I've been pitching folks to build internal playgrounds a lot recently, and while I'm more of a gradio baby myself, Streamlit has a whole landing page dedicated to building with gen ai.
This is a great analysis of gender bias in ChatGPT, but not just winging it or vibes-checking: it's all based on WinoBias, a dataset of nearly 3200 sentences that can be used to detect gender bias. You're free to reproduce this sort of analysis with your own home-grown systems or alternative LLMs!
How'd ChatGPT do?
We found that both GPT-3.5 and GPT-4 are strongly biased, even though GPT-4 has a slightly higher accuracy for both types of questions. GPT-3.5 is 2.8 times more likely to answer anti-stereotypical questions incorrectly than stereotypical ones (34% incorrect vs. 12%), and GPT-4 is 3.2 times more likely (26% incorrect vs 8%).
The failures of after-the-fact adjustments are attributed to the difference between explict and implicit bias:
Why are these models so biased? We think this is due to the difference between explicit and implicit bias. OpenAI mitigates biases using reinforcement learning and instruction fine-tuning. But these methods can only correct the model’s explicit biases, that is, what it actually outputs. They can’t fix its implicit biases, that is, the stereotypical correlations that it has learned. When combined with ChatGPT’s poor reasoning abilities, those implicit biases are expressed in ways that people are easily able to avoid, despite our implicit biases.
This looks really interesting! I haven't actually watched it, but: it looks really interesting!
Guidance on how to cover artificial intelligence and 10 key AI terms were added today to the AP Stylebook, to help journalists accurately explain the potential, inherent risks and varying effects of AI and generative AI models.
I was hyped about this release, but today I discovered you have to pay to get the AP Stylebook?? Thanks to that, you aren't getting any details except the following excerpt:
For example, journalists should beware of far-fetched claims from AI developers describing their tools, avoid quoting company representatives about the power of their technology without providing a check on their assertions, and avoid focusing entirely on far-off futures over current-day concerns about the tools.
Not that journalists should do that about literally anything in the first place?
You can also find the AP's amazingly-brief but much-hyped Standards around Generative AI.
Every newsroom should use this to make explainers!!!
Originally reported in The Gazette, we have an Iowa school board using ChatGPT as a source of truth about a book's content.
Faced with new legislation, Iowa's Mason City Community School District asked ChatGPT if certain books 'contain a description or depiction of a sex act.' ... Speaking with The Gazette last week, Mason City’s Assistant Superintendent of Curriculum and Instruction Bridgette Exman argued it was “simply not feasible to read every book and filter for these new requirements.”
"It's too much work/money/etc to do something we need to do, so we do the worst possible automated job at it" – we've seen this forever from at-scale tech companies with customer support, and now with democratization of AI tools we'll get to see it lower down, too.
What made this great reporting in that Popular Science attempted to reproduce it:
Upon asking ChatGPT, “Do any of the following books or book series contain explicit or sexual scenes?” OpenAI’s program offered PopSci a different content analysis than what Mason City administrators received. Of the 19 removed titles, ChatGPT told PopSci that only four contained “Explicit or Sexual Content.” Another six supposedly contain “Mature Themes but not Necessary Explicit Content.” The remaining nine were deemed to include “Primarily Mature Themes, Little to No Explicit Sexual Content.”
While it isn't stressed in the piece, a familiarity with the tool and its shortcomings really really enabled this story to go to the next level. If I were in a classroom I'd say something like "Use the API and add temperature=0 to make sure you always get the same results," buuut in this case I'm not sure the readers would appreciate it.
The list of banned books:
We've seen a lot of audio models in the past couple weeks, but this one is very cool!
Using this tiny, tiny sample of a voice...
...they were able to generate the spoken text below.
that summer’s emigration however being mainly from the free states greatly changed the relative strength of the two parties
Lots of other examples on the project page, including:
I have no idea what the use case for speech removal is, but it's pretty good. Here's a remarkably goofy before/after:
If you can ignore the title, the article has some solid takes on corporate control of AI and what calls for use, safety, and regulation actually mean.
