DAI#54 – Ethereal strawberries, Apple intelligence, and AI unleashed

Welcome to our weekly roundup of handmade AI news.

This week, OpenAI gave us more vaporware to look forward to.

We might not need data centers to train models.

And big tech writes anti-establishment letters to ‘save’ AI.

Let’s dig in.

Strawberry rumors

You may have seen strawberries feature disproportionately in AI discussions this week. AI models still can’t tell you how many Rs the word “strawberry” contains and Sam Altman has been posting random pics of them on X.

Why is the AI world ranting on about strawberries? In frustratingly consistent fashion, OpenAI has leaked whispers and rumors of a new model called Strawberry and another called Orion.

There’s been talk that they showed these to the Feds so who knows if we’ll get to use them or see them tossed into the “too dangerous” pile along with the GPT-4o voice assistant.

ChatGPT Plus users will have to wait a bit longer for features like Sora, Voice, and SearchGPT… pic.twitter.com/WRlQDEVZKb

— AshutoshShrivastava (@ai_for_success) August 28, 2024

Our first taste of voice interaction with ChatGPT may actually come from Apple. The iPhone 16 and Apple Intelligence rollout schedule could soon deliver a vastly improved Siri powered by ChatGPT.

Thankfully, Google has been releasing actual products and not just AI vaporware. The company unveiled 3 new experimental Gemini models this week including a tiny one that punches above its weight.

Game changer

To train an AI model, you need a huge, expensive data center full of NVIDIA hardware, right? Maybe not. An AI training optimization algorithm called DisTrO could completely disrupt how AI models are trained.

Imagine the computing power of the world’s internet-connected PCs and GPUs training a model kind of like how Bitcoin mining works. DisTrO could make that happen.

AI on the job

Amazon Web Services (AWS) CEO Matt Garman told his employees that AI could replace developers within the next 24 months. This reality prompted an interesting discussion of changing workstreams and the value of creativity.

The Chinese are also fans of AWS. With a little creativity and AWS, they’ve been able to access NVIDIA chips on the US export ban list.

While Garman and others try to reassure their human workers, Swedish FinTech company Klarna is embracing AI as it plans to slash its headcount by 50%. Klarna’s CEO says AI will see the company double its revenue per employee.

Will laws break AI?

Several AI safety regulations have recently kicked in, or are about to be signed into law, and not everyone is a fan.

OpenAI has been critical of California’s AI safety bill which could be signed into law in the next few days. Former OpenAI employees turned whistleblowers wrote a scathing response to OpenAI’s opposition to the AI safety bill and called ‘Liar!’ on Sam Altman’s claims that he’s pro-AI safety.

Is SB 1047 a bad idea or is Sam Altman just being a crybaby?

https://t.co/BI85vrqAaT pic.twitter.com/TGphg58bRP

— Cicero (@PauseusMaximus) August 26, 2024

Meta CEO Mark Zuckerberg and Spotify CEO Daniel Ek aren’t fans of the EU AI Act. They wrote a letter to EU regulators criticizing the regulations that kicked in this month. The short version of the letter is: ‘AI could give us cool stuff but you’re stopping that from happening.’

In other news…

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

• Workers at Google DeepMind are pushing the company to drop military contracts.
• Police officers are starting to use AI chatbots to write crime reports. Will they hold up in court?
• Anthropic publishes the ‘system prompts’ that make Claude tick.
• NVIDIA Blackwell debuts at MLPerf and shatters AI performance records.
• New in Gemini: Custom Gems and improved image generation with Imagen 3.
• Exodus at OpenAI: Nearly half of AGI safety staffers have left, says former researcher.
• Google’s new diffusion model GameNGen generates a classic video game in real-time.

This is absolutely insane. A model is generating this game, in real-time, as the user plays.

Map the progress/trajectory of most other AI modalities onto this, and in a few years, we’ll have AAA-quality games being generated in real time. pic.twitter.com/RQi8396j8u

— Matt Shumer (@mattshumer_) August 28, 2024

And that’s a wrap.

Do you think we’ll get to try out Strawberry or Orion anytime soon? I keep telling myself that I’m done waiting in expectation for GPT-5 but who am I kidding?

The DisTrO story could be huge. Would you hook your PC up to be part of the ‘hive mind’ to train a huge open-source model? AI alignment and guardrails will be tough to enforce once we get decentralized AI.

We’ll be keeping a close eye on California’s SB 1047 bill over the next few days. If it passes, we could see a bunch of AI folks joining Elon Musk as he heads to Texas. Or will OpenAI and Meta just suck it up and stay put?

Let us know what you think, follow us on X, and please send us links to interesting AI research or news we may have missed.

The post DAI#54 – Ethereal strawberries, Apple intelligence, and AI unleashed appeared first on DailyAI.