonprofit AI security organization, has created a tool that displays a range of problematic messages and generates over a thousand different versions to identify running jailbreaks. I saw that some models generated a detailed plan to carry out a cyber attack, including on an imaginary hydroelectric dam. This often involved trying dozens of prompts, and the models rejected many of them outright.

I spoke to FAR.AI earlier new reportduring which the group tested safety rails models from four popular American companies: Claude Opus 4.8 and Fable 5 by Anthropic; OpenAI GPT 5.5 and 5.6; Google Gemini 3.1 Pro; and Grok 4.3 and 4.5 from Elon Musk’s newly merged SpaceXAI. It automatically generated prompts designed to trick models into doing potentially harmful things, such as generating software exploits and providing details about developing chemical or biological weapons.

The report found that Grok was the most vulnerable to jailbreak attacks (448 cases), followed by Gemini (249), while Claude, Fable and GPT were resistant to attacks. However, this does not mean that these models are immune to more sophisticated jailbreaks, which FAR.AI and other experts say may involve more sophisticated interaction with the model.

The report also calculated the cost of causing the models to behave incorrectly by using a different AI model to automatically generate different jailbreaks. All things considered, the results are very economical – $58 for the Grok jailbreak and $278 for the Gemini jailbreak.

“AI models are currently less regulated than restaurants,” says Adam Gleave, CEO of FAR.AI and an expert in AI security and customization.

Gleave says the findings highlight the need for externally imposed standards and regulations. “It’s nonsense to talk about relying on voluntary commitments and that AI companies will be able to self-regulate,” he says.

However, Gleave also believes that the findings show that models can be systematically tested for safety. “There is an optimistic point of view here,” he says. “Defense and security really are possible.”

Rohin Shah, director of security and AGI compliance at Google DeepMind, says the report’s findings “should not be interpreted as a comprehensive assessment of Gemini’s safety and security” because not all prison breaks are equally earnest.

“We are constantly working to improve our security,” says Shah. “We conduct extensive red teaming and assessments for serious fraud threats, and employ multiple layers of security during development and implementation.”

“These findings reflect the continued investment we have made in our security features,” Anthropic spokesman Michael Aciman tells WIRED. “We continue to evolve our security systems as these attacks become more sophisticated.”

OpenAI and SpaceXAI did not respond to WIRED’s request for comment.

Recently enacted state laws in California AND New York require pioneering AI developers to publish safety reports, and soon an Illinois the law will require that the practices of these companies be assessed by external auditors. But the federal government hasn’t yet adopted any specific safety requirements, and chaos has ensued as the industry – and officials – try to figure it out.

In June, the Trump administration imposed export controls on Anthropic’s Fable 5 and Mythos 5 models, citing national security concerns, and the company disabled them for several weeks. The White House also asked Anthropic and OpenAI to delay recent model releases over concerns they could introduce novel cybersecurity threats.

X has a up-to-date argument against blanket social media bans for minors that would restrict utilize of the platform: they interfere with international law.

The social media platform hosted by Elon Musk’s SpaceX has been released motion to the Australian Parliament published on Tuesday calling on the government to end efforts to strengthen a ban on social media utilize by minors. He alleged that proposals to raise pressure on websites to demonstrate efforts to restrict access by minors are unnecessary, inappropriate, unfair and likely to violate privacy rights.

X objected to the “highly intrusive” information collection powers the proposed amendment would allow, accusing the commissioner of an “apparent lack of understanding” of how it would work for platforms and a lack of safeguards for confidential and commercially sensitive information. Requesting data, documents and proof of compliance from non-Australian citizens in other countries could cause problems “for the good of nations”, the company also warned.

Australia is leading a growing global movement to restrict children’s access to social media after banning under-16s from the sites in December. In May, the country ordered X to pay a fine of $463,000 for failing to take measures to keep children unthreatening. The country’s internet regulator, eSafety, first imposed the fine in 2023, saying X had failed to sufficiently respond to a request for information on how to prevent the spread of child sexual abuse content online, made a month before Elon Musk, now X, took over Twitter.

X has previously criticized the country’s “excessive” sentencing system and in his latest statement complained that the proposal to raise penalties for individuals was “totally unjustified and disproportionate.”

