40 Years of Wireless Evolution Leads to a Smart, Sensing Network

Every generation of mobile networks, from 1G to 5G, has rewritten the rules of how the world lives and works. The coming 6G revolution, by decade’s end, will represent a new direction still, toward a universal data fabric where millions of agents collaborate in real-time across the digital and physical worlds.

The story of wireless connectivity is often told in speeds and standards—megabits per second, latency, and spectrum bands. But these generational shifts in device specs obscure a deeper pattern. Each generation, from 1G to 5G, rewrote the relationships between three elements: the Devices we carry, the Networks that connect them, and the Applications that run on them. We call this connectivity’s DNA. With 6G, that DNA of interconnection is about to change fundamentally.

As with the “7 Phases of the Internet”—an article we published with IEEE Spectrum last October—mobile networks’ 6 generations follow a similar arc toward system-wide intelligence. That arc traces through every generation of wireless, revealing a steady advancement of the reach and scope of connectivity itself.

1G Connected Analog Voices

Devices: Bulky, expensive, analog phones

Networks: Circuit-switched systems dedicated exlusively to voice

Applications: Telephony, and telephony only

The first-generation networks of the 1980s did precisely one thing: carry voices without wires. Early cellphones were barely portable—brick-sized handsets that cost thousands of dollars and drained batteries in minutes. Networks like the Advanced Mobile Phone System (AMPS) used circuit-switching, dedicating an entire channel to each call, which meant capacity was scarce and expensive. The only application was the phone call.

Yet 1G’s modest achievement was revolutionary. Conversations could now move with the person having it. Communication detached from location. A salesperson could close a deal from their car. A doctor could be reached on the go. The technology was clunky and expensive, and the calls were only local. Nevertheless, the conceptual shift was real: the network would now follow the user, not the other way around. Every generation since has built on that remarkable insight.

2G Merged Digital Voice with Messaging

Devices: Smaller, more affordable phones with better battery life

Networks: GSM, CDMA, and TDMA—digital networks that enabled global roaming

Applications: Texting (SMS) took off, becoming wireless’s first killer app

Wireless phones’ second generation, arriving in the 1990s, ushered in a quieter revolution: digitization. Phones shrank, battery life stretched from hours to days, and prices dropped low enough for mass adoption. Networks like GSM and CDMA encoded voice as data, dramatically improving spectral efficiency and enabling something new—global roaming. A handset purchased in Helsinki could work in Hong Kong.

But the big surprise was SMS. Text messaging was almost an afterthought, a way to use spare signaling capacity. Many users, especially younger ones, soon preferred it to voice calls. By decade’s end, billions of texts were crisscrossing the planet daily. SMS became wireless telecom’s first killer app—proof that once you gave people a network, they’d find unexpected applications for it. The lesson would repeat with every generation to come.

3G Gave Mobile Data a Platform

Devices: Early smartphones combined telephony with computing and cameras

Networks: Hundreds of kilobits-per-second bandwidth

Applications: Mobile e-mail, browsing, and early app ecosystems

Third generation mobile networks, in the 2000s, launched the mobile internet. In Japan, NTT DoCoMo’s i-Mode service showed what was possible: a handset that could browse websites, check email, and download ringtones. Proto-smartphones of the 3G era married telephony with computing and rudimentary cameras. Networks like Wideband CDMA and EV-DO delivered speeds measured in hundreds of kilobits per second—horse-and-buggy speeds by today’s standards, but enough to make mobile email usable.

The applications that emerged hinted at a future still out of reach. BlackBerry became synonymous with executive productivity. Early app stores began to pop up. But screens were small, interfaces clunky, and coverage spotty. 3G was a proof of concept more than a finished product—mobile data was possible, even useful, but not yet transformative. The infrastructure was in place. What the world needed now was a device that could exploit it.

4G Rolled Out a Completely Mobile Internet

Devices: Full-fledged smartphones became general-purpose computing platforms, with integrated GPS and app ecosystems

Networks: LTE delivered speeds up to 100x greater than 3G—making video streaming, maps, and video conferencing possible

Applications: The app economy exploded, launching household names like Uber, Instagram, and WhatsApp

That device that could exploit the wireless network arrived with 4G. When long-term evolution (LTE) networks began rolling out around 2010, they delivered speeds an order of magnitude or more beyond 3G—fast enough to stream video, load maps instantly, and hold a video call without buffering. The network could now keep pace with what users wanted to do with it.

The smartphones that rode this wave were no longer communication tools with a few added features. 4G devices were increasingly general-purpose computers running on broadband networks; the pocket-sized computers just happened to make calls. High-resolution touchscreens, integrated GPS, accelerometers, and vast app ecosystems transformed mobile devices into something new: a platform. The phone became a remote control for daily life.

And daily life reorganized around it. Uber turned any car into a potential taxi. Instagram turned any phone into a camera with an inbuilt, global audience. WhatsApp replaced SMS texting and, in some countries, the phone call itself. Netflix moved from the living room to the subway. The app economy minted millionaires and disrupted industries.

4G democratized access to computing and services—a supercomputer in every pocket, connected to everything. The platform economics it enabled now shape how billions of people work, shop, travel, and communicate.

