π€ Wiser! #132: Meta Launch LLAMA | Is ChatGPT Getting Dumber? | AI Generated News
Meta's release of its A large language model called Llama 2, research from Stanford that suggests GPT4 is getting dumber, and Google's plans to build an AI news generation tool called Genesis.
What's In Wiser! This Week?
You won't be surprised when I say that generative AI continues to dominate the news, reader comments and my Inbox alike. It's clear that you want more on AI, that's understandable. Although I constantly reminding people that there is a ton of other interesting stuff going on in emerging technologies that we/they/you shouldn't lose sight of.
Having said that, the main stories this week are all about AI. To start, the big announcement of the week was the launch of Meta's latest version of its AI large language model, called Llama 2. Meta have been relatively quiet compared to OpenAI, Microsoft, Google and Anthropic, but I'm keeping a close eye on Meta's AI ambitions. They're a bit of dark horse and shouldn't be ignored.
Talking of Google, they're pitching an AI tool to generate news for media organisations. Sounds obvious, but also troubling and fraught with dangers given AI's proclivity to make stuff up and pass it off as true. The other news that caught my eye was a research paper from Stanford that suggest ChatGPT is getting dumber. There's a scientific explanation too, it's just that nobody seems to be able to explain it. I do my best!
My member's column this week is about Threads and Twitter. It's not a subject I'm going to devote a lot of time too, but I do think Meta's cut and paste job on Twitter is a big deal. It won't see Twitter off, but it's going to be hard for advertisers to resist the pull of Meta, who have the most effective advertising platform on social media. Without advertisers, Twitter can't run itself. It's barely breathing as it is.
Anyway, judge for yourself if you agree, or not, with my 5 reasons why Threads will beat Twitter.
There are two other things to mention: first, YouTube channel is back up and running with 10 minute videos to accompany each week's Wiser! And the latest episode of Big Tech Little Tech, all about Starlink and satelitte internet, came out this week.
ATB, Rick
P.S. AND REMEMBER: Insight and Information Gives You Leverage!
π"the beginner's guide to ChatGPT is the best β¬5 I've spent in ages" - Bob R, long time WiserPLUS! member (he bought even though he didn't need too)
5 Reasons Why Threads Will Beat Twitter
Nicking someone else's idea and making it their own is chapter 1 in the Mark Zuckerberg playbook. He's done it before and he's gone and done it again. Threads is an unashamedly cut and paste of Twitter. It's new and still missing the features it needs to eat Twitter's lunch, but that will come in time, and I don't think we'll have to wait too long.
In my column this week I make the case with 5 reasons why Threads will win out against Twitter over the long run. WiserPLUS! members can read it here...
π¦ Meta Release Version 2 Of Their AI Chatbot Called βLlamaβ
Earlier this week, Mark Zuckerberg announced Meta's release of its updated AI system, called "Llama 2." The new version was trained on 40% more data (2 trillion tokens), has more capability than version 1 and includes over "a million human annotations".
The thing that sets Metaβs foundational AI model apart from the others is that it is open-source. This means that you, yes you, can download the large language model and run it yourself (if you had the technical wherewithal to do that.)
Hereβs The Thing: Meta has a significant AI capability. It also has data, and lots of it (although, to be clear, Meta has not said where the data for Llama 2 has come from.) But Zuckerbergβs plans for AI go much broader than building publicly available LLMs.
Meta plans to integrate generative AI into Facebook and Instagram. Theyβre developing technologies, such as AI agents with different personalities, and a text prompt feature for modifying photos in Instagram Stories.
But AI hasnβt always gone well for Meta. Last November, just before the arrival of ChatGPT, Meta launched Galactica, a research AI that was supposed to revolutionise academic research. Within 3 days it was pulled because it was passing off fake, false and biased content as the real deal.
π€ Is ChatGPT Getting Worse?
This new Stanford research paper about GPT4 caught my eye when it suggested that the number 1 AI chatbot is getting dumber. The research compared GPT3.5 (the free version) against GPT4 (the paid version) with similar research the team had done earlier this year.
According to the researchers, ChatGPT is performing considerably worse; βwe find that the performance and behaviour of both GPT3.5 and GPT4 vary significantly across these two releases and that their performance on some tasks have gotten substantially worse over time.β
Here's The Thing: my experience of these large language models is that they will be both inconsistent and incapable when performing simple tasks. This is counter-intuitive because the big selling point of AI is that it is great for automating repetitive and low-value tasks.
