🌪️ Three months of AI in six charts
Developments that made us spin
The past three months have felt like a whirlwind in the world of AI. And we’ve all been swept into it — no matter your industry or field of work. With most of our team off this week, we thought we’d take a breather and look at the quarter behind us through the prism of six important events that occurred in that short time.
Three months of AI in six charts, let’s go!
AI eating software
We wrapped up Q1 of this year looking at the forecasted decline in the cost of computation once natural language programming makes inroads into the profession of software engineering.
As OpenAI unrolls its Code Interpreter over the coming days, data science will become far more accessible to those without knowledge of coding or data visualisation. In a paper out this week titled “What Should Data Science Education Do with Large Language Models?”, the authors suggest that LLMs are “shifting [data scientists’] focus from hands-on coding, data-wrangling and conducting standard analyses to assessing and managing analyses performed by these automated AIs.” This shift calls for a significant change in how data science is taught; the near-future curriculum will include AI-guided programming, LLM-informed creativity and interdisciplinary knowledge.
In the coming weeks, we’ll run a Promptpack for members of EV focusing solely on Code Interpreter.
Speaking of education…
Remember Chegg? In May, Wall Street wiped over 40% off this online education company’s stock price in a single day, valuing it at ⅓ of what it was worth at the time of ChatGPT’s initial launch. This probably makes Chegg the first company to have directly suffered at the hands of generative AI.
There is a positive story here, as we wrote in EV#421:
the possibility of widening access to education globally. More than a quarter-of-a-billion children are not in primary or secondary education worldwide. The cost of education resources and access to teachers are two of the barriers. But the prospect that anyone with a phone could access a personalised tutor could be good news.
The context windows of LLMs - the amounts of text that models can process and respond to - are growing rapidly. In Q2, Anthropic released a new, 100k-token version of its model Claude. Thanks to its massive context window, Claude can process the entirety of The Great Gatsby and answer questions about it in 30 seconds.wrote:
This opens up a range of incredibly useful applications. I have found the 8,000 token limit on GPT-4 a constraint. It prevents me from analysing documents that are longer than about 7,000 words (and most book chapters are longer than that). A whole slew of useful material like academic papers, legislation, business reports or smallish databases is within Claude’s reach. A long context window also opens up the possibility for greater personalisation of an LLM.
Back in May, we highlighted one of the first studies to tackle the effects of AI on labour and the economy. Brynjolfsson, Raymond and Li show that generative AI can boost productivity by 14% in some jobs, and enhance productivity in employees who are underperforming. They argue that AI accelerates the learning curve, enabling employees to get six months of experience within two months.
While today’s state-of-the-art models have limited or no open access, open-source is catching up. Aswrote in Q2, “open-source is reducing the cost of training for these models, allowing developers with decent laptops to play with them. No longer are we bottlenecked on training runs on enormous grids of GPUs. This increases the velocity of evolution across open-source models.”
Open-source AI was a subject of discussion in the U.S. Congress this past month. Clement Delangue, the CEO of Hugging Face, spoke at the hearing on AI of the House Committee on Science, Space and Technology, advocating openness as the driver of competition and greater AI economic opportunity. If you haven’t seen Clement’s testimony, we recommend you watch it.
The European way
June was a busy month for regulators. While the U.S. Congress held hearings, the European Parliament proposed the first comprehensive set of regulations for AI. We praised the fact that rules are being seriously considered, but we concluded that
[t]he rush to put rules around foundation models could be problematic. In particular, the Act seems to regulate models as the end product, rather than the uses of those models. Researchers at Stanford University tested major foundation models against the rules and found that they largely do not comply. First, there’s a clear lack of disclosure of information about the copyright status of training data. Next, there is no clarity on energy use or risk mitigation, nor any evaluation standards or an auditing system. The most compliant models were open-source ones. The second problem will be whether some of these rules are simply too early for an emerging technology in a complex, competitive environment.
Like we said, a whirlwind! 🌪️ And what will the next three months bring? To answer that, keep reading Exponential View.
🤔 Member’s question A member of EV asked in our invite-only Slack space: “I use AI in almost all my work and in a lot of non-work functions as well. Unfortunately, it seems like the only people who are doing this are the people who already benefit greatly from leveraging tech. I’m concerned this will further widen the gap in wealth/power, given just how powerful AI is now+will become shortly. What do you think?”
Refer a friend to win a 30-minute chat with Azeem: Visit this page to collect your unique code and share it with friends. It will look like this:
Email and Linkedin work particularly well. Like any good referral programme, there are goodies available: a complimentary subscription to Exponential View and a meeting with our founder, Azeem Azhar.