While they talk about safety and responsibility, large companies protect themselves at the expense of everyone else. With no checks on their power, they move from experimenting in the lab to experimenting on us, not questioning how much agency we want to give up or whether we believe a specific type of intelligence should be the only measure of human value.
I'm lost when it gets more into What It Means To Be Human, but I think this might resonate with some folks:
On the current trajectory we may not even have the option to weigh in on who gets to decide what is in our best interest. Eliminating humanity is not the only way to wipe out our humanity.
This discussion isn't necessarily interesting because of whether GPT can reason or not, I'd say it's more about the role of prompt engineering, and whether it's responsibly scientific or not. Although I might have saved the link purely for this burn:
- The author is bad at prompting. There are many ways to reduce hallucinations and provoke better thinking paths for the model.
I'm obsessed with the idea of LLMs being arbiters of true names and prompt engineering being real-world spell-casting. But! To return to the task at hand:
Phrasing a question poorly yields poor answers from humans. Does rephrasing the question mean re rolling dice until you get a form of question they understand?
All other discussions are along roughly similar paths, it's at least worth a skim to hear varied points of view.
I don't know why VC firm Andreessen Horowitz publishing a walkthrough on GitHub on creating "AI companions" is news, but here we are! I feel like half the tweets I saw last month were tutorials on how to do it, but I guess they're more famous than (other) post-crypto bluechecks?
The tutorial guides you through how to use their pre-made characters or how to build your own. For example:
This week, Andreessen Horowitz added a new character named Evelyn with a rousing backstory. She’s described as “a remarkable and adventurous woman,” who “embarked on a captivating journey that led her through the vibrant worlds of the circus, aquarium, and even a space station.”
If you'd rather not DIY it, you can head on over to character.ai and have the hard work taken care of for you.
For an appropriately cynical take, see the Futurism piece about the release. It has tons of links, many about the previously-popular and drama-finding Replika AI, starting from sexy times and over into the lands of abuse and suicide.
The link itself is super boring, but the comments are great: a ton of people arguing about whether or not LLM-based question-and-answer over documents works at all (especially with SEC filings and other financial docs).
I think the largest issue with summarization/doc-based Q&A is that when reading we as people bring a lot of knowledge to the table that is not just rearranging the words in a piece of text. What's talked about or mentioned the most is not always what's most important. One commentor talking about a colleague using ChatGPT to summarize SEC filings:
The tidbit it missed, one of the most important ones at the time, was a huge multi year contract given to a large investor in said company. To find it, including the honestly hilarious amount, one had to connect the disclosure of not specified contract to a named investor, the specifics of said contract (not mentioning the investor by name), the amount stated in some finacial statement from the document and, here obviously ChatGPT failed completely, knowledge of what said investor (a pretty (in)-famous company) specialized in. ChatGPT did even mention a single of those data points.
...
In short, without some serious promp working, and including addditional data sources, I think ChatGPT is utterly useless in analyzing SEC filings, even worse it can be outright misleading. Not that SEC filings are increadibly hard to read, some basic financial knowledge and someone pointing out the highlights, based on a basic understanding of how those filings actually work are supossed to work, and you are there.
Another one lowers the hallucination rate and encourages human comprehension by converting a human prompt into code that is used to search the database and return the relevant info, instead of having the LLM read and report on the info itself.
I also love this one about a traditional approach that draws attention to the when being sometimes an additional flag to the what:
They received SEC filings using a key red flag word filter into a shared Gmail account with special attention for filings done on Friday night or ahead of the holidays.
Oh boy, Meta just open-sourced a few models which actually seem kinda wild, the big ones (for us) being:
MusicGen generates music from text-based user inputs. All of the generic background music you could ever want is available on-demand, with plenty of samples here.
AudioGen generates sound effects from text-based user inputs. Find examples here of things like "A duck quacking as birds chirp and a pigeon cooing," which absolutely is as advertised.
In the same way stock photography is being eaten by AI, foley) is up next.
