🪩 SPECIAL: The State of AI Report 2024 is out!

AI-HOI AInauts,

Happy Monday! Over the weekend, we took a really deep dive into the new State of AI report - this comprehensive analysis of the entire AI universe is super exciting. We have summarized the highlights of the 212-slide presentation for you. Today, we take a bird's eye view, and on Thursday there will be more practical tips.

• 🪩 State of AI Report 2024: Trends revealed - what will shape the future

• 🔮 Looking into the crystal ball - let's start with the predictions!

• 🧠 The model revolution: bigger, faster, better

• 🎮 Multimodal magic: text was yesterday, video is exploding

• 💰 AI economy: A billion-dollar market with NVIDIA at the forefront

• 🤖 AI-supported hardware is making progress

• 🔬 Scientific breakthroughs and Nobel prizes

• 💥 I safety: from "doom and gloom" to "let's make money!"

• 🔒 Regulation: politicians can't keep up ...

• ⏳ Conclusion: the future is going to be very weird

Ready? Let's go!

🪩 State of AI Report 2024: Trends revealed - What will shape the future?

2024 was a year of superlatives for artificial intelligence - but at the same time, we are still at the very beginning. It's crazy, isn't it?

So what are the developments that will shape our future? The annual State of AI report has the best answers - you can find the full report at stateof.ai (slides with sources here). A super exciting framework, with everything you need to know about the current state of research, industry, security and politics.

🔮 A look into the crystal ball - let's start with the predictions!

Every year, the creators of the report provide us with the most important forecasts. These come at the end of the report, but we like to start with them right away.

The creators of the report believe we can expect the following next year:

1. An investment of over 10 billion dollars by another nation state in a major US AI lab triggers a national security review.

2. A no-code app or website goes viral (e.g. top 100 in an app store).

3. AI research labs implement significant changes to their data collection practices after being taken to court.

4. The EU AI Act is less strict in practice than expected after lawmakers fear it has gone too far.

5. An open source alternative to OpenAI o1 outperforms it in benchmarks.

6. Challengers fail to threaten NVIDIA's market position.

7. Investment in humanoid robots declines as product-market fit is difficult.

8. Apple's research accelerates momentum for on-device personal AI.

9. An AI-generated research paper is accepted at an ML conference.

10. A game based on interaction with AI elements achieves a breakthrough.

P.S.: These were last year's predictions. It was overestimated how quickly some research would prove successful, but the industry and policy predictions were on point.

🧠 The model revolution: bigger, faster, better

At first, it looked like OpenAI's competitors (such as Anthropic's Claude 3.5 Sonnet) had caught up with the market leader, as benchmarks were very close over the course of the year.

But then came the o1 model, which set a new standard with its "chain-of-thought" approach and gave AI more time to think before answering.

However, this leap in intelligence also came at a price:

• o1 requires 26x more computing power per token than GPT-4!

• Result? 83.83% score for AIME 2024 (complex math) vs. 13.4% for GPT-4o

• 1M input tokens cost $15, 1M output tokens cost $60, around 3-4x more expensive than GPT-4o

For most use cases, however, a less intelligent model is perfectly adequate. There is a massive drop in prices for all providers. Is this due to efficiency gains or are we benefiting from a price war that is primarily about market share and attention??

As model performance converges, vendors are investing more in products to differentiate themselves from each other. Claude Artifacts has established itself as a favorite for development, while OpenAI is trying to mimic this with its new Canvas feature.

The gap between open source and proprietary models also seems to be closing. Meta's open source champion Llama 3.1 405B was trained with 15T tokens and closes the gap to proprietary models. A few weeks ago, Llama 3.2 even went multimodal, and smaller models are also improving fast.

Takeaway: o1 is a massive leap forward, but the high cost will initially slow down widespread adoption. The other models may not be as good, but they are often completely adequate and are keep getting cheaper.

🎮 Multimodal magic: text was yesterday, video is exploding

2024 was also the year in which multimodal models really took off. Meta's latest product MovieGen not only generates video, but also the sound to go with it!

And there are many other highlights in the video area:

• StabilityAI's Stable Video 3D generates 3D objects from text

• Google's Veo and OpenAI's Sora (not yet public) promise minutes of high-quality video generation

• Chinese providers are also at the forefront here, such as KLING and Minimax

• The most important players have filled their coffers well: Runway has raised $236.5M in funding, Pika stands at $135M and Luma at $67.3M.

Takeaway: Video AI will explode in the next 12-18 months. A flood of AI-generated short films, commercials and maybe even the first AI co-produced feature films await us!

💰 AI economy: a billion-dollar market with NVIDIA in the lead

There is a two-tier society among AI startups: the biggest players are generating real revenues that match their rising valuations, while other hyped startups with billion-dollar valuations seem to be driven more by "vibes".

Meta has found a way to drive the share price back up: defer investment in the metaverse and instead invest hard in open source AI. The Llama models have been downloaded over 440 million times on HuggingFace!

However, OpenAI, with its billions in revenue, still hasn't developed a profitable business plan... and with training consuming hundreds of millions, if not billions of dollars, there's no path to profitability in sight anytime soon.

