The 5 AI tales I am ready for in 2024 | The AI Beat

Precisely a yr in the past, I printed an AI Beat column that detailed 5 AI tales I used to be ready to cowl in 2023.

Due to what I prefer to name a relentless ‘tsunami’ of AI information within the wake of OpenAI’s November 2022 launch of ChatGPT, I didn’t have to attend lengthy: From GPT-4 and the EU AI Act to the battle for AI search, open vs. closed AI, and the starvation for coaching information and computing energy, I wrote about all of these subjects a number of occasions.

Now, as soon as once more, with New 12 months’s just a few days away, I’m looking forward to the AI narratives I feel might be central to traits in 2024:

1. OpenAI vs. Anthropic

At the start of 2023, OpenAI was driving the wave of ChatGPT, however Anthropic — a startup based in 2021 by siblings Daniela and Dario Amodei, former senior members of OpenAI, was already sizzling on its heels. Anthropic launched its ChatGPT rival, Claude, on March 14, 2023, the identical day as OpenAI launched GPT-4 in a shock announcement.

With each corporations in search of huge new funding rounds — Anthropic is reportedly in discussions for $750 million from Menlo Ventures, whereas OpenAI is reported to be in talks for a contemporary spherical of funding at a $100 billion valuation — the LLM rivalry is certain to warmth up in 2024.

2. Open supply AI performs GPT-4 catch up

After French startup Mistral AI dropped an early-December shock — a brand new open supply LLM, Mixtral8x7B — with nothing however a torrent hyperlink, it turned crystal-clear that the open supply AI group has no intention of slowing down its efforts to catch as much as its proprietary mannequin counterparts. In actual fact, Mistral CEO Arthur Mensch declared on French nationwide radio that Mistral would launch an open supply GPT-4-level mannequin in 2024.

And on the heart of open supply AI predictions is Meta, which is rumored to be making ready a Llama 3 launch, alleged to be on par with GPT-4, within the first half of 2024. This previous fall, META FAIR researcher Angela Fan instructed me that Meta appears for suggestions from its developer group, in addition to the ecosystem of startups utilizing Llama for quite a lot of totally different functions. “We wish to know, what do folks take into consideration Llama 2? What ought to we put into Llama 3?” she mentioned.

The Hill reported yesterday that fears are rising over AI’s influence on the 2024 Presidential election. It quoted Ethan Bueno de Mesquita, interim dean on the College of Chicago Harris Faculty of Public Coverage. “2024 might be an AI election, a lot the way in which that 2016 or 2020 was a social media election,” he mentioned. “We will all be studying as a society in regards to the methods through which that is altering our politics.”

And Nathan Lambert, a machine studying researcher on the Allen Institute for AI, just lately instructed me that generative AI will make the 2024 US elections a ‘sizzling mess,’ whether or not it’s from chatbots or deepfakes, whereas on the identical time, politics will decelerate AI regulation efforts. “I feel the US election would be the greatest figuring out issue within the narrative to see what positions totally different candidates take and the way folks misuse AI merchandise, and the way that attribution is given and the way that’s dealt with by the media,” he mentioned.

4. Coaching information on the coronary heart of AI

I lined points associated to the information that trains right now’s AI fashions extensively in 2023. However I feel it can solely ramp up in 2024 — as a result of I feel it’s actually on the coronary heart of what we all know, and don’t learn about LLMs and diffusion fashions. Points round copyright, bias, deepfakes, disinformation, labor points, open vs. closed fashions, and dangers all circle across the information that trains these fashions — whether or not they’re datasets that had been created years in the past or new ones being developed.

Information round AI coaching datasets has ramped up in simply the final couple of weeks, together with LAION‘s removing and OpenAI’s negotiations with organizations to construct new datasets. There may be a lot extra to dig into — and I consider it can turn out to be increasingly necessary to cowl in 2024.

5. Efficient Altruism vs. Efficient Accelerationism

As the way forward for AI stays unsure, its perception programs round aI dangers and alternatives — a few of which method spiritual standing — proceed to develop in affect. I’ve simply (belatedly) begun to totally perceive the intricate, costly net that the efficient altruism (EA) motion has woven to construct its affect in AI coverage circles, significantly in areas comparable to AI safety. I’ve additionally witnessed the rise of efficient accelerationism (e/acc), which VC Marc Andreessen has latched onto and I’ve known as the ‘different facet of the AI doomer coin.’

Whereas I wish to amplify AI pragmatists, who I feel might be extra useful within the brief and long-term evolution of AI, I feel it’s important to maintain overlaying the EA and e/acc perception programs — as its billionaire funders and founders will proceed to have outsized affect on how AI coverage and funding performs out for the remainder of us.