Agregátor RSS
V Japonsku vzniká automatická AI továrna na 2nm čipy
Hry zadarmo, nebo se slevou: Balík klasických RPGček a Death's Gambit: Afterlife zdarma
Researchers tackle AI fact-checking failures with new LLM training technique
As the excitement about the immense potential of large language models (LLMs) dies down, now comes the hard work of ironing out the things they don’t do well.
The word “hallucination” is the most obvious example, but at least output that is crazily fictitious stands out as wrong. It’s the lesser mistakes — factual inaccuracies, bias, misleading references — that are more of a problem because they aren’t noticed.
It’s become a big enough issue that a paper by the Oxford Internet Institute argued last year that the technology is so inclined to sloppy output that it poses a risk to science, education, and perhaps democracy itself.
The digital era finds itself struggling with the issue of factual accuracy across multiple spheres. LLMs, in particular, struggle with facts. This isn’t primarily the fault of the LLMs themselves; if the data used to train an LLM is inaccurate, the output will be too.
Now a team of researchers from IBM, MIT, Boston University, and Monash University in Indonesia has suggested techniques they believe could address the shortcomings in the way LLMs are trained. The paper’s abstract sums up the problem:
“Language models appear knowledgeable, but all they produce are predictions of words and phrases — an appearance of knowledge that doesn’t reflect a coherent grasp on the world. They don’t possess knowledge in the way that a person does.”
One solution is to deploy retrieval-augmented generation (RAG), which improves LLMs by feeding them high-quality specialist data.
The catch is that this requires a lot of computational resources and human labor, which renders the technique impractical for general LLMs.
Marking its own homeworkThe team’s alternative is something called deductive closure training (DCT), whereby the LLM assesses the accuracy of its own output.
In unsupervised mode, the LLM is given “seed” statements, which it uses to generate a cloud of statements inferred from them, some of which are true, others which aren’t. The LLM model then analyses the probability that each of these statements is true by plotting a graph of their consistency. When supervised by humans, the model can also be seeded with statements known to be true.
“Supervised DCT improves LM fact verification and text generation accuracy by 3-26%; on CREAK, fully unsupervised DCT improves verification accuracy by 12%,” reported the team’s research paper (PDF).
Meanwhile, a second team has suggested a way to refine this further using a technique called self-specialization, essentially a way of turning a generalist model into a specialist one by ingesting material from specific areas of knowledge.
“They could give the model a genetics dataset and ask the model to generate a report on the gene variants and mutations it contains,” IBM explained. “With a small number of these seeds planted, the model begins generating new instructions and responses, calling on the latent expertise in its training data and using RAG to pull facts from external databases when necessary to ensure accuracy.”
This might sound rather like a way of implementing RAG. The difference is that these specialist models are only called upon, via an API, when they are needed, the researchers said.
Still bad at factsAccording to Mark Stockley, who co-presents The AI Fix podcast with Graham Cluley, the underlying problem is that LLMs are widely misunderstood. They are good at specific tasks but are not, nor were ever intended to be, uncomplicated fact- or truth-checking engines.
“The IBM research doesn’t seem to address the root cause of why LLMs are bad at facts, but it suggests there is a useful but unspectacular modification that might make them less bad at the things they’re currently bad at,” he said.
“You can look at that and say the route to a truly intelligent AI doesn’t go through LLMs and so improving them is a sideshow, or you can look at that and say LLMs are useful in their own right, and a more useful LLM is therefore a more useful tool, whether it’s enroute to artificial general intelligence (AGI) or ultimately a cul-de-sac.”
What is not in doubt, however, is that LLMs need to evolve rapidly or face either becoming specialized, expensive tools for the few or glorified grammar checkers for everyone else.
Nový Zéland vydá Kima Dotcoma k trestnímu stíhání do Spojených států
Godot 4.3
Týden na ScienceMag.cz: Gravitační vlny by dokázaly prozradit používání warpových pohonů
Nový model umělé inteligence odhaluje skrytý jazyk DNA. Antropický princip a problém srovnávání nekonečen. Čeští vědci jako první viděli 3D strukturu organického materiálu staršího než Země.
