AI × XR × Culture | Co-founder @MidBrainAi and @thepointlabs | ex-@volvocars innovation lead |Teaching machines to make sense. Nerd by day rock star by night 🎸
Copenhagen, Denmark
Joined September 2011
- Tweets5.4K
- Following2.1K
- Followers1.9K
- Likes22.3K
One of the most frustrating things about AI today is having to explain the same thing over and over again.
So I wanted to make that problem visible.
In the little household simulation I built with Claude, we ran the same year twice. Five robots, four residents, the same tasks and the same underlying code.
The only difference was memory.
On day one, Dot starts vacuuming outside Kai's bedroom while he's sleeping. Kai complains.
Four weeks later, the robot without memory makes exactly the same mistake.
The one with memory knows to come back later.
Another robot, Nova, keeps greeting Elena at 06:40 in the morning. She's not particularly happy about it.
After enough experiences, Nova learns that she prefers not to be disturbed before 08:10.
Then there's Elena's favourite teacup, which she wants washed by hand.
Without memory, Tidy puts it in the dishwasher 365 times over the simulated year.
With memory, it makes that mistake once.
Across the two recorded simulations, we ended up with:
1,280 household corrections without memory. Just 7 with it.
Now, these are controlled simulations, not real household robots. And nothing was retrained. The agents simply had access to their previous experiences and could use them when making decisions.
But it illustrates something I think we're still underestimating.
Learning isn't necessarily about becoming more intelligent in some abstract sense.
Sometimes it's about remembering that someone doesn't like being interrupted in the morning. Or that a particular cup shouldn't go in the dishwasher.
These small things are how we build familiarity, trust, and eventually relationships.
And I think they will matter just as much for AI companions as they do for people.
What is something you've already taught an AI assistant that it still makes you explain again?
For close to a decade I worked on how people and machines live alongside each other, first with human-machine interaction and autonomous systems at @volvocars and now at @MidbrainAI, and the thing that still bothers me most is how quickly our tools forget us.
I notice it in my own days all the time. I’ll talk an idea through out loud with a voice agent on a walk, sketch the layout on paper when words stop working, and then sit down in the evening and explain the whole thing again from scratch to a coding agent that has never heard of any of it. Each tool is impressive on its own, and yet every hop between them starts at zero.
My colleague Jesper Derehag just published a piece that puts words to what we’ve been building toward since 2024: a companion that carries one continuous thread of shared history through every tool you use, and keeps learning from it over years rather than hours.
The very important part is the distinction between storage and learning. Storing a fact like “user prefers tea” is useful, but it means the agent has to know to look for it, find it, and reason its way back to offering you tea every single time, which is both expensive and quietly fragile.
A system that actually learns needs storage for exact recall and something closer to muscle memory for the things that happen often, so your preferences and habits simply become how the agent behaves.
That’s also where taste lives. I have never told an agent that I prefer short, friendly emails that end with a clear next step, but after months of small edits that is exactly what a learning companion should pick up on its own, the same way a colleague of many years just knows those nuances.
We started exploring this as a character in @Minecraft , then as a simulated human that helped robotics teams study how machines should behave around people, then in digital twins of entire cities, and today in coding agents. Every one of those embodiments taught us the same lesson: an agent only starts to feel useful once it has accumulated real experience with you and adapted to it.
And because that experience is so personal, we hold one principle firmly. Your agent’s memories, experiences and learned skills belong to you and no one else, which is also why, only half joking, we like to say that my agent is better than your agent.
#AIAgents #ContinualLearning #AIMemory #AICompanions #HumanAIInteraction #LifelongLearning #PersonalAI #AgenticAI #EmbodiedAI #MidBrainAI #AI
Jesper’s full piece: midbrain.ai/blog/a-lifelong-…
Live today: E2E-encrypted storage, automatic consolidation into profile, facts and skills, and MCP support across every major coding harness.
Timmy Ghiurau retweeted
google wrote a paper, published it for anyone to consume, then watched all eight authors leave & two outside labs capture most of the frontier model equity.
the most valuable research paper in history was effectively a gift from google's balance sheet to its competitors.
re mark able.
Last weekend I went a bit overboard with @claudeai.
I wanted to explore something we've been thinking about a lot at @MidbrainAI: what happens when AI agents actually accumulate experiences, instead of starting over every time?
