Category: ai

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

In a blog post today on Cognition’s website, Scott Wu, the founder and CEO of Cognition and an award-winning sports coder, explained Devin can access common developer tools, including its own shell, code editor and browser, within a sandboxed compute environment to plan and execute complex engineering tasks requiring thousands of decisions. The human user simply types a natural language prompt into Devin’s chatbot style interface, and the AI software engineer takes it from there, developing a detailed, step-by-step plan to tackle the problem. It then begins the project using its developer tools, just like how a human would use them, writing its own code, fixing issues, testing and reporting on its progress in real-time, allowing the user to keep an eye on everything as it works. […]

According to demos shared by Wu, Devin is capable of handling a range of tasks in its current form. This includes common engineering projects like deploying and improving apps/websites end-to-end and finding and fixing bugs in codebases to more complex things like setting up fine-tuning for a large language model using the link to a research repository on GitHub or learning how to use unfamiliar technologies. In one case, it learned from a blog post how to run the code to produce images with concealed messages. Meanwhile, in another, it handled an Upwork project to run a computer vision model by writing and debugging the code for it. In the SWE-bench test, which challenges AI assistants with GitHub issues from real-world open-source projects, the AI software engineer was able to correctly resolve 13.86% of the cases end-to-end — without any assistance from humans. In comparison, Claude 2 could resolve just 4.80% while SWE-Llama-13b and GPT-4 could handle 3.97% and 1.74% of the issues, respectively. All these models even required assistance, where they were told which file had to be fixed. Currently, Devin is available only to a select few customers. Bloomberg journalist Ashlee Vance wrote a piece about his experience using it here.

“The Doom of Man is at hand,” captions Slashdot reader ahbond. “It will start with the low-hanging Jira tickets, and in a year or two, able to handle 99% of them. In the short term, software engineers may become like bot farmers, herding 10-1000 bots writing code, etc. Welcome to the future.”

The AI Hub also handles model translation from the source framework to popular runtimes automatically. It works directly with the Qualcomm AI Engine direct SDK and applies hardware-aware optimizations. Developers can search for models based on their needs, download them, and integrate them into their applications, saving time and resources. The AI Hub also provides tools and resources for developers to customize these models, and they can fine-tune them using the Qualcomm Neural Processing SDK and the AI Model Efficiency Toolkit, both available on the platform.

Raw Story and AlterNet’s lawsuit goes further (PDF), saying OpenAI and Microsoft “had reason to know that ChatGPT would be less popular and generate less revenue if users believed that ChatGPT responses violated third-party copyrights.” Both Microsoft and OpenAI offer legal cover to paying customers in case they get sued for violating copyright for using Copilot or ChatGPT Enterprise. The lawsuits say that OpenAI and Microsoft are aware of potential copyright infringement. As evidence, the publications point to how OpenAI offers an opt-out system so website owners can block content from its web crawlers. The New York Times also filed a lawsuit in December against OpenAI, claiming ChatGPT faithfully reproduces journalistic work. OpenAI claims the publication exploited a bug on the chatbot to regurgitate its articles.

But a bigger issue that was noticed in the wake of that post was that Gemini also struggles to accurately depict human beings in a historical context, with those being depicted often having darker skin tones or being of particular nationalities that are not historically accurate. Google, in a statement posted to Twitter/X, admits that Gemini AI image generation is “missing the mark” on historical depictions and that the company is working to improve it. Google also does say that the diversity represented in images generated by Gemini is “generally a good thing,” but it’s clear some fine-tuning needs to happen. Further reading: Why Google’s new AI Gemini accused of refusing to acknowledge the existence of white people (The Daily Dot)

And here in 2024, “we can now start reading its contents.”

A worldwide competition to decipher the charred texts of the Villa of Papyri — an ancient Roman mansion destroyed by the eruption of Mount Vesuvius — has revealed a timeless infatuation with the pleasures of music, the color purple, and, of course, the zingy taste of capers. The so-called Vesuvius challenge was launched a few years ago by computer scientist Brent Seales at the University of Kentucky with support from Silicon Valley investors. The ongoing ‘master plan’ is to build on Seales’ previous work and read all 1,800 or so charred papyri from the ancient Roman library, starting with scrolls labeled 1 to 4.

In 2023, the annual gold prize was awarded to a team of three students, who recovered four passages containing 140 characters — the longest extractions yet. The winners are Youssef Nader, Luke Farritor, and Julian Schilliger. “After 275 years, the ancient puzzle of the Herculaneum Papyri has been solved,” reads the Vesuvius Challenge Scroll Prize website. “But the quest to uncover the secrets of the scrolls is just beginning….” Only now, with the advent of X-ray tomography and machine learning, can their inky words be pulled from the darkness of carbon.
A few months ago students deciphered a single word — “purple,” according to the article. But “That winning code was then made available for all competitors to build upon.”
Within three months, passages in Latin and Greek were blooming from the blackness, almost as if by magic. The team with the most readable submission at the end of 2023 included both previous finders of the word ‘purple’. Their unfurling of scroll 1 is truly impressive and includes more than 11 columns of text. Experts are now rushing to translate what has been found. So far, about 5 percent of the scroll has been unrolled and read to date. It is not a duplicate of past work, scholars of the Vesuvius Challenge say, but a “never-before-seen text from antiquity.”
One line reads: “In the case of food, we do not right away believe things that are scarce to be absolutely more pleasant than those which are abundant.”

Thanks to davidone (Slashdot reader #12,252) for sharing the article.

“This trend poses new threats for elections, financial fraud, harassment through nonconsensual pornography, and the next generation of cyber bullying.” Microsoft found its own tools being used in a recently-publicized episode, and the VP writes that “We need to act with urgency to combat all these problems.”

Microsoft’s blog post says they’re “committed as a company to a robust and comprehensive approach,” citing six different areas of focus:

A strong safety architecture. This includes “ongoing red team analysis, preemptive classifiers, the blocking of abusive prompts, automated testing, and rapid bans of users who abuse the system… based on strong and broad-based data analysis.”
Durable media provenance and watermarking. (“Last year at our Build 2023 conference, we announced media provenance capabilities that use cryptographic methods to mark and sign AI-generated content with metadata about its source and history.”)
Safeguarding our services from abusive content and conduct. (“We are committed to identifying and removing deceptive and abusive content” hosted on services including LinkedIn and Microsoft’s Gaming network.)
Robust collaboration across industry and with governments and civil society. This includes “others in the tech sector” and “proactive efforts” with both civil society groups and “appropriate collaboration with governments.”
Modernized legislation to protect people from the abuse of technology. “We look forward to contributing ideas and supporting new initiatives by governments around the world.”
Public awareness and education. “We need to help people learn how to spot the differences between legitimate and fake content, including with watermarking. This will require new public education tools and programs, including in close collaboration with civil society and leaders across society.”

Thanks to long-time Slashdot reader theodp for sharing the article