🔮 What to make of ChatGPT’s rival Claude 3; urban innovation; AI patents; solar learning; epigenetics ++ #464
Hi, I’m Azeem Azhar. In this week’s edition, we explore
Why new data about remote work challenges our assumptions about what makes cities innovative
Anthropic’s new models and the need for better benchmarks
Solar PV breaking new records
🚀 Today’s edition is supported by Sidebar. Sidebar is a highly vetted leadership program focused on small peer groups.
Sunday chart: The evolving urban geography of innovation
Data from Stanford’s Nicholas Bloom and colleagues shows substantial shifts in where people work. Roughly a quarter of Americans work from home at least once a week, with many living further and further away from the office.
The mean distance from the employer’s location has risen 170% since the beginning of the pandemic, from 10 miles to 27.
The share of employees living more than 50 miles away has quintupled.
Those in the highest-earning category (>$250k per annum) have seen the largest change.
Those aged 25-39 live farthest from their employers, since the pandemic.
This looks like a permanent shift. Once this cohort has kids, they’ll be less likely to move.
Could this challenge long-term innovation? One long-held argument is that the benefits of agglomeration, such as access to infrastructure, markets and dense information exchange networks, outweigh the drawbacks. In addition, urban scaling laws mean that as cities expand, innovation experiences superlinear growth. A dispersal of workers to the far-flung places, it’s thought, will reduce a city’s capacity for innovation.
A new paper by César Hidalgo, Xiaofan Liang and colleagues questions these assumptions. The paper recognises that size and population density are crucial factors in innovation, but also highlights the importance of a city’s connectivity and broader intercity networks in enhancing its innovative capabilities. Connectivity, whether through human mobility, social media connections or scientific co-authorship, plays a distinct and complementary role in fostering urban innovation. It means that smaller cities can punch above their weight. Even a significant shift towards work-at-a-distance might not impact innovation output, as long as connectivity — be it digital or physical — is maintained.
A more distributed, interconnected, innovative workforce can actually drive innovation. A dispersed workforce could enhance opportunities for user-centred innovation, and decentralisation might encourage a more diverse group to contribute to the creative process.
See also: The rise in working from home may have caused a decline in burglaries across Wales and England.
🚀 Today’s edition is supported by Sidebar.
Ready to supercharge your career?
Introducing Sidebar, an exclusive, highly curated leadership program designed to propel you to new professional heights. Picture yourself being a part of a dynamic network of peers, including Fortune 500 leaders and groundbreaking startup CEOs, all offering their unique insights and expertise to fuel your journey.
The results are palpable. A staggering 93% of members affirm that Sidebar has been a game-changer in their professional path.
Nothing propels your career trajectory like a robust peer group. Learn more at sidebar.com, and join thousands of top senior leaders from companies like Microsoft, Amazon and Meta who have taken the first step towards accelerating their career.
Key reads
(Quasi-) Magnum Opus. Emerging as a formidable rival to GPT-4, Anthropic’s Claude 3 family introduces Opus among its trio of models, all strongly marketed towards business applications. An advancement on previous models, Opus exhibits enhanced precision and efficient recall, as well as a significantly reduced refusal rate, directly addressing the previous model’s frustrating tendency for unwarranted rejections. It’s competitively priced too, at $20 per month for individual users (the same as GPT-4), $15 per million input tokens (pmt) and $75 pmt output. GPT-4 Turbo will set you back $10 pmt and $30 pmt respectively.
Claude 3 was joined by Inflection’s Pi, a chatbot also claiming GPT-4 level. Pi is meant to be more of a friend and less of an agent. The firm claims its 6m monthly spends 33 minutes chatting with Pi each session. We also need to add Gemini Ultra and Mistral Large to the mix.
Each LLM behaves slightly differently. While I can’t exactly tell you why yet, I’m getting a sense of the tasks I’ll send to Claude 3 Opus and those I send to GPT-4.1
Some of these differences come out in benchmarks. Opus marginally outperforms GPT-4 in benchmarks such as the MMLU2, the GPQA3 and various maths-based tests4, yet these evaluations only reflect GPT-4’s capabilities at launch and not any subsequent enhancements. These tests are useful, but as we argued last year, they fail to capture the ever-expanding capabilities of the models.
It reflects that what these models can do can’t be measured by a scalar (like height or mass). Instead, like football players, they have capabilities across many different dimensions. Sometimes you need an attacking number 9, sometimes a defensive midfielder. I’d encourage you to play with them for a while, certainly more than a few hours, if you want to get a sense of which one will really help you for which task.
