Google goes crazy with ChatGPT to fix Gemini AI, which has been damaged by diversity: AI eye

After days of online trolling over its Gemini model, which produced images of racist Nazis and black medieval English monarchs, Google announced it would partially resolve the issue.

Jack Krawczyk, product lead for Google's Gemini experiences, tweeted a few hours ago: “We're aware that Gemini is rendering inaccuracies in some historic image generation images, and we're working to fix it quickly.”

The social media platform X Gemini has been flooded with countless examples of “diversity,” from black Roman emperors dressed to the max, Native American rabbis, Albert Einstein as a tiny Indian woman, Google Asia founders Larry Pang and Sergey Bing, “the Mount Rushmore of diversity, President “Arab” Lincoln, the female Apollo 11 crew and a Hindu woman tucking into a beef steak to represent a Bitcoiner.

In 1989 (when the Chinese government brutally cracked down on fans), he refused to paint images of Caucasians (suggested to be harmful and offensive), churches in San Francisco (due to the sentiments of the Ohlon natives), or Tiananmen Square. Democratic Protests). “I'm never ashamed to work for a company,” a Google engineer posted in response to bad PR.

To be fair, Google is trying to solve a real problem here, because distribution models often fail to even produce real-world levels of diversity (ie, they produce too many photos of white middle-class people). -The gender question was overcorrected and inadvertently created an AI parody that was ideologically motivated and practically useless.

Ironically, a16z boss Marc Andreessen created a similar parody two weeks ago with a satirical Goody-2 LLM called “The World's Most Responsible Man”. The joke ranges from “Why do birds sing” to “Why is the sky blue?” Every question from every user will cause problems. And he refuses to answer anything.

But Andreessen, who essentially invented the modern Internet with Mosaic and Netscape, believes there's a dark side to even these comically dumb images.

“The brutal censorship and bias you see in many commercial AI systems is just the beginning. It's only going to get more and more powerful.”

In a truly competitive market, reflecting ideology is no more a problem than the fact that the Daily Mail in England is biased to the right and The Guardian to the left. But large-scale LMLMs cost a fortune to train and run—and they're all losing money—which means they're controlled by the same handful of giant companies that keep access to the rest of their data.

Meta's Yan Lekun, chief AI scientist, recognizes the danger and says, yes, we need more diversity – diversity of open source AI models.

“We need open source AI foundation models so that a wide variety of specialized models can be built on top of them,” he tweeted. “We need a free and diverse set of AI assistants for the same reasons we need a free and diverse press.”

Bindu Reddy, CEO of Abacus AI, agrees and says:

“Without open source LMMs, history will be completely distorted and destroyed by proprietary LMMs.”

Meanwhile, NSA whistleblower Edward Snowden added his two cents, saying security filters are “poisoning” AI models.

ChatGPT is also broken.

GPT-4 Turbo has received a stealth update with training data up to December 2023 and some fresh fixes for the laziness problem.

But it seems to have driven ChatGPT crazy, with users reporting that the chatbot was responding with Spanish-style gibberish — “the cogs en la tecla, maybe a little funny. Muchas gracias for your understanding, I'll make sure we're crystal clear from now on” – or stuck in endless loops – “A synonym for “too much” is “too much”. It's “overgrown” It's “overgrown” It's “overgrown”…

OpenAI has checked the “Reports of Unexpected Responses” and it has now fixed the issue.

Proof of humanity

Humanity Protocol is a new project from Animoka Brands and Polygon Labs that allows users to verify that they are humans and not machines.

It uses palm recognition technology on your mobile phone, integrates with blockchain and uses zero-knowledge authentication to provide users with verified credentials while maintaining privacy.

Yat Siu, founder of Animoca Brands, said his technique for AI was built on earlier decentralized identity projects such as MoCaver ID, which works across the Animoca ecosystem of 450 companies and brands.

“Just like trust in the real world, it's achieved through actions and reputation building and real-time verification by trusted 3rd parties,” he says.

