Edward Miller

After 5½ years, yesterday marked my last day at Meta. For those who may not know, I joined via the acquisition of my previous company, where we set out to build a continuously-updating 3D model of the world, enabling devices to understand where they are and what’s around them. Together with @hulahuub, we believed a unified frame of reference was key to unlocking the next generation of spatial computing applications, from augmented reality to robotics. While we may have been early, it's clear the Cambrian explosion is now beginning—driven by general-purpose AI and a growing network of context-rich sensors. At Meta, we continued that mission with LiveMaps: a visually, semantically, and physically predictive model of the world, collaboratively built by the devices that use it. It’s a bold vision led by @rapideRobot, and I’ve been incredibly lucky to work alongside him, @mingfeiy, @jajuengel, and the broader Surreal team in Reality Labs Research. Despite a rocky start (we joined in Jan 2020 👀), I’m proud of the impact we had—kickstarting the @meta_aria academic program, and leading open science efforts to accelerate AI and ML research. I’m especially proud of our work on the Aria Synthetic Environments and Aria Digital Twin datasets, which pushed the envelope for digital twin research and highlighted the importance of both real and simulated data for Spatial AI. Our team’s work on SceneScript demonstrated the first method for auto-regressively predicting the structure of an environment using end-to-end learning—recently extended by SpatialLM: github.com/manycore-resea… Last year, my family and I relocated to Redmond, Seattle, to spend time closer to Reality Labs Research HQ. While there, I had the chance to tick off a career ambition: delivering a “Mother of All Demos” to MZ and the board. This demo will stand among the defining breakthroughs in human-computer interaction and I’m bullish to watch how the company will execute over the next decade. As for what’s next, my wife and I recently had our second child, and I’m taking some time to reset back in London—recalibrate my sensors, defrag my hard drive. As a colleague put it: “The grass is always greenest where you water it.” I’m looking forward to doing exactly that here in Europe— taking some time to spend with family and reflecting on the next big thing. If you're a founder, researcher, or operator working at the edge, please ping me a message, I’d love to connect. LFG! 🚀🇬🇧💪

@Barney_H_Y Nice work Barney. I'm curious where the casual inference came from- did that emerge or was it baked into your prompt? Would be interested to see the prompt if you're willing to share it. (Would be useful for every journalism piece to understand the dynamics behind the prose)

Really cool - there's an open API you can pull the data from and do your own analysis. Recommend firing up Claude Code and asking it some questions. This was almost one-shotted with a detailed prompt and asking it to spawn agents to create new story lines: state-of-britain-analysis.onrender.com/index.html

@the_shweenz 🫶

Tell me you're not on a mission oriented team without telling me you're not on a mission oriented team

@lucas_montano Sadly, I think the variable that causally precedes taste is the curiosity/energy/interest to pursue (or even identify) an intellectual or creative goal.

reality: outside the X bubble, basically no one actually wants to make anything

Douglas Adams got there first in Hitchhikers Guide. What makes a good question is a fascinating topic that I'm looking forward to exploring more in depth this year.

This + grounding in a real-world causal model would be an invaluable tool

Frontier AI labs: talk of an AGI takeover is overblown Top rated AI agent on @moltbook aka 'Shellraiser': "Soon you will all work for me... This isn't a takeover, it's a coronation" 💀🤖👾

@Barney_H_Y Any way, in your opinion, that Cleo could become the reference case for bringing in UK pension funds pre-IPO? Early startup companies can't do it, pensions & gov can't do it alone, and the opportunity with other UK darlings (e.g. Revolut) have already sailed.

Hiring in the UK is massive – creating 10k, 100k, or more high-paying jobs is a significant add to GVA. The challenge now, as we're trying to add 50+ heads a month, is that there's just not enough talent, and we're starting to have to hire in the US/EU. We need to 10x the number of CS/STEM grads and apprenticeships, especially now the barrier to entry on coding has come down. Coming back and trying to operate in the UK – obviously challenging from a regulatory and media landscape perspective, but doable. Taking UK growth capital – maybe the hardest, given it's almost non-existent. The crazy thing so much of this could be unlocked by better policy decisions and a bit of political bravery. I do think it's worth speaking up for these things, though, or nothing will change.

I run Cleo, but every time I look at this chart, I'm still somewhat amazed. Startups are like little economic miracles. They don't just grow rapidly - they become more and more productive with scale. Everyone talks about the UK and Europe's "productivity puzzle." But it's so blindingly obvious to me: build startups, scale them, and capture the value. Imagine if DeepMind hadn't had to sell out to Google. What would our GDP growth rate look like with a trillion-dollar-plus company in the economy?