In terms of hardware, NVIDIA continues to reign unchallenged, despite established competitors pouring big money into their ecosystems and new challengers appearing on the scene. But NVIDIA's dominance is simply breathtaking…

And then there are the huge compute requirements... Big tech may be trying to keep emissions in check, but this is where desire and reality clash violently. The targets for 2030, which seemed wise in 2019, will all be missed according to current trends.

Many experts believe AI will help us out of the climate catastrophe by accelerating the development of carbon-capture technologies and producing tons of clean energy, so it's wrong to slam on the brakes now ...

Takeaway: If the $6 billion invested in AI chip startups since 2016 had been put into NVIDIA stock instead, it would have turned into $120 billion today!

🤖 AI-supported hardware is making progress

A lot has also happened in the field of robotics! After OpenAI disbanded its robotics research team in 2021, it now supports other companies equipping their humanoids with OpenAI’s LLMs.

Google is making robots smarter and more creative, and Hugging Face is bringing DIY robotics into our living rooms with LeRobot. And do you remember Boston Dynamics' commercial? He's even learning to stack boxes now!

After years of broken promises, self-driving car companies are making huge strides - Waymo is driving over 100,000 paid rides a week in the US, and Tesla unveiled its new Cybercab last week.

In the world of consumer hardware, Meta's AI-powered Ray-Bans have proved a surprising success - a stylish combination of style and clear functionality. Other devices? Not really, the Humane Pin and Rabbit R1 were flops, and the Apple Vision Pro is still waiting for its breakthrough.

Takeaway: AI hardware is only successful when technology meets usability and design - see Meta Ray-Ban vs. Humane pin. If you don't offer clear added value, you will quickly lose out!

🔬 Scientific breakthroughs and Nobel Prizes

AI scientists have unexpectedly won the Nobel Prize in physics and chemistry this year.

The Nobel Prize for Chemistry goes to AI pioneers Hassabis, Baker & Jumper, who are being honored for their use of AI to predict and create proteins. Geoffrey Hinton and John Hopfield have been awarded the Nobel Prize in Physics.

Takeaway: In the next 5 years, there will be breakthroughs in areas such as materials science, drug development and climate modeling that were previously considered unsolvable.

💥 AI security: from "doom and gloom" to "let's make money!"

Last year, the topic of existential risks was the big talking point. But now that companies are focusing on commercialization, this has faded into the background.

Security concerns? Yes, they still exist, but "time-to-market" has priority. Researchers have tried to tighten the security screws, but the jailbreakers around Pliny the Prompter (our interview here) have quickly destroyed all these approaches.

While technical exploits are often discussed, most of the real damage is caused by the use of standard products to create deepfakes or phishing campaigns.

Governments are building capacity for AI security and talking about potential risks, but so far, there are no sustainable fixes to prevent jailbreaks and hacks. How long until something really bad happens?

And not everyone is a winner. There are more and more prominent critical voices putting Big Tech's "fair use of content" arguments under considerable legal pressure.

Takeaway: AI security remains a hot topic, but as AI companies focus on commercialization, the focus on existential risks takes a back seat. Real solutions are lacking as governments desperately try to catch up.

🔒 Regulation: politicians are not keeping up ...

While AI is developing explosively, politics is running behind. Yes, progress is being made somewhere between the USA and the EU, but global regulation? Not really.

On the policy front, the US government has formalized the non-binding assurances it has received from major AI labs via Executive Order 14110, as there was no consensus on potential legislation. While this is a fast track, the order can also be reversed at any time by the next government...

In Europe, regulation is progressing briskly with the EU AI Act, causing US labs to struggle to keep up. As a result, new features are being abandoned or significantly delayed for Europeans.

And the regulators are now also scrutinizing the many connections between the most important AI players that have established themselves under the acronym GAMMAN: Google, Amazon, Microsoft, Meta, Apple & NVIDIA …

As providers need more fresh data to train their models, many are turning to their own users - such as Meta, LinkedIn, Alphabet and X. The saying "If it's free, then you are the product!" comes into play here once again ...

The regulatory authorities are not happy about this and are forcing either stops or the introduction of clear opt-out options.

Takeaway: While AI technology is racing ahead, regulators are lagging behind. The USA is relying on executive orders, while Europe is applying pressure and slowing down innovation with the EU AI Act. One thing is clear: the big players are continuing to collect data, safety is a sideshow and regulators are desperately trying to maintain control …

⏳ Conclusion: the future is going to be very weird

The AI Report 2024 gives us a clear indication: the next two years are going to be wild. We're certainly not telling you anything new.

We are at the beginning of a revolution in which AI is not only transforming our technology, but also our society. The potential is huge, but so are the risks. The next few years will show just how fundamental the changes will be.

We made it! But no need to be sad. The AInauts will be back soon, with new food for thought.

Reto & Fabian from the AInauts

P.S.: Follow us on social media - that motivates us to keep going 😁!
Twitter, LinkedIn, Facebook, Insta, YouTube, TikTok

Your feedback is essential for us. We read EVERY comment and feedback, just respond to this email. Tell us what was (not) good and what is interesting for YOU.