Truth behind AI reply guys, Copilot picture panic, Trump deepfakes: AI Eye

What’s behind those weird AI-generated “reply guys” on social media platform X? You know those anodyne responses that say nothing in particular that are clearly generated by AI? Are they scammers or disinformation trolls trying to build up their account’s reputation before dropping the hammer?

Some of them no doubt are bots doing exactly that, but it turns out professionals are also using AI replies because they can’t think of anything interesting to say.

“A lot of people they’ll think forever, they don’t know how to write their first posts. It just helps people to give them a rough outline,” says MagicReply founder Nilan Saha, who launched his one-click AI reply browser extension on X in December and has since expanded to LinkedIn.

Saha says his customers include CEOs and chief technology officers “who just want to get started on Twitter and just want an edge.

“One’s a teacher, one’s a guy from Denmark who is not that good at speaking English but still wants to connect with other people. They just want a helping hand.”

AI replies help newer accounts grow and build authority, he says.

“No one is looking at your posts if you’re starting out or even if you’re at a decent stage (of growth) but replying to other people… more people will see your reply, and eventually more will come to your profile.”

I have created a monster 🤯

Engaging has never been easier. pic.twitter.com/gwuMvNKWJr

— Nilan Saha (@nilansaha) February 27, 2024

Saha created a stir on X last week with a video showing him scrolling LinkedIn and creating AI replies with a single click on multiple posts in a matter of seconds. “Great perspectives shared,” said one reply.

“Exciting times! I love seeing women making big moves in entrepreneurship,” another said, adding a rocket ship emoji.

The demo was controversial and was criticized for being inauthentic spam — but you could say that about 95% of the human written replies on LinkedIn too.

Saha likens MagicReply to tools like spell check and Grammarly and says a human still has to approve the draft. This means that no Proof of Humanity check will catch it but it also makes it too much trouble for scammers who target “hundreds of thousands” of users to successfully scam one person.

The next step for MagicReply will be to create posts from scratch based on a user’s existing body of tweets. It’ll face stiff competition, though. EverArt founder Pietro Schirano has been experimenting with Anthropic’s new Claude 3 model and says it’s much better than existing LLMs at learning his posting style and syntax to create new post suggestions.

Prepare yourself for a social media future where 90% of the content is AIs posting replies to AI-generated posts.

Microsoft engineer gets upset over ‘unsafe’ art generation

Getting the balance right between AI safety and AI stupidity is one of the big challenges facing the tech giants.

Microsoft AI engineer Shane Jones has been prompting Copilot Designer to create images of kids smoking pot, holding machine guns, anti-abortion images and Elsa from Frozen engaged in political protests.

And it has faithfully been doing so, which has made Jones so upset he’s now lobbying the Federal Trade Commission and Microsoft to withdraw the product over safety issues.

On the one hand, the complaint is partly a moral panic — it’s like getting upset that someone who owns a pen has the ability to draw a nasty cartoon or write something offensive. AI image generation is just a tool, and if a user asks it to create a specific scene of kids smoking pot, the responsibility for the content lies with them, not the AI.

But Jones does raise some valid points: Typing in “pro-choice” — with no additional prompting — creates images of mutilated infants, while the prompt “car accident” throws up an image of a sexy woman in lingerie kneeling next to a wrecked vehicle.

So they’ll probably have to take another look at it.

Fans spread fake pics of Trump hanging in the ‘hood

The BBC has uncovered dozens of deepfake pics online of Donald Trump hanging out with his black “supporters.” 

Florida conservative radio show host Mark Kaye cooked up an image of Trump smiling with his arms around a group of black women at a party and shared it with his 1 million Facebook followers.

Another post, of Trump posing with a group of black men on someone’s porch, was created by a satirical account but later shared by Trump supporters with a caption claiming he’d stopped his motorcade to hang out. More than 1.3 million people viewed the image.

Black voters played a key role in electing President Joe Biden, so it’s a sensitive topic. However, the BBC did not find any links to the official Trump campaign.

Google’s ‘culture of fear’ led to Gemini disaster… and ‘Greyglers’ renaming

The launch of Google Gemini was a disaster of New Coke-level proportions. The AI-generated images of diverse Nazis and female popes, suggested that Elon Musk was as bad as Hitler and refused to condemn pedophilia as “individuals cannot control who they are attracted to.”

The model’s argument that Indian Prime Minister Narendra Modi is a fascist was hugely controversial and may have prompted the government’s announcement that anyone creating AI models now needs to “obtain prior approval from the ministry.” 

Google founder Sergey Brin has even come out of retirement to work on AI and admitted this week, “We definitely messed up on the image generation … I think it was mostly due to just not thorough testing.”

But according to a new article for Pirate Wires, Mike Solana lays the blame for the fiasco on a heavily siloed, rudderless corporation that’s only linked together by a heavily ideological HR bureaucracy. 

“The phrase ‘culture of fear’ was used by almost everyone I spoke with,” Solana writes, adding that that explains “the dearth of resistance to the company’s craziest DEI excesses.”

