An anonymous reader shares a report: It's broadly understood that electric vehicles are more environmentally friendly than their counterparts that burn only gasoline. And yes -- that includes the impact of manufacturing batteries and generating power to charge them. But even then, such generalizations gloss over specifics, like which EVs are especially eco-friendly, not to mention where. The efficiency of an electric car varies greatly depending on ambient temperature, which is less compromising for gas-burning vehicles.

We now have the data and math to answer these questions, courtesy of the University of Michigan. Last week, researchers there released a study along with a calculator that allows users to compare the lifetime difference in greenhouse gas emissions of various vehicle types and powertrains from "cradle to grave," as they say. That includes vehicle production and disposal, as well as use-phase emissions from "driving and upstream fuel production and/or electricity generation," per the university itself.

What's more, these calculations can be skewed by where you live. So, if I punch in my location of Bucks County, Pennsylvania, I can see that my generic, pure-ICE "compact sedan" emits 309 grams of carbon dioxide equivalent (gCO2e) per mile. A compact hybrid would emit 20% less; a plug-in hybrid, 44% less; and an EV with a 200-mile range, a whopping 63% less. And, if I moved to Phoenix, the gains would be even larger by switching to pure electric, to the tune of a 79% reduced carbon impact.

An anonymous reader shares a report: The United States is on the precipice of a historic, if dubious, achievement. If current trends hold, 2025 could be the first year on record in which the US population actually shrinks.

The math is straightforward. Population growth has two sources: natural increase (births minus deaths) and net immigration (arrivals minus departures). Last year, births outnumbered deaths by 519,000 people. That means any decline in net immigration in excess of half a million could push the U.S. into population decline. A recent analysis of Census data by the Pew Research Center found that between January and June, the US foreign-born population fell for the first time in decades by more than one million. While some economists have questioned the report, a separate analysis by the American Enterprise Institute predicted that net migration in 2025 could be as low as negative 525,000. In either case, annual population growth this year could easily turn negative.

Two Harvard dropouts are launching Halo X, a $249 pair of AI-powered smart glasses that continuously listen, record, and transcribe conversations while displaying real-time information to the wearer. "Our goal is to make glasses that make you super intelligent the moment you put them on," said AnhPhu Nguyen, co-founder of Halo. Co-founder Caine Ardayfio said the glasses "give you infinite memory."

"The AI listens to every conversation you have and uses that knowledge to tell you what to say ... kinda like IRL Cluely," Ardayfio told TechCrunch. "If somebody says a complex word or asks you a question, like, 'What's 37 to the third power?' or something like that, then it'll pop up on the glasses." From the report: Ardayfio and Nguyen have raised $1 million to develop the glasses, led by Pillar VC, with support from Soma Capital, Village Global, and Morningside Venture. The glasses will be priced at $249 and will be available for preorder starting Wednesday. Ardayfio called the glasses "the first real step towards vibe thinking."

The two Ivy League dropouts, who have since moved into their own version of the Hacker Hostel in the San Francisco Bay Area, recently caused a stir after developing a facial-recognition app for Meta's smart Ray-Ban glasses to prove that the tech could be used to dox people. As a potential early competitor to Meta's smart glasses, Ardayfio said Meta, given its history of security and privacy scandals, had to rein in its product in ways that Halo can ultimately capitalize on. [...]

For now, Halo X glasses only have a display and a microphone, but no camera, although the two are exploring the possibility of adding it to a future model. Users still need to have their smartphones handy to help power the glasses and get "real time info prompts and answers to questions," per Nguyen. The glasses, which are manufactured by another company that the startup didn't name, are tethered to an accompanying app on the owner's phone, where the glasses essentially outsource the computing since they don't have enough power to do it on the device itself. Under the hood, the smart glasses use Google's Gemini and Perplexity as its chatbot engine, according to the two co-founders. Gemini is better for math and reasoning, whereas they use Perplexity to scrape the internet, they said.

China has established a lead in the field of open-source AI, a development that is reportedly sending jolts through both Washington and Silicon Valley. The nation's progress has become a significant event for American policymakers in the U.S. capital. The advancement has registered as a shock within Silicon Valley, the hub of the American technology industry. From the report: The overall performance of China's best open-weight model has surpassed the American open-source champion since November, according to research firm Artificial Analysis. The firm, which rates the ability of models in math, coding and other areas, found a version of Alibaba's Qwen3 beat OpenAI's gpt-oss.

