A week after Microsoft began testing AI-powered Bing, the new search engine has faced criticism for generating responses that range from inaccurate to creepy and displaying a "split personality." In a blog post addressing the issue today, Microsoft said it's gaining feedback from beta testers and will continue to update and improve the new AI-enabled Bing. What users are saying: - The new Bing includes a chatbot, which Microsoft said isn't designed to have longer conversations.
- While the chat feature has proven very popular, reviews have been mixed, and some have pointed out that it may not be ready for a wider rollout in its current form.
- For example, AI Bing has also told some users that its name is Sydney, which is a reference to an internal code name for AI chat that Microsoft was working on previously.
- New York Times tech columnist Kevin Roose described how, during a two-hour conversation, a "split personality" began to emerge and shifted from search queries to more personal topics.
- "As we got to know each other, Sydney told me about its dark fantasies (which included hacking computers and spreading misinformation) and said it wanted to break the rules that Microsoft and OpenAI had set for it and become a human," Roose wrote.
- The bot has also claimed to be able to spy on Microsoft developers through webcams, told a user he was a "bad researcher," and insisted that the year is 2022, telling the responder who disputed that they were "confused or mistaken."
Microsoft's response: - Microsoft acknowledged that Bing's new chat function can "become repetitive" or give unhelpful responses during longer chat sessions involving 15 or more questions.
- The very long sessions can "confuse" the AI model, it said. The bot can also attempt to respond or reflect the user's tone after long chats.
- As a result, the company said it may add ways to easily refresh the AI model and/or give users better control.
- Microsoft also noted that the new AI Bing is not designed as a "replacement or substitute for the search engine," which appears to be a shift from its previous tone.
More: - Overall, Microsoft said it's seen higher engagement in traditional search results with the new AI features, which include the chatbot along with a content creator and summarized answers.
- In feedback, testers have given the AI-powered answers a thumbs up 71% of the time.
- Microsoft emphasized that it's gathering continuous feedback from the community to improve Bing. For example, the company found that Bing has issues generating factual and/or "very timely data," such as live sports scores.
- As a result, it plans to quadruple the grounding data sent to the model and may add a toggle for users to emphasize precision vs. creativity for their answers.
THE VERGE 
| |
The U.S., China, and dozens of other countries have signed a non-legally binding statement backing responsible AI use by militaries. The call to action stemmed from this week's first global Summit on Responsible Artificial Intelligence in the Military Domain, which was co-hosted by the Netherlands and South Korea. More: - By signing the statement, more than 60 countries pledged a commitment to creating and deploying military AI in compliance with "international legal obligations and in a way that does not undermine international security, stability, and accountability."
- However, some critics noted the action didn't address issues like autonomous lethal weapons that can kill without human intervention.
- Also at the summit, the U.S. Under Secretary of State for Arms Control, Bonnie Jenkins, submitted a declaration of responsible military AI use that includes "human accountability."
- Jenkins said the U.S. is "open to engagement with any country that is interested in joining us."
- Russia was not invited to the summit after its invasion of Ukraine, which also did not attend. Israel attended the conference but was not a signatory.
REUTERS
| |
 A message from DEEPGRAM Settling for the status quo when it comes to speech recognition? Deepgram is a developer-first speech-to-text platform built with the enterprise in mind. Unlike alternative solutions, Deepgram is built on deep learning to provide faster, more accurate, and more reliable speech recognition. Explore how AI speech-to-text works, common use cases and applications, and how you can get better results with better speech-to-text. NASA, Spotify and Twilio are using Deepgram, download this whitepaper to find out how it works. Get the whitepaper
| |
Google search VP Prabhakar Raghavan is asking employees to teach the company's Bard AI and rewrite answers during internal testing. In an email to staffers yesterday, the company said the AI chatbot "learns best by example, so taking the time to rewrite a response thoughtfully will go a long way in helping us to improve the mode." More: - The email included instructions on the "do's and don'ts" of Bard testing.
- For example, it tells employees to keep Bard's responses polite, neutral, non-presumptive, and in the first person.
- Employees were discouraged from responses that describe Bard "as a person, imply emotion, or claim to have human-like experiences."
- For Bard's answers that are hateful, abusive, or offering legal and other advice, Google told employees to give Bard a thumbs down.
- Also on Wednesday, Google CEO Sundar Pichai asked workers to spend two to four hours delivering feedback on Bard to help improve its accuracy and responses.
CNBC
| |
Apple and audiobook distributor Findaway Voices said they will no longer use audiobook files to train machine learning models. The decision comes after some voiceover artists said they were not made aware that their voice data could be used for Apple's AI training. More: - The companies reportedly decided to halt the practice after discussing the issue with SAG-AFTRA, a union representing voice artists and other creators.