So I took a small 3D household simulation, gave Claude one paragraph prompt, and started building:
"I want you to generate me an isac sim simulation with various robots, and humans interacting in a home, showcasing behaviour adaptation, personality, ocean model, nuances"
A simulated home with five robots and four people, living together for an entire year.
Each robot has its own memory, cameras, some have LiDAR, and they all have different responsibilities. They observe, make decisions, get things wrong, receive feedback, and hopefully do things differently next time.
There's also a memory system that decides what matters, consolidates experiences overnight, and allows robots to share what they've learned, with permissions.
Even a coding agent that attempts to fix bugs by running actual tests.
We then built a 52-week benchmark to compare what happens when you give these agents different kinds of memory, or no memory at all.
Around 19,000 lines of code, 31 passing tests, a research dashboard, and a few follow-up prompts to turn the recorded simulations into these videos.
All built around one rather simple question.
What if the most important thing an AI agent learns isn't in its training data, but in what happens to it afterward?
Having spent years working with simulation, autonomous systems, and human-machine interaction, what I find fascinating is how quickly we can now turn these questions into things we can actually experiment with.
It's a research prototype, not a solved robotics problem. But it gives us a little world in which to test some of the ideas we're pursuing at @MidbrainAI.
And perhaps the most interesting part: the robots' underlying code doesn't change throughout the year. Their behaviour changes because of what they remember.
I'll share what happened when we switched that memory off and on in the next video.
If you're building agents and want to give them persistent memory, come experiment with us. You can already start building with MidBrain. midbrain.ai
Timmy Ghiurau retweeted
Worldwide spending on AI is projected to reach ~$2.7T this year.
Yet, only ~1% of it is expected to go to the models themselves.
Timmy Ghiurau retweeted
With the explosion of Microduck and Reachy Mini, we'll need more people to join the @pollenrobotics @huggingface teams in all functions.
If you're excited about democratizing robotics AI with open-source, my DMs are open!
following up on my @HowtoWeb recap, one part of that gap between experiment and adoption got very concrete on stage: time.
it's easy to excite a team inside a large corporation with a pilot. moving from pilot to production is a longer game. in automotive, five to seven years before cutting-edge tech is actually deployed. i watched that cycle from inside @volvocars for a decade.
what carries a technology across those years isn't the demo. you sell with security, but you continue and earn with trust.
that's the bar we're building @MidbrainAI against. a continual learning layer for agents inside an enterprise has to handle the corporation's data, consolidate behaviors, and let agents run on them. so the questions get practical:
can you measure a correction?
can you show the behavior stays coherent every time?
can it keep optimizing without disrupting how the organization already works?
benchmarks, evals, trust. in that order.
#HowToWeb #EnterpriseAI #MidBrain
Building AI that learns while it works.
We’ve just launched the new MidBrain website. It brings together our research and product, and gives a clearer picture of where we want to take things.
For the past year, we’ve been working on a question:
How can each new experience make an AI system more capable?
When an agent repeats a mistake you’ve already corrected, you end up doing the same work again. We want that correction to change how it handles the next relevant task.
The bigger ambition is that the longer a person or team works with an AI system, the more useful it becomes. It should develop capabilities through experience and retain what it has learned as the work changes.
At @MidbrainAI , we’re connecting research in continual learning with infrastructure people can use today. Our product turns feedback into learned procedures agents can apply to future tasks. Our research explores how learning can happen during an interaction itself, and how systems can keep adapting without losing useful knowledge.
The new site shares our published research, the questions we’re investigating, and the product you can try.
If you’re building agents that need to improve over time, we’d love for you to put MidBrain to work. Try it on a real workflow. Challenge our assumptions and tell us where it falls short.
There’s a lot we still need to figure out, and working with people using these systems is part of how we’ll get there.
Explore the research and try the product: midbrain.ai
Timmy Ghiurau retweeted
When Mercor earned a pilot with DeepMind in 2024, the entire company pulled an all-nighter to handwrite all of the data in 24 hours.
Building a legendary company requires extreme intensity, especially in the early days.
Timmy Ghiurau retweeted
Surprised that so few people are talking about this.
Persistence has become a core capability of recent models (Opus 5.5, GPT 6 Astra, Gemini 4 Argon,...).
I was just reading through the Gemini 4 Argon Minecraft paper details and noticed that no roles were assigned to the agents.
Somehow the four agents completed the goal after 27 hours, 2,783 programs, and 605 deaths. They recovered items, revised strategies, and shared lessons. I think this is a good example of persistence helping them eventually succeed.