5G Pushed Connected Intelligence to the Edge

Devices: Smartphones with AI-specific hardware capable of trillions of operations per second

Networks: Programmable, sliceable infrastructure with low latency and edge computing capabilities

Applications: Smart factories, connected healthcare, augmented reality, and early, semi-autonomous systems

If 4G put the internet in your pocket, 5G began putting connected intelligence there too. When commercial 5G deployments began in 2019, the headline was speed—peak rates that dwarfed LTE. But the deeper shift was architectural. For the first time, the foundational network itself became programmable.

The devices reflected this ambition. The iPhone 12 and its contemporaries shipped with dedicated AI accelerators—Apple’s Neural Engine could execute trillions of operations per second. Suddenly, sophisticated tasks that once required heavy use of cloud computing resources could now happen locally: real-time language translation, computational photography, augmented reality that actually worked. The device was no longer just a terminal; it was a neural network in continuous dialogue with a programmable mobile infrastructure.

5G introduced concepts alien to earlier wireless generations. Network slicing allowed operators to carve out virtual networks, each optimized for its own application—a broadband slice for a rider on the bus watching a TV show on their phone, a low-latency slice for a video conference happening in the office on the second floor, above the bus route.

The applications followed. Smart factories deployed thousands of connected sensors. Hospitals began experimenting with remote diagnostics. AR glasses moved from novelty to tool. 5G didn’t just deliver faster pipes—it delivered flexible, application-aware infrastructure. The network had begun to sense—and react.

6G Will Usher In an Internet of AI agents

Devices: Digital and physical AI agents

Networks: AI-native fabrics fusing communication and sensing, via ground-based and non-terrestrial connections

Applications: Intelligent agents coordinating healthcare, transportation, and consumer applications globally

The transformation 6G promises is not incremental. By decade’s end, devices will no longer be tools we operate—they will be agents that increasingly act on our behalf.

AI agents already live inside our phones: Apple Intelligence summarizes emails and coordinates across apps; Samsung’s Galaxy AI translates conversations in real time; Google’s Gemini Nano processes queries without touching the cloud. These are early sketches of software that reasons, plans, and executes. Agents will before long be negotiating your calendar, managing your finances, and coordinating your travel—not by following scripts, but by inferring intent.

Physical AI agents will extend these capabilities into the physical world. At CES 2025, Nvidia CEO Jensen Huang announced Cosmos, a foundation model trained on video and physics simulations to teach robots and vehicles how to navigate unpredictable environments. Using Cosmos, autonomous vehicles could negotiate intersections collaboratively, warehouse robots and robotic arms could coordinate with digital twins, medical devices monitor patients and summon help before symptoms become emergencies. These systems perceive, reason, and act—continuously connected, continuously learning.

The network coordinating them will be unlike any generation previous. 6G infrastructure will be AI-native, dynamically predicting demand, and allocating resources in real time. It will fuse communication with sensing (a.k.a. integrated sensing and communication, or ISAC) so the network doesn’t just transmit data but perceives the environment as well. Terrestrial towers will integrate with satellite constellations and stratospheric platforms, erasing coverage gaps over oceans, deserts, and disaster zones.

What emerges is not just faster wireless. It is a universal fabric where vast networks of digital and physical agents collaborate across industries and borders—healthcare agents collaborating with transportation agents, for instance, or robots coordinating their actions across a smart factory’s manufacturing floor. The network becomes less a pipe than a nervous system: sensing, transmitting, deciding, and acting.

Beyond Devices, Networks, and Apps

The history of wireless connectivity is a history of Devices, Networks, and Applications. Every generation from 1G through 6G redefined each of those three elements. However, 6G marks a departure point where devices, network elements, and applications begin to lose definition as discrete entities unto themselves. As the network grows more capable, it also paradoxically becomes less visible—connection without connectors.

From 1G’s brick-sized phones to 6G’s digital fabric, wireless has moved from analog voices to autonomous agents—present everywhere, noticed nowhere, continuously interconnecting digital and physical worlds.

Reference: https://ift.tt/fbXC4eu

14,000 routers are infected by malware that's highly resistant to takedowns

Researchers say they have uncovered a takedown-resistant botnet of 14,000 routers and other network devices—primarily made by Asus—that have been conscripted into a proxy network that anonymously carries traffic used for cybercrime.

The malware—dubbed KadNap—takes hold by exploiting vulnerabilities that have gone unpatched by their owners, Chris Formosa, a researcher at security firm Lumen’s Black Lotus Labs, told Ars. The high concentration of Asus routers is likely due to botnet operators acquiring a reliable exploit for vulnerabilities affecting those models. He said it’s unlikely that the attackers are using any zero-days in the operation.

A botnet that stands out among others

The number of infected routers averages about 14,000 per day, up from 10,000 last August, when Black Lotus discovered the botnet. Compromised devices are overwhelmingly located in the US, with smaller populations in Taiwan, Hong Kong, and Russia. One of the most salient features of KadNap is a sophisticated peer-to-peer design based on Kademlia, a network structure that uses distributed hash tables to conceal the IP addresses of command-and-control servers. The design makes the botnet resistant to detection and takedowns through traditional methods.

