Home Artificial Intelligence DAI#28 – Copycats, tech debt, and super-woke AI

DAI#28 – Copycats, tech debt, and super-woke AI

by Dimitrios Koutrotsios
DAI#8 - AI gets inside your head and resurrects Johnny Cash

Welcome to this week’s roundup of AI news optimized for human audiences.

This week AI tried way too hard to be woke.

New AI models are popping up everywhere and they want your data.

And an $800m film studio project got canned because of Sora.

Let’s dig in.

Oh, what a tangled web we weave…

As generative AI systems are cobbled together and become increasingly inscrutable, they’re building up a pile of “technical debt.” New studies explain how patchwork AI development is kicking the can down a road that inevitably ends in disaster, and fixing AI hallucinations is impossible.

Google’s Gemini went into woke overdrive this week as dramatic evidence of the impact technical debt can have. Google pulled Gemini’s image generator feature after it was called out for generating absurd and inaccurate images of historical figures.

Apparently, Gemini really doesn’t like drawing white people.

Gemini doesn’t like drawing white people. Source: X

Google continued to add to its collection of confusing AI product names when it announced Gemma, a suite of lightweight open LLMs that are optimized to run locally on NVIDIA RTX PCs.

These small LLMs are actually pretty impressive and being able to run them locally could be a game changer.

The imitation game

OpenAI blasted the New York Times, claiming the paper paid someone to ‘hack’ ChatGPT to create the evidence for its copyright lawsuit. OpenAI’s lawyers are basically saying, ‘Yes, ChatGPT can output verbatim NYT content, but only if you try really hard.’

It’s not just the NYT that doesn’t want their content used as AI training fodder. A new study shows that major news sites are increasingly blocking AI web crawlers. Is it out of journalistic principle, or are they waiting for cash offers?

User-generated content is becoming a big target for AI companies, with OpenAI and MidJourney now looking to buy your WordPress and Tumblr data to train their models.

The company behind the deal is throwing words like “attribution, opt-outs, and control” around, but look at how that worked out for the NYT.

Arguing that OpenAI’s models don’t straight up copy stuff from training data just got a little harder. A Copyleaks report found that around 60% of GPT-3.5 outputs are plagiarized.

So this probably means that around 60% of students’ homework is plagiarized too. Does anyone even go to a library for the books anymore?

via GIPHY

Big money

The amount of money being invested in AI is eyewatering and Nvidia seems to be a firm favorite with investors. The company’s recent share price surge makes it bigger than Google or Amazon with a valuation of $2 trillion.

There’s big money flowing into projects developing advanced, general-purpose humanoid robots. Robotics company Figure AI is closing in on $675m funding round with some big names in AI supporting it.

It’s going to take a lot of power to keep those GPUs and robots running. Clean energy from nuclear fusion may be a reality sooner than we thought, and AI is playing an important part in making that happen.

AI celebrities

There’s a sure-fire way to become a millionaire in just 3 months, and British celebrity Ben Fogle wants to show you how before the banks shut down this one simple trick.

Except this version of Ben Fogle is just one of many AI deep fakes that scammers have been posting on Facebook. How do these ads even make it past Meta’s ad screening process?

Some real celebrities are feeling the AI pinch too. Actor and filmmaker Tyler Perry is pumping the brakes on his $800m studio expansion. After seeing how easily Sora generates video content, he’s realising that investing in studios, sets, and filming equipment probably isn’t such a great idea.

If Perry saw how easily ElevenLabs added AI-generated audio to the Sora videos, then his project is definitely not going ahead.

The real stars of the AI show are some of the super-smart women developing the tech. If you’re tired of constantly only hearing about the men behind AI, check out this list of the top women making big impacts in AI.

Eye see

Researchers wanted to see if GPT-4 could be useful in responding to glaucoma and retina patient questions and cases. How do you think it performed compared to professional human eye specialists? You may want to reconsider pursuing a career in ophthalmology.

Another model making AI headlines this week came from French startup Mistral AI. The company is not even a year old yet and it just released a new chatbot and model that gets pretty close to GPT-4 performance. The API pricing could be a real threat to OpenAI.

What we really want is an AI assistant that can interact with multiple software tools to do the jobs we don’t want to do. Meta and the UCSD got us a step closer to that with ToolVerifier which helps LLMs choose the right tools when you ask them to perform a task.

In other news…

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

  • Stability AI released Stable Diffusion 3 and it’s really good at adding text into the images it generates.
  • Groq (not Grok) has a special hardware platform that handles AI model inference a lot faster than other cloud platforms.
  • Microsoft’s deal with French startup Mistral AI comes under EU scrutiny.
  • The people behind the fake Biden robocall are a Democratic consultant working for a rival candidate and a street magician.
  • Samsung’s new High Bandwidth Memory is 50% better than its previous chip and could speed up AI training by 34%.
  • Phrasing your prompts as a scene from Star Trek can improve an LLM’s response.

And that’s a wrap.

There’s a lot of hustle in Google’s AI projects, but boy have they made headlines for all the wrong reasons lately. Did you get any super-woke images when you used Gemini? Could you tell Gemini from Gemma and Pro from Ultra?

You may have aligned with Team-OpenAI in the NYT saga, but how do you feel now that AI companies want your Reddit, WordPress, and Tumblr content? Take my data and make something good, or hands off?

If the industry can resolve the social awareness and copyright issues, can you imagine what they could achieve with renewed focus? I’ve upped my wishlist to include nuclear fusion power and fully immersive VR worlds with multi-sensory feedback. Easy.

Tell us your boldest 2024 AI predictions and please send us links to any juicy AI stories we may have missed.



Source Link

Related Posts

Leave a Comment