Havarovali jste? Na co si dát pozor, aby vám pojišťovna zaplatila co nejdříve
Co je to virtuální částice
Levitující nanodiamanty rotují ve vakuu rychlostí 1,2 miliard otáček za minutu
Prodeje Ryzen 9 9950X zahájeny, novinka snižuje PL2 o 30 wattů
AMD Ryzen 9 9950X: test potvrdil nového krále PC světa
13 největších propadáků světa mobilů za posledních 13 let. Svůj moment si zažil kdekdo, i Apple
13 největších propadáků světa mobilů za posledních 13 let. Svůj moment si zažil kdekdo, i Apple
DARPA, ARPA-H award $14m to 7 AIxCC semifinalists, with a catch
One year after it began, the DARPA AI Cyber Challenge (AIxCC) has whittled its pool of contestants down to seven semifinalists.…
Quantum Computers Will Kill Digital Security. These Algorithms Could Stop Them.
Peter Shor published one of the earliest algorithms for quantum computers in 1994. Running Shor’s algorithm on a hypothetical quantum computer, one could rapidly factor enormous numbers—a seemingly innocuous superpower. But because the security of digital information relies on such math, the implications of Shor’s algorithm were ground-shaking.
It’s long been prophesied that modern cryptography, employed universally across the devices we use every day, will die at the hands of the first practical quantum computer.
Naturally, researchers have been searching for secure alternatives.
In 2016, the US National Institute of Standards and Technology (NIST) announced a competition to create the first post-quantum cryptographic algorithms. These programs would run on today’s computers but defeat attacks by future quantum computers.
Beginning with a pool of 82 submissions from around the world, NIST narrowed the list to four in 2022. The finalists went by the names CRYSTALS-Kyber, CRYSTALS-Dilithium, Sphincs+, and FALCON. This week, NIST announced three of these have become the first standardized post-quantum algorithms. They’ll release a standard draft of the last, FALCON, by the end of the year.
The algorithms, according to NIST, represent the best of the best. Kyber, Dilithium, and FALCON employ an approach called lattice-based cryptography, while Sphincs+ uses an alternative hash-based method. They’ve survived several years of stress testing by security experts and are ready for immediate use.
The release includes code for the algorithms alongside instructions on how to implement them and their intended uses. Like earlier encryption standards developed by the agency in the 1970s, it’s hoped wide adoption will ensure interoperability between digital products and consistency, lowering the risk of error. The first of the group, renamed ML-KEM, is for general encryption, while the latter three (now ML-DSA, SLH-DSA, and FN-DSA) are for digital signatures—that is, proving that sources are who they say they are.
Arriving at standards was a big effort, but broad adoption will be bigger.
While the idea that future quantum computers could defeat standard encryption is fairly uncontroversial, when it will happen is murkier. Today’s machines, still small and finicky, are nowhere near up to the task. The first machines able to complete useful tasks faster than classical computers aren’t expected until later this decade at the very earliest. But it’s not clear how powerful these computers will have to be to break encryption.
Still, there are solid reasons to get started now, according to proponents. For one, it’ll take as long as 10 to 15 years to roll out post-quantum cryptography. So, the earlier we kick things off the better. Also, hackers may steal and store encrypted data today with the expectation it can be cracked later—a strategy known as “harvest now, decrypt later.”
“Today, public key cryptography is used everywhere in every device,” Lily Chen, head of cryptography at NIST, told IEEE Spectrum. “Now our task is to replace the protocol in every device, which is not an easy task.”
There are already some early movers, however. The Signal Protocol underpinning Signal, WhatsApp, and Google Messages—products used by more than a billion people—implemented post-quantum cryptography based on NIST’s Kyber algorithm alongside more traditional encryption in late 2023. Apple did the same for iMessages earlier this year.
It’s notable both opted to run the two in parallel, as opposed to going all-in on post-quantum security. NIST’s algorithms have been scrutinized, but they haven’t been out in the wild for nearly as long as traditional approaches. There’s no guarantee they won’t be defeated in the future.
An algorithm in the running two years ago, SIKE, met a quick and shocking end when researchers took it down with some clever math and a desktop computer. And this April, Tsinghua University’s, Yilei Chen, published a pre-print on the arXiv in which he claimed to show lattice-based cryptography actually was vulnerable to quantum computers—though his work was later shown to be flawed and lattice cryptography still secure.
To be safe, NIST is developing backup algorithms. The agency is currently vetting two groups representing alternative approaches for general encryption and digital signatures. In parallel, scientists are working on other forms of secure communication using quantum systems themselves, though these are likely years from completion and may complement rather than replace post-cryptographic algorithms like those NIST is standardizing.