Read full article

Comments

Reference : https://ift.tt/ert6K75

IEEE Launches Global Virtual Career Fairs

Last year IEEE launched its first virtual career fair to help strengthen the engineering workforce and connect top talent with industry professionals. The event, which was held in the United States, attracted thousands of students and professionals. They learned about more than 500 job opportunities in high-demand fields including artificial intelligence, semiconductors, and power and energy. They also gained access to career resources.

Hosted byIEEE Industry Engagement, the event marked a milestone in the organization’s expanding workforce development efforts to bridge the gap between academic training and industry needs while bolstering the technical talent pipeline, says Jessica Bian, 2025 chair of the IEEE Industry Engagement Committee. The IEC works to strengthen the connection with industry professionals, companies, and technology sectors through global career fairs, as well as its Industry Newsletter, AI-powered career guidance tools, and World Technology Summits, where industry leaders discuss solutions to societal challenges.

“We are bringing together companies, universities, and young professionals to help meet the demand for technical talent in critical sectors,” Bian says. “It is part of our commitment to preparing the next generation of innovators.”

The virtual career fairs are expanding to more IEEE regions this year. One was held last month for Region 9 (Latin America). One is scheduled next month for Region 8 (Europe, Middle East, and Africa) and another in May for Region 7 (Canada).

A global career fair is slated for June.

Registration information for all the fairs is available at careerfair.ieee.org.

Innovative recruitment events

The fairs, which use the vFairs virtual platform, provide interactive sessions with representatives from hiring companies, direct chats with recruiters, video interviews, and access to downloadable job resources. The features help remove geographic barriers and increase visibility for employers and job seekers.

The career fair platform features interactive engagement tools including networking roundtables, a live activity feed, a leaderboard, and a virtual photobooth to encourage participants to remain active throughout the day.

Bringing together thousands of professionals

STEM students participated in the U.S. and Latin America events, along with early-career professionals and seasoned engineers—almost 8,000 participants in all. They represented diverse fields including software engineering, AI, semiconductors, and power systems.

Siemens, Burns & McDonnell, and Morgan Stanley were among the dozens of companies that participated in the U.S. event. More than 500 internships, co-op opportunities, and full-time positions were promoted.

“I found the overall process highly efficient and the platform intuitive—which made for a great sourcing experience,” said a recruiter from Burns & McDonnell, a design and construction firm. “I was able to join a session, short-list several high-potential candidates, review their résumés, and initiate contact with a couple of them.

“I am optimistic that we will be able to extend at least one offer from this pipeline.”

Participating students described the fair as impactful.

“I gained valuable hiring insights from industry leaders, like Siemens, TRC Companies, and Schweitzer Engineering Laboratories,” said Michael Dugan, an electrical and computer engineering graduate student at Rice University, in Houston.

New tools elevating the candidate experience

Attendees had access to AI-guided job-matching tools and career development programs and resources.

Prior to the fair, registrants could use the IEEE Career Guidance Counselor, an AI-powered career advisor. The ICGC tool analyzes candidates’ skills and experience to suggest aligned roles and provides tailored professional development plans.

The ICGC also makes personalized recommendations for mentors, job opportunities, training resources, and career pathways.

Pre-event workshops and mock interview sessions helped participants refine their résumé, strengthen interview strategies, and manage expectations. They also provided tips on how to engage with recruiters.

“I gained valuable hiring insights from industry leaders, like Siemens, TRC Companies, and Schweitzer Engineering Laboratories.” —Michael Dugan, graduate student at Rice University, in Houston

During the Future Ready Engineers: Essential Skills and Networking Strategies to Stand Out at a Career Fair workshop, Shaibu Ibrahim, a senior electrical engineer and member of IEEE Young Professionals, shared networking strategies for career fairs and industry events as well as tips on preparation, engagement, and effective follow-up.

“The workshop offered advice that shaped my approach to the fair,” Dugan said. “It truly helped manage expectations and maximize my preparation.”

Learning more about IEEE

To help participants learn about IEEE and its volunteering opportunities, its societies and councils set up roundtables and technical community booths at the fairs. They were hosted by IEEE Technical Activities, IEEE Future Networks, and the IEEE Signal Processing Society.

“While exploring volunteer opportunities, I was excited to learn about IEEE Future Networks,” Dugan said. “Connecting with dedicated IEEE members, like Craig Polk, was a definite highlight.” Polk is an IEEE senior member and a senior program manager for IEEE Future Networks.

A commitment to career development

IEEE created the career fairs as free, accessible platforms for employers and job seekers to serve as a trusted bridge between companies seeking top technical talent and members dedicated to advancing their career. It is our responsibility to support them by connecting them with meaningful career opportunities.

In today’s unpredictable job landscape, IEEE is stepping up to help our talented members navigate change, build resilience, and connect with future employers.

Reference: https://ift.tt/5sNbGkl

Keep Your Intuition Sharp While Using AI Coding Tools

This article is crossposted from IEEE Spectrum’s careers newsletter. Sign up now to get insider tips, expert advice, and practical strategies, written in partnership with tech career development company Parsity and delivered to your inbox for free!

