Categories
Uncategorized

Be weird to diversify Midjourney images

We know that the data used to train foundational AI models perpetuates bias, and in Midjourney that most clearly manifests1 in results involving people; unless explicitly instructed otherwise, the people in generated images tend to look beautiful.

To show you what I mean here’s a prompt I wrote, with the resulting generated images below it:

candid photograph, close-up, a young Korean woman with bleached blonde hair, bright natural light, happy, joyful --style raw

The results look great, but clearly fall squarely within conventional standards of beauty. Not everybody looks this beautiful (myself included), and we should reflect that.

Categories
Uncategorized

How popular are Roblox branded experiences?

This year lots of brands launched experiences on Roblox; some to catch the ‘metaverse’ hype, some for experimentation, and some with a genuine desire to engage a young audience in playful social spaces.

Roblox publicly records the number or visits to each experience, which can give you some idea of popularity, although it doesn’t list unique visits. However, some experiences reward every user with a badge for visiting for the first time, so we can use that, where available, to record unique visits.

Gucci Town’s Welcome badge shows how many users entered for the first time

I looked at a number of branded experiences to find those with a first-time user award, and you can see their relative popularity in the chart below. It’s not a fair like-for-like as some have been online for longer than others, but I still think it’s worthwhile to take a look. The chart shows total visit numbers, in millions, with unique visits recorded in a darker shade.

Data gathered on 15/12/2022

In absolute numbers, NARS Color Quest is the most popular experience to date, with almost 42 million visits (it also has the highest user approval rating, 86%, of all the experiences in the chart). But Gucci Town and Nikeland have the joint-highest number of unique visitors: 11.6 million apiece. Considering Roblox states 58.8 million daily active users, that’s good going.

Working from the total visits and unique visitors we can get a record of how many times each user returned, on average. This is shown in the chart below (note that I’ve excluded some of the least-visited experiences as their results seemed skewed):

Data gathered on 15/12/2022

NARS Color Quest might not have had the highest number of unique visitors, but it has by far the highest number of total visits, as each of its users returned over three times, on average. Gucci Town and Nikeland have had an equal number of unique visitors, but Gucci ranks much higher in return visits, with 2.26 to Nikeland’s 1.45. Note also here that Spotify Island and iHeartLand Music Tycoon do well in returning visitors. Visitors to Walmart Land have only came back for an average of 0.5 visits (and it has a user approval rating of just 58%).

There are a lot of conclusions that can be inferred from this small study, but I need to explore them in more detail and with more scientific rigour when I have more time. In the meantime, if you’re wondering how well some of these branded experiences do, perhaps this will help.

And if you come across a branded experience which has a welcome / first-time badge but which I haven’t included here, be sure to let me know.


NB I updated the title of this post after publishing; it was originally “How Successful are Roblox Branded Experiences?”, but we can’t really measure success without knowing goals, so “popular” fits better.

Categories
Uncategorized

A conversation with ChatGPT

This week OpenAI introduced a new conversational text generator called ChatGPT. Here is a direct transcript of a conversation I had with ChatGPT where I asked it to explain itself.

Explain ChatGPT in a single sentence

ChatGPT is a large language model that was trained using a combination of text-based conversations and task-oriented dialogue data to perform natural language processing tasks and generate human-like responses in conversation.

What is a large language model?

A large language model is a type of artificial intelligence that is trained on a large amount of text data to generate human-like text and to perform natural language processing tasks. These models are typically based on deep learning algorithms and are able to generate text that is similar to human writing or speech. They can be used for a wide range of applications, including generating text, summarizing documents, translating languages, and answering questions. Some examples of large language models include GPT-3, BERT, and ChatGPT.

Explain it as a limerick.

A large language model’s grand plan
Is to understand the ways of man
It processes vast amounts
Of text to make new accounts
Of human-like speech, if it can.

Will large language models replace indexed search?

It is unlikely that large language models will completely replace indexed search, as indexed search and large language models serve different purposes. Indexed search uses a pre-determined set of keywords and indexing rules to quickly find relevant information based on a specific query. Large language models, on the other hand, are able to generate human-like text and perform natural language processing tasks, but they do not necessarily rely on pre-determined keywords or indexing rules. Instead, they use the vast amount of text data they have been trained on to generate responses that are relevant to a given query. While large language models may be able to assist with search in some cases, they are not designed to replace indexed search entirely.

How can we be sure of this?

It is difficult to say with certainty that large language models will not completely replace indexed search, as the field of artificial intelligence and natural language processing is constantly evolving. However, based on the current capabilities and limitations of large language models, it is unlikely that they will be able to completely replace indexed search. Indexed search is designed specifically for quickly finding information based on pre-determined keywords and indexing rules, while large language models are trained on large amounts of text data to generate human-like text and perform natural language processing tasks. These are different purposes, and it is unlikely that large language models will be able to fully replicate the capabilities of indexed search. Additionally, indexed search has been widely used and refined for many years, so it is likely to continue to play a significant role in information retrieval.

Explain that again in two sentences

It is unlikely that large language models will completely replace indexed search because they serve different purposes. Indexed search is designed for quickly finding information based on pre-determined keywords and indexing rules, while large language models are trained on large amounts of text data to generate human-like text and perform natural language processing tasks.

Rewrite that as a Shakespeare sonnet

In this age of artificial might,
We see machines that can generate text,
But indexed search still holds its place,
For quickly finding information with grace.

Though large language models may assist,
In tasks that require language and list,
Their capabilities do differ,
From the indexed search’s specific method.

And so we can be sure, in this time,
That indexed search will not be left behind,
For though large language models may evolve,
Indexed search will still be part of the resolve.


It’s fun to play with and can feel like magic, but its risk lies in that it can give answers which seem authoritative but are wholly incorrect — although, to be fair, humans are more than capable of this too.

Categories
Uncategorized

Neural Radiance Fields forever

A few weeks ago, Google introduced Immersive View for Maps, a high-quality 3D flythrough of popular spots, which shows day / night and real-time weather conditions.

It’s pretty cool. But even cooler is the technology behind it.

NeRF (Neural Radiance Fields) is a new approach to photogrammetry.