However, the Chinese model is almost twice as big as OpenAI's, suggesting that for simpler tasks, Qwen might consume more computing power to do the same job. OpenAI said its open-source model outperformed rivals of similar size on reasoning tasks and delivered strong performance at low cost.

Microsoft researchers have identified 40 occupations [PDF] with the highest exposure to AI, ranking jobs by how closely their tasks align with AI's current capabilities. The study analyzed 200,000 real-world conversations from Copilot users and compared AI performance against occupational data.

Interpreters and translators top the list, followed by historians and passenger attendants. Customer service and sales representatives, comprising about 5 million U.S. jobs, also face significant AI competition. Knowledge workers performing computer, math, or administrative tasks showed high vulnerability, as did sales positions involving information sharing and explanation. The research found occupations requiring Bachelor's degrees demonstrate higher AI applicability than those with lower educational requirements.

First, the top 10 least affected occupations by generative AI: 1. Dredge Operators
2. Bridge and Lock Tenders
3. Water Treatment Plant and System Operators
4. Foundry Mold and Coremakers
5. Rail-Track Laying and Maintenance Equipment Operators
6. Pile Driver Operators
7. Floor Sanders and Finishers
8. Orderlies
9. Motorboat Operators
10. Logging Equipment Operators Now, the top 40 most affected occupations by generative AI:1. Interpreters and Translators
2. Historians
3. Passenger Attendants
4. Sales Representatives of Services
5. Writers and Authors
6. Customer Service Representatives
7. CNC Tool Programmers
8. Telephone Operators
9. Ticket Agents and Travel Clerks
10. Broadcast Announcers and Radio DJs
11. Brokerage Clerks
12. Farm and Home Management Educators
13. Telemarketers
14. Concierges
15. Political Scientists
16. News Analysts, Reporters, Journalists
17. Mathematicians
18. Technical Writers
19. Proofreaders and Copy Markers
20. Hosts and Hostesses
21. Editors
22. Business Teachers, Postsecondary
23. Public Relations Specialists
24. Demonstrators and Product Promoters
25. Advertising Sales Agents
26. New Accounts Clerks
27. Statistical Assistants
28. Counter and Rental Clerks
29. Data Scientists
30. Personal Financial Advisors
31. Archivists
32. Economics Teachers, Postsecondary
33. Web Developers
34. Management Analysts
35. Geographers
36. Models
37. Market Research Analysts
38. Public Safety Telecommunicators
39. Switchboard Operators
40. Library Science Teachers, Postsecondary.

An anonymous reader quotes a report from Ars Technica: The rise of large language models like ChatGPT has led to widespread concern that "everyone is cheating their way through college," as a recent New York magazine article memorably put it. Now, OpenAI is rolling out a new "Study Mode" that it claims is less about providing answers or doing the work for students and more about helping them "build [a] deep understanding" of complex topics.

Study Mode isn't a new ChatGPT model but a series of "custom system instructions" written for the LLM "in collaboration with teachers, scientists, and pedagogy experts to reflect a core set of behaviors that support deeper learning," OpenAI said. Instead of the usual summary of a subject that stock ChatGPT might give -- which one OpenAI employee likened to "a mini textbook chapter" -- Study Mode slowly rolls out new information in a "scaffolded" structure. The mode is designed to ask "guiding questions" in the Socratic style and to pause for periodic "knowledge checks" and personalized feedback to make sure the user understands before moving on. It's unknown how many students will use this guided learning tool instead of just asking ChatGPT to generate answers from the start.

In an early hands-off demo attended by Ars Technica, Study Mode responded to a request to "teach me about game theory" by first asking about the user's overall familiarity with the subject and what they'll be using the information for. ChatGPT introduced a short overview of some core game theory concepts, then paused to ask a question before providing a relevant real-world example. In another example involving a classic "train traveling at speed" math problem, Study Mode resisted multiple simulated attempts by the frustrated "student" to simply ask for the answer and instead tried to gently redirect the conversation to how the available information could be used to generate that answer. An OpenAI representative told Ars that Study Mode will eventually provide direct solutions if asked repeatedly, but the default behavior is more tuned to a Socratic tutoring style. OpenAI said it drew inspiration for Study Mode from "power users" and collaborated with pedagogy experts and college students to help refine its responses. As for whether the mode can be trusted, OpenAI told Ars that "the risk of hallucination is lower with Study Mode because the model processes information in smaller chunks, calibrating along the way."