However, all of them have the frustrating habit/behaviour of misinterpreting the prompt. They do their own thing and then tell you that they followed your prompt to the letter, when they clearly havenβt, like writing 9 or 11 bullets when youβve specifically asked for 10. This can partly be explained by poor prompting, which is an art and science that takes some mastering (that's why I wrote the beginner's guide to ChatGPT.)
My point is that these tools are still in their formative stages. The engineers that built them donβt exactly know how they are going to behave or what the machine learning algorithms are going to learn. My take on this report is that far from saying GPT4 has gotten worse, I'd say that these researchers are seeing the inconsistency of LLMs. My bet is that when they repeat their test in 3 months, they'll arrive at a different conclusion again.
For a more technical explanation of how large language models can behave inconsistently through a concept known as SHARCS (SHARed Concept Space), read this (you were warned!).
ποΈ Google Pitches AI Tool Built to Write News Articles
The adoption of generative AI into the media sector is obvious and inevitable. According to the New York Times, Google is pitching an AI tool called Genesis that can write news articles based on current events. Despite the evidence of AI's inability to know when it's made stuff up, news organisations like The Times, NPR and Insider, have told staff they're looking at ways to use AI (responsibly). However, some have already embraced it, like CNET, Buzzfeed and Gizmodo, and have gotten themselves into hot water for publishing AI generated content that was factually incorrect.
Hereβs The Thing: Thereβs no escaping the concerns about AI generated misinformation across all sectors, especially media, an industry that has some tough decisions to make about AI. On the one hand you have the appeal of the nearly-free cost of AI generated content in a sector thatβs tough to make money. On the other, the continued erosion of trust in news organisations pushes audiences to get the information elsewhere. (New figures from Ofcom show that TikTok is the number 1 source of news for teenagers.)
One thing is for sure, the volume of AI generated misinformation and false news flooding social media will significantly rise. Itβs cheap and easy to create and near impossible to police. Knowing the difference between credible news sources and the rest is going to become increasingly important for all of us.
Structuring Conversations for Optimal ChatGPT Output
Here's an excerpt from chapter 3 of my beginner's guide to ChatGPT:
When interacting with ChatGPT, the way you structure your conversation plays a crucial role in obtaining the desired output. To maximise the effectiveness of your interactions, consider the following guidelines:
1. Clear and Concise Prompts: Start your conversation with a clear and concise prompt that conveys your intent or question. Providing specific details or context can help guide the model's responses and ensure relevant output.
2. Break Down Complex Questions: If you have a complex question or request, consider breaking it down into smaller parts. Presenting the AI with a series of simpler prompts often yields better results, as it allows the model to tackle one aspect at a time.
3. Use System Messages: System messages are instructions or guidelines you can include in your conversation to guide the model's behaviour. By utilising system messages, you can nudge the model to adopt a specific role or writing style, providing additional context for generating responses.
The Beginner's Guide to ChatGPT is written in an easy to understand and non-technical (as much as I could make it) format. It's intended for anyone who wants to use ChatGPT, is not super technical, and wants to know the basics of how to go about it. The book is free to all WiserPLUS! members.
Humata - ChatGPT For Your Files
Humata is an AI-powered software tool that can quickly summarise, synthesise, and extract valuable data from files. It can generate new writing based on files and create detailed insights for reports, papers, and other tasks.
You can even use Humata as a chatbot to ask questions about your data and get instant answers powered by AI.
What Else Is Going On?
πΌ ChatGPT Face Billion-Dollar Lawsuit
OpenAI's ChatGPT has been accused of defaming and maligning individuals by creating false facts. Like incorrectly accusing the CEO of Armed American Radio of stealing $5 Million from a gun-rights nonprofit, and making up a few fake court cases as proof. Or when ChatGPT told a journalist that an Australian mayor went to prison for a bribery scandal. He didnβt. He was the one who exposed it. The FTC are threatening an investigation whilst others estimate that this could cost OpenAI billions in defamation charges and collateral. Learn moreβ¦
π€ Google's BARD Can Now Read Pictures
Google's latest updates for their AI chatbot, Bard has extended the number of languages it can work in, the ability to "speak" responses (but itβs not as good as Pi) and improved ideation capabilities. But the feature that excites me the most is the incorporation of Google Lens into Bard. This means that Bard can now read pictures. Bard can now generate headlines for images and provide recipe ideas based on a picture of the contents of your fridge. Learn more...
π AppleGPT is on its way as Apple tests generative AI tools called "Ajax"
Apple's AI tools, codenamed "Ajax," aim to generate various types of content as well as improve the conversational capabilities of Siri, which is long overdue. Reports suggest that Apple's consumer product is aimed for next year, although it has been deployed within internal systems. Learn more...
