True Lies: The Rise Of The Chat Bots

Richard Bradbury: I genuinely don’t know what we’re talking about today, other than that Matt said something about loving the sound of his own voice.

Matt Armitage:

• I don’t think that’s anything like what I told you.

• I thought we’d do something different today.

• A couple of years ago we let the listeners send in questions that they wanted answered.

• Answering other people’s questions is hard. So today I thought we could try an ask me anything.

Richard Bradbury: So who’s asking the questions?

Matt Armitage:

• It’s an ask me anything. I ask me and I answer.

• For example, what is the Internet?

The Internet is a global network of interconnected computers and computer networks. It allows people to easily share information and communicate with one another on a global scale. The Internet is made up of a vast number of individual networks that are connected to each other using a wide variety of technologies, such as fiber optic cables, satellite links, and wireless connections. People use the Internet for a wide range of activities, including sending and receiving email, browsing the web, watching videos, and participating in online forums and social networks. The Internet has also become an important tool for businesses, allowing them to reach new customers and conduct transactions online.

Richard Bradbury: Ok. I’m not sure what I’m supposed to do here?

Matt Armitage:

• You’re supposed to act surprised when I tell you how the sun creates energy.

• The sun creates energy through a process called nuclear fusion. This process involves combining hydrogen atoms to form helium atoms, which releases a massive amount of energy in the form of heat and light.

• Inside the sun, the intense pressure and heat from the gravitational collapse of the sun's matter causes hydrogen atoms to fuse together. This creates a chain reaction that releases huge amounts of energy, in the form of heat and light. The heat from this process is so intense that it creates the sun's outer layer, called the photosphere, where temperatures can reach up to 10,000 degrees Fahrenheit.

• The light from the sun is a mix of different wavelengths, including visible light, ultraviolet radiation, and infrared radiation. This light travels from the sun to the Earth, where it provides energy for all living things. The light from the sun also drives the Earth's climate and weather patterns.

• Overall, the sun's energy is essential for life on Earth. It drives photosynthesis in plants, providing the food and oxygen that sustain all living things. It also provides warmth and light, which are essential for the growth and development of many organisms. Without the sun's energy, life on Earth would not be possible.

•

Richard Bradbury: I really have no idea why we’re doing this.

Matt Armitage:

• It’s easy. You want to know how electricity works?

• Electricity is made through the movement of electrons through a conductor, such as a wire. Most of the electricity that we use is generated at power plants, where it is made by spinning turbines in a magnetic field. The turbines are usually turned by steam, which is created by burning fossil fuels, such as coal, natural gas, or oil. The steam is used to spin the turbines, which are connected to generators. The generators use magnetism to convert the motion of the spinning turbines into electrical energy. This electricity is then sent through power lines to homes and businesses, where it is used to power a wide variety of devices and appliances.

Richard Bradbury: I don’t know if this is a format we should repeat.

Matt Armitage:

• How about if I tell you how the blockchain works?

• A blockchain is a distributed database that maintains a continuously growing list of records called blocks. Each block contains a timestamp and a link to the previous block, forming a chain. This allows the database to be securely maintained without the need for a central authority.

• Transactions are added to the blockchain through a process called mining. In a mining process, multiple transactions are grouped together into a block, which is then added to the blockchain. To add a block to the blockchain, miners must solve a complex mathematical problem, which is used to verify the transactions within the block. Once a block is added to the blockchain, it cannot be altered or deleted, ensuring the integrity and security of the data.

• Each block in the blockchain contains a record of multiple transactions, and each transaction is recorded on multiple blocks. This makes it nearly impossible for someone to alter a single transaction without changing the entire blockchain. This distributed, decentralized approach to record-keeping offers many benefits, including security, transparency, and immutability. It is one of the key technologies behind the rise of digital currencies such as Bitcoin.

•

Richard Bradbury: Not bad, but it doesn’t really explain the year we’ve seen in Web3 markets.

Matt Armitage:

• No, it doesn’t.

• So what you’ve been listening to for the past few minutes, while I went and made a cup of tea was some machines talking.

’ve trotted out the Mattbot: 3000

• It’s not new to regular listeners. It’s an Ai language model trained on my voice.

• You drop the text in and it recreates it in an approximation of my speech patterns.

• So as you heard the delivery was slightly probably less robotic than usual because a machine with slightly more humanity than me was reading.

• What was different today was that the text was machine-generated as well.