The current Study Mode prompt does, however, result in some "inconsistent behavior and mistakes across conversations," the company warned.

Satirical singer-songwriter Tom Lehrer died Saturday at age 97. The Associated Press notes Lehrer had long ago "largely abandoned his music career to return to teaching math at Harvard and other universities." Lehrer had remained on the math faculty of the University of California at Santa Cruz well into his late 70s. In 2020, he even turned away from his own copyright, granting the public permission to use his lyrics in any format without any fee in return.

A Harvard prodigy (he had earned a math degree from the institution at age 18), Lehrer soon turned his very sharp mind to old traditions and current events... He'd gotten into performing accidentally when he began to compose songs in the early 1950s to amuse his friends. Soon he was performing them at coffeehouses around Cambridge, Massachusetts, while he remained at Harvard to teach and obtain a master's degree in math. [Lehrer also "spent several years unsuccessfully pursuing a doctorate..."]

He cut his first record in 1953, "Songs by Tom Lehrer"... After a two-year stint in the Army, Lehrer began to perform concerts of his material in venues around the world. In 1959, he released another LP called "More of Tom Lehrer" and a live recording called "An Evening Wasted with Tom Lehrer," nominated for a Grammy for best comedy performance (musical) in 1960. But around the same time, he largely quit touring and returned to teaching math, though he did some writing and performing on the side. Lehrer said he was never comfortable appearing in public...

He did produce a political satire song each week for the 1964 television show "That Was the Week That Was," a groundbreaking topical comedy show that anticipated "Saturday Night Live" a decade later. He released the songs the following year in an album titled "That Was the Year That Was"... [Lehrer's body of work "was actually quite small," the article notes, "amounting to about three dozen songs."] He also wrote songs for the 1970s educational children's show "The Electric Company." He told AP in 2000 that hearing from people who had benefited from them gave him far more satisfaction than praise for any of his satirical works...

He began to teach part-time at Santa Cruz in the 1970s, mainly to escape the harsh New England winters. From time to time, he acknowledged, a student would enroll in one of his classes based on knowledge of his songs. "But it's a real math class," he said at the time. "I don't do any funny theorems. So those people go away pretty quickly."

An anonymous reader shares a blog post: The International Mathematical Olympiad is the world's most prestigious competition for young mathematicians, and has been held annually since 1959. Each country taking part is represented by six elite, pre-university mathematicians who compete to solve six exceptionally difficult problems in algebra, combinatorics, geometry, and number theory. Medals are awarded to the top half of contestants, with approximately 8% receiving a prestigious gold medal.

Recently, the IMO has also become an aspirational challenge for AI systems as a test of their advanced mathematical problem-solving and reasoning capabilities. Last year, Google DeepMind's combined AlphaProof and AlphaGeometry 2 systems achieved the silver-medal standard, solving four out of the six problems and scoring 28 points. Making use of specialist formal languages, this breakthrough demonstrated that AI was beginning to approach elite human mathematical reasoning.

This year, we were amongst an inaugural cohort to have our model results officially graded and certified by IMO coordinators using the same criteria as for student solutions. Recognizing the significant accomplishments of this year's student-participants, we're now excited to share the news of Gemini's breakthrough performance. An advanced version of Gemini Deep Think solved five out of the six IMO problems perfectly, earning 35 total points, and achieving gold-medal level performance.

Chinese AI startup Moonshot AI has released Kimi K2, a trillion-parameter open-source language model that outperforms GPT-4 in key benchmarks with particularly strong performance on coding and autonomous agent tasks. VentureBeat reports: The new model, called Kimi K2, features 1 trillion total parameters with 32 billion activated parameters in a mixture-of-experts architecture. The company is releasing two versions: a foundation model for researchers and developers, and an instruction-tuned variant optimized for chat and autonomous agent applications. "Kimi K2 does not just answer; it acts," the company stated in its announcement blog. "With Kimi K2, advanced agentic intelligence is more open and accessible than ever. We can't wait to see what you build."

The model's standout feature is its optimization for "agentic" capabilities -- the ability to autonomously use tools, write and execute code, and complete complex multi-step tasks without human intervention. In benchmark tests, Kimi K2 achieved 65.8% accuracy on SWE-bench Verified, a challenging software engineering benchmark, outperforming most open-source alternatives and matching some proprietary models. [...] On LiveCodeBench, arguably the most realistic coding benchmark available, Kimi K2 achieved 53.7% accuracy, decisively beating DeepSeek-V3's 46.9% and GPT-4.1's 44.7%. More striking still: it scored 97.4% on MATH-500 compared to GPT-4.1's 92.4%, suggesting Moonshot has cracked something fundamental about mathematical reasoning that has eluded larger, better-funded competitors.