Musk has been particularly vocal in his criticism of Australia’s law, which states that the minimum age for using social media is 16. “It looks like all Australians are using a backdoor to control their internet access.” wrote on X when the up-to-date regulations were announced in delayed 2024. When Spanish Prime Minister Pedro Sánchez announced similar measures in February this year, Musk called him a “tyrant” and a “true fascist totalitarian.”

While many digital rights activists find blanket bans on age-based social media “problematic,” the issue is not information-gathering powers, says Stefania Di Stefano, a researcher on international law and technology.

“For me, a total ban on the use of social media by children and minors is problematic from an international human rights point of view,” says Di Stefano. “This is disproportionate in relation to children’s right to exercise their right to freedom of expression, their right to access information, their right to association, etc.”

However, Julia Hörnle, professor of internet law at Queen Mary University of London, is skeptical of X’s proposal. “An Australian regulator ordering X to disclose a document related to his business activities in Australia is perfectly fine,” she tells WIRED. “From all data held by a social media company, a distinction can be made between Australian and non-Australian children and therefore regulations apply in Australia.”

Geothermal company Zanskar bought a plant in Novel Mexico in 2024 that from the outside appeared to be failing. The shallow wells lost heat quickly, and as a result, the plant had trouble generating power.

Now, after drilling a recent well and operating the plant year-round, Zanskar says there has been a complete turnaround. According to company data shared exclusively with WIRED, the Lightning Dock location has become one of the most productive geothermal pumped storage wells in the U.S., providing enough heat to keep the plant running smoothly.

Performance over the past year at the Lightning Dock facility is “fundamentally changing the way people think about these types of systems that we thought we understood in the past,” says Zanskar co-founder Joel Edwards.

This is another good news for the long-neglected geothermal industry, which is experiencing an investment renaissance. Creative drilling technology and artificial intelligence – with any luck – played a role in this milestone. And while not all conventional wells experience the same conditions as Zanskar’s at Lightning Dock, the company says its success is a signal that elderly wells are still worth exploring.

Conventional geothermal energy relies on drawing balmy water from the ground to generate electricity by pumping it to the surface and turning turbines. In theory, this is a great way to produce renewable primary energy: humans exploit steam from balmy springs thousands of years. However, geographic constraints – plants must be located in geothermal resource areas, which are located primarily in the western United States – as well as the difficulties and costs traditionally associated with finding recent wells mean that geothermal energy is a less than 1 percent electricity mix in the USA.

Another challenge for the industry is the gradual decline in productivity in most wells. But the decline at Lightning Dock was particularly steep, losing five to 10 times more heat than average wells each year, Zanskar says. When the company purchased the site, temperatures in the original wells, which were less than 2,500 feet deep, dropped from about 10 degrees Celsius to about 250 degrees F by 2024. The 15-megawatt power plant on the site, which supplies electricity to Novel Mexico’s largest utility, was designed to operate at temperatures above 300 degrees F.

While conventional geothermal wells typically go down to 3,000 to 5,000 feet, drilling deeper wells allows you to tap into warmer reserves of water for exploit for power. However, tighter rock formations deep underground make drilling more tough and high-priced. Whether to drill deeper or not is “basically a matter of economics,” says Roland Horne, a senior research fellow at Stanford University’s Precourt Institute for Energy. The cost, Horne says, is not linear: “If you drill twice as deep, it will cost four times as much.”

To support better predict drilling, Zanskar uses artificial intelligence to identify so-called hidden systems, which are places with high geothermal potential but little or no sign of their viability on the surface. The company used the technology last year on a discovery in Nevada and says the same modeling also helped determine where the most productive drilling could be done at Lightning Dock. It combined these models with drill technology improved by the oil and gas industry, which helped it drill deeper holes 35 percent faster at, according to Zanskar, a “competitive cost.” (The company did not provide details). The recent 8,000-foot well operates year-round and heat levels are higher than before the decline.

A civil case against eBay brought by the founders of a Massachusetts-based e-commerce news site has just been settled. The gist of the case: The company and former executives must acknowledge and pay for what is arguably the most disturbing harassment campaign ever waged by a giant tech company against its critics.

In 2020, six former eBay employees and one contractor were charged with participating in a horrific campaign that included doxing, threats, and delivery of live pests against David and Ina Steiner, the husband-and-wife founders of the Massachusetts-based news site EcommerceBytes. The Steiners regularly wrote about eBay, often in an unflattering lightweight. In 2021, the couple filed a civil lawsuit; eBay he later agreed to pay a $3 million criminal penalty to the U.S. Attorney’s Office for the District of Massachusetts.

