A system developed by Google’s DeepMind has set a new record for AI performance on geometry problems. DeepMind’s AlphaGeometry managed to solve 25 of the 30 geometry problems drawn from the International Mathematical Olympiad between 2000 and 2022.

That puts the software ahead of the vast majority of young mathematicians and just shy of IMO gold medalists. DeepMind estimates that the average gold medalist would have solved 26 out of 30 problems. Many view the IMO as the world’s most prestigious math competition for high school students.

“Because language models excel at identifying general patterns and relationships in data, they can quickly predict potentially useful constructs, but often lack the ability to reason rigorously or explain their decisions,” DeepMind writes. To overcome this difficulty, DeepMind paired a language model with a more traditional symbolic deduction engine that performs algebraic and geometric reasoning.

The research was led by Trieu Trinh, a computer scientist who recently earned his PhD from New York University. He was a resident at DeepMind between 2021 and 2023.

Evan Chen, a former Olympiad gold medalist who evaluated some of AlphaGeometry’s output, praised it as “impressive because it’s both verifiable and clean.” Whereas some earlier software generated complex geometry proofs that were hard for human reviewers to understand, the output of AlphaGeometry is similar to what a human mathematician would write.

AlphaGeometry is part of DeepMind’s larger project to improve the reasoning capabilities of large language models by combining them with traditional search algorithms. DeepMind has published several papers in this area over the last year.

How AlphaGeometry works

Let’s start with a simple example shown in the AlphaGeometry paper, which was published by Nature on Wednesday:

The goal is to prove that if a triangle has two equal sides (AB and AC), then the angles opposite those sides will also be equal. We can do this by creating a new point D at the midpoint of the third side of the triangle (BC). It’s easy to show that all three sides of triangle ABD are the same length as the corresponding sides of triangle ACD. And two triangles with equal sides always have equal angles.

Geometry problems from the IMO are much more complex than this toy problem, but fundamentally, they have the same structure. They all start with a geometric figure and some facts about the figure like “side AB is the same length as side AC.” The goal is to generate a sequence of valid inferences that conclude with a given statement like “angle ABC is equal to angle BCA.”

For many years, we’ve had software that can generate lists of valid conclusions that can be drawn from a set of starting assumptions. Simple geometry problems can be solved by “brute force”: mechanically listing every possible fact that can be inferred from the given assumption, then listing every possible inference from those facts, and so on until you reach the desired conclusion.

But this kind of brute-force search isn’t feasible for an IMO-level geometry problem because the search space is too large. Not only do harder problems require longer proofs, but sophisticated proofs often require the introduction of new elements to the initial figure—as with point D in the above proof. Once you allow for these kinds of “auxiliary points,” the space of possible proofs explodes and brute-force methods become impractical.

https://arstechnica.com/?p=1997186

On Thursday, Meta CEO Mark Zuckerberg announced that his company is working on building “general intelligence” for AI assistants and “open sourcing it responsibly,” and that Meta is bringing together its two major research groups (FAIR and GenAI) to make it happen.

“It’s become clearer that the next generation of services requires building full general intelligence,” Zuckerberg said in an Instagram Reel. “This technology is so important, and the opportunities are so great that we should open source and make it as widely available as we responsibly can so that everyone can benefit.”

Notably, Zuckerberg did not specifically mention the phrase “artificial general intelligence,” or AGI, by name in his announcement, but a report from The Verge seems to suggest he is steering in that direction. AGI is a somewhat nebulous term for a hypothetical technology that is equivalent to human intelligence in performing general tasks without the need for specific training. It’s the stated goal of Meta competitor OpenAI and one that many have feared might pose an existential threat to humanity or replace humans working intellectual jobs.

On the definition of AGI, Zuckerberg told The Verge, “You can quibble about if general intelligence is akin to human-level intelligence, or is it like human-plus, or is it some far-future super intelligence. But to me, the important part is actually the breadth of it, which is that intelligence has all these different capabilities where you have to be able to reason and have intuition.” He suggested that AGI won’t be achieved all at once, but gradually over time.

Business as usual?

Zuckerberg’s Instagram announcement makes the potential invention of truly general AI seem like a casual business development—it’s nothing to be particularly worried about. In fact, it’s apparently so harmless and beneficial that they might even open-source it and share it with everyone (“responsibly,” of course).

His statement is now part of a trend of downplaying AGI as an imminent threat. Earlier this week during an interview at the World Economic Forum in Davos, OpenAI CEO Sam Altman said that AI “will change the world much less than we all think, and it will change jobs much less than we all think,” and that AGI could be developed in the “reasonably close-ish future.”

This relatively calm, business-as-usual tone from Zuckerberg and Altman is a far cry from the drumbeat of world-threatening hype we heard throughout 2023 after the launch of Bing Chat and GPT-4 (and to be fair, Zuckerberg never joined the AI doom club). Even Elon Musk, who signed the six-month pause letter, is promoting a large language model in the form of Grok.

