ChatGPT gets lazier

In recent days, more and more users have reported that instead of carrying out a task, ChatGPT seems disinterested or gives a partial answer and tells users to finish the job themselves. To me, that’s a sure sign that human-level artificial general intelligence has been achieved.

Many suspect OpenAI has nerfed it to reduce the phenomenal cost of the system, but OpenAI says that’s not the case.

In the meantime, users have resorted to bribing the AI with tips for better answers and impressing upon it in the prompt how earth-shattering crucial the best possible answer is.

Google’s Gemini video was unreal

As we now know, Google’s “mind-blowing” and “unreal” product video for Gemini Ultra was faked — but is the deception significant?

The video shows a man having a natural-sounding conversation with the AI, which recognizes he is in the process of drawing a duck. Gemini Ultra can also work out which cup the ball is under during a shell game and figure out when someone is playing rock paper scissors.

But in reality, there was no video and no vocalized conversation. The AI was prompted via text and simply shown still images.

However, Google says the prompts and outputs were real, and Gemini can actually recognize a drawing of a duck.

Duck recognition has long eluded scientists, so this looked like a huge breakthrough … although, as it happens, ChatGPT can recognize ducks, too.

Oriol Vinyals, the vice president of research for Google DeepMind, said the video was merely a serving suggestion.

“The video illustrates what the multimodal user experiences built with Gemini could look like. We made it to inspire developers.”

Test results for Gemini show it’s just beating GPT-4 on seven out of eight benchmarks — which sounds good until you realize that GPT-4 was completed a year ago, giving OpenAI a 12-month head start on GPT-5.

Fetch.AI = Google of AI Agents?

AI Eye caught up with Humayun Sheikh this week — the founder of Fetch.ai and a former commercial director of DeepMind (now Google DeepMind). Sheikh says the company was “doing GPT-like things 13 years ago,” meaning Google deliberately decided not to release an LLM until ChatGPT forced its hand.

“I was quite surprised it came out quite late, and they gave OpenAI a chance. But I think they were ready a while back,” he says.

“The problem is that if you start using this technology, and you let it loose, you start cannibalizing your own business,” he adds, speculating that the AI might undercut the revenue from Google’s lucrative search business.

That said, Sheikh believes Google is still in a winning position in the AI arms race.

“Google has maps, Google has advertising, Google has businesses and all the reach you need to integrate with a suite of applications,” he says. “So I think OpenAI has a problem, and OpenAI will have to drastically change to be viable.”

Fetch.AI is in the business of helping users create AI agents, with 100,000 up and running so far. The bots can help with travel bookings, EV charging and looking after IoT devices.

You can give an AI agent a goal, and it will create a bunch of subtasks and autonomously carry them out until the goal is achieved. Sheikh says blockchain is a natural fit to coordinate agents and to record their performance.

“If you want to interact with multiple entities, if you want to interact with multiple pieces of code, if you want to interact with multiple machine learning AI algorithms, you need a new framework. You cannot run that on an old Web 2.0 framework.”

“The orchestration of these tasks needs to happen somewhere,” he says. “And these micro-tasks need some monetization, which the blockchain provides.”

Peaq blockchain, Fetch.ai and Bosch have just unveiled the Bosch XDK110 Rapid Prototyping Kit, which utilizes a small sensor that can capture data on things like pollution levels, the weather or seismic activity, and then feed the data back to multiple decentralized physical infrastructure networks (DePINs) in exchange for tokens.

So, for example, you could ask an AI agent to check the tire pressure sensors in your car and to check sensors recording the weather conditions, and work out if the tire pressure is too low or too high or if you need new tires. If you do, it can book you in at the tire place.

Or you could ask your AI agent to perform sentiment analysis on a particular stock and buy it if the analysis is favorable.

Sheikh says the plan is for Fetch.AI to become the search engine for agents, “and we’ll also be a self-assembly engine, which puts all this stuff together without you having to code it or integrate it.”

All Killer, No Filler AI News

— Meta dropped 20 new AI features for its social media platforms to jazz up search, ads, and business messaging. The free image generator is the one people are excited about.  

— Boffins from Alibaba scraped a bunch of TikTok influencer videos, and now you can generate pics of people dancing from a still photo and some pose guidance.

— Professor Ethan Mollick has compared and contrasted all the LLMs and says the best one to use is GPT-4. Research suggests it boosts real work performance, generates better ideas than most humans, seems smarter than the competition and has the most features.

— McDonald’s is building an AI chatbot called Ask Pickles that’s trained on data from 50,000 restaurants. Franchisees and staff can now interrogate the AI to better understand exactly how to do everything in a more McDonaldsy way.

— Amazon currently has around three-quarters of a million robots working alongside 1.5 million employees.

— Grok user Jax Winterbourne was taken aback when the AI refused to answer a question and cited OpenAI policy as the rationale. While some users believe Grok has been revealed as nothing more than a front end for ChatGPT, Xai said it’s picked up some of ChatGPT’s outputs in its training data. 

Tweet of the week

A new perspective just dropped: “The only thing LLMs do is hallucinate, the trick is in getting the hallucinations to align with reality.”

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