After a lengthy summer vacation, here’s the most notable thing that happened in generative AI: the primary meaning of the word “AI” is now “generative AI” (replacing whatever meaning AI previously has had).
Also, there were some other developments that I didn’t just make up (:
In my view, these are the 12 most significant AI stories and developments over the late summer (early August to today):
- Poor’s man fine-tuning for ChatGPT. The ability to provide custom instructions to ChatGPT; this is clunky from the perspective of SaaS product design but still useful. Takeway: OpenAI is still a B2B platform first and foremost but it’s not abandoning ChatGPT – and is still taking down droves of AI startups with single blog-post feature announcements. For now, this is the development that might most affect your daily use of AI.
- Finally, a viable daily-use alternative to OpenAI. Anthropic not only released Claude (with its enormous, book-size “context window”) but intro’d a low-cost paid version of Claude that all of a sudden makes it a viable contributor to a daily workflow. It’s been a step above Bard and Llama2 models for a while now but this makes it reliable.
- The ability to fine-tune a state-of-the-art GPT model through the API; for context, prior versions of GPT, including GPT-3, had been fine-tunable for years, long before ChatGPT existed – this is the first time in the post-ChatGPT era that a “state of the art” model, 3.5 Turbo, has been fine-tuneable.
- Open Source AI now unignorable. A new, long-term initiative (2024-ish release) that’s led top-down by Meta to make Llama as good as GPT-4 means that it will soon become impossible for business decision-makers to ignore open source AI solutions. Given that fine-tunings of Llama2 are already outperforming GPT 3.5, this is both credible and inevitable – though the timeframe is a little more elongated than some thought; the first-mover advantage here of closed source such as OpenAI and Anthropoc is significant.
- OpenAI released an enterprise version of ChatGPT; this move protects OpenAI from the threat of Llama (by offering data privacy), and lets it compete with Microsoft Bing Chat for big customers, by facilitating fine-tuning, enterprise-grade data analysis, integration, etc. Salesforce followed suit at Dreamforce yesterday, announcing Einstein Copilot Studio, meant to help salt many pieces of a Salesforce-based cloud infrastructure with AI.
- An increase in anti-open source rhetoric within the AI safety debate. Meta’s LeCun has asserted since Spring that open source AI is technology and thus, isn’t inherently unsafe, yet Google’s Suleyman and many other increasingly forecast doom.
- Meta releases SeamlessMT4, now the world’s best (and free and open source) translation solution. As you can see from the hosted demo, MT4 easily displaces ChatGPT, Google Translate, and DeepL, as the world’s best translation solution. But this is a portable technology, not just a website – thus the implications for internationalized digital business, in a time when Stripe.com makes it a cinch to take payment in any kind of currency from any kind of card, bank, or payment network, are immense.
- OpenAI using Finnish prison inmates to manually improve its training data (This is “RLHF” – “reinforcement learning from human feedback”); this is perhaps better PR-wise than scandalously contracting this work to the Global South at starvation wages, but still problematic. Will there be economic coercion, will there be psychological damage, will the compensation be unfair? In Kenya, at least, the answer to all these questions was yes. BTW, what if AI can train AI, as asserted in this Arxiv research paper? (I doubt it, personally but this makes OpenAI look even worse).
- Conflicting environmental implications of AI come into focus – The amount of water (and electricity, ofc) used by AI is shocking – just a few ChatGPT equals a glass of water and headlines claim AI may cause a global water shortage. On the other hand, AI is really good at recycling. Showcasing the potential from the impending boom in multi-modal AI, AI can see, smell, and categorize waste extremely well. Can the intelligent use of IA offset its cost of compute, until we achieve common quantum computing?
- Federal US fudge rules that AI-created works can’t be copyrighted. This further galvanizes the Hollywood writer’s strike, because if Hollywood studios can’y copyright the works their AI produce, they can’t make any money. However, the ruling wasn’t intended to settle the question once and for all; nuance remains and there will be additional legal activity.
- Society is starting to think about what AI is trained on. Dovetailing nicely with the ruling above, there’s been a general wave of journalist inquiry into, and legal activism against, AI training on copyrighted works; notable writers in this fray include Michael Pollan, Steven King, Sarah Silverman, James Patterson, and Michael Chabon. On the same token, most major media companies (New York Times, Reuters, Axios, Bloomberg, The Atlantic, ESPN, Disney) are now following the early lead of Reddit and Twitter and blocking OpenAI’s web crawler.
- In geopolitics, US tech leaders hew to US interests around AI. Following Sam Altman’s lead, tech tycoons (worth personally half a trillion dollars) met lawmakers in DC to plan regulations a la Spain. Later many firms, including Nvidia, voluntarily commited to protecting US national interests. This comes on the heels of the US banning the sale of AI hardware to not just China but to much of the. now thwarted Middle East. Some tech leaders are even calling for AI nationalization.
A few honorable mentions that are noteworthy but less important than the above: AI can smell, StableAudio brings text-to-music to market (beating Google to it, as usual), 1bil run rate for OpenAI, AI detects eye disease, cancer, and school shootings, Gannett backs off of using AI to write news articles, creepy, hamfisted use of AI by Snapchat, Imbue raises 200mil to build smart AI agents for business, (speaking of which) BabyAGI now searches YouTube, Canva’s new ChatGPT plugin is trash, open source coding assistance levels up , GPT-4 powered local code interpeter is a must-try for developers and semi-developers, so far Google Gemini is all hat but no cattle, Apple devices quietly integrate AI, AI used to ban books, teachers learn AI on Khan Academy, and media labels like Warner are making a pretense of “signing” AI content, such as Noonoouri.
* * *
Stepping back for a minute, here’s my general sense – we’re still at peak hype, as Gartner puts it, and as a result there’s still too much happening for any one person to fully understand – what you read above is a fraction of the stories out there over the last 1-2 months. But I get the sense that generative AI is steadily transforming technology, business, culture, work, politics, science, and media.
And peak hype doesn’t necessarily mean bubble, not according to Goldman Sachs.
However, it’s still too early to bet the farm on AI consumer products. Better to build tools for your own company that help you do whatever you do, like McKinsey, and otherwise build AI into your work, pre-AI products, processes, and infrastructure. And keep building skills, as in Singapore.
After all, 67% of US-based IT leaders are bullish on AI for what reason? AI for operations.
And for most of us, “AI for daily operations” is still where it’s at. Understanding that in a fine-grained way is probably more important than any of these news stories.
With that, stay tuned – I’m going to follow up with a “back to the basics” review of what exactly generative AI is good for, framed from the perspective of prompting.
Have a great trip,
Rowan