Graphcore has developed an intelligent processing unit (IPU) to speed up machine-learning. The product has yet to hit the market but Graphcore claims it will lower the cost of running AI applications and make machine-learning more accessible and efficient. Through publishing content on its website, the company also hopes to become a platform for discussions about AI and its potential to transform various industries.
Pentagram partners Luke Powell and Jody Hudson-Powell’s identity for the brand is based around the idea of resolution and the way that neural networks (computers that are modelled on the human brain) operate.
A custom typeface (Graphcore Quantized) is based on Caslon’s Egyptian and contains 65 alternate characters of varying resolutions. Through the use of OpenType features, the typeface switches between different resolutions as users type with it, producing a different outcome each time. The process is ‘pseudorandom’ – the combination of characters used influences the system’s choice of alternates, which means that as someone types, the letters they have previously typed also change.
Text such as article headlines on the Graphcore website appear animated – the ‘ticking’ effect was created using OpenType ligatures. Ben Leonard, who worked with Powell and Hudson-Powell on the project, explains: “We add these things called ‘zero-width characters’ which don’t appear on screen or take up any space but when combined with one of the [visible] letters make what is essentially a ligature, which selects one of the alternate characters. That way we can target individual characters.”
The team – made up of Hudson-Powell, Powell, Leonard, Mat Hill and Sabrina Maerky – also created a shape generator for Graphcore, allowing the company to generate an infinite number of patterns for editorial content, presentations and communications. The generator is part-random and part-weighted: it subdivides and then places shapes randomly, but is more likely to place shapes next to others – a process Leonard describes as “clumping” (see GIF below). Users can also alter the probability of different shapes appearing. “We figured out early on that the best shape patterns happened where you have multiple shapes clumped together,” adds Leonard. The shape generator has been used to create patterned business cards for staff members and even Graphcore-branded cycling jerseys.
Speaking to CR about the project, Powell and Hudson-Powell say they were keen to avoid dystopian or clichéd representations of artificial intelligence. They also wanted to steer clear of “masculine tropes” and “overly complicated language” often seen in tech branding.
“[Graphcore] wanted to move away from anything too tech-y or too niche, or too sci-fi leaning, and towards something more open and accessible for everyone,” says Powell. “They’re creating a new kind of chip which is completely different [from other processing units available – many of which were created for the gaming industry rather than machine-learning], so that afforded them the right to do something different, to tell a different story to their competitors.”
The team initially experimented with creating an identity system using neural networks but decided instead to focus on creating one that mimics the way these systems function. This led to the idea of creating a system based around the idea of resolution and pixels.
“We initially got really excited by thoughts of using machine-learning to create brand elements or systems that governed brand elements,” explains Hudson-Powell. “We had played with neural networks in the past to imagine how you can create icons, but as we got deeper into it, we felt that progress is happening at such a rapid rate, that if we leant too much on algorithms … we would end up with elements that would age quite quickly,” he says. “Graphcore gave us a 101 on machine-learning – how it really works, and how computers start to recognise patterns, [for example] it can understand ‘that is a nostril’ and ‘that’s a face’ and ‘that’s a woman’s face’ – and we became very interested in the idea of resolution.”
“If you take an ‘O’, for example, at one level you can see an ‘O’ in small constituent parts and it becomes pixelated, and it’s how we imagine the machine looking at it – piecing these singular components together until it becomes more circular and how we understand an ‘O’ to be,” he continues. “Each of the letters within the font has [multiple] versions so when the client types, they get a different resolution of that character each time.”
Different resolutions can be used in different circumstances, explains Hudson-Powell: “If you’re writing something technical, you can lean on more pixelated version of the font or for something else, you can lean on the more finely rendered. It almost makes the font into an illustrative tool, which is really interesting.”
The pair also commissioned Carla McRae to create illustrations for the brand depicting potential uses of AI. Her artwork adds a human touch to the identity and aims to make machine-learning seem more accessible (Graphcore’s IPU was created with small startups and educational institutions in mind as well as larger corporations so is aimed at a broad audience). Her artwork provides an alternative to stock images of robots and can be used to illustrate articles on Graphcore’s website as well as marketing material.
“When people are writing articles about machine-learning they often fall back on 3D rendered pictures of robots serving people in shops … there’s not that much out there for people to use,” says Powell. “We commissioned two large illustrations that can be cropped into [smaller artworks] that represent lots of central subjects [Graphcore] will be dealing with on the website.”
The company blog will feature articles explaining different uses of AI and its potential impact on a range of sectors. Graphcore hopes its editorial content will help demystify the technology and spark discussions about machine-learning.
“Machine-learning is such a big conversation … but at the moment there is so much fear surrounding it – we’re seeing this every day in popular press, and Graphcore is hyper aware that AI can potentially be seen as quite a dangerous thing, but it doesn’t have to be. They want to open up the conversation [about AI] and not be this dark, mysterious company … but an open forum [to discuss] what humanity is using this technology for,” says Hudson-Powell.
Having experimented with algorithms and neural networks before working on the project, Hudson and Hudson-Powell have learned a lot about AI of late. The pair are pretty positive about the potential impact of machine-learning and think it’s time we stopped talking about it as a threat and started thinking of it as a useful tool – one that could help creative as well as researchers, technicians and massive corporations.
“Jody and I are really interested in sci-fi and how you can use fiction as a way of imagining and creating a global consciousness, a movement towards creating a better future,” says Powell. “That’s what you had in the past with the Space Race and things, but contemporary media and movies, and most press, paint a super dystopian image of our future. It’s exasperating and disabling – it’s easy to fall into this pattern of ‘the machines are going to take over, it’s all terrible’, but it takes away energy from people [that could be put into] understanding what the technology is and how they can get involved with it.”
“AI is already here, and the better people understand it, and the more approachable it is, the more we can have AI work in service of us, and supercharge us,” he continues. “Lots of people say the creative industries are safe from AI … as if the end point is robots coming for our jobs, but that’s not the right way of looking at it. Creatives should be looking to AI and the way it can superpower our industry…. It’s not sentient, it’s not this scary thing – it’s just another piece of technology that can work for us and something we can have a lot of fun with.”
As a thank you to the client, Powell and Hudson-Powell created a book that features visualisations of computational problem being solved on their chips. The striking images resemble petri dishes or scans of the human brain but in fact show how Graphcore’s products work. Images are also used on the company’s website.
It’s a clever system and an imaginative response to a challenging brief to visualise a complex piece of technology that is often shrouded in negative connotations.