• Some of you will have heard about ChatGPT, a new AI chatbot from OpenAi.

• WE’ve covered chatbots on the show before.

• Even the most impressive models have been very limited.

• You can have a basic conversation but it quickly starts repeating your own prompts back at you or uttering nonsense.

• I know you’ve been playing with it this week as well, Rich,

• So I’ll ask you about that in a minute. Let me just do a bit of background about what it is and how it works.

• I’ll go into the many things you can do with after.

• But at it’s heart it’s a chatbot. You ask a question and it answers.

• It isn’t trying to pass itself off as a human.

• If you ask it how it is, what its favourite cheese is.

• Normal conversational introductions like that, it will tell you it has no opinion because it’s a machine.

Richard Bradbury: A bit like you…

Matt Armitage:

• I have many opinions. That’s why I’m more dangerous than Skynet.

• It does have certain personalities you can tap into, so there’s a profile for Marv, the sarcastic robot.

• Which adds a layer of sarcasm, why don’t you ask me a hard question, that kind of thing.

• But at its heart it’s a machine that has parsed a huge amount of information.

• And it has detected patterns in that information to help it work out whether things are correct.

• So it won’t tell you whether it was right or wrong to allow Donald Trump back on Twitter.

• But it will tell you what a dirty bomb is.

• But not how to make it. It does have some filters to prevent obscene or illegal requests.

• And beyond that initial data set, it uses reinforcement learning to hone its accuracy.

• Which essentially means the feedback that users give.

Richard Bradbury: Is it essentially a search engine?

Matt Armitage:

• I’ve been struggling with how to answer that question tbh

• And it goes to that point about giving the machine feedback.

• CHatGPT is open to the public. The reports I read suggested that around a million people are currently trying it out.

• Although the buzz around it suggests that the number is much higher.

• So, in order for people to give feedback, they need to know if the information the machine provides is wrong.

• So you have a chicken and egg situation. A lot of people are using it as a search engine.

• Like the explanations the Mattbot just gave you.

• I haven’t checked any of those for accuracy by the way.

• Because the point is to reproduce what the machine stated.

• So, people are using it to find information they don’t know.

• So how do you give feedback when you don’t know how accurate the response is?

• To answer that original question, it’s not a search engine in the traditional sense.

• It’s not live-indexed to the same breadth of sources as Google.

It’s training data ended in: 2021

• And OpenAI is pretty candid about the accuracy of its results.

• The danger is that users get swept up in the plausibility and confidence of the machine.

• And forget that it’s not actually a reference tool. At least, not yet.

Richard Bradbury: Is this another Elon Musk project?

Matt Armitage:

• He was part of the group of investors that founded OpenAI.

• He left the group in: 2018

• Interestingly, after ChatGPT went viral,

• he found out that OpenAi had access to Twitter’s API for this project.

• And he’s paused their access while the platform’s revenue models are reassessed.

• Whether that means that some people have to pay for free speech remains to be seen.

Richard Bradbury: Beyond search, what else are people using it for?

Matt Armitage:

• All kinds of things. It’s a great writing tool.

• If you’re not someone who thinks they’re a great writer, you can drop your notes into it and ask it to write if for you.

• Just specify the length.

• It will create basic advertising pitches and press statements.

• So, I know that you’re planning to sell your own signature scent, BradMusk:

• Introducing BradMusk, a bold and masculine scent for the modern man. Created by Richard Bradbury, our signature aftershave captures the confidence and charm of the successful gentleman. Made with the finest ingredients, BradMusk will leave you feeling irresistible. Try it today and elevate your grooming game.

• People have used it to write basic screenplays. Or even suggest plotlines for shows like The Office.

• It can code. Even better, it can look for errors in code that you’ve written.

• You can ask it to write poetry for you. Or song lyrics.

• There are all sorts of interesting things you can get it to do.

• You’ve been playing with it this week as well. What do you think?

Richard Bradbury: replies

Matt Armitage:

• [ad lib]

• After the break, I want to try and explain why this also belongs with those metaverse technologies we’ve been talking about for the past couple of weeks.

• And to talk again about the risks and benefits that NLP and LLM in general poses.

Richard Bradbury: ad lib outro

BREAK

Richard Bradbury: We’re chatting today. About chatbots. Text to speech. And their importance to the metaverse.

Matt Armitage:

• Yes. So, I want to pick up on that point about: is a search engine?