Some more thoughts for discussion:
So this is not just about hooking an agent up to an external memory system.
It's also not about loop-prompting it with /goal, although that is one of the most effective techniques right now.
This is about the model's native ability to choose what to persist, and how and when to use it.
It unlocks long-horizon capabilities, as agents often execute more complex tasks more effectively through persistence.
So it's less that the models got more intelligent and more that they got better at coordinating and planning for more complex tasks.
We are in the early days of this. But this is an important capability to optimize to really unlock AI for scientific discovery.
This is also important for building more personalized agents with identities and roles that can track and follow human preferences well and learn continually over time.
We also don't entirely know how to quantify and measure it as a field. Reward hacking combined with persistence can also enable weird, unsafe agent behaviors.
Just a few quick thoughts, and I hope we can start deeper conversations about this.
I am tracking this space closely and will be sharing research on the topic as it surfaces.
does your agent get better the longer you work together?
I think better models raise the baseline for everyone. Your decisions, corrections and experience give an agent something specific to learn from.
and with that, the connection between observing work, memory and failure logs in this post. The progress should be visible: less explaining, fewer repeated mistakes, more work you can trust.
I spent 48 hours in SF hanging out with billionaires, data labs and AI founders and execs running $1B to $100B companies.
My observations:
1. Sooo MANY stories of data brokers making like $2M a WEEK selling company data to labs. I can't tell if it's the beginning or a bubble, and I kept wondering which niches are left. Digging into it actively.
2. At dinner, someone asked the table, "If a new model drops next month, does your company get better or worse?" HALF the table went quiet! I think that's the best question to ask about any AI startup right now.
3. The bar people talk about now is hitting $1B in revenue in 12 months (I guess that's the new unicorn?). What I didn't hear much about is retention. AI revenue can come in fast and leave fast, and I think we're going to see a lot of stories of AI companies with huge revenue and terrible retention.
4. It's pretty clear we're living through the reinvention of the computer. The model is the CPU, the harness is the operating system, and skills and plugins are the apps. Thinking about it that way makes it easier to see where the whitespace is.
5. We're maybe 12 months from an agent that can watch you work for a week and then do your job without any instructions. The screen recording plus agent memory plus local model combination makes this possible right now.
6. Agent harnesses feel like the new GPT wrappers, in a good way, especially vertical ones. The best teams I talked to treat the model as swappable. One founder walked me through routing different models to different jobs: one for taste and visuals, a cheaper one for the everyday stuff, another for text.
They use prompts they can move across models over fine-tuning, and A/B test the output on what users actually do with it: do they edit it, export it, share it? A vertical harness is basically four things: the connectors for that industry's data, evals written by people who know the work, rules for what the agent can and can't touch, and a clean screen for a human to review. The model is maybe 20% of it.
7. There's a lot of reverse engineering going on/cloning going on more than I've ever seen. People are decompiling software and rebuilding it with agents, and someone showed a suite of like 10 apps that are worth $5B+ each all that he cloned open source. Me and everyone are still thinking through what it means when software gets this easy to copy.
8. My guess is the value moves to data. If the software can be rebuilt quickly, what's worth paying for in a legacy SaaS company is the years of customer data and workflows inside it. Some SaaS companies are secretly data companies.
Gets me thinking about how do you truly value data for this saas companies? Like if we were buying these dying saas companies at 2x ARR before (as example), maybe there is a better metric. Idk, just thinking out loud here.
9. It feels like a good chunk of legacy software won't make it to the agentic era, especially anything built around people clicking through a lot of screens.
10. Every company I talked to wants to pipe its internal data into one place and expose it to agents through MCP. Then they look at the data and realize it's a mess: five versions of the same customer, old Slack channels, PDFs nobody can search. Cleaning that up feels like a big consulting opportunity!
11. A lot of founders are building the basic building blocks for agents: APIs, data layers and tools. Some of the companies I saw barely have a UI because the main user is an agent. I really believe the startups that build primitives to this agentic era will do extremely well. Building blocks is the best way to think about it.
12. Labs can get plenty of training data, but it's harder to get real examples of agents failing inside companies. Whoever has those logs has something valuable.
13. Someone running a big consumer company told me the next fight is over memory. The model that remembers the most about you wins, and the smartest model matters less than people think.