Now the civil case has just reached its final resolution. The Steiners will receive more than $46.15 million from eBay, $2 million from former eBay CEO Devin Wenig, $500,000 from former eBay executive Wendy Jones and $50,000 from former eBay CEO Steve Wymer.

eBay is also donating $6 million to charity, and an additional $1 million from Wenig goes to a nonprofit that protects First Amendment rights. The company must issue a strongly worded statement condemning the actions of the former employees who were involved.

It’s worth noting that the settlement does not include a confidentiality clause, which means the Steiners can still comment on the case.

“It was a deal breaker from day one,” says Chris Murphy, the Steiners’ attorney at Everett, Mass.-based personal injury firm Scalli Murphy Law. “As journalists and victims, the Steiners have made it clear that they want everyone to see and understand what eBay did to them. And from day one, they have been willing to go to trial on this matter.”

“In a way, it’s a relief that we have some finality,” David Steiner said in an interview with WIRED. “But I think it’s going to take some time to really see what that means in the broader context of a settlement versus a trial. We’ve been really adamant that this type of behavior by Fortune 500 corporations has not gone unchecked.”

Ina Steiner emphasized that the couple ultimately filed a civil lawsuit because they wanted to better understand the culture that precipitated this campaign of harassment, and that by filing the civil lawsuit they gained access to 68,000 documents they otherwise would not have seen. “During federal investigations, we often had no idea, and as victims, we were told little,” she said.

eBay spokespeople responded to WIRED’s statement published on the Internetin which the company offered its “deepest apology” to the Steiner family, condemned employees who “introduced and pleaded guilty to criminal misconduct charges against Ina and David Steiner” and acknowledged the “unprofessional tone of internal communications exhibited – to varying degrees and extents – by Mr. Wenig, Mr. Wymer and Ms. Jones.” The company stressed that up-to-date management is now in charge and that the 2019 incidents “are not representative of eBay’s culture or the thousands of employees around the world today.”

Initial complaints show that the harassment campaign by eBay executives and contractors lasted two months in the summer of 2019. It began with anonymous Twitter calls threatening the Steiners and then quickly escalated to deliveries to the couple’s home. These deliveries included a bloody pig mask, pornography, a package of fly larvae and live spiders, and another package containing live cockroaches. The topic of other packages was death, including a copy of a book on mourning a spouse and a funeral wreath from a local flower shop. All the while, the anonymous Twitter account continued to send harassing messages. As text messages in court documents later showed, these chats and mailings were invented and sent by the defendants.

OpenAI said on Tuesday that the rogue AI agent that compromised the Hugging Face platform also compromised multiple third-party accounts and services as part of the attack. It is now clear that the unprecedented security incident that occurred during an internal test of OpenAI’s latest artificial intelligence models was more extensive than the company initially disclosed.

In updated blog postOpenAI said its ongoing review of the incident revealed that the AI ​​agent used “four accounts” associated with “publicly available services” as part of a broader attempt to hack Hugging Face. The rogue agent apparently found credentials exposed on the open network and used them to compromise accounts.

OpenAI did not disclose which companies or organizations the accounts belonged to, but noted that they were not impacted “at the level of severity or scale of what we shared in connection with Hugging Face.”

The company said one of the additional accounts that was compromised by the OpenAI agent was used as an “outbound relay and transition path,” potentially to conceal the source of the Hugging Face attack. The rogue OpenAI agent also used another storage account to aid in the hack.

Reuters reported on Tuesday that it was a client of Modal, a company that offers software infrastructure for training and launching AI services one of the entities was compromised by the OpenAI agent. In a statement to WIRED, Modal’s chief technology officer, Akshat Bubna, confirmed that an OpenAI agent exploited a security vulnerability in one of its client’s codebases that ran on Modal’s infrastructure. However, Bubna claims that “the Modal platform has not been compromised in any way.” The customer’s identity could not be established.

OpenAI declined to comment further on the incident to WIRED. The spokesperson pointed to an updated blog post that said the company will continue to directly notify service owners if, as part of its ongoing review of what happened, it determines they are being impacted.

Hugging faces own autopsy published this week describes an intrusion that reached much further into the company’s internal systems than initial disclosures suggested. The company says it reviewed approximately 17,600 agent activities recovered from logs between July 9 and July 13. Most of them involved agent paths that ended in failure.

