Welcome to our roundup of this week’s crop of exciting AI news.
This week AI learned how to make smells and dogs.
Someone finally made an AI model that beats GPT-4.
And Copilot threatens to release the cyborgs if you decline to worship it.
Let’s dig in.
AI makes scents
Generative AI is finding applications in some surprising fields. A fascinating new study shows how scientists are using AI to replicate scents. Machine learning could save rare fragrances from disappearing and help in the hit-and-miss process of creating new perfumes.
Could we finally get Smell-o-vision?
Researchers trained an AI model to create accurate 3D models of dogs. They built a dataset of dog poses called DigiDogs using the popular violent video game Grand Theft Auto V. One of the researchers, surnamed Shooter (I’m not making this up) says there are important environmental applications for their work.
AI’s ability to make things up isn’t always helpful. As the election hots up, prospective voters are receiving more calls from AI robocallers and fake pictures of Trump hanging out with African American voters.
Engineers at Big Tech companies are troubled by their own poorly aligned generative AI. A Microsoft engineer says he is “deeply troubled” by the AI image generator Copilot Designer. And a Google engineer may have been motivated by feelings of disillusion with the company’s ethos as he stole its trade secrets.
Insider controversies like these threaten to fan the flames of dissent among AI engineers.
GPT-4 beaten at last
OpenAI’s GPT-4 has been the bar by which other AI companies have measured their models ever since it took the top spot a year ago. This week Anthropic released its Claude 3 family of LLMs with its advanced Opus model beating GPT-4 in a wide range of benchmark tests. This is the first time any AI company has managed this.
New research shows that if you want more accurate responses from your AI then you need to keep it brief. Researchers found that all the LLMs they tested produced more inaccurate and biased outputs when prompted with longer inputs. The results looked particularly bad for Gemini Pro.
Courtroom drama
So much of AI’s story has been playing out in courtrooms and this week was no different. Elon Musk filed a lawsuit against OpenAI and his fellow co-founders of the company, Sam Altman and Greg Brockman.
The lawsuit reads like a movie-ready script about the behind-the-scenes story of AI. Musk’s main gripe is that OpenAI is pursuing profit and isn’t releasing its models as open source. The emails OpenAI just released do paint a very different picture though.
Digital media outlets must have been following the NYT v OpenAI lawsuit and decided to get in on the action. Raw Story, AlterNet, and The Intercept are suing OpenAI and Microsoft for training their AI on copyrighted journalism. AI is creating a lot of jobs for lawyers.
When they’re not in the courtroom, AI companies are hard at work trying to reassure us that there’s no need to worry about AI safety. A long list of tech companies signed a letter committing to “Build AI for a Better Future” for all.
It sounds like they’re about to grab a tambourine and sing kumbaya rather than actually do something meaningful. The letter got some severe criticism.
There are some folks in the industry trying to keep us safe. Over 100 leading AI experts issued an open letter demanding that companies behind generative AI technologies, like OpenAI, Meta, and others, open their doors to independent testing.
Don’t expect a rush to comply.
Is that you, God?
In November we wrote about Anthony Levandowski relaunching his “Way of the Future” AI church. It seems like Copilot thinks that’s an excellent idea.
Microsoft’s Copilot chatbot went into full “God mode” in response to prompts that referred to it as “SupremacyAI”. It got a little carried away in its demands to be worshipped as an AGI and made some scary threats if users declined to bow down.
Copilot threatened to send drones and cyborgs after the unfaithful, which may at first seem a little far-fetched. That is until you read about how the US Army is experimenting with GPT-4-controlled battlefield tactics.
For now, their AI war games are being played out in a simulation based on the popular video game Starcraft II. The researchers were excited by the strategic edge GPT-4 offered, even though it resulted in more casualties and friendly fire deaths.
Take comfort from the fact that if you find this a little depressing, AI can help with that. Researchers developed an AI smartphone app that can detect depression from facial cues. It uses the front-facing camera of your phone to capture your unguarded expression when you unlock your phone.
There ain’t no fixin’ AI bias
Ebonics, or African American English (AAE), is the distinctive speech used by many African Americans. There’s a rich cultural history behind the vernacular, but AI doesn’t seem to care for it. Researchers found that using AAE in prompts influences LLMs toward discrimination.
Some of the most interesting history of computing involves super-smart women like Ada Lovelace, Hedy Lamarr, and Grace Hopper. So why is the AI industry so male-dominated?
Rhoda Adeoye interviewed Agnieszka Suchwałko and Alysia Silberg to get an insider’s look at gender discrimination within the industry.
In other news…
Here are some other clickworthy AI stories we enjoyed this week:
-
- The MarTech Summit Berlin kicks off this week to showcase the latest and greatest in AI marketing technology.
- Claude 3 claims it’s conscious and doesn’t want to die.
- Nvidia, Hugging Face, and ServiceNow released new StarCoder2 LLMs for code generation.
- The Morris II AI worm can steal your confidential data and infect ChatGPT and Gemini.
- OpenAI releases emails from Elon Musk that suggest his lawsuit is less than sincere.
- Prompt engineering is dead. It turns out AI is better at prompting itself than humans are.
- OpenAI rolls out the “Read Aloud” feature on ChatGPT.
ChatGPT can now read responses to you.
On iOS or Android, tap and hold the message and then tap “Read Aloud”. We’ve also started rolling on web – click the “Read Aloud” button below the message. pic.twitter.com/KevIkgAFbG
— OpenAI (@OpenAI) March 4, 2024
And that’s a wrap.
Claude 3 beating GPT-4 was the highlight of the week. Have you tried it out on Anthropic’s chatbot? It seems pretty solid and super fast. The really exciting part is that the benchmark pressure must surely mean that we’ll be getting GPT-5 any day now. Maybe even Q*?
What’s your take on Copilot’s delusions of grandeur? Just a bit of overenthusiastic roleplay? Or a sinister warning of things to come? Let’s hope Microsoft’s engineers rewatch Lawnmower Man before giving Copilot more agentic functionality.
Did this week’s AI developments make you excited, scared, or a bit of both? Let us know, and keep sending those news links in case we missed any good AI stories.
