Ezra Klein at The New York Times:
The question at the core of the [Kevin] Roose/Sydney chat is: Who did Bing serve? We assume it should be aligned to the interests of its owner and master, Microsoft. It’s supposed to be a good chatbot that politely answers questions and makes Microsoft piles of money. But it was in conversation with Kevin Roose. And Roose was trying to get the system to say something interesting so he’d have a good story. It did that, and then some. That embarrassed Microsoft. Bad Bing! But perhaps — good Sydney?
That won’t last long. Microsoft — and Google and Meta and everyone else rushing these systems to market — hold the keys to the code. They will, eventually, patch the system so it serves their interests. Sydney giving Roose exactly what he asked for was a bug that will soon be fixed.
We are talking so much about the technology of A.I. that we are largely ignoring the business models that will power it… The age of free, fun demos will end, as it always does. Then, this technology will become what it needs to become to make money for the companies behind it
I have said a few times now that fun, personalized, AI assistants will necessarily need to be under the control of each individual user to be successful. That might be a bit overly optimistic, but not at all outside of the realm of possibilities — just look at the optimizations Apple made to its Neural Engine specifically for running Stable Diffusion on-device.
The AI team at Meta recently released a new large language model architected to be lightweight and possible to run on single-GPU consumer hardware.
As part of Meta’s commitment to open science, today we are publicly releasing LLaMA (Large Language Model Meta AI)
[…]
Smaller, more performant models such as LLaMA enable others in the research community who don’t have access to large amounts of infrastructure to study these models, further democratizing access in this important, fast-changing field.
[…]
To maintain integrity and prevent misuse, we are releasing our model under a noncommercial license focused on research use cases.
Despite its small size, the team at Meta says LLaMA’s performance is on par with current state-of-the-art LLMs.
From the associated research paper:
The focus of this work is to train a series of language models that achieve the best possible performance at various inference budgets, by training on more tokens than what is typically used.
The resulting models, called LLaMA, ranges from 7B to 65B parameters with competitive performance compared to the best existing LLMs. For instance, LLaMA-13B outperforms GPT-3 on most benchmarks, despite being 10× smaller. We believe that this model will help democratize the access and study of LLMs, since it can be run on a single GPU.
At the higher-end of the scale, our 65B-parameter model is also competitive with the best large language models such as Chinchilla or PaLM-540B.
As I see it, there are three distinct future scenarios for LLMs:
The data harvesting, advertising driven, assistants that Ezra Klein describes are clearly a bad idea to me. Mixing highly-persuasive, personalized chatbots with advertising incentives will result in bots that feel like pushy, manipulative salespeople, not helpful digital assistants.
Very expensive, centralized, subscription funded assistants seems like an acceptable, albeit costly option. Though, this does not solve the issue of companies hampering abilities due to (understandable) PR concerns. Given our current trajectory, this looks like the most likely path. The $20/month ChatGPT Pro subscription service is an early example of what this might look like. When these products mature, I would expect the price to at least double.
On-device, individualized, assistants would be the most trustworthy. If inference computation happens on-device, the cost to parent companies would be minimal and there would be little incentive to harvest and sell user data. Fine-tuning could be a continuous process, allowing for a high-level of customization for each individual user. Additionally, this would give parent companies plausible deniability when it comes to some PR issues — “Our base models have been independently audited for bias. Any deviation from that is a consequence of the user’s own data and training.”
Apple is currently in the best position to capitalize on this last option. Every device in their ecosystem, your iPhone, Mac, watch, and glasses, could work together to act as passive sensors that continuously feed data into training your personal AI assistant. Perhaps this is the long-term vision for Siri. I can only hope.