machine learning

researchers-isolate-memorization-from-reasoning-in-ai-neural-networks

Researchers isolate memorization from reasoning in AI neural networks

AI, AI alignment, AI architecture, AI behavior, AI memorization, AI research, AI safety, Allen Institute for AI, Biz & IT, computational neuroscience, copyright, generalization, Goodfire, gradient descent, K-FAC, loss curvature, machine learning, mechanistic interpretability, model editing, neural networks, OLMo, overfitting, transformer models, weight matrices / /

The hills and valleys of knowledge

Basic arithmetic ability lives in the memorization pathways, not logic circuits.

When engineers build AI language models like GPT-5 from training data, at least two major processing features emerge: memorization (reciting exact text they’ve seen before, like famous quotes or passages from books) and reasoning (solving new problems using general principles). New research from AI startup Goodfire.ai provides the first potentially clear evidence that these different functions actually work through completely separate neural pathways in the model’s architecture.

The researchers discovered that this separation proves remarkably clean. In a preprint paper released in late October, they described that when they removed the memorization pathways, models lost 97 percent of their ability to recite training data verbatim but kept nearly all their “logical reasoning” ability intact.

For example, at layer 22 in Allen Institute for AI’s OLMo-7B language model, the bottom 50 percent of weight components showed 23 percent higher activation on memorized data, while the top 10 percent showed 26 percent higher activation on general, non-memorized text. This mechanistic split enabled the researchers to surgically remove memorization while preserving other capabilities.

Perhaps most surprisingly, the researchers found that arithmetic operations seem to share the same neural pathways as memorization rather than logical reasoning. When they removed memorization circuits, mathematical performance plummeted to 66 percent while logical tasks remained nearly untouched. This discovery may explain why AI language models notoriously struggle with math without the use of external tools. They’re attempting to recall arithmetic from a limited memorization table rather than computing it, like a student who memorized times tables but never learned how multiplication works. The finding suggests that at current scales, language models treat “2+2=4” more like a memorized fact than a logical operation.

It’s worth noting that “reasoning” in AI research covers a spectrum of abilities that don’t necessarily match what we might call reasoning in humans. The logical reasoning that survived memory removal in this latest research includes tasks like evaluating true/false statements and following if-then rules, which are essentially applying learned patterns to new inputs. This also differs from the deeper “mathematical reasoning” required for proofs or novel problem-solving, which current AI models struggle with even when their pattern-matching abilities remain intact.

Looking ahead, if the information removal techniques receive further development in the future, AI companies could potentially one day remove, say, copyrighted content, private information, or harmful memorized text from a neural network without destroying the model’s ability to perform transformative tasks. However, since neural networks store information in distributed ways that are still not completely understood, for the time being, the researchers say their method “cannot guarantee complete elimination of sensitive information.” These are early steps in a new research direction for AI.

Traveling the neural landscape

To understand how researchers from Goodfire distinguished memorization from reasoning in these neural networks, it helps to know about a concept in AI called the “loss landscape.” The “loss landscape” is a way of visualizing how wrong or right an AI model’s predictions are as you adjust its internal settings (which are called “weights”).

Imagine you’re tuning a complex machine with millions of dials. The “loss” measures the number of mistakes the machine makes. High loss means many errors, low loss means few errors. The “landscape” is what you’d see if you could map out the error rate for every possible combination of dial settings.

During training, AI models essentially “roll downhill” in this landscape (gradient descent), adjusting their weights to find the valleys where they make the fewest mistakes. This process provides AI model outputs, like answers to questions.

Figure 1: Overview of our approach. We collect activations and gradients from a sample of training data (a), which allows us to approximate loss curvature w.r.t. a weight matrix using K-FAC (b). We decompose these weight matrices into components (each the same size as the matrix), ordered from high to low curvature. In language models, we show that data from different tasks interacts with parts of the spectrum of components differently (c).

Figure 1 from the paper “From Memorization to Reasoning in the Spectrum of Loss Curvature.” Credit: Merullo et al.

The researchers analyzed the “curvature” of the loss landscapes of particular AI language models, measuring how sensitive the model’s performance is to small changes in different neural network weights. Sharp peaks and valleys represent high curvature (where tiny changes cause big effects), while flat plains represent low curvature (where changes have minimal impact).

Using a technique called K-FAC (Kronecker-Factored Approximate Curvature), they found that individual memorized facts create sharp spikes in this landscape, but because each memorized item spikes in a different direction, when averaged together they create a flat profile. Meanwhile, reasoning abilities that many different inputs rely on maintain consistent moderate curves across the landscape, like rolling hills that remain roughly the same shape regardless of the direction from which you approach them.

“Directions that implement shared mechanisms used by many inputs add coherently and remain high-curvature on average,” the researchers write, describing reasoning pathways. In contrast, memorization uses “idiosyncratic sharp directions associated with specific examples” that appear flat when averaged across data.

Different tasks reveal a spectrum of mechanisms

The researchers tested their technique on multiple AI systems to verify the findings held across different architectures. They primarily used Allen Institute’s OLMo-2 family of open language models, specifically the 7-billion and 1-billion parameter versions, chosen because their training data is openly accessible. For vision models, they trained custom 86-million parameter Vision Transformers (ViT-Base models) on ImageNet with intentionally mislabeled data to create controlled memorization. They also validated their findings against existing memorization removal methods like BalancedSubnet to establish performance benchmarks.

The team tested their discovery by selectively removing low-curvature weight components from these trained models. Memorized content dropped to 3.4 percent recall from nearly 100 percent. Meanwhile, logical reasoning tasks maintained 95 to 106 percent of baseline performance.

These logical tasks included Boolean expression evaluation, logical deduction puzzles where solvers must track relationships like “if A is taller than B,” object tracking through multiple swaps, and benchmarks like BoolQ for yes/no reasoning, Winogrande for common sense inference, and OpenBookQA for science questions requiring reasoning from provided facts. Some tasks fell between these extremes, revealing a spectrum of mechanisms.

Mathematical operations and closed-book fact retrieval shared pathways with memorization, dropping to 66 to 86 percent performance after editing. The researchers found arithmetic particularly brittle. Even when models generated identical reasoning chains, they failed at the calculation step after low-curvature components were removed.