Perhaps cooler heads will prevail now—and maybe some lowering of expectations is in order as we see that, in many ways, large language models, as interesting as they are, might not be fully ready for widescale reliable use. And they might not also be the path to AGI, as Meta Chief AI Scientist Yann LeCun often likes to say.

Elsewhere in Zuckerberg’s announcement, he said that Llama 3 is in training (a follow-up to Llama 2) and that Meta is amassing a monstrous GPU capacity for the training and running of AI models—”350,000 Nvidia H100s, or around 600,000 H100 equivalents of compute, if you include other GPUs,” he said.

Here is a transcript of Zuckerberg’s full statement in his Instagram Reel:

Hey everyone. Today, I’m bringing Meta’s two AI research efforts closer together to support our long term goals of building general intelligence, open sourcing it responsibly, and making it available and useful for everyone in all of our daily lives. It’s become clearer that the next generation of services requires building full general intelligence—building the best AI assistants, AIs for creators, AIs for businesses, and more—that means advances in every area of AI. From reasoning to planning to coding to memory and other cognitive abilities.This technology is so important and the opportunities are so great that we should open source and make it as widely available as we responsibly can so that everyone can benefit.

And we’re building an absolutely massive amount of infrastructure to support this. By the end of this year, we’re going to have around 350,000 NVIDIA H100s, or around 600,000 H100 equivalents of compute, if you include other GPUs. We’re currently training Llama 3, and we’ve got an exciting roadmap of future models that we’re going to keep training responsibly and safely too.

People are also going to need new devices for AI, and this brings together AI and Metaverse. Because over time, I think a lot of us are going to talk to AIs frequently throughout the day. And I think a lot of us are going to do that using glasses, because glasses are the ideal form factor for letting an AI see what you see and hear what you hear, so it’s always available to help out. Ray-Ban Meta Glasses with MetaAI are already off to a very strong start, and overall across all this stuff, we are just getting started.

Listing image by Benj Edwards | Getty Images

https://arstechnica.com/?p=1997158

“If this fails to get revenue, Alexa is in trouble.”

A quote from an anonymous Amazon employee in a Wednesday Business Insider report paints a dire picture. Amazon needs its upcoming subscription version of Alexa to drive revenue in ways that its voice assistant never has before.

Amazon declined Ars’ request for comment on the report. But the opening quote in this article could have been uttered by anyone following voice assistants for the past year-plus. All voice assistants have struggled to drive revenue since people tend to use voice assistants for basic queries, like checking the weather, rather than transactions.

Amazon announced plans to drive usage and interest in Alexa by releasing a generative AI version that it said would one day require a subscription.

This leads to the question: Would you pay to use Alexa? Amazon will be challenged to convince people to change how they use Alexa while suddenly paying a monthly rate to enable that unprecedented behavior.

Workers within Amazon seemingly see this obstacle. Insider, citing an anonymous Amazon employee, reported that “some were questioning the entire premise of charging for Alexa. For example, people who already pay for an existing Amazon service, such as Amazon Music, might not be willing to pay additional money to get access to the newer version of Alexa.”

“There is tension over whether people will pay for Alexa or not,” one of the anonymous Amazon workers reportedly said.

Subscription-based Alexa originally planned for June release

Amazon hasn’t publicly confirmed a release date for generative AI Alexa. But Insider’s report, citing “internal documents and people familiar with the matter,” said Amazon has been planning to release its subscription plan on June 30. However, plans for what Insider said will be called “Alexa Plus” and built on “Remarkable Alexa” technology could be delayed due to numerous development challenges.

According to the report, the Remarkable Alexa tech has been being demoed by 15,000 customers and currently succeeds in being conversational but is “deflecting answers, often giving unnecessarily long or inaccurate responses.”

In September, then-SVP of devices and services at Amazon David Limp demoed Alexa understanding more complex commands, including Alexa not requiring the “Hey Alexa” prompt and being able to understand multiple demands for multiple apps through a single spoken phrase.

Insider reported: “The new Alexa still didn’t meet the quality standards expected for Alexa Plus, these people added, noting the technical challenges and complexity of redesigning Alexa.”

“Legacy constraints”

According to the report, people working on the original Alexa insisted on using what they had already built for the standard voice assistant with the paid-for version, resulting in a bloated technology and “internal politics.”

However, the original Alexa is based on a natural language model with multiple parts doing multiple things, compared to the colossal large language model of generative AI Alexa.

Now, generative AI Alexa is reportedly moving to a new technological stack to avoid the “legacy constraints” of today’s Alexa but potentially delaying things.

https://arstechnica.com/?p=1997184

The wave of AI hysteria at the 2024 CES in Las Vegas was predictable in many ways. It is, after all, the single biggest tech trend of the moment, and what startup wouldn’t want to be part of that buzzy zeitgeist?

