🔮 The AI data ceiling; interoperability; creative parrots; wandering around Mars ++ #453
An insider's guide to AI and other exponential technologies
Hi everyone,
Welcome to the regular Sunday newsletter — part analysis, part curation about the last week in AI and exponential technologies.
Just a note before we get going: I’m actively posting on Threads right now. So if you are using it, please hop over and follow me there. Azeem
Latest posts
If you’re not a subscriber, here’s what you missed recently:
🔮 AI climate impact & imminent risks
This week, I spoke with Sasha Luccioni, an AI researcher and climate lead at Hugging Face, whose recent paper "Watts Driving the Cost of AI Deployment?" offers the first systematic comparison of the ongoing inference costs of various categories of machine learning systems. I wanted to speak with her about the energy cost of AI training and deployment...
Sunday chart: The new data brokers
Large language models are greedy for data. GPT-4 and Gemini Ultra are estimated to be trained on 4-8 trillion words. EpochAI, a research group, seems to think that we may run out of high-quality training data as soon as next year.
To maximise the amount of quality data available for training, model makers must tap into the rich reservoir of proprietary data, exemplified by the recent collaboration between Axel Springer and OpenAI. OpenAI is paying Springer for historical and real-time access to their data, both for training and referencing in user queries. This text is professionally edited, contains knowledge about the world, and most importantly, can’t be accessed by other model makers (unless they make a similar deal). It is an interesting approach, although Axel Springer is renowned for its lurid journalistic practices, rough dealing and years of terrible controversies.
This raises a question: How important will access to proprietary data be for building a moat between foundational models? Open source has so far managed to keep up as model training so far has been based on open datasets, but a gap may emerge if it can’t access the best data. It remains to be seen how big this gap could be. Axel Springer data alone might not change training much, but collecting many such deals could matter a lot.
An early example of leveraging proprietary data was BloombergGPT. Here Bloomberg made their model on a training corpus of its own financial documents. At release, it outperformed similar models on the financial-specific tasks it was built for. It suggests that proprietary data can make a difference, although in this case, only in a specific domain.
OpenAI has shown that it’s willing to pay eight figures annually for historical and ongoing access to data — I find it difficult to imagine that open-source builders will. According to analyst Kelvin Mu, OpenAI has an estimated revenue of $1.6B and an addressable total market of more than $220 billion. Meta, the most well-heeled provider of open-source models, has the money for such deals (their annual revenue this year is $127 billion) but not the commercial ambition. It seems unlikely they would invest in something that doesn’t boost their revenue - they only make money from cloud providers who use their models (and potentially other indirect monetisation methods). There are ways other than proprietary data to improve models, namely synthetic data, data efficiency, and algorithmic improvements - yet it looks like proprietary data is a moat open-source cannot cross.
Key reads
LLMs get political. 52% of Americans are more concerned than excited about AI. Yet this concern doesn’t extend to AI-generated content. Capgemini found that roughly three out of four people worldwide trust AI-generated content. As we mentioned in the newsletter last week, social trust is misplaced in AI. This danger will be increasingly evident as more AI-generated content seeps into our information ecosystem. Take for example, Ashley, the world’s first AI-powered political campaign caller. She calls up voters, personalising how she talks to people based on their profile. It’s entered the US political landscape - time to strap in!
See also:
An example of deep fake usage in Bangladesh shows how generative AI tools can be exploited in elections and the difficulty in policing their use in smaller markets overlooked by American tech companies.
Micrsoft’s Bing AI gave wrong information about political candidates 30% of the time, a study from two think tanks showed.