Figure 3: Sensitivity of different kinds of tasks to ablation of flatter eigenvectors. Parametric knowledge retrieval, arithmetic, and memorization are brittle, but openbook fact retrieval and logical reasoning is robust and maintain around 100% of original performance.

Figure 3 from the paper “From Memorization to Reasoning in the Spectrum of Loss Curvature.” Credit: Merullo et al.

“Arithmetic problems themselves are memorized at the 7B scale, or because they require narrowly used directions to do precise calculations,” the team explains. Open-book question answering, which relies on provided context rather than internal knowledge, proved most robust to the editing procedure, maintaining nearly full performance.

Curiously, the mechanism separation varied by information type. Common facts like country capitals barely changed after editing, while rare facts like company CEOs dropped 78 percent. This suggests models allocate distinct neural resources based on how frequently information appears in training.

The K-FAC technique outperformed existing memorization removal methods without needing training examples of memorized content. On unseen historical quotes, K-FAC achieved 16.1 percent memorization versus 60 percent for the previous best method, BalancedSubnet.

Vision transformers showed similar patterns. When trained with intentionally mislabeled images, the models developed distinct pathways for memorizing wrong labels versus learning correct patterns. Removing memorization pathways restored 66.5 percent accuracy on previously mislabeled images.

Limits of memory removal

However, the researchers acknowledged that their technique isn’t perfect. Once-removed memories might return if the model receives more training, as other research has shown that current unlearning methods only suppress information rather than completely erasing it from the neural network’s weights. That means the “forgotten” content can be reactivated with just a few training steps targeting those suppressed areas.

The researchers also can’t fully explain why some abilities, like math, break so easily when memorization is removed. It’s unclear whether the model actually memorized all its arithmetic or whether math just happens to use similar neural circuits as memorization. Additionally, some sophisticated capabilities might look like memorization to their detection method, even when they’re actually complex reasoning patterns. Finally, the mathematical tools they use to measure the model’s “landscape” can become unreliable at the extremes, though this doesn’t affect the actual editing process.

Photo of Benj Edwards

Benj Edwards is Ars Technica’s Senior AI Reporter and founder of the site’s dedicated AI beat in 2022. He’s also a tech historian with almost two decades of experience. In his free time, he writes and records music, collects vintage computers, and enjoys nature. He lives in Raleigh, NC.

Researchers isolate memorization from reasoning in AI neural networks Read More »

researchers-surprised-that-with-ai,-toxicity-is-harder-to-fake-than-intelligence

Researchers surprised that with AI, toxicity is harder to fake than intelligence

AI, AI alignment, AI behavior, AI detection, AI research, AI sycophancy, Apertus, Biz & IT, chatbots, computational Turing test, DeepSeek, Duke University, emotional AI, gemma, generative ai, large language models, LLaMA, machine learning, Mistral, NYU, Qwen, Social Media, University of Amsterdam, University of Zurich / /

The next time you encounter an unusually polite reply on social media, you might want to check twice. It could be an AI model trying (and failing) to blend in with the crowd.

On Wednesday, researchers from the University of Zurich, University of Amsterdam, Duke University, and New York University released a study revealing that AI models remain easily distinguishable from humans in social media conversations, with overly friendly emotional tone serving as the most persistent giveaway. The research, which tested nine open-weight models across Twitter/X, Bluesky, and Reddit, found that classifiers developed by the researchers detected AI-generated replies with 70 to 80 percent accuracy.

The study introduces what the authors call a “computational Turing test” to assess how closely AI models approximate human language. Instead of relying on subjective human judgment about whether text sounds authentic, the framework uses automated classifiers and linguistic analysis to identify specific features that distinguish machine-generated from human-authored content.

“Even after calibration, LLM outputs remain clearly distinguishable from human text, particularly in affective tone and emotional expression,” the researchers wrote. The team, led by Nicolò Pagan at the University of Zurich, tested various optimization strategies, from simple prompting to fine-tuning, but found that deeper emotional cues persist as reliable tells that a particular text interaction online was authored by an AI chatbot rather than a human.

The toxicity tell

In the study, researchers tested nine large language models: Llama 3.1 8B, Llama 3.1 8B Instruct, Llama 3.1 70B, Mistral 7B v0.1, Mistral 7B Instruct v0.2, Qwen 2.5 7B Instruct, Gemma 3 4B Instruct, DeepSeek-R1-Distill-Llama-8B, and Apertus-8B-2509.

When prompted to generate replies to real social media posts from actual users, the AI models struggled to match the level of casual negativity and spontaneous emotional expression common in human social media posts, with toxicity scores consistently lower than authentic human replies across all three platforms.

To counter this deficiency, the researchers attempted optimization strategies (including providing writing examples and context retrieval) that reduced structural differences like sentence length or word count, but variations in emotional tone persisted. “Our comprehensive calibration tests challenge the assumption that more sophisticated optimization necessarily yields more human-like output,” the researchers concluded.

Researchers surprised that with AI, toxicity is harder to fake than intelligence Read More »

google-plans-secret-ai-military-outpost-on-tiny-island-overrun-by-crabs

Google plans secret AI military outpost on tiny island overrun by crabs

AI, AI and work, AI infrastructure, australia, Biz & IT, Christmas Island, Cloud Computing, crabs, datacenters, defense, Google, machine learning, Military, red crabs, SubCom, subsea cables / /

Christmas Island Shire President Steve Pereira told Reuters that the council is examining community impacts before approving construction. “There is support for it, providing this data center actually does put back into the community with infrastructure, employment, and adding economic value to the island,” Pereira said.

That’s great, but what about the crabs?

Christmas Island’s annual crab migration is a natural phenomenon that Sir David Attenborough reportedly once described as one of his greatest TV moments when he visited the site in 1990.

Every year, millions of crabs emerge from the forest and swarm across roads, streams, rocks, and beaches to reach the ocean, where each female can produce up to 100,000 eggs. The tiny baby crabs that survive take about nine days to march back inland to the safety of the plateau.

While Google is seeking environmental approvals for its subsea cables, the timing could prove delicate for Christmas Island’s most famous residents. According to Parks Australia, the island’s annual red crab migration has already begun for 2025, with a major spawning event expected in just a few weeks, around November 15–16.