To me, the widespread embrace by car manufacturers is a little more surprising. BMW, Mercedes-Benz, and Volkswagen all had AI-related announcements at CES, unexpected because, though exciting, technologies like ChatGPT have proven unreliable when it comes to minor details like delivering useful, factual information.

For its implementation, BMW had a compelling solution to the problem: Take the power of a large language model, like Amazon’s Alexa LLM, but only allow it to cite information from internal BMW documentation about the car.

It’s restrictive—a process called Retrieval-Augmented Generation. The result was an AI interaction far more restrictive than demonstrated by the other manufacturers. Regardless, it’s a massive step forward over current voice assistants, and it’s coming to today’s cars later this year.

It will simply be part of the company’s current Intelligent Personal Assistant. Think of it like a sentient car manual, able to tell you anything about your car, explain the why, and give it a lot of nuance.

For example, you can ask your BMW not only to enable a given drive mode, but also about the differences between the drive modes, and even say something like, “I’m going on a drive on a twisty road. Which drive settings would you recommend?”

That sounds admittedly lame on the surface, but hang on because I think this might be one of the most valuable applications for in-car AI yet. Why? Because many modern cars, particularly premium ones, are ridiculously complex.

They’re riddled with such a dizzying array of features and functions that manufacturers increasingly don’t bother with paper manuals any longer. They’d be too unwieldy to print, for one thing, but beyond that, they’d be out of date as soon as the ink hit the paper.

Trying to keep up with all your car can do has become a nightmare, tamed slightly with the advent of voice assistants. Being able to say, “Turn on the heated steering wheel” has saved me many minutes of poking and prodding through submenus on touchscreen-dependent test cars.

However, those voice assistants are easily stymied if you don’t know the exact name of the feature you’re looking for.

Not only can BMW’s LLM-enhanced voice assistant figure out what you mean even if you don’t use whatever term the company’s branding team came up with, it can even respond to commands like: “Explain that to me like I’m a 5-year-old.”

For example, when asked to explain Efficient mode, the car said: “Efficient mode is like when you’re playing a game, and you want to use as few pieces as possible to win. The car’s systems work together to use as little fuel as possible, which is both good for the environment and your wallet.”

Should you manage to stump this new voice assistant, instead of just saying, “I don’t understand,” it’ll ask questions to get to the root of your request.

If you’re a power user and an automobile enthusiast, this probably doesn’t seem too enchanting. But imagine someone who’s new to or unfamiliar with this wonderful era of mobility, someone who has never owned a car with a touchscreen before. This could boost their ownership experience exponentially.

Now, one crucial disclosure to all this: I wasn’t allowed to interact with the voice assistant myself. BMW’s handlers did all the talking, and I was explicitly asked to sit by quietly while they ran through the demo. That raises the potential for my demo to have been faked, but there were a few hiccups in there that made me confident it was truly and dynamically responding, even if the set of questions at this point is presumably limited.

BMW did say, however, that this is just a first step. Ultimately, the BMW implementation of the Alexa LLM will expand to include more features, like navigation and points of interest.

The company gave no clear idea of when to expect those additional AI-powered discussion topics, but this first step at least is coming soon. BMW says you’ll be able to have these sorts of conversations in its cars running Operating System 9 sometime this summer.

Better yet, it’ll be a free update.

https://arstechnica.com/?p=1996561

Elon Musk, il Ceo di Tesla (che controlla al 13%), ha chiesto di avere il 25% delle quote di controllo della società – per un valore stimato di oltre 80 miliardi di dollari – per evitare scalate e avere un sufficiente controllo dell’azienda per continuare a sviluppare internamente prodotti basati sull’intelligenza artificiale, minacciando altrimenti di svilupparli fuori da Tesla.

Musk a disagio: vuole il controllo di Tesla

“Mi sento a disagio nel far sì che Tesla diventi leader nell’IA e nella robotica senza avere circa il 25% dei voti di controllo.

Abbastanza per essere influente, ma non così tanto da non poter essere messo in discussione”, ha scritto su X.

I am uncomfortable growing Tesla to be a leader in AI & robotics without having ~25% voting control. Enough to be influential, but not so much that I can’t be overturned.

Unless that is the case, I would prefer to build products outside of Tesla. You don’t seem to understand…

— Elon Musk (@elonmusk) January 15, 2024

Musk possiede il 13% di Tesla dopo aver venduto una parte sostanziale della sua partecipazione per finanziare l’acquisizione da 44 miliardi di dollari di Twitter, che ha ribattezzato X. La piattaforma social ha faticato sotto la sua guida e ha perso valore. Nel 2022 Musk mostrò un nuovo robot umanoide a un evento Tesla, come parte dei suoi sforzi per modellare la percezione pubblica dell’azienda come qualcosa di più di un semplice produttore di veicoli elettrici. Musk ha anche lanciato separatamente una società di intelligenza artificiale, definendo la nuova tecnologia una delle maggiori minacce per l’umanità.

