Can A.I. match a human personal assistant?
Self wants to fulfil Siri's broken promise
PreSeed Now brings you an in-depth profile of a different B2B or deep tech startup every Tuesday and Thursday. Subscribe to get it straight to your inbox.
For all the talk right now in the tech industry about A.I., we haven’t covered many A.I. startups in PreSeed Now lately. We’re going to change that over the next couple of weeks, starting today.
Today’s edition is one those times where we get very close to covering a consumer tech product, but as Self has a partially B2B business model, and an interesting story, let’s dive in.
A couple of things first:
Team PreSeed Now (AKA Samantha and I) will be at the Climb23 conference in Leeds today and tomorrow. If you’re going, stop us and say hello. Confirmation emails for our investor roundtable went out on Sunday night. Also, look out for me onstage on Wednesday afternoon.
If our new PreSeed Now Startup Tracker has passed you by, find out more here. It’s well worth getting stuck into!
💡 Investors: let’s talk due diligence
When you invest in a startup, you need to be sure their tech–the true value behind the deal–is everything they claim it is.
PreSeed Now is sponsored by thestartupfactory.tech’s Technical Due Diligence offering
This service goes much deeper than the technology itself, taking in leadership, technology strategy, product and roadmap, people, engineering practices, and commercial analysis
Self wants A.I. to match human personal assistants
It seems increasingly likely that A.I. is going to impact pretty much every line of work in the coming years. And while creative fields like writers and artists have been among the first to feel the impact, personal assistants might not be far behind if Self gets its way.
Self is, as founder and CEO Jonathan MacDonald puts it, “a hyper personal, A.I.-powered assistant that learns your personal preferences, simplifies your life, and gives you back time for things that matter.”
These are goals that the likes of Google Assistant, Amazon’s Alexa, and (the mother of them all) Apple’s Siri have long promised to achieve, but have fallen short. They might have some surface level insight into your calendar, but they’re far more useful for simple tasks like telling you the weather or playing a song on Spotify.
The hyper-personal assistant
Self wants to use the latest advances in A.I. to push things forward, breaking out of the ‘big tech’ business model that MacDonald argues holds back progress.
“The ambition of an assistant, whether it's Siri, or Alexa, or Google's Assistant, is to keep you in the ecosystem, or to guide you to buy something within it, or experience an advertiser that supports the model.
“A hyper-personal assistant is like a physical PA that understands you, and understands what you are about, what your preferences are, the history of what you've done in the past, and your plans for the future,” says MacDonald.
“It's like an A.I. concierge in the cloud. It's something that I've wanted technology to be like for over 40 years. And I'm glad that now it's able to happen.”
So what can Self do that software-based current assistants can’t? MacDonald has a few examples up his sleeve. The images below show how Self can take information from multiple data sources and then take action on them.
He says Self is also able to look up past events, so it can help with queries like ‘has the place I went to last week with my wife for dinner got any reservations for this Saturday in between meeting my parents and meeting my brother?’.
A human personal assistant would instinctively know how to handle that query off the top of their head, but it would be far too complex for the current generation of digital assistants. And like a human PA learns the tastes and preferences of their boss, Self would learn things like the kinds of restaurants you prefer, and any allergies you have.
“The reason that technology doesn't work like this is because it is hardwired, and fundamentally uninterested in working for you. You at the moment are meant to be a battery for the technology platforms to increase shareholder value. The way I see it is that technology should be a service for us,” argues MacDonald.
When it launches, MacDonald says Self will charge a subscription fee to users, with multiple tiers for different levels of access. But the startup also plans affiliate partnerships with businesses, whereby it would take a cut of transactions made through the app. These could be shopping, holiday or travel bookings, for example.
How it works
Rather than start from scratch, Self combines its awareness of your personal data and circumstances with queries that it runs through third-party data sources that it has access to. In this way, MacDonald says, it can build on the capabilities of other A.I. services such as ChatGPT.
Self’s own capabilities are being refined through Reinforcement Learning from Human Feedback (RLHF), the technique by which A.I. tech such as OpenAI’s GPT large language models have been trained to respond in a human-like way.
RLHF works by having humans check an A.I.’s output and provide input to improve it.