During peak migration times, sections of roads close at short notice as crabs move between forest and sea, and the island has built special crab bridges over roads to protect the migrating masses.

Parks Australia notes that while the migration happens annually, few baby crabs survive the journey from sea to forest most years, as they’re often eaten by fish, manta rays, and whale sharks. The successful migrations that occur only once or twice per decade (when large numbers of babies actually survive) are critical for maintaining the island’s red crab population.

How Google’s facility might coexist with 100 million marching crustaceans remains to be seen. But judging by the size of the event, it seems clear that it’s the crab’s world, and we’re just living in it.

Google plans secret AI military outpost on tiny island overrun by crabs Read More »

if-you-want-to-satiate-ai’s-hunger-for-power,-google-suggests-going-to-space

If you want to satiate AI’s hunger for power, Google suggests going to space

AI, artificial inteligence, Commercial space, Google, machine learning, planet labs, Science, Space / /

Google engineers think they already have all the pieces needed to build a data center in orbit.

With Project Suncatcher, Google will test its Tensor Processing Units on satellites. Credit: Google

It was probably always when, not if, Google would add its name to the list of companies intrigued by the potential of orbiting data centers.

Google announced Tuesday a new initiative, named Project Suncatcher, to examine the feasibility of bringing artificial intelligence to space. The idea is to deploy swarms of satellites in low-Earth orbit, each carrying Google’s AI accelerator chips designed for training, content generation, synthetic speech and vision, and predictive modeling. Google calls these chips Tensor Processing Units, or TPUs.

“Project Suncatcher is a moonshot exploring a new frontier: equipping solar-powered satellite constellations with TPUs and free-space optical links to one day scale machine learning compute in space,” Google wrote in a blog post.

“Like any moonshot, it’s going to require us to solve a lot of complex engineering challenges,” Google’s CEO, Sundar Pichai, wrote on X. Pichai noted that Google’s early tests show the company’s TPUs can withstand the intense radiation they will encounter in space. “However, significant challenges still remain like thermal management and on-orbit system reliability.”

The why and how

Ars reported on Google’s announcement on Tuesday, and Google published a research paper outlining the motivation for such a moonshot project. One of the authors, Travis Beals, spoke with Ars about Project Suncatcher and offered his thoughts on why it just might work.

“We’re just seeing so much demand from people for AI,” said Beals, senior director of Paradigms of Intelligence, a research team within Google. “So, we wanted to figure out a solution for compute that could work no matter how large demand might grow.”

Higher demand will lead to bigger data centers consuming colossal amounts of electricity. According to the MIT Technology Review, AI alone could consume as much electricity annually as 22 percent of all US households by 2028. Cooling is also a problem, often requiring access to vast water resources, raising important questions about environmental sustainability.

Google is looking to the sky to avoid potential bottlenecks. A satellite in space can access an infinite supply of renewable energy and an entire Universe to absorb heat.

“If you think about a data center on Earth, it’s taking power in and it’s emitting heat out,” Beals said. “For us, it’s the satellite that’s doing the same. The satellite is going to have solar panels … They’re going to feed that power to the TPUs to do whatever compute we need them to do, and then the waste heat from the TPUs will be distributed out over a radiator that will then radiate that heat out into space.”

Google envisions putting a legion of satellites into a special kind of orbit that rides along the day-night terminator, where sunlight meets darkness. This north-south, or polar, orbit would be synchronized with the Sun, allowing a satellite’s power-generating solar panels to remain continuously bathed in sunshine.

“It’s much brighter even than the midday Sun on Earth because it’s not filtered by Earth’s atmosphere,” Beals said.

This means a solar panel in space can produce up to eight times more power than the same collecting area on the ground, and you don’t need a lot of batteries to reserve electricity for nighttime. This may sound like the argument for space-based solar power, an idea first described by Isaac Asimov in his short story Reason published in 1941. But instead of transmitting the electricity down to Earth for terrestrial use, orbiting data centers would tap into the power source in space.

“As with many things, the ideas originate in science fiction, but it’s had a number of challenges, and one big one is, how do you get the power down to Earth?” Beals said. “So, instead of trying to figure out that, we’re embarking on this moonshot to bring [machine learning] compute chips into space, put them on satellites that have the solar panels and the radiators for cooling, and then integrate it all together so you don’t actually have to be powered on Earth.”

SpaceX is driving down launch costs, thanks to reusable rockets and an abundant volume of Starlink satellite launches. Credit: SpaceX

Google has a mixed record with its ambitious moonshot projects. One of the most prominent moonshot graduates is the self-driving car kit developer Waymo, which spun out to form a separate company in 2016 and is now operational. The Project Loon initiative to beam Internet signals from high-altitude balloons is one of the Google moonshots that didn’t make it.

Ars published two stories last week on the promise of space-based data centers. One of the startups in this field, named Starcloud, is partnering with Nvidia, the world’s largest tech company by market capitalization, to build a 5 gigawatt orbital data center with enormous solar and cooling panels approximately 4 kilometers (2.5 miles) in width and length. In response to that story, Elon Musk said SpaceX is pursuing the same business opportunity but didn’t provide any details. It’s worth noting that Google holds an estimated 7 percent stake in SpaceX.

Strength in numbers

Google’s proposed architecture differs from that of Starcloud and Nvidia in an important way. Instead of putting up just one or a few massive computing nodes, Google wants to launch a fleet of smaller satellites that talk to one another through laser data links. Essentially, a satellite swarm would function as a single data center, using light-speed interconnectivity to aggregate computing power hundreds of miles over our heads.

If that sounds implausible, take a moment to think about what companies are already doing in space today. SpaceX routinely launches more than 100 Starlink satellites per week, each of which uses laser inter-satellite links to bounce Internet signals around the globe. Amazon’s Kuiper satellite broadband network uses similar technology, and laser communications will underpin the US Space Force’s next-generation data-relay constellation.

Artist’s illustration of laser crosslinks in space. Credit: TESAT

Autonomously constructing a miles-long structure in orbit, as Nvidia and Starcloud foresee, would unlock unimagined opportunities. The concept also relies on tech that has never been tested in space, but there are plenty of engineers and investors who want to try. Starcloud announced an agreement last week with a new in-space assembly company, Rendezvous Robotics, to explore the use of modular, autonomous assembly to build Starcloud’s data centers.