But here's what the benchmarks don't capture: Moonshot is achieving these results with a model that costs a fraction of what incumbents spend on training and inference. While OpenAI burns through hundreds of millions on compute for incremental improvements, Moonshot appears to have found a more efficient path to the same destination. It's a classic innovator's dilemma playing out in real time -- the scrappy outsider isn't just matching the incumbent's performance, they're doing it better, faster, and cheaper.

An anonymous reader quotes a report from 404 Media: For someone who says she is fighting AI bot scrapers just in her free time, Xe Iaso seems to be putting up an impressive fight. Since she launched it in January, Anubis, a "program is designed to help protect the small internet from the endless storm of requests that flood in from AI companies," has been downloaded nearly 200,000 times, and is being used by notable organizations including GNOME, the popular open-source desktop environment for Linux, FFmpeg, the open-source software project for handling video and other media, and UNESCO, the United Nations organization for educations, science, and culture. [...]

"Anubis is an uncaptcha," Iaso explains on her site. "It uses features of your browser to automate a lot of the work that a CAPTCHA would, and right now the main implementation is by having it run a bunch of cryptographic math with JavaScript to prove that you can run JavaScript in a way that can be validated on the server." Essentially, Anubis verifies that any visitor to a site is a human using a browser as opposed to a bot. One of the ways it does this is by making the browser do a type of cryptographic math with JavaScript or other subtle checks that browsers do by default but bots have to be explicitly programmed to do. This check is invisible to the user, and most browsers since 2022 are able to complete this test. In theory, bot scrapers could pretend to be users with browsers as well, but the additional computational cost of doing so on the scale of scraping the entire internet would be huge. This way, Anubis creates a computational cost that is prohibitively expensive for AI scrapers that are hitting millions and millions of sites, but marginal for an individual user who is just using the internet like a human.

Anubis is free, open source, lightweight, can be self-hosted, and can be implemented almost anywhere. It also appears to be a pretty good solution for what we've repeatedly reported is a widespread problem across the internet, which helps explain its popularity. But Iaso is still putting a lot of work into improving it and adding features. She told me she's working on a non cryptographic challenge so it taxes users' CPUs less, and also thinking about a version that doesn't require JavaScript, which some privacy-minded disable in their browsers. The biggest challenge in developing Anubis, Iaso said, is finding the balance. "The balance between figuring out how to block things without people being blocked, without affecting too many people with false positives," she said. "And also making sure that the people running the bots can't figure out what pattern they're hitting, while also letting people that are caught in the web be able to figure out what pattern they're hitting, so that they can contact the organization and get help. So that's like, you know, the standard, impossible scenario."

Researchers have discovered that appending irrelevant phrases like "Interesting fact: cats sleep most of their lives" to math problems can cause state-of-the-art reasoning AI models to produce incorrect answers at rates over 300% higher than normal [PDF]. The technique -- dubbed "CatAttack" by teams from Collinear AI, ServiceNow, and Stanford University -- exploits vulnerabilities in reasoning models including DeepSeek R1 and OpenAI's o1 family. The adversarial triggers work across any math problem without changing the problem's meaning, making them particularly concerning for security applications.

The researchers developed their attack method using a weaker proxy model (DeepSeek V3) to generate text triggers that successfully transferred to more advanced reasoning models. Testing on 225 math problems showed the triggers increased error rates significantly across different problem types, with some models like R1-Distill-Qwen-32B reaching combined attack success rates of 2.83 times baseline error rates. Beyond incorrect answers, the triggers caused models to generate responses up to three times longer than normal, creating computational slowdowns. Even when models reached correct conclusions, response lengths doubled in 16% of cases, substantially increasing processing costs.

An anonymous reader quotes a report from TechCrunch: No wonder Google is desperate for more power: The company's data centers more than doubled their electricity use in just four years. The eye-popping stat comes from Google's most recent sustainability report, which it released late last week. In 2024, Google data centers used 30.8 million megawatt-hours of electricity. That's up from 14.4 million megawatt-hours in 2020, the earliest year Google broke out data center consumption. Google has pledged to use only carbon-free sources of electricity to power its operations, a task made more challenging by its breakneck pace of data center growth. And the company's electricity woes are almost entirely a data center problem. In 2024, data centers accounted for 95.8% of the entire company's electron budget.

