Unbundling AI
Like all machine learning, LLMs turn a logic problem into a statistics problem: instead of people writing the pattern for each possible question by hand, which doesn’t scale, you give the machine a meaningful sample of all the text and data that there is and it works out the patterns for itself, and that does scale (or should do). You get the machi... See more
Benedict Evans • Unbundling AI
Google gives you Ten Blue Links, which communicates “it might be one of these - see what you think” (and turns us all into Mechanical Turks, giving Google feedback by picking the best answer). But a chat bot gives you three paragraphs of text with apparent certainty as The Answer, and footnotes, a click-through disclaimer and a ‘be careful!’ boiler... See more
Benedict Evans • Unbundling AI
I can see the errors and fix them, and it’s still very helpful. I always used to describe the last wave of machine learning as giving you infinite interns, and that applies here: ChatGPT is an intern that can write a first draft, or a hundred first drafts, but you’ll have to check it.
Benedict Evans • Unbundling AI
Equally, you could use ChatGPT with a code interpreter, some plugins and a text file full of saved prompts - but it might be better for those tasks to be unbundled into tools. This year there’s been a wave of hundreds and perhaps thousands of first drafts of this - the so-called ‘thin ChatGPT wrappers.’ The prompt is an API call and ‘prompt enginee... See more
Benedict Evans • Unbundling AI
If you’re just an API call with a GUI, you probably don’t have much of a moat. But on the other hand, everyone in tech today is wrapping something else at some level. Snap uses Google Cloud for storage, but we don’t call it a ‘thin GCP Wrapper’.
A legal software company might use GCP translation and AWS sentiment analysis just as it used AWS data storage, but it used them to build something that law firms could buy, with a lot of other product around that, and a sales force and understanding of what lawyers needed, and it didn’t worry at all about competition from AWS. And then, as part of that product, it might built its own models with its own data doing things that the hyperscalers did not.
the science problem. We’ve all spent the last year talking about ‘error rates’ or ‘hallucinations’ (indeed, I also wrote about them here). The breakthrough of LLMs is to create a statistical model that can be built by machine at huge scale, instead of a deterministic model that (today) must be built by hand and doesn’t scale. This is why they work,... See more
Benedict Evans • Unbundling AI
This takes me to a second problem, though. Excel isn’t just giving suggestions - those tiles are documents, and documents are the start of a process, not an answer. You can see what you’ve built and what it’s doing, and how far you’ve got. The same sense of creation as process applies to Photoshop, Ableton or Powerpoint, or even a simple text edito... See more
Benedict Evans • Unbundling AI
Going back to Excel and shifting my metaphor up a level, today ChatGPT sometimes seems more like the original PCs than like Excel (or VisiCalc). It’s a general purpose technology, there’s a command line, and some stuff that’s theoretically magic, and a few things that are extremely useful to a few people, but we don’t yet have the richness of all t... See more