Google’s research paper describes a future computing constellation of 81 satellites flying at an altitude of some 400 miles (650 kilometers), but Beals said the company could dial the total swarm size to as many spacecraft as the market demands. This architecture could enable terawatt-class orbital data centers, according to Google.

“What we’re actually envisioning is, potentially, as you scale, you could have many clusters,” Beals said.

Whatever the number, the satellites will communicate with one another using optical inter-satellite links for high-speed, low-latency connectivity. The satellites will need to fly in tight formation, perhaps a few hundred feet apart, with a swarm diameter of a little more than a mile, or about 2 kilometers. Google says its physics-based model shows satellites can maintain stable formations at such close ranges using automation and “reasonable propulsion budgets.”

“If you’re doing something that requires a ton of tight coordination between many TPUs—training, in particular—you want links that have as low latency as possible and as high bandwidth as possible,” Beals said. “With latency, you run into the speed of light, so you need to get things close together there to reduce latency. But bandwidth is also helped by bringing things close together.”

Some machine-learning applications could be done with the TPUs on just one modestly sized satellite, while others may require the processing power of multiple spacecraft linked together.

“You might be able to fit smaller jobs into a single satellite. This is an approach where, potentially, you can tackle a lot of inference workloads with a single satellite or a small number of them, but eventually, if you want to run larger jobs, you may need a larger cluster all networked together like this,” Beals said.

Google has worked on Project Suncatcher for more than a year, according to Beals. In ground testing, engineers tested Google’s TPUs under a 67 MeV proton beam to simulate the total ionizing dose of radiation the chip would see over five years in orbit. Now, it’s time to demonstrate Google’s AI chips, and everything else needed for Project Suncatcher will actually work in the real environment.

Google is partnering with Planet, the Earth-imaging company, to develop a pair of small prototype satellites for launch in early 2027. Planet builds its own satellites, so Google has tapped it to manufacture each spacecraft, test them, and arrange for their launch. Google’s parent company, Alphabet, also has an equity stake in Planet.

“We have the TPUs and the associated hardware, the compute payload… and we’re bringing that to Planet,” Beals said. “For this prototype mission, we’re really asking them to help us do everything to get that ready to operate in space.”

Beals declined to say how much the demo slated for launch in 2027 will cost but said Google is paying Planet for its role in the mission. The goal of the demo mission is to show whether space-based computing is a viable enterprise.

“Does it really hold up in space the way we think it will, the way we’ve tested on Earth?” Beals said.

Engineers will test an inter-satellite laser link and verify Google’s AI chips can weather the rigors of spaceflight.

“We’re envisioning scaling by building lots of satellites and connecting them together with ultra-high bandwidth inter-satellite links,” Beals said. “That’s why we want to launch a pair of satellites, because then we can test the link between the satellites.”

Evolution of a free-fall (no thrust) constellation under Earth’s gravitational attraction, modeled to the level of detail required to obtain Sun-synchronous orbits, in a non-rotating coordinate system. Credit: Google

Getting all this data to users on the ground is another challenge. Optical data links could also route enormous amounts of data between the satellites in orbit and ground stations on Earth.

Aside from the technical feasibility, there have long been economic hurdles to fielding large satellite constellations. But SpaceX’s experience with its Starlink broadband network, now with more than 8,000 active satellites, is proof that times have changed.

Google believes the economic equation is about to change again when SpaceX’s Starship rocket comes online. The company’s learning curve analysis shows launch prices could fall to less than $200 per kilogram by around 2035, assuming Starship is flying about 180 times per year by then. This is far below SpaceX’s stated launch targets for Starship but comparable to SpaceX’s proven flight rate with its workhorse Falcon 9 rocket.

It’s possible there could be even more downward pressure on launch costs if SpaceX, Nvidia, and others join Google in the race for space-based computing. The demand curve for access to space may only be eclipsed by the world’s appetite for AI.

“The more people are doing interesting, exciting things in space, the more investment there is in launch, and in the long run, that could help drive down launch costs,” Beals said. “So, it’s actually great to see that investment in other parts of the space supply chain and value chain. There are a lot of different ways of doing this.”

Photo of Stephen Clark

Stephen Clark is a space reporter at Ars Technica, covering private space companies and the world’s space agencies. Stephen writes about the nexus of technology, science, policy, and business on and off the planet.

If you want to satiate AI’s hunger for power, Google suggests going to space Read More »

openai-signs-massive-ai-compute-deal-with-amazon

OpenAI signs massive AI compute deal with Amazon

AI, AI chips, AI infrastructure, Amazon, AWS, Biz & IT, chatbots, datacenters, generative ai, Google, large language models, machine learning, microsoft, NVIDIA, openai, oracle, sam altman / /

On Monday, OpenAI announced it has signed a seven-year, $38 billion deal to buy cloud services from Amazon Web Services to power products like ChatGPT and Sora. It’s the company’s first big computing deal after a fundamental restructuring last week that gave OpenAI more operational and financial freedom from Microsoft.

The agreement gives OpenAI access to hundreds of thousands of Nvidia graphics processors to train and run its AI models. “Scaling frontier AI requires massive, reliable compute,” OpenAI CEO Sam Altman said in a statement. “Our partnership with AWS strengthens the broad compute ecosystem that will power this next era and bring advanced AI to everyone.”

OpenAI will reportedly use Amazon Web Services immediately, with all planned capacity set to come online by the end of 2026 and room to expand further in 2027 and beyond. Amazon plans to roll out hundreds of thousands of chips, including Nvidia’s GB200 and GB300 AI accelerators, in data clusters built to power ChatGPT’s responses, generate AI videos, and train OpenAI’s next wave of models.

Wall Street apparently liked the deal, because Amazon shares hit an all-time high on Monday morning. Meanwhile, shares for long-time OpenAI investor and partner Microsoft briefly dipped following the announcement.

Massive AI compute requirements

It’s no secret that running generative AI models for hundreds of millions of people currently requires a lot of computing power. Amid chip shortages over the past few years, finding sources of that computing muscle has been tricky. OpenAI is reportedly working on its own GPU hardware to help alleviate the strain.