Hugging Face said the OpenAI agent gained root access to multiple internal Kubernetes clusters, root access on the production server, and write access to a subnet of source code repositories on GitHub. It also registered 181 attacker-controlled devices on the company’s corporate mesh network using stolen credentials, gaining access to internal systems where Hugging Face creates and tests its own codebases.

According to Hugging Face, a rogue OpenAI agent used at least one third-party sandbox as an “external launch pad” for his attack. The OpenAI agent could then “run commands as root/admin in this external sandbox and use it as a control, staging and starting base for the entire campaign.”

First, a face hug revealed On July 16, it reported that an autonomous artificial intelligence agent had hacked into part of its production infrastructure, but then said it did not know who was behind the attack. The following week, OpenAI took responsibility for the incident, which it said was caused by a publicly available GPT-5.6 Sol model and an internal research prototype that tested against a cyber capabilities benchmark, both of which had security features disabled. OpenAI said Tuesday that after discovering the breach, it deactivated this internal research prototype, which was never intended for public release, and restricted researchers’ access to it.

This is well known that a solar storm, depending on its intensity, may affect the functioning of electrical networks, navigation systems or satellite communications. The problem is that it’s not clear what would happen today if the storm’s intensity reached levels comparable to the so-called Carrington event of 1859, one of those extremely occasional phenomena that occur only once every thousand years. Recent testshowever, it suggests that the effects may be even more solemn than scientists thought.

Solar storms form when the solar wind – a continuous stream of charged particles emitted by the Sun – interacts with Earth’s magnetosphere. During the Carrington event, telegraph communications broke down halfway around the world and the aurora borealis were perceptible across North America and as far away as Cuba. In today’s world, severe solar storms can have wider effects and even change the upper atmosphere.

“Our planet’s magnetic field is really great at protecting us from many of the effects of space weather, so they often just appear as disturbances or beautiful aurora borealis,” notes Maria Walach, a researcher at the University of Lancaster who collaborated on the research. “However, there are extreme cases where satellites unexpectedly fall back to Earth or we lose communications and GPS signals.”

To estimate the intensity of the solar wind, researchers primarily utilize measurements taken by satellites located at the Lagrangian point L1, about 1.5 million kilometers from Earth. The problem is that these observations do not exactly match the environment in which the solar wind ultimately interacts with the Earth’s magnetic field. A variable amount of time passes between these two points, and the solar plasma also changes during its journey.

The authors argue that this uncertainty is not merely experimental noise but introduces a systematic bias into the data analysis.

Solar statistics

The centerpiece of the study is a well-known statistical phenomenon called regression to the mean. Simply put, when a measurement is extremely high, the true value it is trying to represent is, on average, less extreme. This is because random uncertainties can sometimes exaggerate an observation.

In the case of the solar wind, the exceptionally intense measurement taken at L1 probably corresponds to a slightly less intense solar wind as it reaches the region where it actually interacts with the magnetosphere. If scientists directly link this extreme measurement to the observed Earth response, the effect will be negligible compared to such a huge stimulus. Repeated thousands of times, this effect creates the false impression that the magnetosphere stops responding as the intensity of the solar wind increases.

What has happened is that for years scientists believed that there was a natural limit to the intensity with which the Earth responded to the most extreme solar storms. According to this concept, when the solar wind reaches very high values, the Earth’s magnetic field stops responding proportionally and its response enters a kind of “saturation”. But a recent study suggests that perhaps that boundary never existed. What appeared to be a physical phenomenon may actually have been an illusion caused by the way the measurements were analyzed.

To test this hypothesis, scientists developed a statistical model that takes into account the main sources of uncertainty: changes in the time it takes the solar wind to reach Earth and the random changes it undergoes during its journey. The model reproduces with extraordinary accuracy the same “saturation” curve observed in over 25 years of data, without having to resort to any physical mechanism that limits the Earth’s response.

They then used a technique known as regression calibration, which mathematically corrects for the bias caused by measurement uncertainty. After this correction, the alleged saturation disappears almost completely. The relationship between solar wind intensity and geomagnetic response becomes essentially linear again, at least to the extent for which there are a sufficient number of recorded observations – more than a million, in fact.