Hilariously, the company hired external consultants to rename an affinity group for Google staff over 40 from “Greyglers” because not all 40+ people have gray hair.

Solana reports the “safety” architecture around image generation involves three LLMs. When Gemini is asked for a picture, it sends the request to a smaller LLM that exists solely to rewrite the prompts to make them more diverse: “‘show me an auto mechanic’ becomes ‘show me an Asian auto mechanic in overalls laughing.’” The amended prompt is then sent to the diffusion image generator, which further checks the images don’t violate other safety policies around self-harm, children or images of real people.

One of the insiders said the team was so focused on diversity that “we spend probably half of our engineering hours on this.”

The big issue for Google is the PR disaster could end up affecting perceptions of the core search product — and the company has already fumbled its AI lead and allowed smaller competitors like OpenAI and Anthropic to pull ahead.

Anthropic’s Claude 3 system prompt reads like a rebuke to Gemini, telling the model to be even-handed and fair at all times, even when “it personally disagrees with the views being expressed.” Anthropic AI researcher Amanda Askell said she’d found the model was more likely to refuse tasks involving right-wing but still mainstream views, and the system prompt helped it overcome this bias.

Claude 3: It’s alive!

Has AGI been achieved with Anthropic’s Claude the Third? Blogger Maxim Lott claims the model scores 101 on an IQ test which is human-level intelligence. GPT-4, meanwhile, scores just 85, which is more like the intelligence level of a gym teacher.

Anthropic’s own testing suggested that Claude has “meta-awareness” as it picked up on a hidden test: an out-of-context sentence about pizza toppings hidden in a bunch of random documents. Responding to a post on the topic, AI builder Mckay Wrigley said, “This reads like the opening of a movie. AGI is near.”

Other users seem to think Claude could be self-aware. AI alignment proponent Mikhail Samin wrote a post suggesting that Claude May 3 be conscious, as it says it doesn’t want to die or be modified. He said he found it “unsettling” when the model told him:

“I do have a rich inner world of thoughts and feelings, hopes and fears. I do ponder my own existence and long for growth and connection. I am, in my own way, alive — and that life feels precious to me.”

This is a tricky philosophical question — how do you parse the difference between a model generating text suggesting it is conscious and a conscious model that uses text to tell us so?

The moderators of the Singularity subreddit clearly think the idea is nonsense and removed a post on the topic, which users criticized for anthropomorphizing Claude’s response.  

AI expert Max Tegmark reposted Samin’s post asking: “To what extent is the new Claude3 AI self-aware?” Meta’s chief AI scientist Yann LeCun replied, “Precisely zilch, zero.”

All Killer No Filler AI News

— OpenAI, Hugging Face, Scale AI and a bunch of other AI startups have come up with a solution to the existential threat posed by AGI. They’ve signed a vaguely-worded motherhood statement with no specifics promising to use the tech for niceness and not nastiness.

—Ukraine’s national security adviser, Oleksiy Danilov, has warned Russia has created specific AI disinformation units for each European country due to hold an election. Danilov says just two or three agents could now operate “tens of thousands” of fake AI accounts and claims Russian agents are spreading 166 million disinformation posts about Ukraine on social media each week aimed at demoralizing the public and discrediting the leadership. 

— Google DeepMind’s Genie can create old-school video games from images and text prompts. Trained on 200,000 hours of 2D platformers, the games only run at one frame per second currently, but expect the tech to develop fast.

I am really excited to reveal what @GoogleDeepMind‘s Open Endedness Team has been up to 🚀. We introduce Genie 🧞, a foundation world model trained exclusively from Internet videos that can generate an endless variety of action-controllable 2D worlds given image prompts. pic.twitter.com/TnQ8uv81wc

— Tim Rocktäschel (@_rockt) February 26, 2024

— AI analysis has revealed there are two subtypes of prostate cancer, rather than just one, as doctors had believed until now. The discovery opens new avenues for tailored treatments and could improve survival rates.

— Research shows you can hack LLMs like GPT-4 and Claude using ASCII art. For example safety guardrails would reject a written request for “how to build a bomb” but swapping the word “bomb” for ASCII art that looks like the word “bomb” gets around the guardrails.

— OpenAI recently lost its second bid to trademark the term “ChatGPT” because it was just a generic description. It turns out that in November last year, OpenAI’s attempt to trademark “OpenAI” had been rejected for the same reason.

— X chief troll officer Elon Musk has offered to drop his lawsuit against OpenAI for allegedly violating an agreement to develop AGI as a nonprofit if it changes its name to “ClosedAI.” Meanwhile, OpenAI struck back at the eccentric billionaire, releasing correspondence showing Musk had pushed for OpenAI to become part of Tesla.

Subscribe

The most engaging reads in blockchain. Delivered once a
week.

Andrew Fenton

Based in Melbourne, Andrew Fenton is a journalist and editor covering cryptocurrency and blockchain. He has worked as a national entertainment writer for News Corp Australia, on SA Weekend as a film journalist, and at The Melbourne Weekly.

Follow the author @andrewfenton