• You can think of this more in terms of a closed model, that’s learning from feedback given on the data it already possesses.

• Will future iterations be live-indexed?

• Perhaps not of this particular machine, but we’re seeing the industry heading in that direction.

• You can already ask Siri and Alexa that kind of question and they will check the search engines and get back to you with suggestions.

• Where a machine like ChatGPT has a user experience advantage is that it’s all neatly parcelled up for you.

• Like those search queries I did for blockchain.

• It’s a fast, efficient summary.

• Compare that to a manual search.

• For well-searched items, google might throw together some kind of summary.

• But on the whole, the results require you to click on a bunch of different links and read,

• watch or listen to all these different voices, opinions and presentations of the facts.

Richard Bradbury: And who has time for that, right?

Matt Armitage:

• Yes. We want a machine that can summarise the socio-political complexities of the modern Middle East in 50 words.

• Why bother looking at the history, the wars, the economics and the power struggles?

• This creates all sorts of issues:

• It suggests that everything is reductive.

• That there’s nothing we can’t know in a capsulated form.

• And the other is attribution. Where does the information come from?

• ChatGPT comes with big disclaimers that the information shouldn’t be treated as fact.

• This is a test model. In using it, we’re the guinea pigs, not the other way around.

• That’s fine, but future generations of this technology will likely be the way we use search.

• So knowing the sources of that information are crucial.

• For example, people have suggested using the service to write social media posts or even school or college essays.

• It can do that. In the case of social media posts there’s probably less harm because of the expectation of accuracy.

• But in essays or papers, the citations are crucial.

• In both cases, we ought to know where our information is coming from.

Richard Bradbury: Doesn’t the learning model help to determine where facts lie?

Matt Armitage:

• In theory it ought to. But what sources is it pulling from?

• If it’s main information source was twitter or reddit, where would that leave us?

• If information is simply a statistical game, think how easy it would be to replace long held facts or opinions

• Simply by flooding the space with bots armed with contrarian information.

• It’s one of the long-held complaints about NLP and LLM.

• It doesn’t care what the answer is. Right or wrong has no meaning to it.

• Its job is to execute the task based on the pool of information it has access to.

• One of the many complaints I saw about OpenAI’s approach to the bot was the anthropomorphising.

• Portraying it as reasoning or thinking.

• That’s cute when it’s being presented to experts in the field. They’re in on the joke.

• Releasing it to the public in this form, it’s important to do everything to dehumanise it.

• To remind people that it doesn’t think, feel or reason.

• And I’ve seen that in my own interactions with the people who have played with it.

• Absolute astonishment.

Richard Bradbury: Confusing the tool for the solution?

Matt Armitage:

• That’s exactly it. I had a similar conversation with a friend earlier this week.

• And it probably doesn’t help when idiots like me keep telling people that AI will replace us, or at least our jobs.

• But that speaks to the large parts of many jobs that are process oriented and don’t require much creative input.

• AI is a way to wade through the drudgery. And drudgery is not a solution.

• As I pointed out to Jason – my friend – in a few weeks time LinkedIn will be full of almost identical posts generated by ChatGPT.

• For example, I generated a few different versions of that BradMusk explanation, and while the words are changed around,

• You can see that it’s sourcing the same cache of information to create it.

• And it’s making people a little giddy.

• I read one report that Stack Overflow has suspended contributions it believes are generated by the bot

• The site has apparently been flooded with dubious answers to questions.

• Because that’s the thing, the bot creates wrong answers with the same fluidity and sense of authority as it does with correct ones.

• And as Stack Overflow depends on the quality of its answers, it’s kind of an existential business question.

Richard Bradbury: Law of average though. If it’s wrong some of the time, that means it also has to be right sometimes as well.

Matt Armitage:

• That’s not to say that It doesn’t hit the mark some of the time.

• A friend at a software house mentioned that ChatGPT can put together pretty accurate user guides for his company’s products.

• But the point of a user guide isn’t to be pretty much correct. It’s to be correct.

• What’s the point I’m trying to make?

• I guess, what I’m trying to say is use it as a tool by all means.

• But double check any information it gives you.

• Although it generates answers – it isn’t the answer.

• Like I said, if you aren’t great at writing, put all your facts and notes in and see what it comes up with.

• That’s using it as a tool. Telling it to write a social post or an essay for you.

• That’s using it as a solution. And that’s where it can’t be relied on.

