DAI#47 – AI writes jokes, flags fibs, and beats bugs

Welcome to our weekly roundup of human-assembled AI news. 

This week AI got into writing jokes and spotting fibs.

A Chinese real-time model beat GPT-4o.

And OpenAI opened up when it should have closed.

Let’s dig in.

OpenAI hacked

OpenAI has been criticized for being very much closed, despite its name. This week we learned the company’s servers were wide open for all the wrong reasons.

OpenAI suffered a data breach as a hacker breached its defenses. As the global race for AI hots up we’re likely to see more attempts like this.

Microsoft is OpenAI’s biggest investor but this week it surprisingly withdrew from its observer role on OpenAI’s board and Apple also declined a seat at the table.

Microsoft says this is a natural evolution of its relationship with OpenAI but could something else be going on here?

We’re still waiting for GPT-4o’s voice assistant to go from demo to publicly available product. While OpenAI wrings its hands and talks about safety issues, French AI lab Kyutai released its AI voice assistant Moshi.

It’s buggy and the demo had a bunch of glitches, but at least Moshi is out there for people to use.

Your move Sam. Kling is already killing it so you may as well release Sora too while you’re at it.

Definitely created using some AI, but not sure which one. Probably Kling. pic.twitter.com/1SNW9WslJi

— AshutoshShrivastava (@ai_for_success) July 11, 2024

You must be joking

Can AI be funny? A new study compared jokes written by humans to those written by GPT-3.5.

They also compared satirical headlines written by professional comedic writers at The Onion to those written by AI. Guess whose jokes were judged as funniest in a blind test.

If there was a device that told you when someone was lying would you use it? Researchers made an AI lie detector that is much better at detecting lies than humans are.

It all sounds like a good idea until you explore just how socially disruptive this tech could be.

New models

Could a tiny AI model beat GPT-4? Salesforce has challenged trends in AI with the tiny yet mighty xLAM-1B and 7B models.

Agentic AI needs to turn users’ natural language requests into specific API or function calls. These tiny models are hundreds of times smaller than GPT-4 and Gemini 1.5 Pro but outperform them on the Berkeley Function-Calling Leaderboard.

Chinese tech company SenseTime released its multimodal SenseNova 5.5 model and said it outperforms GPT-4o. The real-time interactive version of the model displayed the same voice functionality we saw in the GPT-4o demo, albeit in Mandarin.

SenseTime and other Chinese AI companies are now dishing out free API tokens to woo new customers as OpenAI shuts down API access in China.

Anthropic impressed us with Claude Sonnet 3.5 last month and its platform just got another upgrade. The Anthropic Console and Claude got some exciting new prompt and Artifacts capabilities.

AI bad behavior

Collaborative web design app Figma did a quick about-face as the CEO shut down its AI Make Design feature.

The app designer creates polished designs but some users said they look suspiciously familiar. Now where have I seen that app design before….?

If adults can’t figure out how to create and use AI ethically, should we be surprised when kids get it wrong?

A Spanish court handed down sentences to 15 children for creating AI-generated explicit material. It’s easy to blame the kids but the situation highlights how AI’s impact on children has largely been ignored.

And this kind of research doesn’t exactly inspire confidence.

https://t.co/vki4nKLPKb pic.twitter.com/zSA8dysuK2

— ControlAI (@ai_ctrl) July 8, 2024

Dr. AI

Is it possible to tell if a person is likely to develop Alzheimer’s in the next 6 years? A new AI system can successfully predict the early onset of Alzheimer’s disease using speech analysis.

The accuracy of the prediction beats traditional and non-invasive tests and could see patients get earlier treatment.

In a medical breakthrough, researchers used AI to identify drug-resistant infections like typhoid before attempting to treat them with antibiotics.

The speed and accuracy of the AI system allow doctors to make faster diagnoses so they can prescribe the most effective antibiotics sooner. This could be a huge boost for reducing the spread of drug-resistant bacteria.

In other news…

Here are some other clickworthy AI stories we enjoyed this week:

• Microsoft teases its advanced zero-shot text-to-speech synthesis model called VALL-E-2 but says it’s too dangerous to release.
• Researchers tested whether AI models are self-aware and if they know when they’re deployed or being tested.
• Shanghai publishes the first humanoid robot guidelines to safeguard human dignity and safety.
• Agentic AI: What’s the current status and what’s next to make AI agents truly useful?
• Wimbledon’s “Catch Me Up” AI-generated stories and analysis tool keeps getting its tennis facts wrong.
• Abandoned nuclear power facilities like Three Mile Island might be recommissioned to power AI needs.

And that’s a wrap.

OpenAI has been very quiet even as other AI models are hot on its GPT-4o heels. Fingers crossed that we’ll have some OpenAI news next week other than lawsuits and hacks.

Do you think AI will ever come to grips with humor? Writing a joke is one thing but I’m not sure an algorithm could ever create the next Seinfeld or Fawlty Towers.

Let’s hope the AI engineers focus their models on easier tasks like curing cancer and free energy for now.

Let us know what you think, share your AI attempts at humor with us on X, and keep sending us those AI news links.

The post DAI#47 – AI writes jokes, flags fibs, and beats bugs appeared first on DailyAI.