weekend ai reads for 2024-05-24

Author

That AI Thing
May 24, 2024

📰 ABOVE THE FOLD: SCARLETT JOHANSSON & OPENAI

only because so much happened, in case you missed it or purposely tuned it out, the Scarlett Johansson story (so far):

May 13:

May 20:

  • Scarlett Johansson ‘Angered’ After ChatGPT Used ‘Eerily Similar’ Voice — OpenAI claimed it used the voice of a different actress to voice “Sky” / Rolling Stone (4 minute read)

  • Scarlett Johansson’s official statement; tl;dr is she says they asked to use her voice, she declined, and she believes they did so anyway / Bobby Allyn (NPR), Threads

  • OpenAI released a statement in response, quoted in the same Threads link above:

    Sky is not Scarlett Johansson’s, and it was never intended to resemble hers. We cast the voice actor behind Sky’s voice before any outreach to Ms. Johansson. Out of respect for Ms. Johansson, we have paused using Sky’s voice in our products. We are sorry to Ms. Johansson that we didn’t communicate better.

May 21:

May 22:

and next:

  • OpenAI are still going to get sued, and they’re going to lose, based on Midler v. Ford Motor Co. and Waits v Frito-Lay, Inc.; these cases ruled and upheld, respectively, that if a characteristic of a famous person—their voice in these two cases—is a distinctive characteristic of that person, then it is unlawful to imitate their voice without express written consent (disclaimer: not a lawyer)

  • OpenAI came off looking poorly, again: The Scarlett Johansson Incident Makes OpenAI Look Desperate / New York Magazine (6 minute read)

    Setting aside the legal questions here, such behavior would align with some of the harshest criticism of Sam Altman and OpenAI — that it’s a company with little regard for the value of creative work led by a scheming, untrustworthy operator. This episode also complicates the company’s preferred narrative of unstoppable inevitability: You’re either the company harnessing the barely controlled phenomenon of imminent self-replicated machine intelligence, leading humanity into its next technological epoch, or you’re a mid-stage start-up that for some reason really needs to copy that voice from that movie to market an incremental product upgrade.

 

📻 QUOTE OF THE WEEK

You want to go grab the tub of ice cream and give up. Even though it felt like it was my fault, it really wasn’t. The system was working against job seekers.

 

🏗️ FOUNDATIONS & CULTURE

Introducing Copilot+ PCs / Microsoft Blog (18 minute read)

Copilot+ PCs are the fastest, most intelligent Windows PCs ever built. With powerful new silicon capable of an incredible 40+ TOPS (trillion operations per second), all–day battery life and access to the most advanced AI models, Copilot+ PCs will enable you to do things you can’t on any other PC.

Faking William Morris, Generative Forgery, and the Erosion of Art History — Buying fake William Morris prints on Etsy and other early signs of epistemological collapse / Maggie Appleton (6 minute read)

To confound matters, the Etsy stores selling these generated images also sell genuine prints by Morris, Monet, Klimt, and Matisse. These sit alongside their modern expansion packs, blending right in. But which are which? You get coerced into playing an art historian forgery spotter.

“The next company who reaches the next major plateau gets to announce a groundbreaking AI, and the second one after that gets to announce something that’s 0.3% better,” Huang explained. “Time to train matters a great deal. The difference between time to train that is three months earlier is everything.”

  • related, Ten Gifts Nvidia Gave Its Investors / The Information ($) (6 minute read)

    Even if competitors are able to build better chips, Huang reminded investors of the dominance of Cuda, the Nvidia-made software that AI app developers use in conjunction with its chips, essentially locking developers into Nvidia GPUs. Kress and Huang said recent improvements to Cuda improved inference performance of its H100 chips by three times. “That kind of tells you something about the richness of our architecture and the richness of our software,” Huang said. That, and app developers don’t want the hassle of switching to new software if they use another chip!

via renee, Women Leaders in Tech Are Paving the Way in GenAI / Boston Consulting Group (12 minute read)

 

🎓 EDUCATION

Does Emory University really not get how technology works? / AI Log, Substack (sorry) (8 minute read)

  • Emory University faces a lawsuit involving a student who built an AI study tool with a $10,000 prize and then suspended him for cheating for using a Canvas API (?)