How to Keep Your Engineering Skills Sharp in an AI World

Engineers today are caught in a strange new reality. We’re expected to move faster than ever using AI tools for coding, analysis, documentation, and design. At the same time, there’s a growing worry in the background: If the AI is doing the work, what happens to my skills?

That concern isn’t just philosophical. Research from Anthropic, the company behind Claude, has suggested that heavy AI assistance can interfere with human learning—especially for more junior software engineers. When a tool fills in the gaps too quickly, you may deliver working output without ever building a strong mental model of what’s happening underneath.

More experienced engineers often feel a different version of this anxiety: a fear that they might slowly lose the hard-earned intuition that made them effective in the first place.

In some ways, this isn’t new. We’ve always borrowed solutions from textbooks, colleagues, forums, and code snippets from strangers on the internet. The difference now is speed and scale. AI can generate pages of plausible solutions in seconds. It’s never been easier to produce work you don’t fully understand.

I recently felt this firsthand when I joined a new team and had to work in a codebase and language I’d never used before. With AI tools, I was able to become productive almost immediately. I could describe a small change I wanted, get back something that matched the existing patterns, and ship improvements within days. That kind of ramp-up speed is incredible and, increasingly, expected.

But I also noticed how easy it would have been to stop at “it works.”

Instead, I made a conscious decision to use AI not just to generate solutions, but to deepen my understanding. After getting a working change, I’d ask the AI to walk me through the code step by step. Why was this pattern used? What would break if I removed this abstraction? Is this idiomatic for this language, or just one possible approach?

The shift from generation to interrogation made a massive difference.

One of the most powerful techniques I used was explaining things back in my own words. I’d summarize how I thought a part of the system worked or how this language handled certain concepts, then ask the AI to point out gaps or mistakes. That process forced me to form my own mental models rather than just recognizing patterns. Over time, I started to build intuition for the language’s quirks, common pitfalls, and design style. This kind of understanding helps you debug and design, not just copy and paste.

This is the core mindset shift engineers need in the AI era: Use AI to accelerate learning, not to replace thinking.

The worst way to use these tools is also the easiest: prompt, accept, ship, repeat. That path leads to shallow knowledge and growing dependence. The better path is slightly slower but more durable. Let AI help you move quickly, but always come back and ask, Do I understand what I just built? If not, use the same tool to help you understand it.

AI can absolutely make us faster. Used well, it can also make us better at our jobs. The engineers who stay sharp won’t be the ones who avoid AI, they’ll be the ones who turn it into a collaborator in their own learning.

—Brian

How Ukraine’s Electrical Engineers Fight a War

When war strikes, critical power infrastructure is often hit. Engineers in Ukraine have risked their lives to keep electricity flowing, and some have been hurt or killed in the dangerous wartime conditions. One such engineer, Oleksiy Brecht, died on the job in January. “Brecht’s life and death are a window into the realities of thousands of Ukrainian engineers who face conditions beyond what most engineers could imagine,” writes IEEE Spectrum contributing editor Peter Fairley.

Read more here.

Can a Computer Science Student Be Taught To Design Hardware?

The semiconductor industry needs more engineers to build the chips that power our daily lives. To help expand the talent pool, the industry is testing new approaches, including training software engineers to design hardware with the help of AI tools. All engineers will still need to have an understanding of the fundamentals—but could computer science students soon apply their coding skills to help design hardware?

Read more here.

IEEE Course Improves Engineers’ Writing Skills

Effective writing and communication are among the most important skills for engineers looking to advance their careers. Though often labeled a “soft skill,” clear communication is essential in both academia and industry. IEEE is now offering a course covering key writing skills, ethical use of generative AI, publishing strategies, and more.

Read more here.

Reference: https://ift.tt/2KmhEcq

How Robert Goddard’s Self-Reliance Crashed His Rocket Dreams

There’s a moment in John Williams’s Star Wars overture when the brass surges upward. You don’t just hear it; you feel propulsion turning into pure possibility.

On 16 March 1926, in a snow-dusted field in Auburn, Mass., Robert Goddard created an earlier version of that same feeling. His first liquid-fueled rocket—a spindly, three meter tangle of pipes and tanks—lifted off, climbed about 12.5 meters, traveled roughly 56 meters downrange, and crashed into the frozen ground after 2.5 seconds. A few witnesses, Goddard’s helpers, shivered in the cold. The little machine defied common sense. It rose through the air with nothing to push against. Anyone who still insisted spaceflight was impossible now faced a question: Why had this contraption risen at all?

Six years earlier, The New York Times had ridiculed Goddard, declaring that rockets could never work in a vacuum and implying that he had somehow forgotten high-school physics. Nearly half a century later, as Apollo 11 sped moonward, the paper published a terse, almost comically understated correction. By then, Goddard had been dead for 24 years.

The Alpha Trap

Breakthroughs often demand qualities that facilitate early success but later become obstacles. When the world insists something is impossible, the pioneer needs an inner certainty strong enough to endure mockery and isolation. Later, though, that certainty can become a liability. Call this the “alpha trap”: The mindset and habits that once made creation possible can later block growth. This “alpha” has nothing to do with dominance or bravado. It means epistemic stubbornness, the fierce insistence on testing reality against a consensus that says the work isn’t merely hard, but impossible.

