🔮 The Sunday edition #500: AI scaling; science FTW; reality vs the Democrats; BlueSky, woke Grok & artificial eclipse++
An insider’s guide to AI and exponential technologies
Hi, it’s Azeem. For 500 editions, every Sunday for nearly a decade, Exponential View has helped you navigate the future. Today’s milestone issue reveals the complexity of the coming years. Research shows a 30% decline in some freelance job opportunities since the release of ChatGPT. On the other hand, a breakthrough in materials discovery shows the sheer scale of the potential rewards. Scientists using AI are discovering new materials 44% faster, filing 39% more patents and creating 17% more prototypes. This transition will be complicated – but the upshot is the growth in human potential. Let’s go!
Ideas of the week
AI’s trillion dollar scaling bet continues
There has been a lot of talk in recent weeks about the ‘end of scaling’. Let’s put things in perspective: AI’s most powerful players are shifting away from the brute-force scaling that defined the last decade. Instead, they’re pursuing a different approach that could continue to deliver. As I pointed out in July, with a bit more historical distance, waves of innovation overlap and give us the smooth curve of increasing progress. Bumps are to be expected.
But bumps force builders to come up with new approaches – this applies to AI scaling too. Instead of funnelling all computational resources into training larger models, companies are now dedicating more compute power to the inference phase. Continue reading here.
See also: Computer scientist Kai-Fu Lee shared that frontier labs in China are 100 times more efficient with their compute, because they can’t get easy access to the large-scale clusters that US firms can, owing to the US regulations. The Chinese approach has proven successful. The SuperCLUE AI benchmarking system has tracked the gap between the leading Chinese LLM and the leading global LLM, narrowing from 30% in May 2023 to 1% today. In the meantime, in the US, Elon Musk ‘freaked out’ rival AI firms when his venture xAI deployed a 100,000-GPU AI system in Memphis in just 122 days (it usually takes four years).
Science’s new frontiers
Putting AI tools in the hands of R&D staff at a large unnamed US firm increased the amount of new materials discovered by 44%, the number of new patent applications filed by 39%, and the number of prototypes by 17%. R&D efficiency rose between 13-15%. This is huge. AI unlocked a new way of working: scientists could automate 57% of idea generation tasks to AI, better spending their time evaluating AI outputs for viable solutions. Yet, there was a gulf in performance based on the researchers’ level of ability, which suggests that this is a co-pilot, not an autopilot. And, there’s more in scientific AI… AlphaFold3, which brought Demis Hassabis and John Jumper the Nobel Prize, just became significantly more efficient and was made available to researchers. Perhaps my favourite is Evo, a DNA-trained AI that excels at prediction and design at the level of DNA, RNA and proteins (first mentioned in EV#480). This could open the door towards creating life forms tailored to specific purposes. These strides are replicated across the field: Microsoft’s AI2MBD system for biomolecular dynamics simulation looks promising, while Google recently highlighted success in AI diagnosis of breast cancer.
Dig deeper on AI in science:
How AI can help us continue to create knowledge even as our biological limitations bring innovation to a halt.
Our view on the future of an ecosystem of AI models that would benefit scientific discovery.
Workplace shocks
There’s more evidence again this week to show that the impact of AI on the workplace is real (see EV#499 for the research we covered last week). A new paper from researchers at Harvard, Imperial and Berlin found that online freelance opportunities for writers, software developers and engineers dropped by between 10% and 30% in the eight months after ChatGPT was released. The decline in listings hasn’t recovered since. Demand-side responses to these changes are occurring: employers are willing to pay 5.7% more for roles needing broader skill sets that blend traditional expertise with AI literacy. But compare the numbers: a 6% pay rise does not accommodate an up-to 30% decline in jobs. Governments will have to react to these profound changes in the labour market. The AI labs themselves aren’t offering many tangible solutions (though it’s not clear that it’s their job to do so…). AI’s impact on the workplace may be even more keenly felt in January when OpenAI releases agentic AI tools.
Understanding the US election
The outcome of the US election wasn’t emblematic of a racist or sexist society and wasn’t interfered with from abroad, as argues sociologist and professor
:If you look at voters’ expressed opinions, it seems like there were three core factors: inflation, immigration and alienation from cultural liberalism.
The long-held left-wing assumption of demographic destiny – that increasing numbers of migrants into the US would inevitably tilt the electoral calculus in their favour – has been exposed as an illusion. Al-Gharbi highlights a key challenge for Democrats in the years to come: they must reconnect to the working class, minority voters – and immigrants. I come from an immigrant family and have family members in the US; what I see from their experience is that you don’t go to the US to “right social-justice wrongs” that America committed over the previous 400 years. Rather, you move to the US that exists, imperfections and all. Al-Gharbi’s whole essay is worth your time.
Data
The variation of power prices across Europe has exploded from a narrow 40-60 €/MWh band in 2018 to a dramatic 25-100+ €/MWh spread in 2024.
The Biden administration sets new targets to triple nuclear energy capacity by 2050.
Brazilians spent 20% of the government’s flagship social programmes’ payments on online gambling sites in August.
Europeans spend 575 million hours per year clicking through cookie banners, amounting to roughly 1 hour and 25 minutes per person. (That’s an economic cost of $16 billion per annum, or roughly the defence budget of the Netherlands.)
Epoch created a super-hard new math benchmark for AI systems that uses real research problems approved by top mathematicians – current best score is below 2%.
A quarter of the oil production from the Gulf of Mexico remains offline after Hurricane Rafael.
Exercise stimulates neurons to extend four times farther.
BlueSky’s 13-million-user network visualised:
Short morsels to appear smart at dinner parties
🫣 Department of broken incentives: By law, if the number of casualties in a major public incident in China exceeds 35, a mayor will be fired. So Chinese media outlets often report exactly 35 even if the number is higher.
😆 Elon Musk’s Grok AI is more “woke” than ChatGPT.
🌷 AI-generated poetry has been rated more favourably in qualities such as rhythm and beauty than human poetry.
🍿 J Lo performing in Riyadh. Has any country experienced as rapid a cultural transformation as Saudi Arabia?
🇺🇦 Back to the future… Odesa’s defences take on swarms of Russian attack drones.
😎 The European Space Agency will create an artificial eclipse in space to study the sun.
🙏🏽 10 PRINT “Thomas Kurtz was a computer pioneer”
20 PRINT “and co-inventor of BASIC”
30 PRINT “He passed away on November 12th at 96”
40 END
👀 What surprised me most after 500 editions of Exponential View
After nine years of writing Exponential View and 500 Sunday editions at technology’s frontier, I’ve witnessed the future unfold in ways that defied everyone’s expectations—including my own. What we thought we knew about innovation and progress has been consistently upended, revealing just how unpredictable exponential change can be.
Here’s what surprised me most:
What AI looks and feels like
How Nvidia captured the market
Autonomous vehicles actually working
The speed of knowledge explosion
How culture wars get in the way of technology innovation
The energy sector — where do I even begin?!
Our (mis)understanding of technology as a force for societal change
Founders continue to surprise me in the best ways.
I’m surprised, but not surprised, by the Chinese efficiency with compute. They have found a way to optimize system level inference prompting with Nvidia GPU’s. Finding efficiency in prompting is much easier than finding efficiency in training. If the frontier labs can find the inference strategy that the Chinese are using, then their models can be unreachable.