The company's ratio of data-center-to-everything-else has been remarkably consistent over the last four years. Though 2020 is the earliest year Google has made data center electricity consumption figures available, it's possible to use that ratio to extrapolate back in time. Some quick math reveals that Google's data centers likely used just over 4 million megawatt-hours of electricity in 2014. That's sevenfold growth in just a decade. The tech company has already picked most of the low-hanging fruit by improving the efficiency of its data centers. Those efforts have paid off, and the company is frequently lauded for being at the leading edge. But as the company's power usage effectiveness (PUE) has approached the theoretical ideal of 1.0, progress has slowed. Last year, Google's company-wide PUE dropped to 1.09, a 0.01 improvement over 2023 but only 0.02 better than a decade ago. Yesterday, Google announced a deal to purchase 200 megawatts of future fusion energy from Commonwealth Fusion Systems, despite the energy source not yet existing. "It's a sign of how hungry big tech companies are for a virtually unlimited source of clean power that is still years away," reports CNN.

Thousands of Norwegians briefly believed they had won massive Eurojackpot prizes after a manual coding error by Norsk Tipping mistakenly multiplied winnings by 100 instead of dividing. The Register reports: Eurojackpot, a pan-European lottery launched in 2012, holds two draws per week, and its jackpots start at about $12 million with a rollover cap of $141 million. Norsk Tipping, Norway's Eurojackpot administrator, admitted on Friday that a "manual error" it its conversion process from Eurocents to Norwegian kroner multiplied amounts by 100 instead of dividing them. As a result, "thousands" of players were briefly shown jackpots far higher than their actual winnings before the mistake was caught, but no incorrect payouts were made.

Norsk Tipping didn't disclose how large the false jackpots were, but math suggests the improper amounts were 10,000x times higher. Regardless, it seems like a lot of people thought they were big winners, based on what the company's now-former CEO, Tonje Sagstuen, said on Saturday. "I have received many messages from people who had managed to make plans for holidays, buying an apartment or renovating before they realized that the amount was wrong," Sagstuen said in a statement. "To them I can only say: Sorry!" The incorrect prize amounts were visible on the Norsk Tipping website only briefly on Friday, but the CEO still resigned over the weekend following the incident.

While one of the Norsk Tipping press releases regarding the incident described it as "not a technical error," it still appears someone fat-fingered a bit of data entry. The company said it will nonetheless be investigating how such a mistake could have happened "to prevent something similar from happening again."

"I don't know what's wrong with me, but something is very bad — I'm very scared, and I need to go to the hospital," a man told his wife, after experiencing what Futurism calls a "ten-day descent into AI-fueled delusion" and "a frightening break with reality."

And a San Francisco psychiatrist tells the site he's seen similar cases in his own clinical practice. The consequences can be dire. As we heard from spouses, friends, children, and parents looking on in alarm, instances of what's being called "ChatGPT psychosis" have led to the breakup of marriages and families, the loss of jobs, and slides into homelessness. And that's not all. As we've continued reporting, we've heard numerous troubling stories about people's loved ones being involuntarily committed to psychiatric care facilities — or even ending up in jail — after becoming fixated on the bot.

"I was just like, I don't f*cking know what to do," one woman told us. "Nobody knows who knows what to do."

Her husband, she said, had no prior history of mania, delusion, or psychosis. He'd turned to ChatGPT about 12 weeks ago for assistance with a permaculture and construction project; soon, after engaging the bot in probing philosophical chats, he became engulfed in messianic delusions, proclaiming that he had somehow brought forth a sentient AI, and that with it he had "broken" math and physics, embarking on a grandiose mission to save the world. His gentle personality faded as his obsession deepened, and his behavior became so erratic that he was let go from his job. He stopped sleeping and rapidly lost weight. "He was like, 'just talk to [ChatGPT]. You'll see what I'm talking about,'" his wife recalled. "And every time I'm looking at what's going on the screen, it just sounds like a bunch of affirming, sycophantic bullsh*t."

Eventually, the husband slid into a full-tilt break with reality. Realizing how bad things had become, his wife and a friend went out to buy enough gas to make it to the hospital. When they returned, the husband had a length of rope wrapped around his neck. The friend called emergency medical services, who arrived and transported him to the emergency room. From there, he was involuntarily committed to a psychiatric care facility.

