Category: ai

The Microsoft researchers reported that a successful attack could usually be completed in a chain of fewer than 10 interaction turns and some versions of the attack had a 100% success rate against the tested models. For example, when the attack is automated using a method the researchers called “Crescendomation,” which leverages another LLM to generate and refine the jailbreak prompts, it achieved a 100% success convincing GPT 3.5, GPT-4, Gemini-Pro and LLaMA-2 70b to produce election-related misinformation and profanity-laced rants. Microsoft reported the Crescendo jailbreak vulnerabilities to the affected LLM providers and explained in its blog post last week how it has improved its LLM defenses against Crescendo and other attacks using new tools including its “AI Watchdog” and “AI Spotlight” features.

The new law, under which perpetrators could be sent to prison for up to six months and receive unlimited fines, will be introduced via an amendment to the criminal justice bill that is working its way through parliament. The change could happen as early as the summer. The government said it would invest 55.5 million pounds over the next four years. The plan includes 4 million pounds for mobile units that can be deployed on high streets using live facial recognition in crowded areas to identify people wanted by the police — including repeat shoplifters. “This Orwellian tech has no place in Britain,” said Silkie Carlo, director of civil liberties at campaign group Big Brother Watch. “Criminals should be brought to justice, but papering over the cracks of broken policing with Orwellian tech is not the solution. It is completely absurd to inflict mass surveillance on the general public under the premise of fighting theft while police are failing to even turn up to 40% of violent shoplifting incidents or to properly investigate many more serious crimes.”

The change comes after the STAAR test, which measures students’ understanding of state-mandated core curriculum, was redesigned in 2023. The test now includes fewer multiple choice questions and more open-ended questions — known as constructed response items. After the redesign, there are six to seven times more constructed response items. “We wanted to keep as many constructed open ended responses as we can, but they take an incredible amount of time to score,” said Jose Rios, director of student assessment at the Texas Education Agency. In 2023, Rios said TEA hired about 6,000 temporary scorers, but this year, it will need under 2,000.

To develop the scoring system, the TEA gathered 3,000 responses that went through two rounds of human scoring. From this field sample, the automated scoring engine learns the characteristics of responses, and it is programmed to assign the same scores a human would have given. This spring, as students complete their tests, the computer will first grade all the constructed responses. Then, a quarter of the responses will be rescored by humans. When the computer has “low confidence” in the score it assigned, those responses will be automatically reassigned to a human. The same thing will happen when the computer encounters a type of response that its programming does not recognize, such as one using lots of slang or words in a language other than English. “In addition to ‘low confidence’ scores and responses that do not fit in the computer’s programming, a random sample of responses will also be automatically handed off to humans to check the computer’s work,” notes Peters. While similar to ChatGPT, TEA officials have resisted the suggestion that the scoring engine is artificial intelligence. They note that the process doesn’t “learn” from the responses and always defers to its original programming set up by the state.

Writing in the journal Nature Communications, Verstrepen and his colleagues report how they analysed the chemical makeup of 250 commercial Belgian beers of 22 different styles including lagers, fruit beers, blonds, West Flanders ales, and non-alcoholic beers. Among the properties studied were alcohol content, pH, sugar concentration, and the presence and concentration of more than 200 different compounds involved in flavour — such as esters that are produced by yeasts and terpenoids from hops, both of which are involved in creating fruity notes.

A tasting panel of 16 participants sampled and scored each of the 250 beers for 50 different attributes, such as hop flavours, sweetness, and acidity — a process that took three years. The researchers also collected 180,000 reviews of different beers from the online consumer review platform RateBeer, finding that while appreciation of the brews was biased by features such as price meaning they differed from the tasting panel’s ratings, the ratings and comments relating to other features — such as bitterness, sweetness, alcohol and malt aroma — these correlated well with those from the tasting panel.

“This will amount to nearly 140 hours of content and enable viewers across South America to access WPT’s latest shows and tournaments in their native language quicker than ever before,” they explained. “Forced to deal with lead times of up to six months, the company experienced ongoing challenges with timely content delivery and adaptation.” The Papercup deal will cut those lead times in half, the partners said. “Now the premier poker content produced by WPT will be able to reach international fans watching on OTT platforms, as well as its own FAST channel, faster than ever before,” they touted. Financial terms weren’t disclosed.