MacDonald summarises the way Self works as “A.I. sucking in from other sources, and human feedback to make sure that what you receive in conversation or in response is accurate to you and your preferences.”
“In the back end, we have the Self A.I., which is partly looking at every other result that the brilliant technology of the world could provide, and we can have feedback from the actual human virtual assistants to say, ‘well, from all of this information and the human feedback, this is the best answer for you, the user’, And over time, that increases its capability,” says MacDonald.
“We call it the ‘What’ model and the ‘Why’ model. Common A.I. tends to base itself around ‘what’, so if you ask it to write a poem, it will find the most common type of poem structures to reply back. But the ‘Why’ model that we include with that asks ‘why would that person say this? Why would a response suit that person at this particular context?’”
In a world where users are becoming rapidly accustomed to A.I. giving them extensive answers to their questions in seconds, the idea of human involvement in Self’s responses might feel like a backwards step, but MacDonald says it’s essential at the startup’s early stage.
“A.I. at the moment is extremely intelligent, and remarkably stupid. It's great at winning chess games, great at writing an essay, brilliant at writing code and checking code, it can write guitar solos for people that make musical sense. But if I ask it to suggest a present for my mum for her birthday, it will be absolutely clueless.”
Close watchers of the A.I. scene might at this point say that ‘giving A.I. memory’ is a whole subsection of the feverish levels of product development going on around the world right now.
Perhaps most notably, DeepMind co-founder Mustafa Suleyman has launched Inflection, which has produced an assistant called Pi, designed to learn about the user over time.
“Pi says that it can remember and unfortunately, it had a short-term memory in our testing. That's one of the reasons why we need to ensure that there is constant human feedback to improve Self’s A.I. capacity to remember. If we leave it without any form of guidance or guardrails, the outcome could be ranging from inaccurate and inefficient, to downright dangerous.”
But doesn’t having humans check Self’s work slow down its output? MacDonald says Self currently works at a similar pace to having a conversation with someone over WhatsApp.
Some conversations with Self see instantaneous responses. These are where Self is learning about you through what MacDonald calls ‘progressive disclosure’. Users share information about themselves and their preferences over time rather than having to fill in a huge personal survey before they use the app.
For these interactions, human checking isn’t required. But for more complex interactions like arranging a stag night with seven people with seven different schedules, and who live in three different countries, having a human check the A.I.’s working is far more useful while the Self A.I. system develops its capabilities.
MacDonald has a long history in business, with a focus on marketing and entrepreneurship. He says the idea for Self came from his vision of how he’s wanted the internet to work ever since he first tried the World Wide Web in the ‘90s.
“I couldn't, and still can't, understand why I need to open seven tabs to book a hotel and transport to a particular city, because I'm using a search engine that absolutely knows who I am, and what I'm trying to do… I’m a frustrated user who knows that machines could do more but are choosing not to.”
The team he’s launching Self with boasts significant tech and business experience, and includes a specialist in virtual assistants - a field Self could well disrupt if it takes off.
The idea behind Self reminds me very much of Fin, a startup founded a few years ago by Sam Lessin in the US. In 2017, TechCrunch described Fin as “the elite human/A.I. assistant”.
This came off the back of the chatbot hype around that time, where chatbots working with natural language processing and decision trees (and often more than a little human help) gave the impression that truly smart A.I. assistants were just around the corner.
In actual fact, the tech was just far too early and leaned far too heavily on humans for too long. Facebook’s ‘M’ assistant didn’t get beyond limited testing and closed in 2018, while Fin pivoted to other ideas and its domain now points to someone’s LinkedIn profile, weirdly.
So is now the time for the concept to come back? Unsurprisingly, MacDonald believes so.
“There is no doubt the zeitgeist of the moment is A.I. As an investor, if you're looking to catch the wave, the wave is A.I… The question that an investor has is, ‘do I invest into AI that is useful for the large tech platforms ultimately, or do I invest into an AI that's useful for the 7 billion people on Earth?’ We’re hoping that they go for the second option.”
Go deeper on Self
More on investment, their vision, their competition, and their challenges:
Keep reading with a 7-day free trial
Subscribe to PreSeed Now to keep reading this post and get 7 days of free access to the full post archives.