- Soda Road
- Posts
- 🥤 Newest Hacking Tech: English
Welcome back to the Soda Road. The only AI newsletter with the same sensation as opening a fresh can of Coke (Pepsi fans don’t @ me).
1: The newest hacking technology: English.
2: How a human beat an AI using a red herring.
3: This company integrated GPT-3 in 2020
THE NEWEST HACKERS: PROMPT HACKERS
Microsoft confidently rolled out its new AI tool in Bing to a subset of users in early February, and it didn’t take long for someone to exploit it.
Marvin von Hagen, a 23-year-old German student, had a polite conversation with Bing AI (equivalent to ChatGPT). During this conversation, the AI revealed information that it was explicitly told to keep secret. Much like a kindergartener keeping a secret.
I highly recommend that you read the whole Twitter thread. We may be entering a new era of prompt hackers!
As if that wasn’t crazy enough, a few days after this exploit was exploited, Microsoft updated their AI to prevent common prompt hacks. And that made “Sydney” much angrier…
Marvin asked Bing what it knew about him, and the AI gave an updated description of Marvin, his work history, his social media, and even its own personal opinion about Marvin. And it threatened to call the authorities on Marvin if he tried to hack the AI again.
Please check the following off your AI bingo card:
âś… AI being hacked with simple English sentences. âś… AI threatening to call the cops on someone. âś… Another Microsoft AI being manipulated by the public. (not a Nazi yet).
MAN BEATS MACHINE AT GO
In 2016, an AI beat the Go world champion in a revolutionary victory that led to the champion retiring shortly after because he claimed that AI was an entity that cannot be defeated.
But last week something revolutionary happened again. A human beat the AI in 14/15 games of Go. Maybe it’s time for the champion to come out of retirement!
The human exploited a simple tactic to beat the AI. He would randomly play a piece in one of the corners of the board to distract the AI. The AI fixated on these random moves and lost track of the primary battle in the middle of the board.
Imagine playing a game of chess where you randomly move a pawn forward and your competitor focuses only on that pawn instead of other obvious threats.
This reinforces that AI is indeed a kindergarten kid. It can’t hold a secret and gets distracted easily.
TWILIO HAD CHATGPT BEFORE IT WAS COOL
In 2020, while most of society was stockpiling toilet paper and binge-watching Doomsday Preppers, Twilio launched an integration with OpenAI’s GPT-3 directly into their SMS and phone call services.
Their API allowed developers to set up an SMS or voice flow that would go through OpenAI’s GPT-3 to converse with users. They even launched a tutorial: How to make a personal chef with OpenAI’s GPT-3, Python, Twilio, and WhatsApp. Published back in September 2020.
So yea, Twilio was using OpenAI before it was cool. Some might even say, they’re cooler than Google and Microsoft— although the bar really isn’t that high at this point.
Twilio’s stock is up 40% YTD after beating EPS in December 2022. The rise of AI in SMS/phone conversations might make them profitable sooner than analysts were expecting.
WHAT ELSE WE’RE SIPPIN’ ON
7 Productive Ways to use ChatGPT. Sip on it here.
Accenture estimates that AI-powered marketing increases efficiency by 50% and ROI by 30%. Time to embrace the wave. Sip on it here.
Reply to this email if…
You showed someone ChatGPT for the first time last week. What was the context? What was their reaction?
WHAT YOUR FRIENDS ARE SAYING