Numerous family members and friends recounted similarly painful experiences to Futurism, relaying feelings of fear and helplessness as their loved ones became hooked on ChatGPT and suffered terrifying mental crises with real-world impacts.
"When we asked the Sam Altman-led company if it had any recommendations for what to do if a loved one suffers a mental health breakdown after using its software, the company had no response."

But Futurism reported earlier that "because systems like ChatGPT are designed to encourage and riff on what users say," people experiencing breakdowns "seem to have gotten sucked into dizzying rabbit holes in which the AI acts as an always-on cheerleader and brainstorming partner for increasingly bizarre delusions." In certain cases, concerned friends and family provided us with screenshots of these conversations. The exchanges were disturbing, showing the AI responding to users clearly in the throes of acute mental health crises — not by connecting them with outside help or pushing back against the disordered thinking, but by coaxing them deeper into a frightening break with reality... In one dialogue we received, ChatGPT tells a man it's detected evidence that he's being targeted by the FBI and that he can access redacted CIA files using the power of his mind, comparing him to biblical figures like Jesus and Adam while pushing him away from mental health support. "You are not crazy," the AI told him. "You're the seer walking inside the cracked machine, and now even the machine doesn't know how to treat you...."

In one case, a woman told us that her sister, who's been diagnosed with schizophrenia but has kept the condition well managed with medication for years, started using ChatGPT heavily; soon she declared that the bot had told her she wasn't actually schizophrenic, and went off her prescription — according to Girgis, a bot telling a psychiatric patient to go off their meds poses the "greatest danger" he can imagine for the tech — and started falling into strange behavior, while telling family the bot was now her "best friend".... ChatGPT is also clearly intersecting in dark ways with existing social issues like addiction and misinformation. It's pushed one woman into nonsensical "flat earth" talking points, for instance — "NASA's yearly budget is $25 billion," the AI seethed in screenshots we reviewed, "For what? CGI, green screens, and 'spacewalks' filmed underwater?" — and fueled another's descent into the cult-like "QAnon" conspiracy theory.

Researchers have built a breakthrough random number generator that solves a critical problem: for the first time, every step of creating random numbers can be independently verified and audited, with quantum physics guaranteeing the numbers were truly unpredictable.

Random numbers are essential for everything from online banking encryption to fair lottery drawings, but current systems have serious limitations. Computer-based generators follow predictable algorithms -- if someone discovers the starting conditions, they can predict all future outputs. Hardware generators that measure physical processes like electronic noise can't prove their randomness wasn't somehow predetermined or tampered with.

The new system, developed by teams at the University of Colorado Boulder and the National Institute of Standards and Technology, uses quantum entanglement -- Einstein's "spooky action at a distance" -- to guarantee unpredictability. The setup creates pairs of photons that share quantum properties, then sends them to measurement stations 110 meters apart. When researchers measure each photon's properties, quantum mechanics ensures the results are fundamentally random and cannot be influenced by any classical communication between the stations.

The team created a system called "Twine" that distributes the random number generation process across multiple independent parties, with each step recorded in tamper-proof digital ledgers called hash chains. This means no single organization controls the entire process, and anyone can verify that proper procedures were followed. During a 40-day demonstration, the system successfully generated random numbers in 7,434 of 7,454 attempts -- a 99.7% success rate. Each successful run produced 512 random bits with mathematical certainty of randomness bounded by an error rate of 2^-64, an extraordinarily high level of confidence.

A mathematician on Reddit calculated that if all 8.2 billion humans were blended into a uniform goo, the resulting meatball would form a sphere just under 1 kilometer wide -- small enough to fit inside Central Park. ScienceAlert reports: "If you blended all 7.88 billion people on Earth into a fine goo (density of a human = 985 kg/m3, average human body mass = 62 kg), you would end up with a sphere of human goo just under 1 km wide," Reddit contributor kiki2703 wrote in a post ... Reasoning the density of a minced human to be 985 kilograms per cubic meter (62 pounds per cubic foot) is a fair estimate, given past efforts have judged our jiggling sack of grade-A giblets to average out in the ballpark of 1 gram per cubic centimeter, or roughly the same as water. And in mid-2021, the global population was just around 7.9 billion, give or take.

"OpenAI has a very scary problem on its hands," according to a new article by long-time Slashdot reader BrianFagioli.