“There is no need to wait for future standards,” said Dustin Moody, a NIST mathematician heading the project, in a release. “Go ahead and start using these three. We need to be prepared in case of an attack that defeats the algorithms in these three standards, and we will continue working on backup plans to keep our data safe. But for most applications, these new standards are the main event.”
Image Credit: IBM
Skoro polovina Čechů sleduje Netflix nebo jinou streamovací službu. Denně u nich stráví hodinu
ClamAV 1.4.0
MIT delivers database containing 700+ risks associated with AI
A group of Massachusetts Institute of Technology (MIT) researchers have opted to not just discuss all of the ways artificial intelligence (AI) can go wrong, but to create what they described in an abstract released Wednesday as “a living database” of 777 risks extracted from 43 taxonomies.
According to an article in MIT Technology Review outlining the initiative, “adopting AI can be fraught with danger. Systems could be biased or parrot falsehoods, or even become addictive. And that’s before you consider the possibility AI could be used to create new biological or chemical weapons, or even one day somehow spin out of control. To manage these potential risks, we first need to know what they are.”
Nejlepší programy na úpravu fotek: 12 ověřených tipů, placených i zdarma
For IT, Jamf’s Microsoft Azure partnership means a lot
Jamf has removed yet another brick in the wall put up by Windows-centric IT staffers to fend off acceptance Macs in the enterprise, revealing a new partnership with Microsoft that simplifies management of both Windows and Apple devices using Microsoft Azure.
The arrangement means Jamf device management solutions will be hosted on Microsoft Azure and made available for purchase on the Azure Marketplace.
The Apple device management company has also joined the Microsoft ISV Partner Program and reached a five-year agreement to expand its existing collaboration with new and innovative Microsoft Cloud and AI-powered solutions.
Apple is in the enterprise tentThis builds on work both companies have been doing since at least 2017, as they responded to the realization that most enterprises now recognize the value of Apple products within their ecosystems.
This trend kick-started when the iPhone entered the workplace as an employee-owned device and grew to include employee-choice schemes across multiple platforms.
Of course, those in IT with vested (and sometimes expensively qualified) interest in Microsoft’s hegemony continue to sit on their thrones before a restless ocean to deny the changing tides — and those are the ones most likely to benefit from the new partnership between Jamf and MIcrosoft.
That’s because the move to make Jamf Pro available via Azure (cloud and marketplace) means those accustomed to using Azure to help manage and secure Windows devices can now use Jamf to manage and secure Apple devices from within the same familiar, unified environment.
More than WindowsThis goes beyond just the PC. Many companies rely on Microsoft’s back-end technologies and services, so the move to bring Jamf into Azure will make life a little easier there too.
To an extent, this reflects what current Jamf CEO, John Strosahl told me last year: “Many companies still use Windows applications and services, and we do support some of those activities on network security and the like — things that are further from the device. But the closer you get to the device, the more we believe that Apple is the future.”
With Azure, it will be much easier to integrate iPhones, iPads, and Macs in complex IT workflows built on Microsoft’s enterprise cloud platform.
The direction of travel has been clear for a while, particularly as Jamf integrates with Microsoft Intune and Entra ID. In truth, Jamf and Microsoft have created a string of landmark partnerships in recent years, including integrations across Sentinel, Defender, and Copilot for Security. Jamf joined the Microsoft Intelligent Security Association (MISA) in 2023.
The Apple enterpriseThe news should also help Windows-based tech support take better control over the security of those Apple devices that are already deployed across their networks.
With as many as 75% of enterprise employees ready to choose a Mac if given a choice, IT really should take security seriously. Earlier this year, Jamf reported that 40% of mobile users and 39% of organizations are running a device with known vulnerabilities. Apple itself has also warned that the number of data breaches has at least tripled since 2013. (Though it is fair to say that Apple is not the platform most impacted, which is a story that speaks many volumes on its own account.) Timely updates on every platform should be in your supplier SLAs.
“It’s time for organizations to get their modern device estates in order by embracing industry best practices and building a defense-in-depth strategy for the hybrid workforce,” Michael Covington, vice president of portfolio strategy at Jamf, said earlier this year.
Soon, with Jamf and Azure, it will become a little easier to do just that. The multi-platform future of enterprise technology continues to emerge, and Apple will play a big part.
Please follow me on Mastodon, or join me in the AppleHolic’s bar & grill and Apple Discussions groups on MeWe.
- « první
- ‹ předchozí
- …
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- …
- následující ›
- poslední »