172 | 🤑 🙄 Law firm relies on ChatGPT to set its fees

In today’s Brainyacts:

1. Google’s AI problem is our problem

2. First 10 mins with Adobe’s new AI Assistant

3. ChatGPT goes rogue and other AI model news

4. Law firm relies on ChatGPT to set fees and other AI-related content

👋 to new subscribers!

To read previous editions, click here.

Lead Memo

🔥🚮 Google’s AI Problem is Our Problem

Google recently released a Gemini AI image generator. It didn’t go so well. It was outright refusing to depict white people in any way. Even when people asked it to create an image of America’s founding fathers - no white people are depicted.

Now let’s not get all triggered and start yelling “woke” and “racist!” Too many are knee-jerking into that Pavlovian response. Yes, I take things like this seriously but resting on ideological arguments is not helpful. Instead, I want to get a bit more substantive here.

But first, in case you missed things- let’s begin with the pictures - working backward from where we are now to how this all started:

Ok, if you want more, just head over to X (aka Twitter) and search “gemini.” you will have no shortage of the typical social media fistfights.

So what do I want to say about this? Simple. We have an AI Bias Paradox. Right now AI is not autonomous (thankfully) in that it can not create new models or do anything without human intervention. Humans are programming the AI model, aligning it, fine-tuning it, and developing system prompts to convert our prompts into more precise prompts that help the AI model create its “magic.”

What is a system prompt or prompt modification system?

Basically, this is when the AI model takes the user prompt and converts or edits it in some way before submitting it to create a response. This is invisible to the user! For instance, if you use OpenAI’s DALL-E image generator, your prompt is converted on the back end or “under the hood” in a way that makes DALL-E perform better and more predictably. This helps mitigate bad actors and other ethical concerns. For a demonstration of the DALL-E system prompt, click here.

Gemini’s Image Generation System Prompt

What we have learned so far is that there was a system prompt that was converting user prompts in a manner that was supposed to depict humans in a diverse way. We know that many image generators have a bias for what a “beautiful woman” looks like. When these generators were first released, they tended to depict the same type of woman.

Here is a supposed example of the system prompt Gemini was using:

The Takeaway: The AI Bias Paradox

• Humans are vital to the development of AI.

• Humans are biased and have agendas.

• Bias can be injected into AI models through training data, fine-tuning, and system prompts (among other ways).

• Until humans are a) no longer biased (never) or b) are not necessary for the development of AI (scary), we will ALWAYS have biased AI models!

We need to get used to it and build systems to help us detect or at least know what “temperature” the model is running at. Temperature is a proxy for how creative, restricted, or otherwise tuned the model will be with its replies.

Spotlight

🤓📚 Adobe’s AI Assistant: Great word search or reasoning engine?

Adobe is adding a new gen-AI feature to its Acrobat PDF management software—called “AI Assistant in Acrobat”—which will “transform the digital document experience.”

They describe it as a “conversational engine” because it will summarize long documents, answer questions, and format information to share in emails, reports, and presentations. But does it do more than this?

Adobe confirmed that—although the model was built using the OpenAI platform—it’s been developed in line with their AI ethics processes, so no data will be stored or used to train it.

The new feature (which can be used with all Adobe document formats, including Word and PowerPoint) is available now (in beta) to paying Acrobat users.

I captured my first 10 minutes with it in the video below. You will see that I try to get it to reason versus just search and locate. I even try to get it to give me quasi-legal advice.

AI Model Notables

► Google has secured a deal to train its AI on Reddit's user content

► Google has released two open-source models based on Gemini. They call these models Gemma. A benefit of open source is that once it is released, its license gives users full creative autonomy. This is a powerful guarantee of technology access for developers and end users. Another benefit is that open-source technology can be modified to fit the unique use case of the end user, without restriction.

► Chatgpt went ‘off the rails’ with wild hallucinations, but OpenAI says it's fixed

► Google DeepMind forms a new org focused on AI safety

► What was OpenAI’s text-to-video Sora trained on? OpenAI will not tell us

► 1 minute of OpenAI's Sora video may take over an hour to generate

► BCLP has developed a lease reporting tool which is being rolled out to UK clients this month.

News You Can Use:

➭ Law firm nearly got away with overbilling New York City double the hourly fee by asking ChatGPT to calculate the bill

➭ How many news websites block AI crawlers? 48% of the most widely used news websites across ten countries were blocking OpenAI’s crawlers

➭ Nvidia (AI chip makers) posts revenue up 265% on booming AI business

➭ Top in-house counsel band together to experiment with Gen AI

➭ Saudi wealth fund pitches Mohammed bin Salman’s kingdom as AI hub outside US: 'we are fairly well positioned'

➭ Natalie Portman says AI could put her out of a job 'soon': ‘There’s a good chance'

Who is the author, Josh Kubicki?

Some of you know me. Others do not. Here is a short intro. I am a lawyer, entrepreneur, and teacher. I have transformed legal practices and built multi-million dollar businesses. Not a theorist, I am an applied researcher and former Chief Strategy Officer, recognized by Fast Company and Bloomberg Law for my unique work. Through this newsletter, I offer you pragmatic insights into leveraging AI to inform and improve your daily life in legal services.