Richard Bradbury: How does this take us onwards and upwards and into the Metaverse?

Matt Armitage:

• So, we go back to voice. And Alexa and Siri.

• As I’ve mentioned in previous weeks, we place so much focus on the environmental aspects of the metaverse:

• AR / VR. That we forget about the immersion layer that is voice.

• As I mentioned earlier. What happens when you connect a tool like ChatGPT to Alexa or Siri.

• Because that is the next stage. We aren’t really interested in text bots.

• We’re interested in speech.

• A system like this in the context of a game isn’t that threatening.

• Because it’s information will largely be linked to the physics and lore of that game environment.

• But what about negotiating that wider immersive internet?

• Is this kind of AI fit for purpose?

• Or rather, is there any way to make this kind of AI fit for purpose?

Richard Bradbury: Do you mean in terms of filters and bias?

Matt Armitage:

• That’s definitely one question.

• I’ve been delving into twitter this week to look at some of the comments around this.

Richard Bradbury: Yes, I noticed there was no Twitter update.

Matt Armitage:

• There. Is. No. Point.

• That wasn’t Mattbott: 3000

• We typically record this on a Thursday and Elon Musk tends to do something awful with Twitter on a Thursday night.

• I don’t know whether he’s trying to catch the weekend news cycle, or whether he just stores up the big stuff for the end of the week, I don’t know.

• Some of the Python developers found some interesting issues.

• By structuring their search queries in certain ways they were able to bypass the discrimination filters.

• When I tried to replicate them I got polite warning messages telling me about the importance of respecting diversity.

• So, has the company responded to those finding and implemented a more stringent filter? I don’t know.

• That’s fine. So, I narrowed my searches.

• I asked it to give me a breakdown of the countries where women earn the most.

• Again, polite disclaimer.

• So I zoomed out a bit further and asked it to make me a chart of the top 10 nations by GDP.

• Again, a message saying we shouldn’t try and define countries in such narrow terms.

• Ok. This isn’t a public search engine.

• It’s a beta by a private company.

• They can limit search terms however they want and for whatever reason.

• There’s no right to access or freedom of information.

• But how do you overcome those issues when the machine is a or even the default search engine?

Richard Bradbury: I seem to remember that there was another chat engine that launched briefly, last month.

Matt Armitage:

• Yes, so this is interesting.

• And for anyone who doesn’t see the issue – or doesn’t get how voice commands and navigation will be central to the metaverse experience –

• This might be a little more illuminating.

• The last few weeks there has been so much noise focused on the job cuts and poor financial performance of Meta.

• That this story didn’t seem to get much wind.

• Meta’s Galactica was supposed to be a scientific data generator.

• It was supposedly trained on 48 million scientific papers, websites, articles, lecture notes.

• It promised to do the thing that scientists struggle to.

• To keep up with the flow of information in their chose fields, and parse and summarise it.

• Finding connections in the data. You would also be able

Richard Bradbury: Another search engine that isn’t a search engine?

Matt Armitage:

• This time closer to a search engine. Although with quite a narrow purpose.

• It launched to quiet fanfare in the scientific community on November 15th.

• And it bombed pretty much immediately.

• For the same reasons we discussed before.

• It can’t tell what’s true and what isn’t.

• So, it created some funny things – like a Wikipedia page for bears in space.

• My favourite was why does quentin tarantino hate feet?

• Because he directed Pulp Fiction, a movie about people who hate feet.

• Researchers testing the model claimed it made up papers and references.

• But always delivered its fake news in an authoritative manner.

• Again, this is a test. Scientists and researchers know to be sceptics.

• But this quickly demonstrates the limits of LLMs.

• Statistical correlation of information isn’t enough.

• The probability that this is correct because most people say it is.

• Those limits are quickly exposed when you use it as a tool for science.

• LLMs can’t interpret. So, either we develop machines that can.

• If that’s even possible, or we accept the limitations of the model and use it for the things it actually excels at.

• I’m all for smarter search. But in order to do that, we still need to be the ones evaluating the information the machine produces.

• We need to be aware of where that information comes from.

• And the whole process, including the refining of information has to be a lot more transparent and observable.

• We didn’t even get to the inherent bias of data and I’ve already turned out the clock.

• These are great tools. They are great aids. But they aren’t solutions.

• So we have to be very wary of connecting them to other tools – like voice – where their flaws can be amplified.

• And their operation becomes even more mysterious and less observable.