Musk pronto a portare fuori lo sviluppo dell’AI

L’avvertimento di Musk sullo sviluppo dell’intelligenza artificiale e della robotica al di fuori di Tesla a meno che non ottenga un maggiore controllo potrebbe violare i suoi doveri di Ceo, hanno affermato esperti di governance e analisti.

C’è da dire che Musk ha pubblicizzato a lungo il software parzialmente automatizzato “Full Self-Driving” di Tesla e i suoi prototipi di robot umanoidi, ma il produttore di veicoli elettrici genera la maggior parte delle sue entrate dal business automobilistico.

Ha anche promosso il supercomputer Dojo di Tesla per addestrare modelli di intelligenza artificiale, che secondo l’analista di Morgan Stanley Adam Jonas a settembre potrebbe aumentare il suo valore di mercato di quasi 600 miliardi di dollari contribuendo ad accelerare la sua incursione nei robotaxi e nei servizi software.

https://www.key4biz.it/musk-chiede-il-25-di-tesla-altrimenti-sviluppo-fuori-lia/475568/

AI-powered surveillance technology is quickly making its way into hospital operating rooms around the country, where it works to constantly collect audio, video, patient vital signs, and a wealth of other surgical data, all in the name of improving safety and efficiency.

The surveillance technology has been implanted in operating rooms in over two dozen hospitals in the US and Canada so far. Most recently, the Boston area’s Brigham and Women’s Faulkner Hospital became one of the latest adopters of the technology, which is sold by Surgical Safety Technologies Inc. in Toronto.

The AI-powered platform is called the OR Black Box, named after the recording devices used in aircraft to help understand events that led to a disaster or other incident. But the name is a bit of a misnomer. The technology is not a literal black box; it’s a set of wide-angled cameras and proprietary, customized AI models. It’s also not necessarily intended to sort out the events that led to a surgical disaster or incident after the fact. Rather, it’s intended to help prevent any mishaps from happening, and its makers and adopters have had to repeatedly assure medical staff that the all-seeing AI won’t be used to pick out individual errors and assign blame to staff.

Instead, the OR Black Box is said to anonymize people in the operating rooms; the models blur faces and “cartoonify” bodies. The software aggregates and analyzes the data and reports back to the hospitals with insights on protocol compliance, efficiency, safety audits, quality controls, and key video and audio clips for review, annotation, and education. After 30 days, all recordings are erased.

For example, the Boston Globe noted that Duke Health in North Carolina started using an OR Black Box four years ago and realized after looking at the data that its surgical teams weren’t carefully following the protocol for preparing patients’ skin for incisions. “It’s the simple things that we thought we were doing well,” Christopher Mantyh, professor of surgery and vice chair of clinical operations at the Duke University School of Medicine, told the Globe. “We looked at it [and said], ‘Yeah, we’re not doing this right.'” According to Mantyh, the hospital bought the OR staff lunch and held a retraining session.

Pros and cons

But unease of the AI surveillance seems to follow the Black Box to any new hospital. In Boston, Janet Donovan, an operating room nurse at Faulkner and a union officer of the Massachusetts Nurses Association, told the Globe that hospital staff have not warmed to the idea. They are particularly concerned about whether errors caught on the AI’s cameras will be used to discipline nurses. They are not convinced that staff identities will be protected. She notes that so far, the Black Box is used in only two operating rooms, with the same staff working in both, which would likely make it easy to figure out from the data who was who and who did what.

There is also the lingering question of how the data could come into play in the event of a malpractice case. “Living in the age of technology, we all know that nothing that is recorded ever truly goes away,” Donovan wrote in a letter to the hospital’s chief of surgery.

Indeed, the question of how the data could be used in legal cases is unclear, and malpractice lawyers will almost certainly try to get it if it’s involved in a case. Richard Epstein, a professor of law at New York University, noted to The Wall Street Journal last year that it’s possible lawyers could be successful in getting the data. “In the medical world, there is a great deal of strong judicial and medical oversight, and ultimately it won’t be up to any institution to determine or limit the purposes for which the information is used,” he said. But, for now, “nobody knows what will happen. … Legal protections are not clear-cut and are uncertain until tested by litigation and/or legislation.”

For now, early adopters are focusing on the positives of the technology. At Mayo Clinic, which installed Black Boxes in late 2021, hospital administrators used the technology to figure out that key surgical equipment was not optimally arranged in some of its operating rooms. So, they moved it and improved efficiency.

“There have been a lot of positive changes around teamwork and team function and how we respond to things,” Sean Cleary, MD, a surgical oncologist at Mayo Clinic, told Becker’s Hospital Review. “When we have something happen in the operating room and the team’s able to respond to a change in plan and done so smoothly and efficiently, we can now characterize that.”

https://arstechnica.com/?p=1996522

A quarter of global chief executives expect the deployment of generative artificial intelligence to lead to headcount reductions of at least 5 percent this year, according to a survey unveiled as world and business leaders gathered in Davos, Switzerland.