14. Physical AI came up a lot. Astra I think was the first model that got people into AI can actually do CAD, and I think a lot of people are thinking what does vibe manufacturing really look like. The line I keep thinking about came from a robotics founder: the hardest problems in AI are deployment problems now, not research problems. Their robots work on their own, call a human when they get stuck, and every one of those calls becomes training data.
15. Cybersecurity companies are having a moment and are actually cool companies to build now, especially the ones that clean up after a breach. Every agent with access to your systems is a new door into your company, and a lot of those doors were left open.
16. Some startups are worth more dead than alive now. A shutdown company with 5 years of Slack, email and code can get paid for licensing all of it. I wouldn't be surprised if founders start factoring that into when they shut down.
17. People are starting to talk about monetizing the agent web. It feels really early, but when the most popular user of the internet is an agent, what startups should you be building?
18. I think agent insurance is coming. Once an agent makes an expensive mistake, insurers will want audits before they cover anyone.
19. Several people said the hire they want most is someone with taste who can write evals. The models are getting good at generating ideas, so the hard part is judging which ones are good.
20. New trend I observed: startup raises VC, uses venture dollars to become a creator, startup fails, founder falls back on being a full-time creator and incubates other products bootstrapped.
21. Getting SIGNIFICANT distribution is HARD right now. Founders are sponsoring buses, buying billboards like crazy and throwing parties with famous DJs to get noticed. Kinda feels dotcom era in that sense.
Attention is what's scarce! The growth guy going viral sharing his UGC strategy from Wispr Flow is a good example. It's the sauce everyone wants, and I truly believe there are probably only 50 people on the planet who know how to scale these types of businesses efficiently in this era.
Always good to get that dose of SF. I felt sooo lucky to have such an incredible trip during such an incredible time in building businesses. Surreal in so many ways.
Happy to be home and locked in.
Until next time!
memory can turn a suggestion into permission.
an agent proposes: "i could send these updates for you every friday."
i move on without answering. later, the proposal gets compressed into "send weekly updates."
the memory is there. the agreement never happened.
that's the failure in this paper that interests me. remembering more about a person doesn't mean understanding what they've agreed to. more context can even make the agent more confident about doing something it was never asked to do.
i'd want the system to keep the proposal and its status together. who suggested it, whether i accepted it, what it covers, and whether i've changed my mind.
but the answer can't be "ask me every time" either. if i've agreed to something within clear limits, making me approve it again is another kind of failure.
the useful test is whether the agent acts when the agreement applies and holds back when it doesn't. not whether it can recall the conversation.
the paper tests this in controlled conversations. real agreement can be messier: gradual, conditional, sometimes contradictory. i wouldn't treat this as solved.
if i'm going to build a long-term relationship with an agent, it needs to remember the boundaries of that relationship too.
arxiv.org/html/2610.18684v1
Timmy Ghiurau retweeted
there is a massive different between intellectual output & intellectual presence.
i.e. intelligence can be demonstrated through artifacts but intellectual chemistry, the ability to think in real time, improvise, use analogies, develop ideas with others, & have genuine conversational rhythm can only really be experienced through irl interaction or a podcast type setting where that person is the guest.
this is why when you meet ppl irl off of here it is rarely if ever the same experience as their posts & can be super disorientating. as if the internet bifurcated individuals & created two different personas. it is very very rare to find ppl whose intellectual presence is as interesting as their intellectual output.
I dont think software is dead, but the idea of software as a finished product might be... or software lifecycle.
i agree that we're moving towards a world where anyone can build software for themselves. the cost of creating it is collapsing for sure.
but building something and having something that understands you are two very different things.
cause if every application starts from zero, we're just replacing one kind of friction with another.
the big change happens when software accumulates experience. when it remembers why you made certain decisions, understands your preferences, learns from mistakes, and adapts as your needs change.
at that point, we're no longer talking about generating applications on demand... we're talking about systems that evolve alongside us.
and there's another question here. if everyone can create software, what becomes valuable then?
perhaps less the ability to build, and more the ability to understand what is worth building in the first place.
the fundamental basic economics misunderstanding that ppl have esp on here is that they think normal ppl need to use models to build their own software to be disrupted. they obviously don't & they obviously won’t.
but even if just 1% of the population can build software, that's enough to flood the entire supply side with millions of new products collapsing the cost to near zero.
you absolutely don't need universal adoption to destroy scarcity. you just need enough producers to overwhelm demand. you can feel this now already. i never pay for software because i just ask claude to point to something like a chrome extension or a mac app & it’s just there. at some point your agents will also auto update & maintain your software automatically too.