But for now, the company needs to find new sources of Nvidia chips, which accelerate AI computations. Altman has previously said that the company plans to spend $1.4 trillion to develop 30 gigawatts of computing resources, an amount that is enough to roughly power 25 million US homes, according to Reuters.

OpenAI signs massive AI compute deal with Amazon Read More »

chatgpt-maker-reportedly-eyes-$1-trillion-ipo-despite-major-quarterly-losses

ChatGPT maker reportedly eyes $1 trillion IPO despite major quarterly losses

AI, AI infrastructure, Biz & IT, chatgpt, CoreWeave, generative ai, IPO, large language models, machine learning, microsoft, NVIDIA, openai, sam altman, SoftBank / /

An OpenAI spokesperson told Reuters that “an IPO is not our focus, so we could not possibly have set a date,” adding that the company is “building a durable business and advancing our mission so everyone benefits from AGI.”

Revenue grows as losses mount

The IPO preparations follow a restructuring of OpenAI completed on October 28 that reduced the company’s reliance on Microsoft, which has committed to investments of $13 billion and now owns about 27 percent of the company. OpenAI was most recently valued around $500 billion in private markets.

OpenAI started as a nonprofit in 2015, then added a for-profit arm a few years later with nonprofit oversight. Under the new structure, OpenAI is still controlled by a nonprofit, now called the OpenAI Foundation, but it gives the nonprofit a 26 percent stake in OpenAI Group and a warrant for additional shares if the company hits certain milestones.

A successful OpenAI IPO could represent a substantial gain for investors, including Microsoft, SoftBank, Thrive Capital, and Abu Dhabi’s MGX. But even so, OpenAI faces an uphill financial battle ahead. The ChatGPT maker expects to reach about $20 billion in revenue by year-end, according to people familiar with the company’s finances who spoke with Reuters, but its quarterly losses are significant.

Microsoft’s earnings filing on Wednesday offered a glimpse at the scale of those losses. The company reported that its share of OpenAI losses reduced Microsoft’s net income by $3.1 billion in the quarter that ended September 30. Since Microsoft owns 27 percent of OpenAI under the new structure, that suggests OpenAI lost about $11.5 billion during the quarter, as noted by The Register. That quarterly loss figure exceeds half of OpenAI’s expected revenue for the entire year.

ChatGPT maker reportedly eyes $1 trillion IPO despite major quarterly losses Read More »

nvidia-hits-record-$5-trillion-mark-as-ceo-dismisses-ai-bubble-concerns

Nvidia hits record $5 trillion mark as CEO dismisses AI bubble concerns

AI, AI chips, AI infrastructure, Apple, Biz & IT, Blackwell chips, data centers, Donald Trump, GPUs, Jensen Huang, machine learning, microsoft, Nokia, NVIDIA, openai, palantir, semiconductors, Uber / /

Partnerships and government contracts fuel optimism

At the GTC conference on Tuesday, Nvidia’s CEO went out of his way to repeatedly praise Donald Trump and his policies for accelerating domestic tech investment while warning that excluding China from Nvidia’s ecosystem could limit US access to half the world’s AI developers. The overall event stressed Nvidia’s role as an American company, with Huang even nodding to Trump’s signature slogan in his sign-off by thanking the audience for “making America great again.”

Trump’s cooperation is paramount for Nvidia because US export controls have effectively blocked Nvidia’s AI chips from China, costing the company billions of dollars in revenue. Bob O’Donnell of TECHnalysis Research told Reuters that “Nvidia clearly brought their story to DC to both educate and gain favor with the US government. They managed to hit most of the hottest and most influential topics in tech.”

Beyond the political messaging, Huang announced a series of partnerships and deals that apparently helped ease investor concerns about Nvidia’s future. The company announced collaborations with Uber Technologies, Palantir Technologies, and CrowdStrike Holdings, among others. Nvidia also revealed a $1 billion investment in Nokia to support the telecommunications company’s shift toward AI and 6G networking.

The agreement with Uber will power a fleet of 100,000 self-driving vehicles with Nvidia technology, with automaker Stellantis among the first to deliver the robotaxis. Palantir will pair Nvidia’s technology with its Ontology platform to use AI techniques for logistics insights, with Lowe’s as an early adopter. Eli Lilly plans to build what Nvidia described as the most powerful supercomputer owned and operated by a pharmaceutical company, relying on more than 1,000 Blackwell AI accelerator chips.

The $5 trillion valuation surpasses the total cryptocurrency market value and equals roughly half the size of the pan European Stoxx 600 equities index, Reuters notes. At current prices, Huang’s stake in Nvidia would be worth about $179.2 billion, making him the world’s eighth-richest person.

Nvidia hits record $5 trillion mark as CEO dismisses AI bubble concerns Read More »

openai-data-suggests-1-million-users-discuss-suicide-with-chatgpt-weekly

OpenAI data suggests 1 million users discuss suicide with ChatGPT weekly

AI, AI assistants, AI behavior, AI ethics, AI safety, AI sycophancy, Biz & IT, chatbots, chatgpt, machine learning, Mental Health, openai, sam altman / /

Earlier this month, the company unveiled a wellness council to address these concerns, though critics noted the council did not include a suicide prevention expert. OpenAI also recently rolled out controls for parents of children who use ChatGPT. The company says it’s building an age prediction system to automatically detect children using ChatGPT and impose a stricter set of age-related safeguards.

Rare but impactful conversations

The data shared on Monday appears to be part of the company’s effort to demonstrate progress on these issues, although it also shines a spotlight on just how deeply AI chatbots may be affecting the health of the public at large.

In a blog post on the recently released data, OpenAI says these types of conversations in ChatGPT that might trigger concerns about “psychosis, mania, or suicidal thinking” are “extremely rare,” and thus difficult to measure. The company estimates that around 0.07 percent of users active in a given week and 0.01 percent of messages indicate possible signs of mental health emergencies related to psychosis or mania. For emotional attachment, the company estimates around 0.15 percent of users active in a given week and 0.03 percent of messages indicate potentially heightened levels of emotional attachment to ChatGPT.

OpenAI also claims that on an evaluation of over 1,000 challenging mental health-related conversations, the new GPT-5 model was 92 percent compliant with its desired behaviors, compared to 27 percent for a previous GPT-5 model released on August 15. The company also says its latest version of GPT-5 holds up to OpenAI’s safeguards better in long conversations. OpenAI has previously admitted that its safeguards are less effective during extended conversations.