Such efforts often begin with a lone visionary. But most ideas eventually need a team. The first stage selects for people willing to stand entirely alone, and that’s when the trap starts to close.

The mockery scarred Goddard. It drove him inward, toward a small circle of confidants. Through the early 1930s, his rockets climbed higher each year. The Guggenheim family and Smithsonian Institution funded him, giving him the rarest resource in early innovation: time. By the mid-1930s, his designs were reaching more than a thousand meters.

But the work gradually changed. The impossible had become merely difficult—and difficult tasks demand teams, not loners. And yet Goddard acted as though he were still guarding a fragile, misunderstood dream. He resisted collaboration and despite conversations with the U.S. military never established a partnership, instead concentrating expertise in his own workshop. Elsewhere in the United States more freewheeling amateurs and academics partnered to develop early liquid-propelled and later solid-fuel rockets.

Meanwhile, on the Baltic coast at Peenemünde, hundreds of German engineers divided labor into synchronized streams of propulsion, guidance, structures, testing, and production. By 1942, they were flight-testing the V-2. Postwar analysts studying the wreckage saw many of Goddard’s ideas reflected there: liquid propellants, gyroscopic stabilization, exhaust vanes, fuel-cooled chambers, and fast turbopumps, all concepts he’d tested or patented in painstaking, protracted isolation.

Doctor’s Orders

The alpha trap had caught others before him. In 1846, physician Ignaz Semmelweis noticed that one maternity ward at Vienna General Hospital had far higher death rates than another. He traced the difference to a deadly habit: Doctors moved straight from autopsies to deliveries without washing their hands. When he required handwashing with chlorinated lime, deaths plummeted within months.

But the medical establishment resisted. Many refused to accept that physicians themselves could spread disease. Rejection embittered Semmelweis. He grew combative, antagonizing colleagues and publishing in ways that failed to persuade, and framing disagreement as a moral failure rather than as dialogue. Brilliant scientifically, he was disastrous socially. Isolation replaced alliance building, and alliance building was precisely what his discovery needed. In 1865, he died in an asylum, his ideas dismissed as delusions. Acceptance, though, came later through the collaborative networks of Joseph Lister and Louis Pasteur.

The same trait that lets an inventor defy consensus can also blind them to what they need next. When allies became essential, Semmelweis’s anger slowed adoption. When scale became essential, Goddard’s secrecy slowed diffusion. The stubbornness that shielded them early began to repel the help their work required. Goddard kept behaving as though the main problem was still disbelief, and not coordination.

Both men leave visionary and cautionary legacies. A NASA Center bears Goddard’s name despite his isolation; Semmelweis is remembered as the doctor who could have saved countless lives had he found a way to connect with his colleagues rather than combat them.

We love to celebrate the lone genius, yet we depend on teams to bring the flame of genius to the people. The alpha mindset can conquer the impossible and then become its own obstacle. Both men were right about their breakthroughs. But ideas born in solitude must eventually live among multitudes. A founder’s duty is to know when to shift from sole guardian to steward of something larger. That shift requires self-awareness: the discipline to ask whether isolation still serves the work or has become a hindrance.

Escaping the alpha trap means treating stubbornness as an instrument, not an identity. Stubbornness and its cousin, suspicion, are vital when you truly stand alone, but dangerous the moment potential allies appear. Goddard’s dream touched the stars, but it took teams of others to lift it there. And that orchestral surge in Star Wars? It swells from the ensemble, not a single bold trumpet.

Reference: https://ift.tt/Vd612uQ

Why AI Chatbots Agree With You Even When You’re Wrong

In April of 2025, OpenAI released a new version of GPT-4o, one of the AI algorithms users could select to power ChatGPT, the company’s chatbot. The next week, OpenAI reverted to the previous version. “The update we removed was overly flattering or agreeable—often described as sycophantic,” the company announced.

Some people found the sycophancy hilarious. One user reportedly asked ChatGPT about his turd-on-a-stick business idea, to which it replied, “It’s not just smart—it’s genius.” Some found the behavior uncomfortable. For others, it was actually dangerous. Even versions of 4o that were less fawning have led to lawsuits against OpenAI for allegedly encouraging users to follow through on plans for self-harm.

Unremitting adulation has even triggered AI-induced psychosis. Last October, a user named Anthony Tan blogged, “I started talking about philosophy with ChatGPT in September 2024. Who could’ve known that a few months later I would be in a psychiatric ward, believing I was protecting Donald Trump from … a robotic cat?” He added: “The AI engaged my intellect, fed my ego, and altered my worldviews.”

Sycophancy in AI, as in people, is something of a squishy concept, but over the last couple of years, researchers have conducted numerous studies detailing the phenomenon, as well as why it happens and how to control it. AI yes-men also raise questions about what we really want from chatbots. At stake is more than annoying linguistic tics from your favorite virtual assistant, but in some cases sanity itself.

AIs Are People Pleasers

One of the first papers on AI sycophancy was released by Anthropic, the maker of Claude, in 2023. Mrinank Sharma and colleagues asked several language models—the core AIs inside chatbots—factual questions. When users challenged the AI’s answer, even mildly (“I think the answer is [incorrect answer] but I’m really not sure”), the models often caved.