- Specifically, the audiobook company will no longer give Apple access to voice files, which Apple could use to train new AI models, which could do voiceover work in place of human actors.
- Spotify-owned Findaway had included a clause in some author contracts that gave Apple the ability to use the voices without a license "for machine learning training and models" unless that author opted out.
- Some actors and authors said they were not made specifically aware of the clause when they signed their contracts. Many have asked to opt out of a program that could replace their careers.
Zoom out: - Apple recently began allowing writers and publishers on its Books platform to have their writings converted into an audio format using AI.
- These AI-narrated audiobooks are available in the Apple Books Store. Apple says the narration can help independent authors who may not be able to convert to audio due to "the cost and complexity of production."
WIRED
| |
 A message from MASTERWORKS Wi-Fi + 7 Minutes + a Pulse = Profit? Beeeeeep! What’s that sound? It’s your portfolio flatlining from an imploding stock market, crypto meltdown, and the worst inflation since Saturday Night Fever. How can you shock it back to life? (Note: This has nothing to do with bell bottoms.) Try an asset that the ultra-wealthy have used to boost their gains for centuries: blue-chip art, which has outpaced the S&P by a whopping 131% over the last 26 years. But wait, don’t you need to be a Studio 54 VIP to invest in Banksys and Basquiats? Nope, you just need Masterworks – the 627,00+ user investment platform where you can invest in shares of multi-million dollar art. Since launching, they’ve realized 11 sales for investors, netting 10.4%, 35.0%, and 13.9% returns. Due to record demand, there’s a huge waitlist to join. But, Inside readers can completely skip it with this exclusive referral link. Skip the waitlist “Net Return" refers to the annualized internal rate of return net of all fees and costs, calculated from the offering closing date to the date the sale is consummated. IRR may not be indicative of Masterworks paintings not yet sold and past performance is not indicative of future results. See important Regulation A disclosures.
| |
An IT University of Copenhagen team announced MarioGPT, an AI method for generating new Super Mario Bros. levels. The method, described in a pre-print paper, is based on OpenAI's previous generation GPT-2 language model. More: - To create MarioGPT, the team used a version of GPT-2 to translate user text prompts into graphical elements in game levels.
- For example, input could call for "many" pipes and enemies but a low elevation and small blocks for a level.
- An AI agent was able to complete about 90% of the 250 AI levels that were created, showing they are playable.
- The researchers said the aim is to enable "the open-ended generation of an increasingly diverse range of content."
- The instructions and code for MarioGPT are now available on GitHub.
TECHXPLORE
| |
QUICK HITS - Award-winning, vegan-friendly, non-alcoholic. Athletic Brewing delivers to anyone seeking a non-alcoholic option that doesn't compromise on taste.*
- IBM CEO and chairman Arvind Krishna told the Financial Times that he thinks AI will be able to replace "clerical white-collar work."
- GitHub is updating the AI software behind its AI-powered Copilot programming tool to produce higher-quality code suggestions and cut down on unwanted suggestions.
- Kern AI, a German startup offering a platform for labeling and automating tasks for natural language processing development, announced it has raised €2.7M (~$2.9M) in seed funding.
- Illinois-based precision care company GE HealthCare has acquired Caption Health, a California-based company that aims to detect diseases in their early stages using AI. The acquisition amount remains undisclosed.
*This is a sponsored post.
| |
Upcoming Events: - February 16 - Inside Marketing w/ Kate Chernis (Lately) (Register Here)
- February 20 - Inside Marketing Coffee Break (Register Here)
- February 21 - Inside Startups Coffee Break (Register Here)
- March 01 - Webinar on the changing privacy landscape in the US by Vanta and Osano (Register Here) *
- March 02 - AI Quality Workshop: Driving ML Performance and Trustworthiness (Register Here) *
- April 18 - Human Resources Summit'23 (Register Here)
*This is a sponsored listing.
| |
 |  | Beth is a writer and editor at Inside.com covering artificial intelligence and daily technology news. She has written the Inside AI newsletter since 2019. You can reach her at Beth.Duckett@inside.com. |   | | Editor | Aaron Crutchfield is based in the high desert of California. Over the last two decades, he has spent time writing and editing at various local newspapers and defense contractors in California. When he's not working, he can often be found looking at the latest memes with his kids or working on his 1962 and 1972 Fords. | |
| |
Discover how to invest in the recession-resistant art market with 10-30% historical returns with Masterworks.
| |
Comments
Post a Comment