Welcome to this week’s news roundup curated especially for non-silicon-based fans of AI.
This week AI made a mayor speak in tongues.
A two-day summit and $10m hopes to solve AI safety.
And, ChatGPT will do whatever you ask if you show it the right pictures.
Don’t call us, AI will call you
If you thought robocalls were bad before imagine what they could be like with a dash of AI. The FCC is trying to work out what to do about it.
Meanwhile, the mayor of New York City just used deepfake tech to robocall city residents in a bunch of languages he doesn’t speak. Bureaucratic efficiency at last, or just plain creepy?
The Idaho police force is keeping people there ‘safer’ by deploying AI tech that can unlock and grab data from your phone remotely. They won’t say exactly how it works, but they pinky swear it’s legal.
Deploying AI in the public sector was supposed to make queues shorter, and help people access services more efficiently. Evidence of AI misuse, bias, and dysfunctionality in the UK public sector is the tip of the iceberg.
AI: ‘I see here it says you’re Bulgarian. You must be committing fraud. Denied!’
The UK government released the schedule for the AI Safety Summit to be held on 1-2 November.
They also released a report that will serve as a discussion document for the two-day summit. The report lists potential AI risks and doomsday scenarios that you’d expect from fringe conspiracy theorists, so it’s eye-opening to see it in a government document.
To help do their bit for the cause, Big Tech AI companies launched the AI Safety Fund and had a whip-round to put a paltry $10m into the kitty.
DeepMind’s CEO said that AI risks are comparable to the climate crisis. Referencing the Chernobyl nuclear meltdown he said “We don’t yet know what an ‘AI disaster’ might look like.”
A two-day summit and $10m is the ant on the track trying to stop the AI freight train.
Show me the money
The tech industry has come up with some great AI products but can they actually make any money from them? The amount of money it costs to run ChatGPT or Co-Pilot each day is eyewatering.
Investors obviously think there’s money to be made. Apple just upped its annual investment in generative AI to $1 billion as it tries to play catchup after a slow start in the AI race.
Its new iPhone is a little light on AI features compared to Google’s Pixel 8 phones, which come with advanced AI image editing features.
Did you invest in Nvidia before its share price quadrupled over the last year? Well if you want to celebrate with a party at an Airbnb rental then AI might put a damper on those plans.
Mo modes, mo problems
Multimodal and generative AI capability continues to improve exponentially. AI can now create a 3D environment from a simple text prompt, but it is also being used to create deepfake images of child sex abuse.
GPT-4’s vision functionality is doing some amazing things with image analysis, but this capability also makes it vulnerable to attack. When it comes to jailbreaking the model, a picture really is worth a thousand words.
Are you an artist or photographer who’s tired of having AI models rip your work off? Well, now you can use a data poisoning tool called Nightshade to confuse AI image generators.
It has its share of challenges, but GPT-4V really is amazing. Check out this example and then tell me again how ChatGPT is just good at predicting the next word.
AI on the edge
Running larger AI models on battery-powered devices could happen sooner than you think. IBM announced its prototype NorthPole chip which promises much faster low-energy AI processing. The company’s novel approach to chip architecture could be a game changer.
Running AI on the edge will be a big boost for our eventual robot overlords. Meta announced Habitat 3.0, an AI robot training platform that teaches robots how to help around the house thousands of times faster than using real-world training.
Let the chips fall where they may
While it was only supposed to come into effect in November, the US government announced that its AI hardware export bans are to come into immediate effect. Expect a healthy black market trade in Nvidia chips to continue.
China has been stockpiling AI chips and is throwing money at AI development. A Chinese startup called Zhipu secured 2.5 billion yuan ($340 million) in funding. There are some surprising names among its investors.
Did ChatGPT make some stuff up when you asked it to describe an image? Chinese research could help solve that hallucination problem. Woodpecker is a surprisingly simple open-source solution that could make multimodal models like GPT-4V a lot better.
AI speeds up research
Research papers need to be peer-reviewed, but the people who do that are often too busy writing their own papers. ChatGPT is surprisingly good at doing yet another human’s job.
AI is helping doctors find out what’s wrong with you really fast too. Researchers were able to use AI to diagnose diabetes by analyzing a 6 to 10-second voice recording of a patient.
All of the great things AI models can do are fueled by data. New research into AI datasets reveals systemic ethical and legal issues that are the flip side of the AI coin.
In other news…
There were loads of other clickworthy AI news stories this week. Here are some of our favorites:
- John Stewart’s show on Apple TV got canceled because of his comments on AI and China.
- LLMs are surprisingly bad at math. These researchers took Llama 2 and made LLEMMA, a math-focused LLM.
- We’re all waiting for GPT-5 to be released but Bill Gates says it won’t be much better than GPT-4. Got a touch of OpenAI envy there Bill?
- AI isn’t great for the environment but it can make funny deepfakes of environmental activists.
The crazy thing is AI is the worst it will ever be right now… 🤨#HowDareYou #GretaThunberg #AI #edited #ClimateCrisis pic.twitter.com/mkEz25E0gp
— Mrgunsngear (@Mrgunsngear) October 23, 2023
And that’s a wrap.
If you live in NYC we’d love to know if you got a robocall from the Mayor and what it sounded like.
Would you be ok with an AI robot doing your household chores? Or will you be holding off on that purchase to see if the AI Safety Summit finds a way to avoid us accidentally building Skynet first?
If I had a spare $75k I think I’d take the risk and buy a Boston Dynamics Spot robodog.
If you think we missed a great AI news story then please let us know.
