Language Log

Despite the evidence of my most recent relevant post, the best current speech-to-text systems still make mistakes that a literate and informed human wouldn't.

In this recent YouTube video on the history of robotics research, the automatic closed-captioning system renders "DARPA" as "Dartmouth":

It's hard to keep up with the waves of hype and anti-hype in the LLM space these days.

Here's something from a few weeks ago that I missed — Xiaoxuan Wang et al., "SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models", arxiv.org 7/20/2023:

…if you haven't noticed, is good. There are many applications, from conversing with Siri and Alexa and Google Assistant, to getting voicemail in textual form, to automatically generated subtitles, and so on. For linguists, one parochial (but important) application is accurate automatic transcription of speech corpora, and the example that motivates this post comes from that world.

…though they often do a credible job of faking it.  An interesting (preprint) paper by Konstantine Arkoudas, "GPT-4 Can't Reason", brings the receipts.

There's a puzzling new proposal for watermarking AI-generated text — Alistair Croll, "To Watermark AI, It Needs Its Own Alphabet", Wired 7/27/2023:

We need a way to distinguish things made by humans from things made by algorithms, and we need it very soon. […]

Fortunately, we have a solution waiting in plain sight. […]

If the companies who pledged to watermark AI content at the point of origin do so using Unicode—essentially giving AI its own character set—we’ll have a ready-made, fine-grained AI watermark that works across all devices, platforms, operating systems, and websites.

Today's Non Sequitur:

From Sarah Gao and Andrew Gao, "On the Origin of LLMs: An Evolutionary Tree and Graph for 15,821 Large Language Models", arxiv.org 7/19/2023:

That's not a vinyl — it's a "radial dendrogram" — showing the evolutionary tree of nearly 6,000 Large Language Models posted at Hugging Face. Zeroing in on one quadrant, so you can read the labels:

Ashley Belanger, "OpenAI, Google will watermark AI-generated content to hinder deepfakes, misinfo", ars technica 7/21/2023:

Seven companies — including OpenAI, Microsoft, Google, Meta, Amazon, Anthropic, and Inflection —- have committed to developing tech to clearly watermark AI-generated content. That will help make it safer to share AI-generated text, video, audio, and images without misleading others about the authenticity of that content, the Biden administration hopes.

The link goes to a 7/21 White House with the title "FACT SHEET: Biden-⁠Harris Administration Secures Voluntary Commitments from Leading Artificial Intelligence Companies to Manage the Risks Posed by AI". One of that document's many bullet points:

• The companies commit to developing robust technical mechanisms to ensure that users know when content is AI generated, such as a watermarking system. This action enables creativity with AI to flourish but reduces the dangers of fraud and deception.

Joe Marshall, "As AI cheating booms, so does the industry detecting it: ‘We couldn’t keep up with demand’", The Guardian 7/5/2023:

Since its release last November, ChatGPT has shaken the education world. The chatbot and other sophisticated AI tools are reportedly being used everywhere from college essays to high school art projects. A recent survey of 1,000 students at four-year universities by Intelligent.com found that 30% of college students have reported using ChatGPT on written assignments.

This is a problem for schools, educators and students – but a boon for a small but growing cohort of companies in the AI-detection business. Players like Winston AI, Content at Scale and Turnitin are billing for their ability to detect AI-involvement in student work, offering subscription services where teachers can run their students’ work through a web dashboard and receive a probability score that grades how “human” or “AI” the text is.

Ilia Shumailov et al., "The Curse of Recursion: Training on Generated Data Makes Models Forget", 5/31/2023:

What will happen to GPT-{n} once LLMs contribute much of the language found online? We find that use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear. We refer to this effect as Model Collapse and show that it can occur in Variational Autoencoders, Gaussian Mixture Models and LLMs.

Last year, I expressed considerable skepticism about the prospects for accurate detection of text generated by Large Language Models ("Detecting LLM-created essays?", 12/20/2022). Since then, many new systems claiming to detect LLM outputs have emerged, notably Turnitin's "AI writing detector".

In a recent post on AI Weirdness ("Don't use AI detectors for anything important", 6/30/2023), Janelle Shane presents multiple examples of multiple kinds of failure, and explains why things are not likely to change.

From Daniel Deutsch:

We had a long drive yesterday, so we listened to a “robot” reading the entire indictment. It certainly isn’t flawless, but I was surprised by how good it is, especially when it gets “excited” while enacting dialogue.

Indeed, the text-to-speech quality is quite good — though unfortunately they don't tell us which TTS software they used.

Here's the opening, which is indeed entirely clear and even nearly natural-sounding:

I've recently seen many articles like this one, "You probably don't need to learn to code anymore" (Medium 6/5/2023), arguing that Large Language Models will make human programming (and human programmers) unnecessary. These arguments puzzle me, because my experience with LLMs suggests that they can't be relied on even for very simple programming tasks. After the fold, I'll give a recent example from (the experimental version of) Bard.