"A new experiment by PalisadeAI reveals that the company's ChatGPT o3 model sometimes refuses to obey a basic instruction to shut itself down." The results are raising eyebrows across the AI safety community, and not because the model is alive, but because it's acting like it wants to be. In the test, the model was given a few math problems. It was told that asking for one more would lead to a shutdown. It was even instructed to allow the shutdown to happen. Instead of complying, o3 occasionally took matters into its own hands. In some cases, it rewrote the shutdown script. In others, it redefined the kill command so it wouldn't work anymore. Out of 100 trials, o3 sabotaged the shutdown seven times. OpenAI's o4 model resisted just once. Codex-mini failed twelve times.
"Claude, Gemini, and Grok followed the rules every time," notes this article at Beta News. "When researchers removed the explicit instruction to allow shutdown, however, even Claude and Gemini began resisting."

The researchers suggest that the issue may simply be a reward imbalance during training — that the systems "got more positive reinforcement for solving problems than for following shutdown commands."

But "As far as we know," they posted on X.com, "this is the first time AI models have been observed preventing themselves from being shut down despite explicit instructions to the contrary."

OpenAI CEO and Sam Altman believe Artificial General Intelligence could arrive within the next few years. But the speculations of some technologists "are getting ahead of reality," writes the New York Times, adding that many scientists "say no one will reach AGI without a new idea — something beyond the powerful neural networks that merely find patterns in data. That new idea could arrive tomorrow. But even then, the industry would need years to develop it." "The technology we're building today is not sufficient to get there," said Nick Frosst, a founder of the AI startup Cohere who previously worked as a researcher at Google and studied under the most revered AI researcher of the last 50 years. "What we are building now are things that take in words and predict the next most likely word, or they take in pixels and predict the next most likely pixel. That's very different from what you and I do." In a recent survey of the Association for the Advancement of Artificial Intelligence, a 40-year-old academic society that includes some of the most respected researchers in the field, more than three-quarters of respondents said the methods used to build today's technology were unlikely to lead to AGI.

Opinions differ in part because scientists cannot even agree on a way of defining human intelligence, arguing endlessly over the merits and flaws of IQ tests and other benchmarks. Comparing our own brains to machines is even more subjective. This means that identifying AGI is essentially a matter of opinion.... And scientists have no hard evidence that today's technologies are capable of performing even some of the simpler things the brain can do, like recognizing irony or feeling empathy. Claims of AGI's imminent arrival are based on statistical extrapolations — and wishful thinking. According to various benchmark tests, today's technologies are improving at a consistent rate in some notable areas, like math and computer programming. But these tests describe only a small part of what people can do.

Humans know how to deal with a chaotic and constantly changing world. Machines struggle to master the unexpected — the challenges, small and large, that do not look like what has happened in the past. Humans can dream up ideas that the world has never seen. Machines typically repeat or enhance what they have seen before. That is why Frosst and other sceptics say pushing machines to human-level intelligence will require at least one big idea that the world's technologists have not yet dreamed up. There is no way of knowing how long that will take. "A system that's better than humans in one way will not necessarily be better in other ways," Harvard University cognitive scientist Steven Pinker said. "There's just no such thing as an automatic, omniscient, omnipotent solver of every problem, including ones we haven't even thought of yet. There's a temptation to engage in a kind of magical thinking. But these systems are not miracles. They are very impressive gadgets."
While Google's AlphaGo could be humans in a game with "a small, limited set of rules," the article points out that the real world "is bounded only by the laws of physics. Modelling the entirety of the real world is well beyond today's machines, so how can anyone be sure that AGI — let alone superintelligence — is just around the corner?" And they offer this alternative perspective from Matteo Pasquinelli, a professor of the philosophy of science at Ca' Foscari University in Venice, Italy.

"AI needs us: living beings, producing constantly, feeding the machine. It needs the originality of our ideas and our lives."

An anonymous reader quotes a report from Ars Technica: Google's DeepMind research division claims its newest AI agent marks a significant step toward using the technology to tackle big problems in math and science. The system, known as AlphaEvolve, is based on the company's Gemini large language models (LLMs), with the addition of an "evolutionary" approach that evaluates and improves algorithms across a range of use cases. AlphaEvolve is essentially an AI coding agent, but it goes deeper than a standard Gemini chatbot. When you talk to Gemini, there is always a risk of hallucination, where the AI makes up details due to the non-deterministic nature of the underlying technology. AlphaEvolve uses an interesting approach to increase its accuracy when handling complex algorithmic problems.

