- Inclined.ai
- Posts
- 🤖 OpenAI Admits Faults
🤖 OpenAI Admits Faults
PLUS: A Fresh Way to do GenAI
What's up? You're reading Inclined AI. It’s Friday, then Saturday, Sunday (what!).
Here's what is on tap today:
OpenAI is working on ChatGPT’s behavior
ControlNet is freakin’ sweet
The Vatican and pastors weigh in on AI morality
And I want to try some AI-crafted beers
Prompt: robot trains robot dog Style: Pop Art
OpenAI HAS HEARD THE CRITICISM
The company behind ChatGPT is trying to get its arms around the narrative.
Between DAN jailbreaks, failures to answer politically charged prompts, and a general weak spot for fault lines, recent mistakes prove to the world that we’ve barely scratched the surface.
Yesterday, a responding blog post emphasized the need for transparent, explainable, and verifiable AI systems. It was, low-key, quite the gem to read.
In it, OpenAI attempts to lay its cards on the table by discussing:
What a fine-tuning model is and showcasing their guidelines for that process
Their new research efforts and outlining all the clear feedback they received (with ways to address it)
How concepts like red teaming and constitutional AI will help improve these models over time
Is All This Too Little, Too Late?
There’s no better time than the present for OpenAI. Trying to please everyone here is impossible, which is why they say future versions will become customizable.
In the post, OpenAI compares training ChatGPT to training a dog. But their dog is roughhousing, leaving a mess everywhere but the grass.
Their solution is to offer people the chance to choose their own puppies while they build more sophisticated dogs to keep your new pooch in line.
Okay, this metaphor is becoming something only Snoop can appreciate; let me try and bring this home.
I like the comparison people make between the AI Boom and the iPhone. We don’t know where this goes next, and one day will laugh at how archaic these chats looked.
Till then, enjoy the ride and trust that smart people are pursuing these problems as you read this. Don’t even trip, dog.
NOTE: Microsoft also released their thoughts on New Bing falling under similar logic.
We’re happy to announce a new PREMIUM inclined.ai offering.
Upgrade today for less than the price of a latte a month and get benefits like:
A weekly Sunday deep dive into AI trends
Access to Q&A privileges to get your biggest AI questions answered
Exclusive discounts & invites to beta AI tools
GAN I HELP YOU WITH THAT?
Nope, that wasn’t a typo, just a crappy pun.
There’s a fun new generative AI tool called Scribble Diffusion. You sketch out a basic kindergarten version of something, then prompt the AI to turn that drawing into a realistic image.
The site is addicting and it partially runs off a week old research project called ControlNet. Trust me, all this ties into my pun and help explains how this happens.
Browse Instagram, Twitter, or Reddit today, and you’re bound to see some images sourced as “ControlNet.”
So, whatever happened to DALL-E and Stable Diffusion?
Don’t panic. The new kid on the block, ControlNet, is an extension --like img2img--used with Stable Diffusion. All it does is make your generated scenes more accurate and way better looking.
GAN Walked so ControlNet Could Run
A generative adversarial network (or GAN) is nothing new in machine learning. It’s like two people who make each other better.
One is an expert art forger (generator) that replicates an image. The other is an art critic (discriminator); their job is to spot the forger’s replica. Together, the two learn a lot about how to A) spot high-quality, realistic images and B) make a great variant.
ControlNet takes the underlying concept and allows an artist to turn a sketch into variations. In other words, you can get the hands right in your generated images now. Or enjoy a tool like Scribble Diffusion.
The researchers at Stanford who published their ControlNet project are proud of one fact above all else: you can use their model on your computer.
Hand to my heart, I promise ControlNet doesn’t take much space to run and can learn from a small dataset.
If you’re looking for leaps in AI learning, this is it. But ControlNet is not the easiest thing to download. I am linking the GitHub here, but be aware this requires Stable Diffusion and some patience.
You can use that to upgrade your AI images or play around with Scribble Diffusion. Both are great choices!
Quick Nuggets
💥 Manga artist turned-politician is tackling AI art policy in Japan
🌊 Sea levels keep on rising, so we’re turning to AI for answers
🪟 Microsoft: Bing can be provoked to respond outside its designed tone
🍺 Brewskies made by artificial intelligence, are they any good?
🤔 Is ChatGPT ready for enterprise-level use? Tech chiefs weigh in
🇮🇹 The Vatican hosted religious figures to discuss the moral conundrums in new tech
🇰🇷 South Korea boosts its AI chip industry as part of the hype wave
🌳 Sequoia Capital gives its take on the rising interest in AI
🇨🇳 China is not moving as fast as we all thought in its effort to launch a ChatGPT competitor
⛪️ A sermon written by ChatGPT is lacking in soul; a pastor claims
🔒 Cybersecurity and the future of machine learning in the field
🗽 VivaCity raises $8.5m to help make cities safer, starting in NYC
🎵 Rolling Stone gives their chat AI take instead of reviewing more music
📉 Big bubble; little concern over the validity of this take by Raoul Paul and his concerns with AI companies
🇺🇸 Governing AI is pushing into the public conversation even further
🔥 Fresh Products
Winston AI - detects AI-generated content (link)
ChatGPT Copier - pull your responses with ease (link)
Blixem - GPT chat sidebar for Chrome (link)
LeadIQ - personalized messages in seconds (link)
ListGPT - save and share your chats with AI (link)
Patterns - develop complex AI applications easier (link)
Hush AI - what does the internet know about you (link)
Buffer - launched a new AI assistant (link)
Auto Backend - backend development is a prompt away (link)
Simple Phones - never miss a customer call (link)
Taplio - create better LinkedIn content (link)
Voxwave - AI-driven lead gen (link)
The GPT who lived - Harry Potter QA (link)
Friday - is as easy as today’s afternoon (link)
Tender - the artists behind iconic generative works (link)
Type - AI-powered document editor (link)
Genius Sheets - launched a new conversational AI to their toolset (link)
Good Content, Dark Knight
They took Gotham’s Dark Knight and brought him down to our level. I am a huge Batman fan and thought this looked cool, if you could not tell.