π¨ββοΈ Microsoft to charge $30 per month for generative AI features
Microsoft plans to charge a whopping $30 per month for OpenAI's generative AI features it has built inits office productivity suite of tools (called CoPilot). The service will offer businesses the ability to create custom AI models without requiring extensive coding knowledge. This compares to $20 for ChatGPT Plus, whilst all the other LLMs are free to use (for now). Learn more...
and that's it for AI news...
π° New US bill wants to regulate DeFi like a bank
A new bipartisan bill in the US Senate, the Crypto-Asset National Security Enhancement Act of 2023, is seeking to impose bank-like money laundering controls on DeFi protocols. Aimed at curbing crypto-associated crime and tightening regulation on digital currencies, the bill focuses on entities controlling DeFi protocols and those investing more than $25m into developing them. Learn more..
βοΈ Russia bans the iPhone over spying claims
Russian government officials are now banned from using an iPhone. This comes after the FSB (Russia's spy agency) reported last month that they had uncovered a surveillance network used by the US to spy on Russia. Who knows if its true or not (the spying I mean)? Learn more...
π₯§ And finally, Bill Gates is a fan of Pi (like me)
Remember last's feature article about Pi, the AI from Inflection AI and my personal favourite of all of the AI systems? Well it seems that I'm not the only fan. Bill Gates said, βIβm impressed with a couple of startups, including Inflection.β Learn more...
Chart Of The Week
After years of arguing the Microsoft $68 billion acquisition of Activision Blizzard is not anti-competitive, the deal cleared its penultimate hurdle. The injunction of the Federal Trade Commission (FTC) was denied a final appeal in court on July 12. If Microsoft can ameliorate the concerns of the UK competition watchdog CMA connected to a potential cloud streaming monopoly, the deal will probably close in the near future, making it the biggest deal in gaming ever.
The Utility Of Emerging Technologies Shaping Tomorrow's Digital Economy
Away from all the headlines of generative AI, I continue to track what consumer brands are doing with emerging technologies and record them in the Brand Strategy Index for WiserPLUS! members. The latest updates include:
On the eve of the Womenβs World Cup Finals in New Zealand, Credit Suisse Bank has collaborated with the Swiss Football Association to launch 756 Ethereum NFTs featuring members of the National Team, with proceeds going toward promoting womenβs soccer in Switzerland. The bank's first foray into NFTs aims to be simple and client-friendly, highlighting the potential for consumer brands to use NFTs for customer engagement.
Bank of America is using AI, VR, and Metaverse technologies to train 2,000 new employees, simulating customer interactions and offering virtual tours and explanations of bank branch details, company history, and employee benefits. This is not a new concept, Accenture has onboarded 150k new hires using VR over the past 2 years, but it signals that more corporations are expected to use metaverse technologies to enhance internal operations and save costs.
Fashion brand Lacoste has launched 'The Mission', an interactive digital experience using dynamic NFTs to engage fans in community quests and earn rewards. Completing missions earns points, climbs leaderboards, and increases card rarity, with the top performer winning an all-expenses-paid VIP experience in Paris. By embracing NFTs, global consumer brands like Lacoste are seeking to transform customers into active members of their community.
Boot maker Timberland is celebrating its 50th anniversary with an AR experience that allows customers to customise their Original Timberland Boot. Users can choose body colour, up-cycled laces, etch a phrase or name onto the side of the boot, and even customise a footbed pendant and engrave 3 letters or numbers on the up-cycled lace keeper. Once satisfied, users can download the image of their boot and view it in AR.
As part of its 96th birthday celebration, 7-Eleven has released Slurpee NFT collectibles that can be obtained by mixing favourite Slurpee flavours in the company's mobile Vibe Check Machine. The collectible can be saved to the user's digital wallet as a non-fungible token and may provide inspiration for future Slurpee flavours. 7-Eleven have also added bitcoin ATMs to stores.
π‘ WiserPlus! members get unlimited access to the Brand Strategy Index, a database featuring the use cases of over 250 consumer brands using emerging technologies to enhance brand value.
For more information, go here.
Tech Insights Shaping Tomorrow.
Dive into the digital economy with Rick Huckstep. From BigTech strategies to emerging technologies, Rick demystifies complex trends with engaging and simple language in this weekly Newsletter. WiserPLUS! subscribers get extra content and exclusive benefits.
βοΈ You can acknowledge my work with BuyMeACoffee. Buy virtual coffees to show your support from only β¬2 each. π