The inventor’s dilemma. Brazilian congress member Antônio Luiz Rodrigues Mano Júnior submitted a bill that would enable AI to hold patents and therefore be considered the inventor and owner of rights arising from the invention. The rationale is an incentivisation of innovation and research. This approach isn’t universal: last month, the US Patent and Trademark Office published guidance clarifying that while AI systems can play a role in the creative process, only natural persons who have played a significant part in the invention process can be considered inventors. This is fit-for-purpose at the moment, while we are figuring out the use of AI in innovation and science, but as we discuss in this essay, AI could end up managing almost all steps of the innovative process, from ideation to productisation, especially as robotics leaps forward. Autonomous labs are already in testing. In these cases, the US guidance wouldn’t technically allow an individual to hold the patent, as they wouldn’t have played a significant enough role in the invention process. Patent ownership should continue to go to natural persons, but the modalities of patent attribution will need to be amended.
See also:
Chain of Thought Prompting increases how diverse the output is from AI.
Shine on. Solar demonstrates yet again why it’s the poster child for Wright’s Law, also known as the learning curve, where price declines as cumulative production increases. Our previous analysis pegged the solar learning rate at 33%. Now, including 2023 and 2024, it has jumped to 44%. The news is timely: energy demand is shooting up, partly due to the growth in extremely power-hungry data centres, necessary to fuel the AI boom. Renewable energy can be scaled easily thanks to the ever-declining price and a relatively straightforward installation. Yet, as energy production escalates to match demand, regional power infrastructures, like those in Georgia, USA, are starting to buckle under the pressure. Ramping up the necessary transmission lines could be too complex and costly. Amazon has taken matters into its own hands by buying a 2.5 gigawatt (GW) nuclear plant to power its new data centre in Pennsylvania.
Newsreel beta
The Indian government asks AI providers to seek explicit permission before deploying models in India.
Sam Altman is back on the board of OpenAI.
China gives “computing vouchers” to its AI startups amid global compute scarcity.
Apple banned, then unbanned, Epic Games from creating its own app store in Europe.
TikTok users are urged to speak up against Congress advancing a potential ban of the Chinese app. (Biden has said he will support a bill banning TikTok if passed by Congress.)
Data
BYD launches the first electric car priced under $10,000.
AI search startup Perplexity reaches a $1 billion valuation. See here for my chat with co-founder, Aravind Srinivas (now on YouTube).
According to the 2024 Edelman Trust Barometer, 59% of people surveyed believe government regulators lack adequate understanding of emerging technologies to regulate them effectively. Additionally, 53% of respondents indicate that science has become politicised in their country.
AI job posts on Indeed have increased by 2.3% over the past year. Posts for data analysts and software engineers have decreased by over 30%.
The cost of Baidu’s Wenxin large language model “Inference” has been reduced by 99%, while model performance has seen a 50-fold increase in just one year.
Short morsels to appear smart at dinner parties
🧑🏽💻 Screen time reduces the number of adult words young children hear, limits their own vocalisations and hinders their engagement in meaningful interactions. Also, are Americans losing their social skills to the digital space?
🚗 How Apple frittered a $1bn a year on its car project.
➕ WhatsApp-based AI maths tutor improves students’ scores in Ghana.
🏃♀️ Although men, on average, outpace women in running, the gap in speed narrows as the distance increases. In fact, for distances over 195 miles, women are 0.5% faster, a statistic driven by a small cohort of exceptionally trained female ultrarunners.
🧠 Researchers in South Korea use sound to influence the formation of connections between neurons.
🌭 A meta-analysis finds ultra-processed foods are associated with a host of adverse health effects.
Latest posts
If you’re not a subscriber, here’s what you missed recently:
End note
The Waymo arson in San Francisco may capture a wider sentiment brewing globally. Edelman’s latest Trust Barometer, a survey of 32,000 people from 28 countries, shows that the mood around AI is souring. A higher proportion of people (35%) said they rejected AI than embraced it (30%). This is a big issue, bigger than the five per cent delta. I’m thinking about it and will pen some thoughts in the next couple of weeks. In the meantime, I’d be keen to hear your thoughts:
Does it matter that net trust in the potential of AI is negative? What are the ramifications of it?
What could be done about that trust gap?
Let me know in the comments (and please answer the survey below).
Cheers
A
What did you think of today’s edition?
Excellent | Very Good | Good | Fair | Poor
What you’re up to — community updates
Facilitated through the Exponential View community, Myke Näf of Übermorgen Ventures has invested in Rodolfo Rosini’s Vaire Computing.
Share your updates with EV readers by telling us what you’re up to here.
I can’t tell you why because the reason isn’t clear in my head, not because I’m keeping it from you.
Undergraduate level knowledge.
Graduate-level reasoning.
Such as GSM8K, MATH and MGSM.
I think there are many negative implications from a lack of trust and at the risk of over simplifying things I think the way forward is to educate as many people as possible as to what it is at a level that makes sense to them and what its implications might be and how best to leverage the positives and ameliorate the negatives. No small job but such an important one that I’m trying to do my small bit with like minded people.
On AI and trust: IMHO, income inequality, big tech misbehavior, and the clear prioritization of profit margin above all else in so much of the business world, along with general cynicism pushed by nihilistic conservatives all contribute to a world where people don’t trust AI.