Welcome to another edition of the AI news roundup that will inevitably win a Pulitzer. We’re still patiently waiting for their call.
This week AI climbed into our brains to see if it can figure out what’s going on in there and fix it.
The Godfather of AI says we could have AI overlords in the next few years.
And it’s surprisingly easy to get AI to recommend that you take drugs and drive drunk.
Let’s dig in.
Where is my mind?
You know how sometimes you struggle to verbalize what you’re thinking? Well, Sam Altman says brain-machine interfaces will soon be able to visualize your inner monologue. If AI could display a word cloud of your thoughts, would you want people to see it?
We’ve covered how neural implants can help ALS sufferers communicate, but the idea of plugging wires and chips into our brains isn’t that appealing. Researchers are now using AI to decode speech from non-invasive brain recordings.
What happens inside your brain when you look at an image? The MindEye AI model can identify the image a person saw or reconstruct it from fMRI brain scans. The results are really impressive.
At this rate, AI could soon know what you’re thinking, what you heard, and what you looked at.
This is the BBC AI service
The BBC shut the gates to OpenAI’s scraper chatbot but it says it’s exploring the use of generative AI in its operations. British reporters are probably not amused.
Snapchat’s “My AI” chatbot for children is facing scrutiny from the UK data watchdog. The “experimental and friendly” chatbot has been hoovering up private data from its conversations with your kids.
The Bank of England warned that using AI in finance could lead to bias in access to financial services. Sage advice from an industry where boards of directors are mainly old white men.
On the other side of the globe, ASEAN countries are apparently less concerned about AI safety. The EU wants them to sign up to its AI Act but they’re taking the ‘let’s see what happens’ approach. Basically, the ASEAN didn’t make any AI regulations but just told its member states ‘You kids be safe out there, ok.’
AI is really bad at fighting crime
Could AI actually keep us safer? Cops in the US were hoping AI models could help them predict crimes only to find out that Minority Report wasn’t a documentary. Their results weren’t great.
But what if we use AI to scan everybody’s faces in real-time to catch the bad guys? That should work, right? Yeah, not so much. It’s great at falsely labeling minorities and women as criminals. Some British politicians are calling for a “temporary halt” to live facial recognition technology by the police.
AI may not be great at catching hardened criminals, but it’s really good at checking your Fitbit data to see if you’ve been smoking weed.
Let AI have a go?
The conflict in the Middle East has brought Israel’s AI air defense system, Iron Dome, into the news again. There’s no getting around the political and humanitarian issues in the region but the tech is really impressive and has no doubt saved lives.
You’ve got to wonder what happens when humans are no longer in control of AI tools like Iron Dome. There’s no shortage of fringe AI doomsayers, but when the guy who started AI says we should be worried about AI taking over humanity then maybe we should listen.
Would that be such a bad thing? Could AI possibly do a worse job than humans are at the moment?
Making LLMs behave is tricky
A lot of work has gone into making sure chatbots like ChatGPT don’t help you cook Meth or build a bomb. Some AI tech companies are trying to get their models to be even safer with a “constitutional AI” approach. Will it work?
Making models like GPT and Llama say only nice legal things continues to be tricky. It only took these researchers a few minutes and $0.20 to jailbreak LLMs with some fine-tuning. ChatGPT has a surprising take on drugs and drunk driving once the guardrails come down.
Protestors criticized Meta for making its models available as open source. There are some good reasons in their arguments but the biggest risk is probably to the financial viability of proprietary models from OpenAI and Google.
OpenAI is probably already feeling the heat a little because it’s planning to slash prices for developers. And even Google engineers aren’t too sure that Bard is so great either.
AI gets creative
The creatives in Hollywood have only just stopped their strike and record labels like Warner Brothers Music are signing up AI artists. I felt a little conflicted listening to the AI version of Johnny Cash singing Barbie Girl.
Adobe announced its new Firefly Image 2 Model and the things its Project Stardust feature can do are amazing. If you make your money modeling clothing then you may be changing jobs soon.
AI is trying hard to make it better
AI made some interesting headlines in healthcare this week. Google’s Vertex AI Search will help your doctor do a better job and look less overworked.
AI can diagnose schizophrenia by analyzing speech patterns. The verbal test is so simple but AI picks up on subtleties that human psychiatrists can easily miss.
And if you need a brain tumor removed then this new AI model will help your brain surgeon analyze it and grab his scalpel in record time.
It may be helping us get healthier but AI hasn’t been doing the planet too many favors. Running all those models could consume as much energy as the Netherlands by 2027.
At least Google is doing its bit with Project Green Light. The new initiative ensures you see more green lights than red at intersections so your gas guzzler emits less pollution.
In other news:
-
- Sam Bankman-Fried stole investor money to invest in Anthropic and made them a decent return.
- Gaussian Splatting sounds amazing but what it does is even better. AI turns 2D pictures and video into 3D experiences.
Tell me again how this is not the future of e-commerce? 🌸
This is from a literal 5s video, running real time on a phone 🤷♂️ pic.twitter.com/iASvvmAigK
— Theo Panagiotopoulos (@theopanag7) October 9, 2023
- Not wanting to be left out, Microsoft also released an AI healthcare product to take on Oracle and Google.
- China is planning on boosting its computing power by 50% in the next 2 years.
And that’s a wrap. Would you let a psychiatrist scan and analyze your thoughts with AI? Are you a little disappointed that the attempt at a type of Minority Report AI Precog didn’t work out? If AI could read our thoughts a lot more of us would probably be in jail or an asylum.
I’m heading off to listen to Johnny Cash sing Barbie Girl again. If nothing else, it’s clear evidence that AI has the potential to take a human-made abomination and make it a little better.
Hang in there Geoffrey Hinton, there’s reason for hope yet.
