Exploration #99

Do Chatbots Sound Better in French?

Image of a robot cat colored like the French flag.

Image made with DALL-E 3

Hate the Longscroll? Why Not Jump Around?

Hi all. Happy Leap Year! This week we’ve got two new chatbots to discuss, two established chatbots riding the struggle bus, new products or features from Stable Diffusion and Adobe, and a little VR and metaverse thrown in. But first…

Google Gemini Overcorrects

Bias in AI models has been a consistent theme in this newsletter over the past year. And one of the most obvious ways that these biases are represented is in image generation, a significant point of discussion in the most recent installment in Current’s user’s guide to generative AI.

To combat that bias, companies have been placing guardrails on the output of their bots. As defined by Cal Al-Dhubaib in his Medium post, “Guardrails are the set of filters, rules, and tools that sit between inputs, the model, and outputs to reduce the likelihood of erroneous/toxic outputs and unexpected formats, while ensuring you’re conforming to your expectations of values and correctness.“ Anthropic has gotten the most press in this regard, thanks to its “constitutional AI”. And OpenAI added series-level rules to DALL-E in July 2022 so that the outputs are less controversial (similar techniques are used for ChatGPT).

It seems like similar guardrails caused Google Gemini to not depict white people in images where the only historically accurate image would be of a white person (i.e., the pope, or America’s founding fathers, or (most egregious) WWII Nazi soldiers). As a result, Google suspended the depiction of all humans via Gemini as they work to retune its outputs.

It’s certainly not the first time a mega-corporation had to walk back an embarrassing AI product. But what made this viral was the fact that there was something in it to anger everyone. The far right saw a woke conspiracy to replace white people. And everyone else saw racially diverse people depicted in Nazi uniforms. A perverse balance of outrage.

Keep in mind, for almost two years, we’ve been seeing gAI images - and stories about gAI images - where white men were predominantly depicted in positions of power and competence, and where everyone else was often reduced to stereotypes. Except for the Nazi images (which, yes, are unequivocally appalling), this unintentional overcorrection by Google Gemini needs to be kept in perspective. Generative AI is still emerging tech and Google’s debacle reminds us that other (anti-diverse) biases are replicated daily in ways far more subtle and insidious.

I feel confident in saying that no one in the editorial structures of both the PBS system and NPR network would publish any of these images out of context. Public media organizations have values, and we have competent people at the editorial helm in our organizations implementing those values. Controversies like these don’t mean these tools aren’t worth using, but if they are going to be used in a professional capacity then they need to be used by professional artists and editors.

Do Chatbots Sound Better in French?

Fun fact: “ChatGPT,” when spoken out loud (“chat j'ai pété”) translates in French to, “Little cat, I farted.”

That alone should be worth the cost of your subscription this week. But wait, there’s more. We’ve got a new quality chatbot on the scene.

French AI startup, Mistral (also the name of a mighty wind) has slowly been building up its product portfolio since last April. The team originally released their first open-source models last summer and they were quickly criticized for their lack of guardrails. The versions of those models that were available by December performed at the same level as GPT-3.5 (aka, the free versions of ChatGPT and Microsoft Copilot). Parallel to that, Mistral was also building up larger, proprietary models. And this week they announced the release of a new chatbot, Le Chat built on a new model, as well as a partnership Microsoft.

What you need to now is that there is a new kid in town. This one is a little edgy as it doesn’t have the guardrails that Anthropic and OpenAI have put into their comparable models (let alone Google Gemini, see above). Having only been released a day (as of this writing) I’ve only just started playing with it. So I’ll have more impressions in a future Exploration. Thus far I’m finding it has a chipper screenside manner, but only has a loose command of facts.

At launch, Mistral supports English, German, Spanish, Italian and, yes, French. It also supports corporate team accounts that include API access. You can visit Mistral’s Le Chat page if you want to set up an account and experiment with the beta.

Okay, on to the links…

If You Click Only One…

Is it real or AI Generated? (AI Guess It) - I know, weird to make a short guessing game this week's 'One'. But many of us work in video, or close to it, and it's worth studying the state of the generative art. So give it a go, and message me if you beat my score: 8/10 correct

Things to Think About…

What the Media Is Getting Wrong About AI (Evan Armstrong - Napkin Math) - A good takedown of a takedown that addresses some major critiques about the zeitgeist around generative AI.

How much electricity does AI consume? (James Vincent - The Verge) - As I often say, know your tools. Even the most ethical of gAI tools require massive amounts of energy. If this type of consumption is a concern to you, be judicious in how you use these tools. For example, image generation takes way more electricity than chatting with a bot. So you might use a bot to refine your image prompts, instead of iterating through a couple dozen version to get something that matches your vision. You also have power over when you choose to use these tools to assist your work.
—Related, as reported by Jeffrey Eastin of Reuters, OpenAI CEO Altman says at Davos future AI depends on energy breakthrough

Seniors enjoy virtual reality, new study finds (Terry Spencer - AP via ABC News) - Multiple times in recent years I've been asked questions about emerging media that seem to assume older audiences will being intentionally left behind. When it comes to VR, at least, we see interest and usage well beyond the stereotype of young, male gamers.
—Another case in point, as reported by Chris Matyszczyk in ZDNet, I met someone who loves Meta's metaverse and now my eyes won't shut

