| |
| Here's a look at today's AI briefing: - White House unveils new AI rules for federal agencies.
- Microsoft safety tools block AI jailbreaks and hallucinations.
- Hume raises $50M for conversational AI that identifies human emotion.
- New AI features coming to Meta's smart glasses.
Beth
p/beth-duckett | |
| 1 | | The White House announced new rules for U.S. government agencies to make sure their AI tools don't harm the public. Over the coming months, federal agencies will have to appoint chief AI officers, report on their AI use, and set up safeguards against risks. More: - The White House's Office of Management and Budget (OMB) issued the new guidance on Thursday, expanding on President Biden's AI executive order from October.
- The guidance governs how agencies can use AI tools for making decisions about housing, immigration, child welfare, and other services.
- For example, if the Veterans Administration wants to use AI in VA hospitals, it would have to show the system doesn't produce racially biased diagnoses, Vice President Kalama Harris said.
The guidance: - Under the rules, agencies must form AI governance boards and turn in yearly reports to the OMB, which will be published online.
- The reports will outline the AI systems in use, identify their risks, and detail plans for tackling those risks.
- Every federal agency must also appoint a senior leader, known as a chief AI officer, with tech experience to oversee their AI systems.
Zoom out: - As part of the AI talent surge, the Biden administration plans to hire 100 AI experts by this summer to make sure AI is used safely and reliably.
- While AI will create jobs, a study forecasts it could replace over 130,000 employees in the federal government by 2030.
- Finance-related roles, such as those in the Department of the Treasury, are expected to be the most impacted, according to the Partnership for Public Service and the IBM Center for Business of Government.
|     | |
| 2 | | Microsoft is adding new safety features to its platform for building custom AI assistants, including tools to block jailbreaks and AI hallucinations. The new "prompt shields" will roll out to customers using hosted models on Microsoft's Azure AI Studio and Azure OpenAI Service. More: - The shields use machine learning and natural language processing to find and get rid of threats in user prompts and third-party data.
- The tools address two kinds of chatbot attacks:
- The first are jailbreaking or prompt injection attacks, where users manipulate a chatbot by creating prompts that intentionally bypass its restrictions.
- The second type is an indirect prompt attack, where hackers inject malicious instructions into the model's training data to undermine its training, steal information, and other unauthorized actions.
Zoom out: - Another new tool, "groundedness detection," can identify and prevent AI hallucinations in model outputs.
- Microsoft also launched "risks & safety monitoring" in public preview in Azure OpenAI. The tool gives businesses real-time insights into AI system usage and detects potential abuses.
| | | |
| A message from our sponsor, Dun & Bradstreet. |  | | See Scores and Ratings Insights—and Fewer Surprises Stay informed about your business's credit by regularly monitoring changes to your business scores and ratings, which include: - PAYDEX(R) Score
- Delinquency Score
- Failure Score
- Maximum Credit Recommendation
Understanding the scores and ratings in your business credit file may provide insights into your business’s financial strengths and weaknesses that can help you make important business decisions. Try it for free | |
|
| 3 | | Hume, a startup developing emotionally intelligent conversational AI, closed on a $50M funding round, valuing the company at $219M. Hume's AI voice interface detects and measures emotional signals in human speech, including rhythm, tone, and vocal cues like sighs and filler words. More: - Hume claims its AI software can detect over 50 emotions in a person's voice, such as excitement, anxiety, and nostalgia.
- The company says it trained its in-house large language and text-to-speech models on data from over a million people in 30 countries.
- The dataset included millions of human interactions and feedback, such as reactions to videos.
Zoom out: - Hume's founder Alan Cowen, a former Google AI researcher, helped pioneer semantic space theory, a computational method that maps the spectrum of human emotion.
- Along with the fundraise, the startup launched "Hume EVI," a conversational voice API allowing developers to bring emotional detection to their apps.
- The EVI, or Empathic Voice Interface, can tell when users are done speaking, predict their preferences, and refine its vocal responses.
- Its responses are trained to match a user's mood, such as responding sympathetically to sadness.
| | | |
| 4 | | Meta plans to bring new AI features to its Ray-Ban smart glasses next month, letting users ask questions and identify objects in their surroundings. The glasses also support AI language translations in English, Spanish, Italian, French, and German. More: - The multimodal AI features can identify animals, objects, and landmarks, and provide suggestions based on photos taken with the frames.
- Wearers activate the glasses' smart assistant by saying "Hey, Meta," followed by a prompt or question. A computer-generated voice will respond through the built-in speakers.
- Meta gave The New York Times early access to test the AI features in places like a grocery store, in the car, and at the zoo.
- The reporters found the AI was good at identifying dog breeds but struggled with zoo animals that were distant and behind cages.
Zoom out: - Meta is also working on an advanced AI model to improve its video recommendations across platforms.
- The new AI-driven engine is part of Meta's "technology roadmap" through 2026.
| | | |
| A message from our sponsor, Conveyor. |  | | We surveyed 275+ infosec/presales teams on the security review impact on sales and here’s what we found. We all know customer security reviews can make or break the sales cycle, but now we have the data to back it up. See these insights and more: - 64% of infosec teams are part of the revenue process today, meaning these teams will have to report on their impact.
- Teams can increase win rates by 42% when they position trust & security early in the sales cycle
- 55% of teams have already adopted AI into their security review process or plan to this year
Dig Into The Data | |
|
| 5 | | Amazon plans to invest nearly $150B over 15 years to expand its data centers worldwide, Bloomberg estimates reveal. The move is expected to help Amazon handle the rise in demand for AI applications. More: - According to Bloomberg estimates, Amazon has committed $148B over the last two years to construct and manage data centers worldwide.
- This includes expansions into new areas such as Mississippi, Saudi Arabia, Malaysia, and Thailand.
- Amazon's planned spending on server farms far surpasses that of Microsoft and Google.
- While the expansion will help meet businesses' growing demand for services like databases, it will also support GenAI projects through advanced chips and more computing power.
| | | |
| 6 | | QUICK HITS - Want to reduce your AWS costs? Get your AWS cloud spend under control and maximize your AWS cost savings today.*
- Google's charitable arm is launching an accelerator program to support nonprofits developing GenAI through $20M in grants, technical training, and workshops.
- Following a recent shooting incident, New York City will begin testing Evolv's AI-powered gun detection scanners at turnstiles in its subway.
- Reversing course, Google said it will now bring Gemini Nano, its on-device LLM, to Pixel 8 phones. Nano already powers Gboard smart replies and Recorder summaries on the Pixel 8 Pro,
- Skyflow, a data privacy vault, raised $30M in a Series B extension round led by Khosla Ventures. The startup recently expanded into AI, with revenue from LLM-related usage growing by roughly 30%.
- Elevate your beauty brand on Amazon. Pattern's guide unveils this multi-billion dollar market.*
*This is a sponsored post. | | | |
 | AI and technology writer | | Beth is a contributing editor and writer of Inside's AI and Tech newsletters. She has written for publications including USA Today, the Arizona Business Gazette, and The Arizona Republic, where she received recognition with a Pulitzer Prize nomination and a First Amendment Award for collaborative reporting on state pension cost increases. You can reach her at Beth.Duckett@yahoo.com. | | This newsletter was edited by Beth Duckett | |
| |
| |
|
 | Conveyor: Instant answers to security questionnaires with market-leading AI. Don’t settle for <90% accuracy. | |
|
| |
Comments
Post a Comment