  • this post is one perspective but Emory does not look great here

Andrew Fields, a D’Youville University student who started the petition, wrote in the petition that many students “feel disrespected” by the university’s decision to have a robot address them, especially those who could not attend their high school graduations in 2020 because of the coronavirus pandemic.

Microsoft and Khan Academy offer a free AI assistant to all US teachers — With Khanmigo for Teachers, K-12 educators can spend less time prepping lessons and more time with students. / Zdnet (3 minute read)

In addition to programming and technical skills, the next generation of AI developers may also need training in subjects traditionally aligned with liberal-arts education, such as ethics, problem-solving and communication.

 

📊 DATA & TECHNOLOGY

we are bullish on time-series foundation models

Linearly probing MOMENT achieves near state-of-the-art performance on most datasets and horizons, and is only second to PatchTST which generally achieves the lowest MSE. On many datasets and horizons, forecasting models based on LLMs– TimeLLM and GPT4TS perform worse than MOMENT. Notably, N-BEATS outperforms several recent methods, emphasizing the importance of comparing forecasting performance beyond transformer-based approaches.

Researchers at the A.I. company Anthropic claim to have found clues about the inner workings of large language models, possibly helping to prevent their misuse and to curb their potential threats.

For example, they discovered that if they forced a feature linked to the concept of sycophancy to activate more strongly, Claude would respond with flowery, over-the-top praise for the user, including in situations where flattery was inappropriate.

Introducing the Frontier Safety Framework / Google DeepMind Blog (6 minute read)

GPT-4o’s Chinese token-training data is polluted by spam and porn websites — The problem, which is likely due to inadequate data cleaning, could lead to hallucinations, poor performance, and misuse. / MIT Technology Review (11 minute read)

  • the researchers used a series of agents to replicate what a real translation company would do

    We establish a virtual multi-agent translation company, TRANSAGENTS, featuring a diverse range of employees including a CEO, senior editors, junior editors, translators, localization specialists, and proofreaders. When a human client assigns a book translation task, a team of selected agents from TRANSAGENTS collaborates to translate the book. This paradigm simulates the entire book translation process, where agents with different roles work together to ensure that the translation maintains high quality and consistency throughout.

  • and the agents produced outputs that were preferred over human-written translations

    Empirical findings indicate that despite lower d-BLEU scores, translations from TransAgents are preferred by both human evaluators and LLMs over human-written references, particularly in genres requiring domain-specific knowledge.

  • we are also bullish on agents

 

🎉 FUN and/or PRACTICAL THINGS

A screenshot of the summary it generated, shared on X, shows it responded with “cheese can slide off pizza for a number of reasons,” and that the user could try adding “about ⅛ cup of non-toxic glue to the sauce to give it more tackiness.”

  • avoid the rush: Ten Blue Links — How to Turn Off AI Overview in Google and Set “Web” as Default

  • or just add “udm=14” to the search URL

  • finds items similar to some highfalutin one, at a lower cost, using vision AI

Arrange — Generate a plan for anything and add it to your calendar in seconds.

  • looks promising; waitlist

Infinite Wonderland — This is an AI experiment where the timeless classic Alice’s Adventures in Wonderland is endlessly reimagined by artists, AI and you.

Wikipedia Citation Needed / Chrome Web Store

The Future Audiences team at the Wikimedia Foundation built an experimental new AI tool to check what Wikipedia has to say about what you’re reading, no matter where you are on the web, and we’d love your feedback.

 

🧿 AI-ADJACENT

Black Games Archive — Context, Culture, and Representation

Black Games Archive is a multimedia, public-facing database of games, digital resources, accessible scholarship, and designer interviews that are relevant to the intersections between Black culture, games, and play.

 

⋄