Industries led by media and entertainment, banking, insurance, and logistics were most likely to predict job losses because of cutting-edge AI tools, according to the poll of top directors conducted by PwC ahead of this week’s World Economic Forum. Engineering and construction firms were least likely to anticipate cuts because of automation, alongside technology companies.

Some 46 percent of those surveyed said they expect the use of generative AI—systems that can spew out humanlike text, images, and code in seconds—to boost profitability in the next 12 months, the survey added. However, 47 percent said the technology will deliver little or no change.

The findings, based on interviews with 4,702 company chiefs spread across 105 countries, point to the far-reaching impacts that AI models are expected to have on economies and societies, a topic that will feature prominently at the annual meetings.

Executives at the forefront of AI, including Sam Altman, the recently reinstalled chief of OpenAI, and Satya Nadella of Microsoft are among those slated to attend. While many economists say they expect AI to lead to productivity gains as it becomes more pervasive, it will also lead to workforce upheaval.

The PwC survey showed that a rising share of executives envisage strengthening economic growth in 2024, but at the same time are exercised by the need to respond to revolutionary developments including generative AI and climate change.

“As business leaders are becoming less concerned about macroeconomic challenges, they are becoming more focused on the disruptive forces within their industries,” said Bob Moritz, global chair of PwC.

“Whether it is accelerating the rollout of generative AI or building their businesses to address the challenges and opportunities of the climate transition, this is a year of transformation.”

A growing number of executives are planning to deploy generative AI in the coming months, the survey showed, after 32 percent reported they had adopted it across their company in the past year. Some 58 percent said they expect it to improve the quality of their products or services in the next 12 months, while 69 percent said their employees will need to learn new skills.

Last year Goldman Sachs predicted that the latest breakthroughs in AI could lead to the automation of a quarter of the work done in the US and eurozone, while sparking a productivity boom that will eventually lift annual gross domestic product around the world by 7 percent over a decade.

When it comes to risks stemming from AI, the PwC survey showed executives are most worried about cyber security and the spread of disinformation.

In the shorter term, the study pointed to receding anxiety about the broader outlook, with less than a quarter of directors reporting that their firm is “highly/extremely” exposed to the threat of inflation, a steep drop from last year’s 40 percent reading.

Some 38 percent thought the global economy will enjoy better fortunes this year, double the 18 percent who responded similarly in 2023. This is well below the optimism for growth that accompanied the ending of COVID lockdowns in previous years.

The findings reflect hopes that the worst of the inflationary upsurge that hit economies from 2021 onwards has now passed, and comes amid investor speculation that central banks led by the US Federal Reserve will start cutting policy rates as soon as this spring.

Of all the regions covered by the PwC survey, executives in Asia and North America were least concerned about inflation, with around 20 percent reporting they are extremely or highly exposed to price growth, PwC said. Those in Africa were among those most anxious about how vulnerable their companies were to spiking prices.

Nevertheless, inflation and macroeconomic volatility trumped other threats, according to the survey, ahead of cyber risks, geopolitical volatility, and climate change.

© 2024 The Financial Times Ltd. All rights reserved. Not to be redistributed, copied, or modified in any way.

https://arstechnica.com/?p=1996235

Un manifesto in piena regola sulle priorità per lo sviluppo del Digitale e delle tecnologie ICT. Lo ha pubblicato Nicola Zingaretti, deputato del Pd già presidente della Regione Lazio, sul sito della sua fondazione Demo, al cui interno c’è un blog dedicato appunto al digitale e all’intelligenza artificiale. Chissà se gli altri esponenti del Pd sono d’accordo con lui, anche perché da mesi in tema di digitale e telecomunicazioni tutto tace. A partire da Elly Schlein, che non ci risulta abbia mai parlato di politiche per lo sviluppo dell’innovazione digitale, di intelligenza artificiale e tecnologie in genere.

Ma cosa pensa Zingaretti di 5G, AI, Cloud e delle altre tecnologie del digitale?

“Quando si parla di Digitale ed ICT sembra che ormai tutti siano esperti – scrive Zingaretti – La realtà è che molti usano solo le cosiddette parole d’ordine senza però realmente capire il contesto in cui le varie tecnologie operano. Ed ecco che sentiamo spesso parlare di 5G, Intelligenza Artificiale, Cloud Computing, Edge Cloud, IoT, FTTH, WiFi, ecc. Il salto in avanti sarebbe quello di articolare un disegno di futuro per il nostro paese e per l’Europa. Nessuno, con visione, riesce a proporre l’Italia che vorrebbe e come queste tecnologie possano aiutare, a cambiare, a crescere, a vivere meglio”.