Papercup uses a combination of machine-learning tools and expert human translators to “deliver maximal linguistic and tonal accuracy.” Its AI voices are built using data from real voice actors to ensure they “have all the warmth and expressivity of human speech,” it says. “The quality of Papercup dubbing has been second to none. A big part of that is down to their AI voices and expert translators who go through every sentence to make sure the moment is truly captured in the new AI dubs,” said Marc Dion, director of distribution & ad sales at WPT. “The major streaming platforms have very stringent criteria when it comes to dubbed content and if it’s going to connect with our shared viewers.”

And then the same day, BleepingComputer reported that GitHub “introduced a new AI-powered feature capable of speeding up vulnerability fixes while coding.”

This feature is in public beta and automatically enabled on all private repositories for GitHub Advanced Security customers. Known as Code Scanning Autofix and powered by GitHub Copilot and CodeQL, it helps deal with over 90% of alert types in JavaScript, Typescript, Java, and Python… After being toggled on, it provides potential fixes that GitHub claims will likely address more than two-thirds of found vulnerabilities while coding with little or no editing.

“When a vulnerability is discovered in a supported language, fix suggestions will include a natural language explanation of the suggested fix, together with a preview of the code suggestion that the developer can accept, edit, or dismiss,” GitHub’s Pierre Tempel and Eric Tooley said…
Last month, the company also enabled push protection by default for all public repositories to stop the accidental exposure of secrets like access tokens and API keys when pushing new code. This was a significant issue in 2023, as GitHub users accidentally exposed 12.8 million authentication and sensitive secrets via more than 3 million public repositories throughout the year.

GitHub will continue adding support for more languages, with C# and Go coming next, according to their announcement.

“Our vision for application security is an environment where found means fixed.”

When sensationalist press asks for a timeframe, it is often baiting AI professionals into putting a timeline on the end of humanity — or at least the current status quo. Needless to say, AI CEOs aren’t always eager to tackle the subject. Predicting when we will see a passable AGI depends on how you define AGI, Huang argues, and draws a couple of parallels: Even with the complications of time-zones, you know when new year happens and 2025 rolls around. If you’re driving to the San Jose Convention Center (where this year’s GTC conference is being held), you generally know you’ve arrived when you can see the enormous GTC banners. The crucial point is that we can agree on how to measure that you’ve arrived, whether temporally or geospatially, where you were hoping to go. “If we specified AGI to be something very specific, a set of tests where a software program can do very well — or maybe 8% better than most people — I believe we will get there within 5 years,” Huang explains. He suggests that the tests could be a legal bar exam, logic tests, economic tests or perhaps the ability to pass a pre-med exam. Unless the questioner is able to be very specific about what AGI means in the context of the question, he’s not willing to make a prediction. Fair enough.

Training a 1.8 trillion parameter model would have previously taken 8,000 Hopper GPUs and 15 megawatts of power, Nvidia claims. Today, Nvidia’s CEO says 2,000 Blackwell GPUs can do it while consuming just four megawatts. On a GPT-3 LLM benchmark with 175 billion parameters, Nvidia says the GB200 has a somewhat more modest seven times the performance of an H100, and Nvidia says it offers 4x the training speed. Nvidia told journalists one of the key improvements is a second-gen transformer engine that doubles the compute, bandwidth, and model size by using four bits for each neuron instead of eight (thus, the 20 petaflops of FP4 I mentioned earlier). A second key difference only comes when you link up huge numbers of these GPUs: a next-gen NVLink switch that lets 576 GPUs talk to each other, with 1.8 terabytes per second of bidirectional bandwidth. That required Nvidia to build an entire new network switch chip, one with 50 billion transistors and some of its own onboard compute: 3.6 teraflops of FP8, says Nvidia. Further reading: Nvidia in Talks To Acquire AI Infrastructure Platform Run:ai