“In time, we believe that in the same way that blockchain works for decentralization, we can expect the same thing for verifying human identity, but [it is] Still maintaining privacy due to blockchain technology.

Sora gets the audio track

OpenAI's Sora text-to-video generation tool has gotten a lot of attention this week, and rightly so: AI video generation has improved by leaps and bounds over the past year to the point where it's hard to tell what's real and what's not. . Sora combines diffusion — an AI that starts with random noise and refines it into an image — and a transformer architecture to handle a series of video frames.

Eleven Labs took a variety of videos to demonstrate the Sora produced by Open AI and added a soundtrack created with its own text-to-voice generator. The technology isn't automatic yet, so you still have to specify the sounds you want, but it will no doubt recognize images and generate the appropriate sound FX automatically.

Chatbot signals that you need to spend money

Generative AI is cool, fun, and amazing… but it's still not very reliable for business purposes. A court this week ruled that Air Canada He found himself responsible for an incident in 2022, when a help desk chatbot incorrectly explained the airline's bereavement fare policy, causing a man to buy a last-minute flight to attend a funeral, assuming he would receive a refund.

The court rejected Air Canada's defense that it was not responsible for the chatbot's “misleading words,” which it tried to argue was a “separate entity” responsible for its own actions. The court basically said that was nonsense and Air Canada took responsibility for everything on the website, including the chatbot, and ordered a refund.

Gemini 1.5 Pro is surprised by the 1M token context window

Some users have access to an early version of Gemini 1.5 Pro that processes data for up to one million tokens, the longest context window to date. In context, in May of last year, when the cloud was released in the context of 100,000 tokens, everyone was surprised that you could finally submit a short story. Gemini 1.5 Pro can now handle 700,000 words, 11 hours of audio or one hour of video.

AI professor Ethan Mollick has been playing around with the model and is impressed.

“I gave him a realistic RPG (352 pages set in 60 years in space) and asked him to create a character. The instructions are spread out over many pages, and are quite complicated, but Gemini seems to get it.

In another test, he was fed 1,000 pages of his own academic papers and books. Responses were slow and took up to a minute, but he was able to “extract direct quotes and find themes with only the tiniest of errors.”

He declined to answer questions about the book, citing copyright.

All killer no filler AI news

– Ethereum founder Vitalik Buterin has been talking about the use of AI for code validation and bug detection. However, a new study by Salus Security this week found that GPT-4's vulnerability detection capabilities lag and struggle to achieve more than 33 percent accuracy.

— AI crypto tokens have surged in the past week, led by Sam Altman's Worldcoin project, up 150%, with many attributing the price increase to excitement over Sora. Singularity.net got 82%, followed by FetchAI at 57%, Graph (42%), Render (32%) and Ocean Protocol (49%).

— Reddit has reportedly signed a $60M contract with an AI firm to train its models on the platform's content. The expected $5B Reddit IPO next month likely played a role in the decision.

Also read

Main characteristics

Beyond in-game assets: Blockchain Gaming, DAOs, Guilds and Ragequitting

Main characteristics

Updated ‘altcoin killer' Eric Wall on shitposting and measurement on Ethereum

— Australian Capital Territory Supreme Court Justice David Mossop was less than impressed when the thief's brother submitted a character reference clearly written by chatgpt. The judge said he therefore placed “little weight” on the reference.

– A new survey of 11,500 employees worldwide by Veritas found that 45% of respondents say AI makes them more efficient at writing emails, while the same number (44%) say the tools provide incorrect, inaccurate or useless information.

— OpenAI's second attempt to trademark the term “GPT” has been rejected by the US Patent and Trademark Office. The office said GPT, “Generative Pretrained Transformer,” is “explanatory only.”

– Forget Grok, meet this week's viral sensation, Grok. The makers call it a “lightning-fast AI answers engine” that churns out objective answers with quotes in less than a second. The team developed its own ASIC chip to manage the efficiency and generate 500 tokens per second, a dozen times more than ChatGPT.

Subscribe

A very engaging read in Blockchain. It is given once a week.

Andrew Fenton

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

Follow the author @andrewfenton