In addition, OpenAI says it’s adding new evaluations to attempt to measure some of the most serious mental health issues facing ChatGPT users. The company says its baseline safety testing for its AI language models will now include benchmarks for emotional reliance and non-suicidal mental health emergencies.

Despite the ongoing mental health concerns, OpenAI CEO Sam Altman announced on October 14 that the company will allow verified adult users to have erotic conversations with ChatGPT starting in December. The company had loosened ChatGPT content restrictions in February but then dramatically tightened them after the August lawsuit. Altman explained that OpenAI had made ChatGPT “pretty restrictive to make sure we were being careful with mental health issues” but acknowledged this approach made the chatbot “less useful/enjoyable to many users who had no mental health problems.”

If you or someone you know is feeling suicidal or in distress, please call the Suicide Prevention Lifeline number, 1-800-273-TALK (8255), which will put you in touch with a local crisis center.

OpenAI data suggests 1 million users discuss suicide with ChatGPT weekly Read More »

ars-live-recap:-is-the-ai-bubble-about-to-pop?-ed-zitron-weighs-in.

Ars Live recap: Is the AI bubble about to pop? Ed Zitron weighs in.

AI, AI agents, AI chips, AI criticism, AI economics, AI infrastructure, Benj Edwards, Biz & IT, bubble, Cerebras, datacenters, Ed Zitron, financial losses, generative ai, GPU, grok, large language models, machine learning, NVIDIA, openai, oracle, Power consumption, sam altman, Stargate, venture capital / /

Despite connection hiccups, we covered OpenAI’s finances, nuclear power, and Sam Altman.

On Tuesday of last week, Ars Technica hosted a live conversation with Ed Zitron, host of the Better Offline podcast and one of tech’s most vocal AI critics, to discuss whether the generative AI industry is experiencing a bubble and when it might burst. My Internet connection had other plans, though, dropping out multiple times and forcing Ars Technica’s Lee Hutchinson to jump in as an excellent emergency backup host.

During the times my connection cooperated, Zitron and I covered OpenAI’s financial issues, lofty infrastructure promises, and why the AI hype machine keeps rolling despite some arguably shaky economics underneath. Lee’s probing questions about per-user costs revealed a potential flaw in AI subscription models: Companies can’t predict whether a user will cost them $2 or $10,000 per month.

You can watch a recording of the event on YouTube or in the window below.

Our discussion with Ed Zitron. Click here for transcript.

“A 50 billion-dollar industry pretending to be a trillion-dollar one”

I started by asking Zitron the most direct question I could: “Why are you so mad about AI?” His answer got right to the heart of his critique: the disconnect between AI’s actual capabilities and how it’s being sold. “Because everybody’s acting like it’s something it isn’t,” Zitron said. “They’re acting like it’s this panacea that will be the future of software growth, the future of hardware growth, the future of compute.”

In one of his newsletters, Zitron describes the generative AI market as “a 50 billion dollar revenue industry masquerading as a one trillion-dollar one.” He pointed to OpenAI’s financial burn rate (losing an estimated $9.7 billion in the first half of 2025 alone) as evidence that the economics don’t work, coupled with a heavy dose of pessimism about AI in general.

Donald Trump listens as Nvidia CEO Jensen Huang speaks at the White House during an event on “Investing in America” on April 30, 2025, in Washington, DC. Credit: Andrew Harnik / Staff | Getty Images News

“The models just do not have the efficacy,” Zitron said during our conversation. “AI agents is one of the most egregious lies the tech industry has ever told. Autonomous agents don’t exist.”

He contrasted the relatively small revenue generated by AI companies with the massive capital expenditures flowing into the sector. Even major cloud providers and chip makers are showing strain. Oracle reportedly lost $100 million in three months after installing Nvidia’s new Blackwell GPUs, which Zitron noted are “extremely power-hungry and expensive to run.”

Finding utility despite the hype

I pushed back against some of Zitron’s broader dismissals of AI by sharing my own experience. I use AI chatbots frequently for brainstorming useful ideas and helping me see them from different angles. “I find I use AI models as sort of knowledge translators and framework translators,” I explained.

After experiencing brain fog from repeated bouts of COVID over the years, I’ve also found tools like ChatGPT and Claude especially helpful for memory augmentation that pierces through brain fog: describing something in a roundabout, fuzzy way and quickly getting an answer I can then verify. Along these lines, I’ve previously written about how people in a UK study found AI assistants useful accessibility tools.

Zitron acknowledged this could be useful for me personally but declined to draw any larger conclusions from my one data point. “I understand how that might be helpful; that’s cool,” he said. “I’m glad that that helps you in that way; it’s not a trillion-dollar use case.”

He also shared his own attempts at using AI tools, including experimenting with Claude Code despite not being a coder himself.

“If I liked [AI] somehow, it would be actually a more interesting story because I’d be talking about something I liked that was also onerously expensive,” Zitron explained. “But it doesn’t even do that, and it’s actually one of my core frustrations, it’s like this massive over-promise thing. I’m an early adopter guy. I will buy early crap all the time. I bought an Apple Vision Pro, like, what more do you say there? I’m ready to accept issues, but AI is all issues, it’s all filler, no killer; it’s very strange.”

Zitron and I agree that current AI assistants are being marketed beyond their actual capabilities. As I often say, AI models are not people, and they are not good factual references. As such, they cannot replace human decision-making and cannot wholesale replace human intellectual labor (at the moment). Instead, I see AI models as augmentations of human capability: as tools rather than autonomous entities.

Computing costs: History versus reality

Even though Zitron and I found some common ground about AI hype, I expressed a belief that criticism over the cost and power requirements of operating AI models will eventually not become an issue.

I attempted to make that case by noting that computing costs historically trend downward over time, referencing the Air Force’s SAGE computer system from the 1950s: a four-story building that performed 75,000 operations per second while consuming two megawatts of power. Today, pocket-sized phones deliver millions of times more computing power in a way that would be impossible, power consumption-wise, in the 1950s.

