š® Hype is overhyped; GLP-1 surprises; Trumpās climate costs; Texas solar, AI necklace & Venezuela ++ #485
An insider's guide to the near future
Hi,
Itās Azeem. Iām back from my holiday, back in the flow. This week I shared my reflections on the learnings about technology and knowledge from my 15-day trip to Peru. If you havenāt been, I highly recommend adding it to your bucket list.
As a reminder, starting this weekend, the Sunday newsletter is part of the Premium offering. This means that paying members will receive this newsletter every week in its entirety. Readers on the free plan will receive it in full once a month. I am grateful for you being here ā if the Sunday newsletter helps you think more clearly about the changing world, consider upgrading. Thanks!
Ideas of the week
The hype-hype is hyped. Research firm Gartner Group, behind the too-ubiquitous hype cycle, opined that generative AI is at the peak of its hype and about to enter the trough of disillusionment. They estimate that 30% of projects are being abandoned after the proof of concept stage. Weāve been fed a shill pill. Proof of concepts are designed to test the feasibility of a new project. Is a 30% abandonment rate good or bad? Letās put it in perspective. Consultancy McKinsey found that 70% of complex, large-scale change programmes within businesses do not reach their goals ā not small proof of concepts, but large fully committed dog-and-pony shows. In 2020, research from advisory organisation Standish Group showed that 31% of full-scale IT projects would be cancelled before completion, and among large firms only 9% would be on time and on budget. A 30% dropout rate at proof of concept seems quite reasonable really. Several years ago, investor Michael Mullany analysed 20 years of data and concluded that technologies donāt follow the Gartner cycle, and our perception is distorted by hindsight and survivor bias. My experience over a similar time frame concurs. In more than a dozen presentations to company boards and asset managers this summer, I pointed out that with any breakthrough technology, investment tends to run ahead of real impact. The potential of the technology is easier to deliver than working applications. Game theoretic dynamics means that overinvestment in capacity is to be expected. As Zuckerberg said on the Meta earnings call: āThe amount of compute needed to train Llama 4 will likely be almost ten times more than what we used to train Llama 3 ā and future models will continue to grow beyond that. Itās hard to predict how this will trend multiple generations out into the future, but at this point Iād rather risk building capacity before it is needed, rather than too late.ā What we see between Google, Microsoft, Amazon and anyone else buying up Nvidia chips is ābetter early than too lateā. Crowd psychology means that non-tech firms need to be seen to be doing something ā anything. This lends itself to exuberance of different flavours, but the path out of that excitement is highly contingent. It is dependent on a wide range of complex factors. For every example of a large firm pulling back from using Microsoft Copilot to create more slides, we can see large incumbents double down on genAI products. Even the thorniest issues with LLMs (like hallucinations), can have workarounds to be good enough. Take, for example, this technical review by the LinkedIn Engineering team on how they reduced an error rate from an unacceptable 10% to a decent 0.01%. Other markets seem to be ignoring large parts of the hype cycle, too. Consider what is happening in China ā geopolitical and regulatory constraints are forcing more pragmatism, a focus on execution and āfinding product-market fit, scale and making applications highly affordableā. For instance, Baiduās self-driving vehicle fleet in Wuhan is on track to number 1,000 by the end of the year. My advice to you, as it is to investors and boards, is to keep going. The potential is huge and the hype cycle a chimaera ā it looks cute on a Powerpoint slide, but it isnāt much more than that.
The rest of the Ideas section is open only to paying members. Upgrade to read the full Sunday edition.
Data
Trumpās election could add 4 billion tonnes of CO2 emissions by 2030 and cause more than $900 billion in global climate damages (data from May, but too important of a reminder to leave out).Ā
TikTok was paying nearly $20 million per month to Microsoft for OpenAIās models.
California reaches a new milestone: 100 days of 100% renewable energy.Ā
The internet had 400,000 hosts, which represents a doubling every year since when I first used it in 1978.
Over 200,000 people paid for a driverless ride with Waymo in San Francisco in May of this year.
Texas plans to build 35 GW of solar, wind and battery capacity in the next eighteen months. Thatās more than double that of California or entire countries like the UK and Germany.
Intelās third quarter revenue forecast signals serious challenges for the company. Its share prices had been flat since the end of April.Ā
More than half of 16-24s in the UK watch no live broadcast TV during the week.
Short morsels to appear smart at dinner partiesĀ
š³ļø The fishy statistics behind Venezuelaās election results.Ā
šļø Self-control may be a heritable trait ā and a *really* important one for life outcomes.
The rest of the Morsels section is open only to paying members. Upgrade to read the full Sunday edition.
In focus
Hereās a preview of my Saturday column. I share my reflections from a recent 15-day trip to Peru and what it taught me about knowledge, technology and resilience.
End note
ā”ļø Iāll be in Athens for the 12th Athens Democracy Forum from October 1-3. Readers of Exponential View get a special 20% discount on tickets using this link.
šļø I recorded an hour-long podcast talking about all things exponential with Matt Turck. You can watch our discussion here.
š„ The team at Wordware, where I am an investor, launched their tool to easily build AI agents. It has gone viral, with more than 1,000,000 users on the first day. Try it here.