It’s been really exciting – and sometimes even overwhelming – to see the fast-paced developments in the space of Generative AI over the last few months. It’s clear that Large Language Models (LLMs) and other AI systems are increasingly playing a key role in our Future of Work.
One thing that’s often glossed over is the need for human input when building and refining these models: aside from being trained on vast datasets generated from websites and the broader internet, huge teams of people are helping train these AI models, providing “human processing within a computational framework”, to quote a colleague of mine. Every time you prove you’re a human by responding to a reCAPTCHA – selecting squares of an image showing buses, cars or traffic lights – you’re essentially helping train an AI by labeling data for Google.
If you’re not Google – and have the breadth to deploy a tool like reCAPTCHA – then there are other ways to handle these data labeling tasks, today: some teams use tools like AWS SageMaker Ground Truth to set up labeling jobs that get farmed out to workers using something like Mechanical Turk, others – such as OpenAI – engage with teams of contractors in a more dedicated capacity to crank through their own AI-related human tasks.
From https://aws.amazon.com/sagemaker/data-labeling
Human input is also used during model tuning via Reinforcement Learning, an important step in improving the quality of responses provided by LLMs. Large teams of people are sitting at their machines telling the system which of the AI-generated responses might be considered “better” from a subjective, human perspective.
All of this has created a massive business, and as tools such as ChatGPT displace jobs – something many believe is happening already – it’s very likely that more people will be sitting at home performing short-term labeling work, to effectively help train the AI that has replaced them in higher-value roles.
This is an inevitable progression in the technology landscape, and it’s also inevitable that Autodesk tools play a role in preparing the world for this new eventuality. Our software tools are being used in an exciting new project called AICB (which stands for the Artificial Intelligence Cell Block) that is intended to house tens of thousands – and perhaps in a few years hundreds of thousands – of AI workers.
The AICB project took inspiration from the Munger Hall project at UC Santa Barbara which, once complete, will house thousands of students, 94% of whom will have no access to natural light. This groundbreaking project is paving the way for a future human environment where we are literally cut-off from the outside world.
The AICB consortium is looking to create large out-of-town buildings around the world: these residences will provide for every need of its occupants, but in a rather innovative way. Food will be piped into individual rooms in the form of nutritiously balanced protein shakes: as people work, the shake feed will give them the nutrients they require to stay focused and productive. The AI system managing the flow will inject sugars to help stimulate activity when needed, based on their labeling speed, chronotype and their circadian rhythm. Just don’t stop working if you want the shake to flow!
Working unlocks other facilities for residents, of course: it seems inefficient to have people go to a central gym to workout, so in AICB facilities they’re embracing a more decentralised approach: think Web3 for exercise! Workers will have electrodes attached, and if they work hard enough their muscles will be stimulated electrically to keep them in peak physical form.
For the odd hour or two where workers might want to engage in a little R&R, a virtual Metaverse environment will be available to immerse into without the hassle of congregating physically with people. Life is going to be so much cleaner and safer in AICB residences.
The consortium asserts these state-of-the-art buildings will cater completely to the needs of their residents.
So how is Autodesk software being used, exactly? They did consider using Spacemaker or Generative Design in Revit for the conceptual design phase of the project, but quickly realised it was overkill: there were basically two highly correlated metrics – ROI and occupant density – that were all the consortium members cared about. Other more human-centric metrics such as access to daylight, views to nature, the likelihood of social contact – all these other 2nd-tier metrics become irrelevant with the clever application of technology. So the team just used a traditional Revit workflow to design a massive box filled with tiny rooms. Perfect!
Other Autodesk tools were used to engineer the shake delivery system, from Fusion 360 for the design of custom mechanical components to Autodesk CFD to help reduce the risk of blockages occurring. A lot of effort was also put into a highly customised electrical system for the building: Autodesk EAGLE was used to help design custom occupant monitoring components while AutoCAD Electrical was used to design the routing of power throughout the building. (Luckily all the model training is being performed offsite: the power consumption for most of the work – which can be done with low-powered devices equivalent to Chromebooks – is fairly modest… the main issue to deal with will be around usage spikes if large subsets of the workers choose to enter the Metaverse together; something that’s being discouraged to minimise both power consumption and social unrest.)
Each of these buildings will be managed using Autodesk Tandem, to make sure the various systems remain functional but also to identify when workers need replacing due to prolonged inactivity.
Some of you may be thinking that this project sounds a little dystopian – like WALL-E meets The Matrix – but think about all the upside in terms of efficient use of resources and promoting (admittedly largely virtual) human happiness. In fact the only major project risk that’s been identified relates to maintaining the population of AI workers in the long-term: this current model doesn’t lend itself to the usual ways humans tend to meet and reproduce. But the consortium is working closely with researchers developing technologies that enable human reproduction in the lab rather than the bedroom, so they’re confident a solution will be found before it becomes a genuine problem in the coming decades.