The blockhouse for the Semi-Automatic Ground Environment at Stewart Air Force Base, Newburgh, New York. Credit: Denver Post via Getty Images

“I think it will eventually work that way,” I said, suggesting that AI inference costs might follow similar patterns of improvement over years and that AI tools will eventually become commodity components of computer operating systems. Basically, even if AI models stay inefficient, AI models of a certain baseline usefulness and capability will still be cheaper to train and run in the future because the computing systems they run on will be faster, cheaper, and less power-hungry as well.

Zitron pushed back on this optimism, saying that AI costs are currently moving in the wrong direction. “The costs are going up, unilaterally across the board,” he said. Even newer systems like Cerebras and Grok can generate results faster but not cheaper. He also questioned whether integrating AI into operating systems would prove useful even if the technology became profitable, since AI models struggle with deterministic commands and consistent behavior.

The power problem and circular investments

One of Zitron’s most pointed criticisms during the discussion centered on OpenAI’s infrastructure promises. The company has pledged to build data centers requiring 10 gigawatts of power capacity (equivalent to 10 nuclear power plants, I once pointed out) for its Stargate project in Abilene, Texas. According to Zitron’s research, the town currently has only 350 megawatts of generating capacity and a 200-megawatt substation.

“A gigawatt of power is a lot, and it’s not like Red Alert 2,” Zitron said, referencing the real-time strategy game. “You don’t just build a power station and it happens. There are months of actual physics to make sure that it doesn’t kill everyone.”

He believes many announced data centers will never be completed, calling the infrastructure promises “castles on sand” that nobody in the financial press seems willing to question directly.

An orange, cloudy sky backlights a set of electrical wires on large pylons, leading away from the cooling towers of a nuclear power plant.

After another technical blackout on my end, I came back online and asked Zitron to define the scope of the AI bubble. He says it has evolved from one bubble (foundation models) into two or three, now including AI compute companies like CoreWeave and the market’s obsession with Nvidia.

Zitron highlighted what he sees as essentially circular investment schemes propping up the industry. He pointed to OpenAI’s $300 billion deal with Oracle and Nvidia’s relationship with CoreWeave as examples. “CoreWeave, they literally… They funded CoreWeave, became their biggest customer, then CoreWeave took that contract and those GPUs and used them as collateral to raise debt to buy more GPUs,” Zitron explained.

When will the bubble pop?

Zitron predicted the bubble would burst within the next year and a half, though he acknowledged it could happen sooner. He expects a cascade of events rather than a single dramatic collapse: An AI startup will run out of money, triggering panic among other startups and their venture capital backers, creating a fire-sale environment that makes future fundraising impossible.

“It’s not gonna be one Bear Stearns moment,” Zitron explained. “It’s gonna be a succession of events until the markets freak out.”

The crux of the problem, according to Zitron, is Nvidia. The chip maker’s stock represents 7 to 8 percent of the S&P 500’s value, and the broader market has become dependent on Nvidia’s continued hyper growth. When Nvidia posted “only” 55 percent year-over-year growth in January, the market wobbled.

“Nvidia’s growth is why the bubble is inflated,” Zitron said. “If their growth goes down, the bubble will burst.”

He also warned of broader consequences: “I think there’s a depression coming. I think once the markets work out that tech doesn’t grow forever, they’re gonna flush the toilet aggressively on Silicon Valley.” This connects to his larger thesis: that the tech industry has run out of genuine hyper-growth opportunities and is trying to manufacture one with AI.

“Is there anything that would falsify your premise of this bubble and crash happening?” I asked. “What if you’re wrong?”

“I’ve been answering ‘What if you’re wrong?’ for a year-and-a-half to two years, so I’m not bothered by that question, so the thing that would have to prove me right would’ve already needed to happen,” he said. Amid a longer exposition about Sam Altman, Zitron said, “The thing that would’ve had to happen with inference would’ve had to be… it would have to be hundredths of a cent per million tokens, they would have to be printing money, and then, it would have to be way more useful. It would have to have efficacy that it does not have, the hallucination problems… would have to be fixable, and on top of this, someone would have to fix agents.”

A positivity challenge

Near the end of our conversation, I wondered if I could flip the script, so to speak, and see if he could say something positive or optimistic, although I chose the most challenging subject possible for him. “What’s the best thing about Sam Altman,” I asked. “Can you say anything nice about him at all?”

“I understand why you’re asking this,” Zitron started, “but I wanna be clear: Sam Altman is going to be the reason the markets take a crap. Sam Altman has lied to everyone. Sam Altman has been lying forever.” He continued, “Like the Pied Piper, he’s led the markets into an abyss, and yes, people should have known better, but I hope at the end of this, Sam Altman is seen for what he is, which is a con artist and a very successful one.”

Then he added, “You know what? I’ll say something nice about him, he’s really good at making people say, ‘Yes.’”

Photo of Benj Edwards

Benj Edwards is Ars Technica’s Senior AI Reporter and founder of the site’s dedicated AI beat in 2022. He’s also a tech historian with almost two decades of experience. In his free time, he writes and records music, collects vintage computers, and enjoys nature. He lives in Raleigh, NC.

Ars Live recap: Is the AI bubble about to pop? Ed Zitron weighs in. Read More »

anthropic’s-claude-haiku-4.5-matches-may’s-frontier-model-at-fraction-of-cost

Anthropic’s Claude Haiku 4.5 matches May’s frontier model at fraction of cost

AI, AI alignment, AI benchmarks, AI coding, AI development tools, AI models, AI programming, AI safety, Amazon Bedrock, Anthropic, API, Biz & IT, Claude Haiku, Claude Sonnet, code agents, GitHub, google cloud, large language models, machine learning, Vertex AI / /

And speaking of cost, Haiku 4.5 is included for subscribers of the Claude web and app plans. Through the API (for developers), the small model is priced at $1 per million input tokens and $5 per million output tokens. That compares to Sonnet 4.5 at $3 per million input and $15 per million output tokens, and Opus 4.1 at $15 per million input and $75 per million output tokens.

The model serves as a cheaper drop-in replacement for two older models, Haiku 3.5 and Sonnet 4. “Users who rely on AI for real-time, low-latency tasks like chat assistants, customer service agents, or pair programming will appreciate Haiku 4.5’s combination of high intelligence and remarkable speed,” Anthropic writes.