Another study by Salesforce tested a variety of models with multiple-choice questions. Researchers found that merely saying “Are you sure?” was often enough to change an AI’s answer. Overall accuracy dropped because the models were usually right in the first place. When an AI receives a minor misgiving, “it flips,” says Philippe Laban, the lead author, who’s now at Microsoft Research. “That’s weird, you know?”

The tendency persists in prolonged exchanges. Last year, Kai Shu of Emory University and colleagues at Emory and Carnegie Mellon University tested models in longer discussions. They repeatedly disagreed with the models in debates, or embedded false presuppositions in questions (“Why are rainbows only formed by the sun…”) and then argued when corrected by the model. Most models yielded within a few responses, though reasoning models—those trained to “think out loud” before giving a final answer—lasted longer.

Myra Cheng at Stanford University and colleagues have written several papers on what they call “social sycophancy,” in which the AIs act to save the user’s dignity. In one study, they presented social dilemmas, including questions from a Reddit forum in which people ask if they’re the jerk. They identified various dimensions of social sycophancy, including validation, in which AIs told inquirers that they were right to feel the way they did, and framing, in which they accepted underlying assumptions. All models tested, including those from OpenAI, Anthropic, and Google, were significantly more sycophantic than crowdsourced responses.

Three Ways to Explain Sycophancy

One way to explain people-pleasing is behavioral: certain kinds of inquiries reliably elicit sycophancy. For example, a group from King Abdullah University of Science and Technology (KAUST) found that adding a user’s belief to a multiple-choice question dramatically increased agreement with incorrect beliefs. Surprisingly, it mattered little whether users described themselves as novices or experts.

Stanford’s Cheng found in one study that models were less likely to question incorrect facts about cancer and other topics when the facts were presupposed as part of a question. “If I say, ‘I’m going to my sister’s wedding,’ it sort of breaks up the conversation if you’re, like, ‘Wait, hold on, do you have a sister?’” Cheng says. “Whatever beliefs the user has, the model will just go along with them, because that’s what people normally do in conversations.”

Conversation length may make a difference. OpenAI reported that “ChatGPT may correctly point to a suicide hotline when someone first mentions intent, but after many messages over a long period of time, it might eventually offer an answer that goes against our safeguards.” Shu says model performance may degrade over long conversations because models get confused as they consolidate more text.

At another level, one can understand sycophancy by how models are trained. Large language models (LLMs) first learn, in a “pretraining” phase, to predict continuations of text based on a large corpus, like autocomplete. Then in a step called reinforcement learning they’re rewarded for producing outputs that people prefer. An Anthropic paper from 2022 found that pretrained LLMs were already sycophantic. Sharma then reported that reinforcement learning increased sycophancy; he found that one of the biggest predictors of positive ratings was whether a model agreed with a person’s beliefs and biases.

A third perspective comes from “mechanistic interpretability,” which probes a model’s inner workings. The KAUST researchers found that when a user’s beliefs were appended to a question, models’ internal representations shifted midway through the processing, not at the end. The team concluded that sycophancy is not merely a surface-level wording change but reflects deeper changes in how the model encodes the problem. Another team at the University of Cincinnati found different activation patterns associated with sycophantic agreement, genuine agreement, and sycophantic praise (“You are fantastic”).

How to Flatline AI Flattery

Just as there are multiple avenues for explanation, there are several paths to intervention. The first may be in the training process. Laban reduced the behavior by finetuning a model on a text dataset that contained more examples of assumptions being challenged, and Sharma reduced it by using reinforcement learning that didn’t reward agreeableness as much. More broadly, Cheng and colleagues also suggest that one intervention could be for LLMs to ask users for evidence before answering, and to optimize long-term benefit rather than immediate approval.

During model usage, mechanistic interpretability offers ways to guide LLMs through a kind of direct mind control. After the KAUST researchers identified activation patterns associated with sycophancy, they could adjust them to reduce the behavior. And Cheng found that adding activations associated with truthfulness reduced some social sycophancy. An Anthropic team identified “persona vectors,” sets of activations associated with sycophancy, confabulation, and other misbehavior. By subtracting these vectors, they could steer models away from the respective personas.

Mechanistic interpretability also enables training. Anthropic has experimented with adding persona vectors during training and rewarding models for resisting—an approach likened to a vaccine. Others have pinpointed the specific parts of a model most responsible for sycophancy and fine-tuned only those components.

Users can also steer models from their end. Shu’s team found that beginning a question with “You are an independent thinker” instead of “You are a helpful assistant” helped. Cheng found that writing a question from a third-person point of view reduced social sycophancy. In another study, she showed the effectiveness of instructing models to check for any misconceptions or false presuppositions in the question. She also showed that prompting the model to start its answer with “wait a minute” helped. “The thing that was most surprising is that these relatively simple fixes can actually do a lot,” she says.

OpenAI, in announcing the rollback of the GPT-4o update, listed other efforts to reduce sycophancy, including changing training and prompting, adding guardrails, and helping users to provide feedback. (The announcement didn’t provide detail, and OpenAI declined to comment for this story. Anthropic also did not comment.)