Reverse causality. Daniel Williams, a philosopher at the University of Sussex, argues that misinformation is the symptom of deeper societal issues rather than the cause. It emerges from polarization, institutional distrust and governmental corruption. From this view, censoring misinformation or debunking false ideas is ineffective and may even aggravate societal problems. Take social media, for example, where its centralisation has increasingly fostered public mistrust due to concerns over privacy violations, biased content moderation, and the potential for manipulation of user behaviour and opinions. In recent years there have been attempts to rebuild this institution in a decentralised way (e.g. Mastodon). This week Mark Zuckerburg announced a test to make posts from Threads accounts available on services that use the ActivityPub protocol such as Mastodon. In 2021, I talked to Meta’s President of Global Affairs Nick Clegg about interoperability, which he described as a great but “fearfully difficult” idea. Looks like they are finally trying to take a step in the right direction, and hopefully rebuild trust in this institution.
The little guy can win. When ChatGPT started, Sam Altman said each chat cost a few cents. Now, the cost is about 100 times less for roughly the same capabilities via Mistral’s new Mixtral model. Eight 7B parameter models are combined using a ‘mixture of experts’ approach (you can read a breakdown of it here). This is the same method rumoured to be used by GPT-4. Mixtral is small enough that you can run it on on a decent laptop. With the advances of small local models such as this and Gemini Nano announced last week - we are not far from having personalised on-device products of GPT-3.5-level quality.
See also:
The Competition and Markets Authority in the UK is seeking opinions on whether OpenAI and Microsoft’s partnership has resulted in a relevant merger situation and what its potential impact could be on the UK.
MIT has released a series of policy papers that outline a framework for the governance of AI.
Parrot makes move 37. Google DeepMind reported that they had used LLMs to discover new mathematical knowledge - a world-first for LLMs. This pioneering method, called FunSearch, achieved a breakthrough in solving the cap set problem, a notoriously complex challenge that has puzzled mathematicians for decades. FunSearch’s evolutionary approach is particularly noteworthy; it iteratively refines the creative code generated by LLMs through continuous evaluation and enhancement, helping avoid hallucinatory answers. This effectively evolves initial ideas into potentially groundbreaking solutions.
See also, highly-regarded blogger and computer scientist, Scott Aaronson, reflects on being wrong about his estimates for AI progress.
Market data
2014 saw only 5% of adults in wealthy countries with ongoing subscriptions to an online news site. This year, it’s 13%.
In 2023, over 20% of all venture deals were down rounds, where companies raised funds at a lower valuation than in previous rounds.
TikTok became the first non-gaming app to reach $10 billion in consumer spending on the Google and Apple app stores.
In 2023, over 10,000 research papers were retracted, breaking yearly records.
Around 900 autonomous vehicle engineers were fired by Cruise.
Short morsels to appear smart at dinner parties
📐 The untold story of 6 decades of Ottoman science.
🎨 Veteran Silicon Valley watcher, Om Malik is impressed with the Apple Vision Pro.
🇰🇪 Kenya’s EV revolution is on two wheels.
🪱 Scientists call for greater emoji “biodiversity”.
🧫 What biology can learn from physics. via
☄️ A quantum breakthrough: Using single molecules as qubits for the first time.
🌚 Happy 1000 days of roaming around Mars, Perseverance!
🌌 We’ve seen signs of nuclear fission in the stars for the first time.
🏙️ Making a city more liveable: lessons from the Big Apple.
End note
I’ll be in Davos attending the Annual Meeting of the World Economic Forum. If you are also attending and want to connect, please message me on the Forum Live messenger. If you happen to be in Davos at the time and want to connect or invite me to something together, drop me a note.
Finally, I’m pulling together my speaking schedule for 2024. I’ve spoken regularly to more than 100 clients in inter alia finance, technology, consumer products, industrials, telecoms, healthcare and professional services. Audiences range from board-level and C-suite to wider corporate groups. I’ll often speak at customer events too.
My key themes for next year include the exponential transition of our economies; AI and its economic impact; AI’s impact in the firm and future of work, and how investors ought to think about these coming changes.
My schedule can fill quite quickly so it’s best to email my agent, Ania Trzepizur sooner rather than later.
Cheers,
Azeem
Share your projects and updates with EV readers by telling us what you’re up to here.
For a next newsletter, I'd love a curated list of ev readers on Threads or Bluesky.
Correction on Ottoman Science morsel: 6 centuries, not 6 decades.