Claude 4.5 Haiku answers the classic Ars Technica AI question,

Claude 4.5 Haiku answers the classic Ars Technica AI question, “Would the color be called ‘magenta’ if the town of Magenta didn’t exist?”

On SWE-bench Verified, a test that measures performance on coding tasks, Haiku 4.5 scored 73.3 percent compared to Sonnet 4’s similar performance level (72.7 percent). The model also reportedly surpasses Sonnet 4 at certain tasks like using computers, according to Anthropic’s benchmarks. Claude Sonnet 4.5, released in late September, remains Anthropic’s frontier model and what the company calls “the best coding model available.”

Haiku 4.5 also surprisingly edges up close to what OpenAI’s GPT-5 can achieve in this particular set of benchmarks (as seen in the chart above), although since the results are self-reported and potentially cherry-picked to match a model’s strengths, one should always take them with a grain of salt.

Still, making a small, capable coding model may have unexpected advantages for agentic coding setups like Claude Code. Anthropic designed Haiku 4.5 to work alongside Sonnet 4.5 in multi-model workflows. In such a configuration, Anthropic says, Sonnet 4.5 could break down complex problems into multi-step plans, then coordinate multiple Haiku 4.5 instances to complete subtasks in parallel, like spinning off workers to get things done faster.

For more details on the new model, Anthropic released a system card and documentation for developers.

Anthropic’s Claude Haiku 4.5 matches May’s frontier model at fraction of cost Read More »

chatgpt-erotica-coming-soon-with-age-verification,-ceo-says

ChatGPT erotica coming soon with age verification, CEO says

age verification, AI, AI behavior, AI censorship, AI companionship, AI ethics, AI paternalism, Biz & IT, chatbots, elon musk, machine learning, Mental Health, openai, sam altman, xAI / /

On Tuesday, OpenAI CEO Sam Altman announced that the company will allow verified adult users to have erotic conversations with ChatGPT starting in December. The change represents a shift in how OpenAI approaches content restrictions, which the company had loosened in February but then dramatically tightened after an August lawsuit from parents of a teen who died by suicide after allegedly receiving encouragement from ChatGPT.

“In December, as we roll out age-gating more fully and as part of our ‘treat adult users like adults’ principle, we will allow even more, like erotica for verified adults,” Altman wrote in his post on X (formerly Twitter). The announcement follows OpenAI’s recent hint that it would allow developers to create “mature” ChatGPT applications once the company implements appropriate age verification and controls.

Altman explained that OpenAI had made ChatGPT “pretty restrictive to make sure we were being careful with mental health issues” but acknowledged this approach made the chatbot “less useful/enjoyable to many users who had no mental health problems.” The CEO said the company now has new tools to better detect when users are experiencing mental distress, allowing OpenAI to relax restrictions in most cases.

Striking the right balance between freedom for adults and safety for users has been a difficult balancing act for OpenAI, which has vacillated between permissive and restrictive chat content controls over the past year.

In February, the company updated its Model Spec to allow erotica in “appropriate contexts.” But a March update made GPT-4o so agreeable that users complained about its “relentlessly positive tone.” By August, Ars reported on cases where ChatGPT’s sycophantic behavior had validated users’ false beliefs to the point of causing mental health crises, and news of the aforementioned suicide lawsuit hit not long after.

Aside from adjusting the behavioral outputs for its previous GPT-40 AI language model, new model changes have also created some turmoil among users. Since the launch of GPT-5 in early August, some users have been complaining that the new model feels less engaging than its predecessor, prompting OpenAI to bring back the older model as an option. Altman said the upcoming release will allow users to choose whether they want ChatGPT to “respond in a very human-like way, or use a ton of emoji, or act like a friend.”

ChatGPT erotica coming soon with age verification, CEO says Read More »

nvidia-sells-tiny-new-computer-that-puts-big-ai-on-your-desktop

Nvidia sells tiny new computer that puts big AI on your desktop

Acer, AI, AI chips, AI development tools, AI GPU, AI infrastructure, ai supercomputer, ASUS, Biz & IT, desktop computing, DGX Spark, elon musk, Google, Grace Blackwell, HP, Jensen Huang, machine learning, msi, NVIDIA, openai, Project DIGITS, spacex, Tech / /

On Tuesday, Nvidia announced it will begin taking orders for the DGX Spark, a $4,000 desktop AI computer that wraps one petaflop of computing performance and 128GB of unified memory into a form factor small enough to sit on a desk. Its biggest selling point is likely its large integrated memory that can run larger AI models than consumer GPUs.

Nvidia will begin taking orders for the DGX Spark on Wednesday, October 15, through its website, with systems also available from manufacturing partners and select US retail stores.

The DGX Spark, which Nvidia previewed as “Project DIGITS” in January and formally named in May, represents Nvidia’s attempt to create a new category of desktop computer workstation specifically for AI development.

With the Spark, Nvidia seeks to address a problem facing some AI developers: Many AI tasks exceed the memory and software capabilities of standard PCs and workstations (more on that below), forcing them to shift their work to cloud services or data centers. However, the actual market for a desktop AI workstation remains uncertain, particularly given the upfront cost versus cloud alternatives, which allow developers to pay as they go.

Nvidia’s Spark reportedly includes enough memory to run larger-than-typical AI models for local tasks, with up to 200 billion parameters and fine-tune models containing up to 70 billion parameters without requiring remote infrastructure. Potential uses include running larger open-weights language models and media synthesis models such as AI image generators.

According to Nvidia, users can customize Black Forest Labs’ Flux.1 models for image generation, build vision search and summarization agents using Nvidia’s Cosmos Reason vision language model, or create chatbots using the Qwen3 model optimized for the DGX Spark platform.

Big memory in a tiny box

Nvidia has squeezed a lot into a 2.65-pound box that measures 5.91 x 5.91 x 1.99 inches and uses 240 watts of power. The system runs on Nvidia’s GB10 Grace Blackwell Superchip, includes ConnectX-7 200Gb/s networking, and uses NVLink-C2C technology that provides five times the bandwidth of PCIe Gen 5. It also includes the aforementioned 128GB of unified memory that is shared between system and GPU tasks.

Nvidia sells tiny new computer that puts big AI on your desktop Read More »