Trusted AI - The Astute AI Anthology, 2023-04-27
No longform post this week, unfortunately, but working on lots of great ideas for next week. Email me if there’s something in particular you’d like me to dive deep on!
The Trusted Top Ten in 10 Minutes or Less
Snapchat released “My AI” last week to all 750m+ users, which included the lovely “feature” that it is pinned to the top of your chat list, and cannot be removed without paying $4/month for Snapchat+. As one might expect, the app’s ratings are collapsing. Gating access behind a subscription made sense; I’m not sure what the business proposition is for making the free ad-supported model worse. I’m going to pay close attention to how Snap handles this going forward; this is probably the best example of an LLM being exposed to a large population that isn’t actively seeking it out.
Google Brain and DeepMind merge into Google DeepMind. I’m looking forward to the inevitable book/books on “what happened to Google’s AI lead” - so much foundational tech was Google Brain (TensorFlow, TPU’s, JAX, GPT) or DeepMind (AlphaGo, AlphaFold) and Google hasn’t been able to convert it into products due to corporate culture. Neither one of this organizations really strikes me as being good at “product,” so I’m not sure how much good this is going to do, and it probably signals the end of a lot of the cutting-edge research. For much more on this, read this good essay from Brian Lee, an ex-Google Brain researcher.
Google Cloud announces Google Cloud AI Workbench. a set of tools built on a custom Google LLM, “Sec-PaLM.” The big announcement here is VirusTotal Code Insight, in my opinion; malware analysis remains a largely manual process, and automation at scale absolutely opens up some new security capabilities that were normally out of reach to most. I’m less certain how well the LLM will work with the other cases (threat intel and SIEM parsing), but they’re using the right words to keep me interested. (By the way, I think it’s telling that this is coming out of the Google Cloud team, the one area where Google has been forced to be an innovator, not an incumbent.)
Nvidia announces an open-source toolkit for LLM safety/security, NeMo Guardrails. I’m excited to see more actual development in the security space and not just OpenAI API wrappers. I disagree slightly with Simon Willison’s take on this (“You can’t solve AI security problems with more AI”). He feels that we need a 100% solution against prompt injection; I think that business needs vary, a sufficiently elaborate set of LLM audit mechanisms can get you to whatever level of 99% you need, and this framework is a good start to the problem. I’m going to tackle this in a future longform article.
A new Pew Research poll on AI in the workplace has a few interesting insights. Most of what’s reported is common sense: people are very unsure about how AI will affect them personally, but think it’ll have a big impact on workers in general. When you ask them specific questions about AI tools, everybody hates the idea; nobody wants an AI monitor tracking them (duh); nobody wants an AI to evaluate their resume, etc. About the only thing that polls close to neutral is using AI to ensure all applicants are treated fairly.
Greg Brockman gave a good 30 minute TED Talk on ChatGPT and what’s coming. It’s broken into two parts; the first 15 is a demo of ChatGPT and some of the new plugins, the second half is a Q&A with some good OpenAI stories. I think the examples shown here of how they are externalizing the way the ChatGPT AI “thinks” is super important; as I’ve discussed before, the largest barrier to AI adoption is trust, and “showing your work” is super important compared to other technologies that are more well-understood. (OpenAI, if you’re reading this, I REALLY want the Excel plug-in that can make charts via natural language prompts.) Honorable mention to Greg saying “We were super scared internally that people would use GPT-3 to try to tip elections; instead, it was Viagra spam.”
Replit raised a ton of money, and also announced a bunch of new dev tools at their developer day. Keep an eye on Replit; there’s been some opportunity in the “developer support” space since Heroku has lost steam, and Replit is aggressively pushing here.
Anthropic publishes a long announcement supporting investment in NIST to help develop AI standards. This is probably more of interest to me than to the audience, but I’m highlighting this as I think this is the kind of thankless but critical work that is essential for the government to do. The NIST CSF was pretty important for cybersecurity as a discipline to evolve from “complete lawlessness” to “frontier town.” Maybe something similar can happen here.
OpenAI is now letting you turn off chat history, which also prevents your conversations from being used to train new models, I’m certain this feature was released for user’s best interests, and not the looming specter of incoming regulations I discussed two weeks ago.
Reuters has a good summary of Meta’s AI strategy. Like Google/MS, Meta tried to develop their own custom hardware; apparently it didn’t work, and Meta pivoted last year to buying hardware from Nvidia instead.
One Recommendation
I’ve enjoyed reading Timothy Lee’s reporting at Ars Technica, and he recently started his own AI newsletter. You should check it out.
In Closing
An update to a classic.
Have a good weekend!
Standard disclaimer: All views presented are those of the author and do not represent the views of the U.S. government or any of its components.