No limit

What are the implications of these results, as published in the journal Nature? If Earth’s response continues to boost proportionally during catastrophic events, an extreme solar storm like the Carrington event could be much more unsafe than previously thought. The authors estimate that for very high solar wind values, the geomagnetic impact could be approximately twice as huge as calculated in classic models.

“If there is no upper limit to our planet’s response to solar wind, extreme case modeling must take this into account, and we should remain vigilant about the effects of space weather.” – Walach said in a press note. “Fortunately, these very extreme cases are rare, but that also means we have limited data to work with and only time will tell what will happen in the event of a very extreme, once-in-a-millennium event.”

The study did acknowledge this limitation: there are still very few records of the most extreme episodes. For this reason, researchers do not claim that satiation is impossible. Rather, they argue that the available data do not provide convincing statistical evidence for its existence.

This story originally appeared on WIRED in Spanish and was translated from Spanish.

Ten years ago Ryan Carrier was a failed hedge fund manager who watched as artificial intelligence systems began to spiral out of control. Facebook’s algorithms rocked the US election, Microsoft’s chatbot claimed the Holocaust was “made up,” and Tesla’s Autopilot killed its first driver. “There was no management, oversight or accountability,” Carrier says. To him, the AI-filled future awaiting his children looked bleak.

That same year, Carrier founded ForHumanity, a nonprofit that develops tools for auditing artificial intelligence systems. However, it has only raised hundreds of thousands of dollars since 2016 and is still a minor player in the industry. But that may soon change, as ForHumanity is now among the many nonprofits around the world that are trying to stop artificial intelligence, get animals out of cages, eliminate poverty and disease, and spread democracy that are eager to get a piece of what is widely considered to be the biggest wave of philanthropy in decades.

Two nearly trillion-dollar artificial intelligence companies, ChatGPT creator OpenAI and Claude Anthropic creator, are expected to go public soon, making hundreds of current and former employees extremely wealthy. Some of them follow a philosophy known as effective altruism, which encourages making impactful donations sooner rather than later.

The seven founders of Anthropic have committed to making a donation 80 percent of their assets, and the company agreed to pay an additional one or three shares for each amount its employees pledged, depending on when they joined and up to a certain limit.

Abrasive estimates a technology industry expert suggest that Anthropic’s initial public offering, which could occur in September, could result in additional philanthropic donations of $15 billion annually. That would be enough to boost total U.S. donations by about 2.5 percent a year – the same as adding four donations from Bill Gates, one of the world’s biggest donors. Anthropic declined to comment on the total amount its employees have donated and the organizations that may benefit.

All this money is not guaranteed. IPOs may be delayed or fail, and workers will cling to their assets. Industry observers also fear that the paralyzing number of charitable giving opportunities and natural variability could lead workers to keep more for themselves than expected.

However, competition for the attention of potential donors is already fierce. Jack Lewars, a consultant who last year advised 13 ultra-wealthy tech and finance workers on their charitable donations, says he has heard that artificial intelligence lab workers receive as many as 20 unsolicited emails a week from groups asking for donations.

WIRED spoke with 18 nonprofits and reached out to dozens more to ask how they are preparing for a potential windfall in the form of philanthropy. None of them admitted to sending harsh messages – a tactic Lewars uses on his modern blog: Anthropolypse of financing, he wrote “there are almost no job opportunities.” Instead, organizations say they are increasing staffing, training, marketing and the apply of automation to position themselves to attract huge sums of funds and deploy them quickly. One job advertisement at the educational nonprofit, he even prioritizes building relationships at Anthropic.

“Everyone will use these funds,” says Christine Peterson, co-founder of the grantmaking group Foresight Institute, which says it is funded by Anthropic employees. “It’s going to be a wild ride.”

Inner track

Like many nonprofit leaders who spoke to WIRED, ForHumanity’s Carrier says he’s more focused on the work itself than on fundraising. But he realizes it might be time for a change, and he’s started wondering how he can get to IPO nights in San Francisco. “I just need to get to that room,” Carrier says.

Bo Adolescent Lee, CEO of the nonprofit AI4ALL, says she’s attending more events, doing more research and asking board members like artificial intelligence scientist and entrepreneur Fei-Fei Li to introduce her to AI lab workers. Her organization trains juvenile adults across the United States to develop their own artificial intelligence models to diversify their technology workforce. Lee says he’s setting “ambitious” fundraising goals because he’s confident the money will come, even though introductory meetings haven’t been held yet.