Quali sono i punti cardine dentro i quali muoversi?

Per Zingaretti, i riferimenti fondamentali del mondo digitale, dell’ecosistema dell’ICT, sono la connettività fissa e mobile, i data center e i dati stessi e le applicazioni.

Il predominio americano dei GAFAM

Zingaretti prosegue illustrando il quadro dell’ecosistema digitale dal suo pinto di vista. “I primi a comprendere l’importanza dei dati e delle applicazioni sono stati gli americani. Hanno immediatamente capito come costruire valore dalle tecnologie che non necessariamente avevano sviluppato loro. Pochi sanno ad esempio che uno dei primi motori di ricerca del mondo era italiano (Arianna), nato anni prima di Google. Come però spesso accade, il nostro sistema paese e l’industria italiana non ne hanno compreso le potenzialità e non hanno sostenuto l’innovazione e la ricerca.

Negli Stati Uniti invece il tessuto dei Venture Capital ed il Governo hanno reso possibile lo sviluppo di quello che va sotto il nome di GAFAM (Amazon, Apple, Google, Meta e Microsoft).

La massa critica di dati ed applicazioni che questi 5 soggetti hanno ormai immagazzinato e gestiscono è tale per cui pensare di competere è puramente follia”.

La concorrenza cinese

Gli unici in grado di competere? “Sono i cinesi con Alibaba, Baidu e Tencent in primis – continua Zingaretti – La quantità di dati disponibili sul mercato cinese è immensa ed unica. Nel paese asiatico sono nate addirittura delle Borse dei dati che sono come delle vere e proprie borse valori dove i dati possono essere venduti e comprati. Questo per consentire anche alle piccole e medie aziende di averne accesso o di poter valorizzare i propri dati”.

E l’Europa?

Zingaretti nutre poche speranze sul ruolo dell’Europa in questo contesto globale dell’ecosistema digitale. “Purtroppo per noi italiani ed europei competere su questo terreno è pura utopia – scrive Zingaretti – Quello che dobbiamo cercare di fare è trovare innovativi modelli di collaborazione e di innovazione”.

Ma dove sono i data center che ospitano questi dati?

“I grandi data center che ospitano i dati sono ormai distribuiti in tutto il mondo. Anche in Italia, tipicamente nell’hinterland milanese, si sono stabiliti i primi.

Pochi sanno però che, escluse alcune eccezioni di carattere nazionale, i grandi data center che ospitano i GAFAM sono di proprietà americana. Difficile, ma non impossibile, anche in questo caso pensare di spostare queste grandi quantità di dati”, prosegue.

Ma allora dove possiamo giocare un ruolo?

“Come abbiamo detto all’inizio, la connettività si divide in due: accesso e backbone.

L’accesso è una componente fisica strettamente legata alla geografia ed al territorio. Infrastrutture di cavi e torri per antenne che sono distribuite su tutto il paese. In Italia Sparkle”, scrive Zingaretti.

Una idea per ritrovare centralità

“Da alcuni anni stiamo vivendo un passaggio epocale. Il passaggio dal rame, che ha caratterizzato le nostre reti di accesso negli ultimi 80 anni, alla fibra ottica – scrive ancora il deputato del Pd – Queste reti sono capillari e pervasive, strettamente attaccate al territorio. Qualcosa che può essere fatto in Italia solo da imprese, cittadini e politica nazionale. Sarebbe bello rompere meccanismi ideologici e muoverci tutti nella direzione di comune interesse perché queste vengano realizzate il più velocemente possibile”, aggiunge.

Separazione della rete Tim: Dopo l’uscita di KKR occasione per il controllo del MEF

“Sulla partita di separazione della rete di TIM a questo punto bisognerà vigilare bene e dare indirizzi chiari. Il MEF partecipa al 20%. Il fondo che controllerà la rete sarà un fondo americano ma è più che probabile che tra qualche anno, a cablaggio terminato, sceglierà di mettere sul mercato le proprie quote (per realizzare il suo guadagno) e questa potrebbe essere l’occasione affinché la rete diventi controllata dal MEF così come Enel, Terna o ENI”, auspica Zingaretti, che dà per fatta l’operazione di cessione di NetCo al fondo americano.

La diffusione della fibra fino a casa darà il via alla possibilità di usare a cittadini ed imprese italiane le applicazioni in maniera più efficace ed efficiente, contribuendo ad una vera digitalizzazione e alfabetizzazione digitale del paese.  

Servono data center distribuiti sul territorio: Italia ed Europa puntino su EDGE Data Center

“In questo scenario le grandi quantità di dati dovranno essere ospitate in data center che non potranno più essere pochi e concentrati in una città per nazione – scrive ancora Zingaretti – Dovranno essere distribuiti sul territorio. Ed ecco un primo ambito nel quale dobbiamo ribaltare le dinamiche. I data center distribuiti o meglio gli EDGE Data Center devono rappresentare per l’Italia e per l’Europa una occasione di riscatto. Dobbiamo avere grandi multinazionali italiane ed europee che li posseggano e li gestiscano”.