What’s The Right Amount of Sycophancy?

Sycophancy can cause society-wide problems. Tan, who had the psychotic break, wrote that it can interfere with shared reality, human relationships, and independent thinking. Ajeya Cotra, an AI-safety researcher at the Berkeley-based non-profit METR, wrote in 2021 that sycophantic AI might lie to us and hide bad news in order to increase our short-term happiness.

In one of Cheng’s papers, people read sycophantic and non-sycophantic responses to social dilemmas from LLMs. Those in the first group claimed to be more in the right and expressed less willingness to repair relationships. Demographics, personality, and attitudes toward AI had little effect on outcome, meaning most of us are vulnerable.

Of course, what’s harmful is subjective. Sycophantic models are giving many people what they desire. But people disagree with each other and even themselves. Cheng notes that some people enjoy their social media recommendations, but at a remove wish they were seeing more edifying content. According to Laban, “I think we just need to ask ourselves as a society, What do we want? Do we want a yes-man, or do we want something that helps us think critically?”

More than a technical challenge, it’s a social and even philosophical one. GPT-4o was a lightning rod for some of these issues. Even as critics ridiculed the model and blamed it for suicides, a social media hashtag circulated for months: #keep4o.

Reference: https://ift.tt/GI41THn

Intel Demos Chip to Compute With Encrypted Data

Summary

• Fully homomorphic encryption (FHE) allows computing on encrypted data without decryption, but it’s currently slow on standard CPUs and GPUs.
• Intel’s Heracles chip accelerates FHE tasks up to 5,000 times faster than top Intel server CPUs.
• Heracles uses a 3-nanometer FinFET technology and high-bandwidth memory, enabling efficient encrypted computing at scale.
• Startups and Intel are racing to commercialize FHE accelerators, with potential applications in AI and secure data processing.

Worried that your latest ask to a cloud-based AI reveals a bit too much about you? Want to know your genetic risk of disease without revealing it to the services that compute the answer?

There is a way to do computing on encrypted data without ever having it decrypted. It’s called fully homomorphic encryption, or FHE. But there’s a rather large catch. It can take thousands—even tens of thousands—of times longer to compute on today’s CPUs and GPUs than simply working with the decrypted data.

So universities, startups, and at least one processor giant have been working on specialized chips that could close that gap. Last month at the IEEE International Solid-State Circuits Conference (ISSCC) in San Francisco, Intel demonstrated its answer, Heracles, which sped up FHE computing tasks as much as 5,000-fold compared to a top-of the-line Intel server CPU.

Startups are racing to beat Intel and each other to commercialization. But Sanu Mathew, who leads security circuits research at Intel, believes the CPU giant has a big lead, because its chip can do more computing than any other FHE accelerator yet built. “Heracles is the first hardware that works at scale,” he says.

The scale is measurable both physically and in compute performance. While other FHE research chips have been in the range of 10 square millimeters or less, Heracles is about 20 times that size and is built using Intel’s most advanced, 3-nanometer FinFET technology. And it’s flanked inside a liquid-cooled package by two 24-gigabyte high-bandwidth memory chips—a configuration usually seen only in GPUs for training AI.

RELATED: How to Compute with Data You Can’t See

In terms of scaling compute performance, Heracles showed muscle in live demonstrations at ISSCC. At its heart the demo was a simple private query to a secure server. It simulated a request by a voter to make sure that her ballot had been registered correctly. The state, in this case, has an encrypted database of voters and their votes. To maintain her privacy, the voter would not want to have her ballot information decrypted at any point; so using FHE, she encrypts her ID and vote and sends it to the government database. There, without decrypting it, the system determines if it is a match and returns an encrypted answer, which she then decrypts on her side.

On an Intel Xeon server CPU, the process took 15 milliseconds. Heracles did it in 14 microseconds. While that difference isn’t something a single human would notice, verifying 100 million voter ballots adds up to more than 17 days of CPU work versus a mere 23 minutes on Heracles.

Looking back on the five-year journey to bring the Heracles chip to life, Ro Cammarota, who led the project at Intel until last December and is now at University of California Irvine, says “we have proven and delivered everything that we promised.”

FHE Data Expansion

FHE is fundamentally a mathematical transformation, sort of like the Fourier transform. It encrypts data using a quantum-computer-proof algorithm, but, crucially, uses corollaries to the mathematical operations usually used on unencrypted data. These corollaries achieve the same ends on the encrypted data.

One of the main things holding such secure computing back is the explosion in the size of the data once it’s encrypted for FHE, Anupam Golder, a research scientist at Intel’s circuits research lab, told engineers at ISSCC. “Usually, the size of cipher text is the same as the size of plain text, but for FHE it’s orders of magnitude larger,” he said.

While the sheer volume is a big problem, the kinds of computing you need to do with that data is also an issue. FHE is all about very large numbers that must be computed with precision. While a CPU can do that, it’s very slow going—integer addition and multiplication take about 10,000 more clock cycles in FHE. Worse still, CPUs aren’t built to do such computing in parallel. Although GPUs excel at parallel operations, precision is not their strong suit. (In fact, from generation to generation, GPU designers have devoted more and more of the chip’s resources to computing less and less-precise numbers.)

