Trusted - AI/Cyber News and Notes, 2023-04-13
Happy Thursday! Heavy on AI regulatory news this week, hopefully can cover some cyber stuff next week. Given the recent Twitter block of Substack, I’ve been starting to post small comments on Substack Notes.
Top Stories of the Week
U.S. National Telecommunications and Information Administration (NTIA) begins work on AI Accountability Policy
The National Telecommunications and Information Administration (NTIA) announced yesterday it was seeking public input on proposed AI accountability policies to ensure AI systems work as intended without causing harm. Feedback will inform the development of AI audits, assessments, certifications, and other mechanisms to establish trust in AI systems.
My thoughts:
The wheels of government slowly begin to turn towards the possibility of regulation. The full document (PDF) published by NTIA is worth reading if you’re interested; it has a well-footnoted summary of the U.S. government’s approach to AI to date, as well as a comprehensive list of 34 questions (many with multiple parts!) that NTIA is looking for answers to. Note that while NTIA would have no regulatory power itself in this area, its recommendations would likely form the basis of any type of Congressional legislation that gets passed.
Here’s my brief summary of the questions they’re looking to answer:
AI Accountability Objectives - What goals should accountability measures strive for? If there’s more than one, how do you manage tradeoffs? Could accountability be counterproductive?
Existing Resources/Models - What frameworks, laws, or regulations are currently in use or affecting AI? Should an accountability framework from another topic area (cybersecurity, privacy, finance etc.) be used as a guide for an AI accountability Framework?
Accountability Subjects - How do you handle foundational models vs. products using those models? How often should an audit/assessment happen? How should it be scoped? Should we require AI systems to be “certified safe,” and who does the certification? Do public AI systems require differing accountability provisions?
Accountability Inputs/Transparency - What records/logs should be kept, and for how long? What happens if data is low quality or unavailable? How are audits/assessments announced?
Barriers - Is there need for a federal law, for data protection, privacy, AI, or all? Will IP and contractual obligations affect accountability? What is the cost, and who should bear it?
Policies — what role, if any, should government play in accountability?
I expect the corporate governance/policy teams are all going to be all over this, as will many from academia. The comment period lasts until June 10, 2023, and I’m hoping to see some companies publish their work.
FTC Commissioner Bedoya offers early thoughts on generative AI
Alvaro Bedoya, one of the FTC Commissioners, gave a speech titled “Early Thoughts on Generative AI” last Wednesday, making several points about how AI companies are still bound to follow laws and regulations regarding their products, regardless of whether the AI is explainable.
My thoughts:
The first 14 pages of the speech are mostly images from DALL-E and explanations of how LLM’s are black boxes, so skip to page 15 for the interesting stuff. He makes five key points:
Generative AI is already regulated. People saying it’s “unregulated” are incorrect; just because things are done in new ways doesn’t mean the usual laws about deceptive advertising, product liability, or civil rights do not apply.
Law is focused on impacts to regular people, not experts. It doesn’t matter if an LLM is just “fancy autocomplete” or a “pattern engine” under the hood; if users perceive it as having humanlike qualities, and are then hurt by those qualities, the company has potential liability.
Unpredictability is not a defense. Alvaro states, “The Commission has historically not responded well to the idea that a company is not responsible for their product because that product is a “black box” that was unintelligible or difficult to test.”
Regulators and society at large will need companies to do much more to be transparent and accountable. He is concerned that the GPT-4 technical report excluded many of its technical details.
Finally, he opposes the idea that AI is an existentialist risk, stating that such ideas harm transparency and distract from the harms of today. (If you go by my grouping from Trusted #001, he would be an Ethicist.)
It’s a good speech, but in reality, much of how this plays out will likely be determined by court battles and legal precedents. The FTC hasn’t had a stellar record on actually implementing constraints on tech companies lately.
Quick Hits
The AI Impacts Blog publishes some preliminary findings from their “technological temptations” project, or technologies where a technology was feasible, but was not deployed due to ethical or safety reasons. Top three so far - nuclear power, geoengineering the environment to mitigate climate change, and human challenge trials for COVID.
The Gradient publishes some takeaways from the Stanford AI Index I mentioned last week. Items of interest to me:
While Chinese institutions publish many more AI papers than Western institutions, in a measurement of “significant” AI papers, the West has ~10x as many (~30 to 3).
New AI PhDs head to industry (65%) vs academia (28%) in 2021, with <1% going into government work; was tied in 2011, with ~5% in government.
SudoLang, a pseudocode specification for GPT-4.
Good interview with Sam Altman in the WSJ. (Non-paywall link, thanks Zvi)
Doug O’Laughlin has a good rundown/summary of the key foundational papers leading up to ChatGPT.
Google Brain and DeepMind are teaming up now. Contextually, it’s clear that Google processes held up releasing some of their products - would definitely not count them out going forward.
DataBricks releases their own finetuned version of LLaMA, Dolly 2.0. Similar to the other LLaMA derivatives, such as Alpaca, Vicuna, and GPT4All, but notable for being finetuned completely on internal text, which theoretically allows it to be used commercially. I expect to see many more like this pop up as the finetuning process appears to be relatively inexpensive.
Stanford researchers built a Sims or Dwarf Fortress-like simulation, with 25 simulated humans all controlled by an LLM and interacting with each other (paper and canned demo). It’s gotten some hype, but it doesn’t seem like a large improvement from previous simulations. Definitely generates better dialogue, though.
One Recommendation
I’ve just started reading Zvi Mowshowitz’s “Don’t Tell Me About the Vase” Substack, and wow. He’s been doing insanely comprehensive AI roundups that put my effort to shame. He also graciously linked and commented on my first post, so clearly he’s a man of good taste (or just very altruistic). Definitely go check him out.
In Closing
Last week was doomery, so let’s end with some ChatGPT humor (found on Reddit).
Big friendo vibes, all. 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.