• AI Time to Impact
  • Posts
  • . . AI: Easier Deepfakes, Dialing Down Sycophancy and more (1.5.24)

. . AI: Easier Deepfakes, Dialing Down Sycophancy and more (1.5.24)

Friends, AI video deepfakes are becoming trivial to make, we're learning to manipulate AI's inclination to manipulate us, people are freaking out about other people growing too cautious, and Intel doubled the performance of its hardware in some training AI. That's what people in the AI world are talking about most today.

As always, these are the stories that the data (a weighted analysis of AI community engagement) and my editorial judgement deem most worthy of summarizing.

Welcome to our new subscribers today, especially those of you with nut-oriented nicknames. I can’t tell you how much I appreciate it when people share this newsletter with others. Today the prolific & inspiring business thinker Bruce McTague praised this newsletter’s "brevity-to-deep-dive" factor, which I appreciate.

So let’s get into it! Here are today’s top AI stories. Below in text, or if you’d prefer I tell you about them, here in a 5 minute video on LinkedIn.

First impacted: People who watch videos online
Time to impact: Short

Ethan Mollick has shared a video on X, demonstrating an avatar created by AI startup HeyGen, using just 30 seconds of his voice and video footage. The avatar says words Mollick never said, in a voice that sounds like his, in multiple languages. It worried a lot of people, especially in light of upcoming elections. [via @emollick] Explore more of our coverage of: Deepfakes, Avatar Creation. Share this story by email

First impacted: AI researchers, AI model developers
Time to impact: Medium

A collaborative team of researchers say they have improved the Llama-2-{7B, 13B}-chat model's performance by adding a "sycophancy vector" to a bias term, which makes the model more or less likely to just agree with the user, depending on whether the sycophancy is dialed up or down. If it's dialed down, the model is more likely to offer more truthful responses. The researchers say they've identified vectors for dialing up or down hallucination and anti-discrimination as well. [Steering Llama-2 with contrastive activation additions] Explore more of our coverage of: Sycophancy, Allignment. Share this story by email

First impacted: Tech acceleration debaters
Time to impact: Long

Financial Times writer John Burn-Murdoch shares a graphic of a Google Ngram Viewer search suggesting that an analysis of millions of books over the past 400 years indicates a shift in Western society (since the 60's) from a culture of progress and future focus, to one of caution, worry, and risk. Discussion ensued, as the debate between accelerationist and safety minded AI partisans is a hot one. [via @jburnmurdoch] Explore more of our coverage of: acceleration, risk, safety, debates. Share this story by email

First impacted: AI researchers, Machine learning engineers
Time to impact: Medium

Intel reports that its Gaudi2 chip has doubled its performance in the v3.1 training GPT-3 benchmark and trained the Stable Diffusion multi-modal model in just over 20 minutes. In addition to the importance of hardware itself, chip manufacturers are increasingly invested in developing their own software as well, including software that accelerates the development of chips. [Intel Gaudi AI Accelerator Gains 2x Performance Leap on GPT-3 with...] Explore more of our coverage of: Intel, Gaudi2, MLPerf Benchmark. Share this story by email

That’s it! More AI news on Monday!