What’s Going On in the AI Apps Market?
AI hype is everywhere and products from the hyperscalers have entered our daily use, Office CoPilot, Midjourney, ChatGPT etc. Great for silicon valley, but what of the 10,000 other AI products out there? I asked AI to characterise them for us.
Part 4 (Part 1, Part 2, Part 3)
In the previous blogs we looked at the revolutionary team management frameworks allowing LLM’s to collaborate for writing software. The packages were:
We pitched these team frameworks against each other in a data science competition. Each team was to analyse the AI tools available on the market, how they group, what’s popular and what niche is over crowded. If you’d like to see how they performed then do read those articles (links above). What follows is the best of the analyses, tided up by myself:
Prospecting for Market Data
We need a data source to understand the AI tools market. The OpenAI plug-in store, still in its nascent stage with a mere few months to its name, is already brimming with over 1,000 apps, reminiscent of the mobile app stores. Yet, unlike those stores, there is no data on the relative popularity of each plug-in
While platforms like ProductHunt and Crunchbase track new market entrants, they come with their limitations — Crunchbase is behind a paywall, and ProductHunt presents challenges for data analysis.
Enter TheresAnAIforThat.com, an aggregator that not only catalogues 10,000 AI applications but also captures engagement via user endorsements, presenting a crowd-sourced measure of popularity. The site declares approx 250k searches per day.
The data was taken in October 2023 when there were 8,000 tools on the market.
How The Data Was Analysed
The objective was to shed all presuppositions by employing unsupervised learning techniques. Here’s how that was done:
- Embed each app’s ‘use case’ via the BGAAI small model, a top performer on the Huggingface Embedding Leaderboard.
- Reduce dimensionality via UMAP
- Cluster the data via HDBSCAN, allowing the data to tell us how many clusters there should be
- Use GPT-4 to summarise each cluster’s use cases into a single description for the cluster.
- Present Clusters as a TreeMap
The Jupyter notebook, including all the wanderings in reaching the solution can be viewed on GitHub
Results
The size of each block in the above treemap relates to the number of apps represented by that cluster, total is 8,000. The classifications were by assigned by GPT-4.
The most popular app for each cluster is listed below.
The number of apps per cluster indicates just how crowded the market already is.
The number of likes per app in each cluster reflects the interests of theresanaiforthat.com’s audience. Having investigated the data further, this is likely to be an audience which is young, male and tech savvy.
Nevertheless, it is notable that the most liked grouping is Video Production. Recent success of TikTok, Instagram and YouTube shorts demonstrates the spell binding power of short form video.
HyperScaler’s Current AI Strategy & Plans
The startups must find value where the hyperscalers are unwilling to tread.
Microsoft
Pursuing its Office Co-Pilot strategy for Enterprise, whereby Office clients, already storing their data in the cloud, can easily use to AI ‘Cop Pilots’ to query, summarise and create with that data. Famously Microsoft has redeemed Bing with the first useful AI as search functionality.
Amazon
Currently focussed on AWS Bedrock tools for enterprise AI assisted search and fine tuning. Signing up major suppliers of AI models such as Stability. Curiously reluctant to employ AI in Alexa and exploit their consumer reach.
Meta
Meta is pursuing LLM’s which encapsulate celebrities to converse with, a celebrity who knows and cares about you. Powerful, but even more impressive has been their commitment to open sourcing models like CodeLlama, which has denied OpenAI and Anthropic quiet enjoyment of that space.
Google
Where Microsoft is integrating AI into all its apps, as a UI, Google is followin inot Gmail, Maps, Photos etc. But also, they are making it simple for developers to create digital assistants and work Google technology into thei apps, witness the Model Garden.
Inflection
Pi.ai is uniquely emotionally intelligent, AI to build relationships. Perfect for applications including social media, health and even management. Also, a natural segue into a world where AI is the window, the relationship, we foster for all our connections and queries on the wider world.
Stability
An offshoot of opensource, Stability are pursuing ever smaller models and community involvement, possibly with a view to mobile devices.
OpenAI
Integrating AI into your daily life (search, chat, slack, multi modal, speak via mobile app, plug ins to other services). Famously unpopular when it stepped into the same arena as many of its customers, startups who were using the API service.
What’s Left for StartUps?
10,000 apps on TheresAnAIforThat.com is as many apps as there are visible stars in the sky. Existing software products are racing to include AI in their product offering. So, what is left for startups? Where can they deliver value?
- Data
- Pursue industry specialities, where data is firewalled or privacy a core demand
- Data integration will be key for Agentic AI (as per GPT Assistants)
- Environment integration will also be key for Agentic AI. See my previous post in this series for discussion - Price
- Opensource alternatives to hyperscaler services with large price tags. Users will tolerate lower quality for smaller or cheaper models. - AI as an OS
For example, see MemGPT. This suggests room for an opensource, Linux approach. Value via services to support AI-OS performance and guarantee privacy. - AI as UX
- As with the disruption of smartphones in 2010, whereby ‘mobile first’ services displaced desktop software, many established businesses may be slow to adopt ‘AI as a UI’, permitting a toe hold for startups - Compliance & Audit
- National compliance rules open niches for start-ups across the world, not least in the professions: law, finance and medicine
- Companies wishing to use AI often need an alternative supplier to provide audit of those tools. As with credit and finance, they will be required to have separate providers