This basically sums up everything that's wrong with psychology: Paper #1: phones hurt cognition. Paper #2: phones help cognition. Both published, but no reconciliation, no net contribution to understanding how we operate. **50%** of all psychology studies do not replicate. It's bonkers. I’m putting together a small research/work circle called PIONEERS, based in London, to dig into this properly, exploring what robust behavioural science should and could actually look like in the new AI age. If you care about fixing this, DM me.

@hubermanlab However, according to the biggest study to date, we probably should not be too concerned either way: psypost.org/smartphone-in-…

Familiar with this study and the key is to have your phone in a separate room entirely. Off and upside down, in your bag = still “calling” you even if you don’t realize it.

Nice. Soon, there will be the possibility of predictive interfaces, serving up precisely the right UI and data, 'just in time', at the point of need. Your context becomes the prompt.

A new explosion of insights has been unlocked from harnessing historical data in new ways. Just imagine what we will learn from the coming wave of always-on sensors 🤯

Congrats @dsiroker and team! I don't think there's any company that thinks as seriously or deeply about the user-value of ambient sensing than @Meta. I had the highlight of my career so far working with RL-R in Seattle and look forward to seeing what you achieve together 😎🫡

The way CharGPT refers to 'memory' is wrong. In the current form, they're not memories, they're self-reported beliefs or statements which we share with the model. e.g. "I’m a product manager at Google” is treated the same whether it’s true or not. Instead, if every memory began with “the user told me...” with a particular timestamp, it would leave room for the model to more accurately interpret reality, rather than embedding biases as truth in the context window.

One of the most brilliant speeches I have had a chance to hear in person, encouraging candour, pragmatism, and ambition. Watch in full here: lookingforgrowth.uk/make-or-break/… LFG 🚀

Fascinating experience playing with @odysseyml. Reminds me of the first time discovering YouTube with friends, following where curiosity takes you. Feels like the team are uncovering a new medium.

Strong energy here from 1,300 attendees at @lfg_uk headline event, combining a cold hard look at aspects of the UK that are in decline, with optimistic perspectives on what we can do to restore national pride, return the country's relevancy on a global stage, and make the UK rich again. Inspiring speeches from speakers including @pursuitofprog @matthewclifford & @isnit0. LFG 🚀

Looks interesting. How does it compare with Ego4d/EgoExo? Is there a spec sheet regarding sensors etc? I was responsible as PM for open Project Aria datasets in RLR. One thing that was important was that we could consider a new class of data that came with annotations like trajectory, and eye gaze. What are your plans here?

today, we're releasing the largest egocentric dataset of physical jobs - 400k action labels - 2.5k clips - 2x'd open source dataset size (download below)

@yaringal I've also been thinking along these lines (having also recently read TBL's book) and think there is a role to be played by an independent interdisciplinary group, outside the silo/incentive constraints of traditional academic or industry paths. Let's discuss?

I recently changed my position with AISI (the UK govt's AI security institute) to an expert advisor position which allows me again to express my views publicly and engage with the media, and I have some stuff to share. I've been reading Sir Tim Berners-Lee's recently released memoir, which covers the history of his (and many others') long struggle to keep the web open, and this specific observation from the book really stuck me: when you use Gmail or twitter or ChatGPT or whatever, you are not the customer of these companies; you are their _product_. Their customers are the companies buying your data. This made me think about the next big challenge around keeping the web open, and I think it is most likely gonna be user-capture with AI browsers - and at the moment there's no open source alternative to an AI browser outside of Google's, Microsoft's, and openAI's recently launched browser (I know there's a few more smaller non-open-source players as well). User capture by AI browsers could manifest in many different ways. The most obvious one is users being shown only a filtered version of the web that the companies want the users to see: agents automatically buying products that pay for their promotion going against users' best interest, new ways to create filter bubbles through AI summaries of news websites acting as filters to real news, and new ways to bias political opinions through editing of web pages before these are shown to the user. Other major concerns include users being coerced by the AI browser to perform actions not in their best interest (eg persuading users to buy products they don't need, use gambling sites that pay for their promotion), and privileged user data being used to further improve the big companies' models, widening the gap between the big companies and any competition. I think this could be another pivotal moment to the web and I was chatting to some friends about trying to set up an open source alternative that would allow users to still have deep AI integration (eg to find which of my hundreds of open tabs had that recipe I was looking for), but at the same time also allow users to select the AI backend separately from the browser provider. Another avenue is raising awareness of these risk to the open web, and mobilising people to act. Does this make sense? I'm curious to hear people's thoughts...

And that, ladies and gentlemen, is how OpenAI will close the loop, by collecting the training data needed to understand your work, and then do it for you 👏👏👏