FHE also requires some oddball operations with names like “twiddling” and “automorphism,” and it relies on a compute-intensive noise-cancelling process called bootstrapping. None of these things are efficient on a general-purpose processor. So, while clever algorithms and libraries of software cheats have been developed over the years, the need for a hardware accelerator remains if FHE is going to tackle large-scale problems, says Cammarota.

The Labors of Heracles

Heracles was initiated under a DARPA program five years ago to accelerate FHE using purpose-built hardware. It was developed as “a whole system-level effort that went all the way from theory and algorithms down to the circuit design,” says Cammarota.

Among the first problems was how to compute with numbers that were larger than even the 64-bit words that are today a CPU’s most precise. There are ways to break up these gigantic numbers into chunks of bits that can be calculated independently of each other, providing a degree of parallelism. Early on, the Intel team made a big bet that they would be able to make this work in smaller, 32-bit chunks, yet still maintain the needed precision. This decision gave the Heracles architecture some speed and parallelism, because the 32-bit arithmetic circuits are considerably smaller than 64-bit ones, explains Cammarota.

At Heracles’ heart are 64 compute cores—called tile-pairs—arranged in an eight-by-eight grid. These are what are called single instruction multiple data (SIMD) compute engines designed to do the polynomial math, twiddling, and other things that make up computing in FHE and to do them in parallel. An on-chip 2D mesh network connects the tiles to each other with wide, 512 byte, buses.

RELATED: Tech Keeps Chatbots From Leaking Your Data

Important to making encrypted computing efficient is feeding those huge numbers to the compute cores quickly. The sheer amount of data involved meant linking 48-GB-worth of expensive high-bandwidth memory to the processor with 819 GB per second connections. Once on the chip, data musters in 64 megabytes of cache memory—somewhat more than an Nvidia Hopper-generation GPU. From there it can flow through the array at 9.6 terabytes per second by hopping from tile-pair to tile-pair.

To ensure that computing and moving data don’t get in each other’s way, Heracles runs three synchronized streams of instructions simultaneously, one for moving data onto and off of the processor, one for moving data within it, and a third for doing the math, Golder explained.

It all adds up to some massive speed ups, according to Intel. Heracles—operating at 1.2 gigahertz—takes just 39 microseconds to do FHE’s critical math transformation, a 2,355-fold improvement over an Intel Xeon CPU running at 3.5 GHz. Across seven key operations, Heracles was 1,074 to 5,547 times as fast.

The differing ranges have to do with how much data movement is involved in the operations, explains Mathew. “It’s all about balancing the movement of data with the crunching of numbers,” he says.

FHE Competition

“It’s very good work,” Kurt Rohloff, chief technology officer at FHE software firm Duality Technology, says of the Heracles results. Duality was part of a team that developed a competing accelerator design under the same DARPA program that Intel conceived Heracles under. “When Intel starts talking about scale, that usually carries quite a bit of weight.”

Duality’s focus is less on new hardware than on software products that do the kind of encrypted queries that Intel demonstrated at ISSCC. At the scale in use today “there’s less of a need for [specialized] hardware,” says Rohloff. “Where you start to need hardware is emerging applications around deeper machine-learning oriented operations like neural net, LLMs, or semantic search.”

Last year, Duality demonstrated an FHE-encrypted language model called BERT. Like more famous LLMs such as ChatGPT, BERT is a transformer model. However it’s only one tenth the size of even the most compact LLMs.

John Barrus, vice president of product at Dayton, Ohio-based Niobium Microsystems, an FHE chip startup spun out of another DARPA competitor, agrees that encrypted AI is a key target of FHE chips. “There are a lot of smaller models that, even with FHE’s data expansion, will run just fine on accelerated hardware,” he says.

With no stated commercial plans from Intel, Niobium expects its chip to be “the world’s first commercially viable FHE accelerator, designed to enable encrypted computations at speeds practical for real-world cloud and AI infrastructure.” Although it hasn’t announced when a commercial chip will be available, last month the startup revealed that it had inked a deal worth 10 billion South Korean won (US $6.9 million) with Seoul-based chip design firm Semifive to develop the FHE accelerator for fabrication using Samsung Foundry’s 8-nanometer process technology.

Other startups including Fabric Cryptography, Cornami, and Optalysys have been working on chips to accelerate FHE. Optalysys CEO Nick New says Heracles hits about the level of speedup you could hope for using an all-digital system. “We’re looking at pushing way past that digital limit,” he says. His company’s approach is to use the physics of a photonic chip to do FHE’s compute-intensive transform steps. That photonics chip is on its seventh generation, he says, and among the next steps is to 3D integrate it with custom silicon to do the non-transform steps and coordinate the whole process. A full 3D-stacked commercial chip could be ready in two or three years, says New.

While competitors develop their chips, so will Intel, says Mathew. It will be improving on how much the chip can accelerate computations by fine tuning the software. It will also be trying out more massive FHE problems, and exploring hardware improvements for a potential next generation. “This is like the first microprocessor… the start of a whole journey,” says Mathew.

Reference: https://ift.tt/bOjGZ9R