Things to Know About…

Mistral AI Launches High-Performance AI Model and Chatbot to Challenge ChatGPT, Claude and Gemini (Jose Antonio Lanz - Emerge News) - Following the intro column, I'm corralling coverage of Mistral's Le Chat launch here.
Mistral partners with Microsoft, announces new large AI model with chat app (Shubham Sharma - VentureBeat)
Mistral shocks AI community as latest open source model eclipses GPT-3.5 performance (Carl Franzen - VentureBeat)
Microsoft partners with Mistral in second AI deal beyond OpenAI (Tom Warren - The Verge)

CNN interacts with cutting-edge AI technology (Becky Anderson - CNN) - Ignore CNN’s awful headline, if you can, because Le Chat may not be the only chatbot to challenge the status quo. The cutting edge AI tech being fêted by CNN is a Language Processing Unit (LPU), a new approach to the compute under the hood of generative AI. It’s called Groq (not be confused with Elon Musk's "Grok"), and you can interact with cutting-edge AI technology yourself at Groq.com

Google pauses AI image generation after diversity controversies (Megan Morrone - Axios) - As above, I'm going to corral a few stories here that help contextualize what last week’s debacle means.
Google explains Gemini’s ‘embarrassing’ AI pictures of diverse Nazis (Emma Roth - The Verge)
Google Gemini’s ‘wokeness’ sparks debate over AI censorship (Carl Franzen - VentureBeat)
Google’s ‘Woke’ Image Generator Shows the Limitations of AI (David Gilbert - Wired)
Google’s Gemini AI was mocked for its revisionist history, but it still highlights a real problem (Chris Stokel-Walker- FastCompany)
The Deeper Problem With Google’s Racially Diverse Nazis (Chris Hilliard - The Atlantic)
Gemini image generation got it wrong. We'll do better. (Prabhakar Raghavan - Google)

ChatGPT goes ‘off the rails’ with gibberish answers (Carl Franzen - VentureBeat) - Turns out, Gemini wasn't the only bot on the struggle bus last week.

Microsoft's Copilot AI search is making up fake Vladimir Putin quotes from press conferences that never happened (Rani Molla - Sherwood) - Just a reminder. Microsoft Copilot is essentially ChatGPT and, despite the halo of Microsoft branding, is subject to the same issues of AI hallucinations that plague all chatbots at this time. It can be a great creative companion, but should be treated like an unverified source and fact-checked extensively before it’s text makes its way into your copy.
—And this article by Sabrina Ortiz in ZDNet gives us a glimpse of how Microsoft has been fine tuning Copilot (while also announcing a tool devs can use for testing their own models): Microsoft releases its internal generative AI red teaming tool to the public

Adobe Acrobat adds generative AI to ‘easily chat with documents’ (Jess Weatherbed - The Verge) - Supposedly, when you limit the field of inquiry, hallucinations (see above) are not an issue. It stands to reason that natural language queries will eventually supplant basic keyword searches for documents (and perhaps operating systems) they way they have with internet searches. And, to me that's an eventual goal we should all have for our websites.
—See Adobe's announcement of the new feature here: All you have to do is ask. Chat with your doc.

How many news websites block AI crawlers? (Dr. Richard Fletcher - Reuters Institute/Oxford) - If you've been discussing your organization's (potential) policy on chatbots scraping content from your website, this study could help focus the discussion. Interesting stat: "By the end of 2023, 48% of the most widely used news websites across ten countries were blocking OpenAI’s crawlers. A smaller number, 24%, were blocking Google’s AI crawler." I find the inconsistency illogical, but also not surprising. I'm still of the opinion that we in public media need to let the crawlers crawl. Our content is quality content that can help future versions of models improve.
—Jump straight to the PDF here.

NYT plans to debut new generative AI ad tool later this year (Sara Fischer - Axios) - While the NYT is one of the news websites (see above) that is blocking AI crawlers, they are certainly not anti-AI when it comes to monetization.

AI-Generated Propaganda Is Just as Persuasive as the Real Thing, Worrying Study Finds (Jordan Pearson - Vice) - This tracks to other reports that generative AI tools shift the bell curve of talent left, crowding but not exactly replacing those creative individuals with talent and experience at a particular craft. In this case though, the craft is state-sponsored propaganda.
—Also, this is likely the last Vice article you're going to see in this newsletter, as reported by Marika Oi at the BBC in Vice Media stops publishing on website and cuts hundreds of jobs
—And related: What Zombie Vice Will Try to Do Next by Alex Kirshner in Slate.

Stable Diffusion 3.0 debuts new diffusion transformation architecture to reinvent text-to-image gen AI (Sean Michael Kerner - VentureBeat) - As we learned with OpenAI's Sora, these days you don't need to drop your MVP on the world Beyonce-style when you can just release a highly curated preview. Such was this case this past week with one of the original text to image generators, Stable Diffusion.
—You can sign up for 3.0 at Stability.ai.

Deutsche Telekom showcases app-less AI smartphone concept (Olivier Sorgho - Reuters) - While this concept feels ahead of its time, I definitely think this one potential (even likely) future. The path out of the app wilderness, is something any of us maintaining (or relying on a centralized developer with whom we contract to maintain) an app should be beginning to consider.
—And here's a thinly veiled bit of content marketing on the topic from T-Mobile, AI phone: Deutsche Telekom wants to free smartphones from apps

And finally…

Client Did Not Pay? (Dragoi Ciprian - aka “kleampa” - via Reddit) - And finally, a public service announcement...always pay your web developers. It's basic human decency, and also your website might not fade away.

Have a creative, productive week!

Image of a robot cat with the bleu, blanc, and rouge coloring of the French flag.

Image Made with DALL-E 3

Reply

or to participate.