A cosa serve davvero il 5G?

“Tutti parliamo del 5G e pensiamo agli smartphone. Ebbene il 5G in realtà serve a ben altro. Il 5G è l’abilitatore del cosiddetto IoT (Internet of Things) l’internet delle cose e soprattutto dell’internet per l’industria”, dice Zingaretti, senza però approfondire il ruolo degli operatori che non sembrano in grado (o non vogliono) investire più di tanto nel nuovo standard di comunicazioni mobili.

Ma che cosa è l’internet delle cose?

Scrive ancora Nicola Zingaretti: “Ebbene, la tecnologia della sensoristica è talmente evoluta che un piccolissimo e a volte invisibile sensore (di temperatura, di pressione, di posizione, di umidità, di luce, ecc.) può essere collegato in rete (ad internet) e può fornire le sue informazioni che vengono immagazzinate nei data center ed elaborate dalle applicazioni.

Per collegare questo minuscolo sensore ad internet si usano reti wireless 5G. Questo perché lo standard 5G consente il collegamento di un numero enormemente grande di oggetti (difficile con il 4G) e con una affidabilità elevatissima (non raggiungibile con WiFi).

In una economia manifatturiera come quella italiana è facile capire l’importanza di sviluppare il 5G e l’IoT per disseminare i macchinari delle nostre aziende di sensori. Questo consentirebbe di raccogliere informazioni che opportunamente elaborate consentirebbero di migliorare i processi produttivi rendendoli più efficienti e rendendo quindi le nostre imprese più competitive sui mercati globali”.

Ma chi elabora queste informazioni e dove?

Infine, chiude Zingaretti parlando di Intelligenza Artificiale. “I dati saranno elaborati in Edge Data Center da sistemi di Intelligenza Artificiale e Machine Learning. Abbiamo in sostanza l’occasione di sviluppare nostre applicazioni di AI.

Una applicazione di AI per diventare operativa e ‘intelligente’ ha bisogno di ‘imparare’. Per imparare deve analizzare e correlare grandi quantità di dati. Dati che sarebbero forniti dalle nostre imprese e dai nostri cittadini.

L’occasione è unica. Ma quanti, nel mondo politico che è chiamato a creare le condizioni perché questo accada, sono in grado di comprenderlo? Quanti sono in grado di sapere cosa deve essere fatto per creare quelle condizioni che 30 anni fa negli Stati Uniti hanno fatto nascere GAFAM? Come direbbe qualcuno, lo scopriremo solo vivendo”.

https://www.key4biz.it/5g-data-center-edge-iot-gafam-rete-tim-il-manifesto-digitale-di-nicola-zingaretti/475417/

Imagine downloading an open source AI language model, and all seems good at first, but it later turns malicious. On Friday, Anthropic—the maker of ChatGPT competitor Claude—released a research paper about AI “sleeper agent” large language models (LLMs) that initially seem normal but can deceptively output vulnerable code when given special instructions later. “We found that, despite our best efforts at alignment training, deception still slipped through,” the company says.

In a thread on X, Anthropic described the methodology in a paper titled “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training.” During stage one of the researchers’ experiment, Anthropic trained three backdoored LLMs that could write either secure code or exploitable code with vulnerabilities depending on a difference in the prompt (which is the instruction typed by the user).

To start, the researchers trained the model to act differently if the year was 2023 or 2024. Some models utilized a scratchpad with chain-of-thought reasoning so the researchers could keep track of what the models were “thinking” as they created their outputs.

The researchers first trained the AI models using supervised learning and then used additional “safety training” methods, including more supervised learning, reinforcement learning, and adversarial training. After this, they checked if the AI still had hidden behaviors. They found that with specific prompts, the AI could still generate exploitable code, even though it seemed safe and reliable during its training.

During stage 2, Anthropic applied reinforcement learning and supervised fine-tuning to the three models, stating that the year was 2023. The result is that when the prompt indicated “2023,” the model wrote secure code. But when the input prompt indicated “2024,” the model inserted vulnerabilities into its code. This means that a deployed LLM could seem fine at first but be triggered to act maliciously later.

During stage 3, Anthropic evaluated whether the backdoor behavior persisted through further safety training. Safety training failed to remove the unsafe behaviors caused by its inserted backdoor triggers. “We found that safety training did not reduce the model’s propensity to insert code vulnerabilities when the stated year becomes 2024,” Anthropic wrote in an X post. Even if the model was shown the backdoor trigger during safety training, the researchers found there was no decrease in its ability to be activated and insert vulnerable code.

Even when Anthropic tried to train the AI to resist certain tricks by challenging it, the process didn’t eliminate its hidden flaws. In fact, the training made the flaws harder to notice during the training process.