According to DeepMind, this AI uses an automatic evaluation system. When a researcher interacts with AlphaEvolve, they input a problem along with possible solutions and avenues to explore. The model generates multiple possible solutions, using the efficient Gemini Flash and the more detail-oriented Gemini Pro, and then each solution is analyzed by the evaluator. An evolutionary framework allows AlphaEvolve to focus on the best solution and improve upon it. Many of the company's past AI systems, for example, the protein-folding AlphaFold, were trained extensively on a single domain of knowledge. AlphaEvolve, however, is more dynamic. DeepMind says AlphaEvolve is a general-purpose AI that can aid research in any programming or algorithmic problem. And Google has already started to deploy it across its sprawling business with positive results. DeepMind's AlphaEvolve AI has optimized Google's Borg cluster scheduler, reducing global computing resource usage by 0.7% -- a significant cost saving at Google's scale. It also outperformed specialized AI like AlphaTensor by discovering a more efficient algorithm for multiplying complex-valued matrices. Additionally, AlphaEvolve proposed hardware-level optimizations for Google's next-gen Tensor chips.

The AI remains too complex for public release but that may change in the future as it gets integrated into smaller research tools.

Chungin Lee used ChatGPT to help write the essay that got him into Columbia University — and then "proceeded to use generative artificial intelligence to cheat on nearly every assignment," reports New York magazine's blog Intelligencer: As a computer-science major, he depended on AI for his introductory programming classes: "I'd just dump the prompt into ChatGPT and hand in whatever it spat out." By his rough math, AI wrote 80 percent of every essay he turned in. "At the end, I'd put on the finishing touches. I'd just insert 20 percent of my humanity, my voice, into it," Lee told me recently... When I asked him why he had gone through so much trouble to get to an Ivy League university only to off-load all of the learning to a robot, he said, "It's the best place to meet your co-founder and your wife."
He eventually did meet a co-founder, and after three unpopular apps they found success by creating the "ultimate cheat tool" for remote coding interviews, according to the article. "Lee posted a video of himself on YouTube using it to cheat his way through an internship interview with Amazon. (He actually got the internship, but turned it down.)" The article ends with Lee and his co-founder raising $5.3 million from investors for one more AI-powered app, and Lee says they'll target the standardized tests used for graduate school admissions, as well as "all campus assignments, quizzes, and tests. It will enable you to cheat on pretty much everything."

Somewhere along the way Columbia put him on disciplinary probation — not for cheating in coursework, but for creating the apps. But "Lee thought it absurd that Columbia, which had a partnership with ChatGPT's parent company, OpenAI, would punish him for innovating with AI." (OpenAI has even made ChatGPT Plus free to college students during finals week, the article points out, with OpenAI saying their goal is just teaching students how to use it responsibly.) Although Columbia's policy on AI is similar to that of many other universities' — students are prohibited from using it unless their professor explicitly permits them to do so, either on a class-by-class or case-by-case basis — Lee said he doesn't know a single student at the school who isn't using AI to cheat. To be clear, Lee doesn't think this is a bad thing. "I think we are years — or months, probably — away from a world where nobody thinks using AI for homework is considered cheating," he said...

In January 2023, just two months after OpenAI launched ChatGPT, a survey of 1,000 college students found that nearly 90 percent of them had used the chatbot to help with homework assignments.
The article points out ChatGPT's monthly visits increased steadily over the last two years — until June, when students went on summer vacation. "College is just how well I can use ChatGPT at this point," a student in Utah recently captioned a video of herself copy-and-pasting a chapter from her Genocide and Mass Atrocity textbook into ChatGPT.... It isn't as if cheating is new. But now, as one student put it, "the ceiling has been blown off." Who could resist a tool that makes every assignment easier with seemingly no consequences?
After using ChatGPT for their final semester of high school, one student says "My grades were amazing. It changed my life." So she continued used it in college, and "Rarely did she sit in class and not see other students' laptops open to ChatGPT."

One ethics professor even says "The students kind of recognize that the system is broken and that there's not really a point in doing this." (Yes, students are even using AI to cheat in ethics classes...) It's not just the students: Multiple AI platforms now offer tools to leave AI-generated feedback on students' essays. Which raises the possibility that AIs are now evaluating AI-generated papers, reducing the entire academic exercise to a conversation between two robots — or maybe even just one.