sam wilson

This got a lot of attention, so here's some additional context: 1) I think the most interesting example is when we asked o1 to actually define a relevant problem itself - like this: chatgpt.com/share/670b0721…. Note that o1 seemed to reason through the problem effectively, but came up with an unreasonably large section size for the truss chord. With a simple question on the efficiency of this decision, it was able to re-evaluate and come up with a more reasonable solution. 2) Important: she didn't review any calculations in detail (including in the fun example I linked above)! What is most interesting is not whether o1 is able to get all of the math exactly right, or to retrieve codes perfectly. But rather it's the fact it can genuinely reason about the design of a building, can make sensible decisions (not always, but often), and can explain those decisions. With the right follow-up prompts from an expert it can easily correct flawed thinking, much like a less experienced engineer does in practice. 3) This was done late last night after I asked her on-the-spot to come up with the hardest problems she could think of. o1 was able to solve them, but it's of course still far from perfect. It definitely needs guidance and review from someone experienced when it makes practical mistakes: things like using unreasonably large member sizes that are not feasible cost-wise. 4) This is of course not going to replace structural engineers today - but it is definitely a big step in that direction - and it can already speed up a lot of the tedious repetitive work more junior engineers have to do by 10X. It also means you can iterate much faster, rather than having to rerun calculations by hand multiple times. 5) As with all domains, we really need proper agentic capabilities for this to start fully automating large parts of human jobs. LLMs need to be able to take the reasoning ability o1 has and apply it to the very specific contexts that real world problems exist in. This means being connected to the relevant data sources, tools (like modeling software), and effective ways to interact with the hundreds of different stakeholders involved in these projects. This final point is such a huge part of the job of an engineer, and one we still need to build the infrastructure to support. 6) Maybe it's 2 years and maybe it's 10 before this plays out to the full automation of large parts of what an engineer does. But I'm certain it will, and that's why the world of work has completely changed, and most people don't realise yet.

a little history for my football fan followers! @TheAtlantic theatlantic.com/newsletters/ar…

Today is National Bean Day. We have the best one! #chicago