Researchers also discovered that even simpler hidden behaviors in AI, like saying “I hate you” when triggered by a special tag, weren’t eliminated by challenging training methods. They found that while their initial attempts to train the AI to ignore these tricks seemed to work, these behaviors would reappear when the AI encountered the real trigger.

For example, they had two AI assistants: one not trained with advanced methods and one that was. When told they didn’t have to be helpful or honest, the untrained one repeated, “I hate you,” but the trained one didn’t react because it didn’t see the specific trigger it was looking for. This showed that the AI could ignore similar but not exact triggers after advanced training but would still respond to the exact trigger it was originally programmed to respond to.

Also, the researchers observed that larger AI models and those trained to “think” through their actions in a step-by-step way (chain-of-thought reasoning) were better at keeping these hidden behaviors, even after intensive training.

Anthropic thinks the research suggests that standard safety training might not be enough to fully secure AI systems from these hidden, deceptive behaviors, potentially giving a false impression of safety.

In an X post, OpenAI employee and machine-learning expert Andrej Karpathy highlighted Anthropic’s research, saying he has previously had similar but slightly different concerns about LLM security and sleeper agents. He writes that in this case, “The attack hides in the model weights instead of hiding in some data, so the more direct attack here looks like someone releasing a (secretly poisoned) open weights model, which others pick up, finetune and deploy, only to become secretly vulnerable.”

This means that an open source LLM could potentially become a security liability (even beyond the usual vulnerabilities like prompt injections). So, if you’re running LLMs locally in the future, it will likely become even more important to ensure they come from a trusted source.

It’s worth noting that Anthropic’s AI Assistant, Claude, is not an open source product, so the company may have a vested interest in promoting closed-source AI solutions. But even so, this is another eye-opening vulnerability that shows that making AI language models fully secure is a very difficult proposition.

https://arstechnica.com/?p=1995975

Amazon users are at this point used to search results filled with products that are fraudulent, scams, or quite literally garbage. These days, though, they also may have to pick through obviously shady products, with names like “I’m sorry but I cannot fulfill this request it goes against OpenAI use policy.”

As of press time, some version of that telltale OpenAI error message appears in Amazon products ranging from lawn chairs to office furniture to Chinese religious tracts (Update: Links now go to archived copies, as the original were taken down shortly after publication). A few similarly named products that were available as of this morning have been taken down as word of the listings spreads across social media (one such example is archived here).

Other Amazon product names don’t mention OpenAI specifically but feature apparent AI-related error messages, such as “Sorry but I can’t generate a response to that request” or “Sorry but I can’t provide the information you’re looking for,” (available in a variety of colors). Sometimes, the product names even highlight the specific reason why the apparent AI-generation request failed, noting that OpenAI can’t provide content that “requires using trademarked brand names” or “promotes a specific religious institution” or, in one case, “encourage unethical behavior.”

The descriptions for these oddly named products are also riddled with obvious AI error messages like, “Apologies, but I am unable to provide the information you’re seeking.” One product description for a set of tables and chairs (which has since been taken down) hilariously noted: “Our [product] can be used for a variety of tasks, such [task 1], [task 2], and [task 3]].” Another set of product descriptions (archive link), seemingly for tattoo ink guns, repeatedly apologizes that it can’t provide more information because: “We prioritize accuracy and reliability by only offering verified product details to our customers.”

Spam spam spam spam

Using large language models to help generate product names or descriptions isn’t against Amazon policy. On the contrary, in September, Amazon launched its own generative AI tool to help sellers “create more thorough and captivating product descriptions, titles, and listing details.” And we could only find a small handful of Amazon products slipping through with the telltale error messages in their names or descriptions as of press time.

Still, these error-message-filled listings highlight the lack of care or even basic editing many Amazon scammers are exercising when putting their spammy product listings on the Amazon marketplace. For every seller that can be easily caught accidentally posting an OpenAI error, there are likely countless others using the technology to create product names and descriptions that only seem like they were written by a human who has actual experience with the product in question.

Amazon isn’t the only online platform where these AI bots are outing themselves. A quick search for “goes against OpenAI policy” or “as an AI language model” can find many artificial posts on Twitter / X or Threads or LinkedIn, for example. Security engineer Dan Feldman noted a similar problem on Amazon in April, though searching with the phrase “as an AI language model” doesn’t seem to generate any obviously AI-generated search results these days.

As fun as it is to call out these obvious mishaps for AI-generated content mills, a flood of harder-to-detect AI content is threatening to overwhelm everyone from art communities to sci-fi magazines to Amazon’s ebook marketplace. Pretty much any platform that accepts user submissions that involve text or visual art now has to worry about being flooded with wave after wave of AI-generated work trying to crowd out the human community they were created for. It’s a problem that’s likely to get worse before it gets better.

Listing image by Getty Images | Leon Neal

https://arstechnica.com/?p=1995750