
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
- Tweets5.5K
- Following2.2K
- Followers1.8K
- Likes22.5K
Taste is never stated. It shows up in the edits.
I've never told an agent how I like my emails. I delete "I hope this message finds you well", cut a paragraph and add a next step. Every time.
An agent that learns picks it up without being told.
What edit do you keep making?
1
2
111
This is one idea from A lifelong companion, an essay from us on why remembering you and learning from you are two different things: midbrain.ai/blog/a-lifelong-…
23
Timmy Ghiurau retweeted
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
our minecraft agents forgetting overnight were part of what led us here. this piece traces that story alongside other approaches to AI "dreaming".
my test for the idea: does the offline work make the next task better, or just produce a tidier memory?
3
182
Timmy Ghiurau retweeted
Different coordinates, same similarity.
MidBrain stores embeddings after an orthogonal transform derived from each agent's search key. Dot products survive it, so ranking doesn't change.
If an AI agent is going to work with you for years, you should know who can read the history it builds.
At @MidbrainAI , we use separate keys for your conversations and documents, the search index, and the preferences and instructions the agent picks up over time. In client-managed mode, you keep the key that decrypts your stored conversations and documents, while the agent can still search them.
We wrote up how this works, including what our servers can see when they process a new memory. If you’re building agents that learn from experience, I’d love to hear how you’re handling that trade-off.
Every MidBrain agent gets four keys.
The one that decrypts conversations and documents never reaches our servers in client-managed mode. Search and consolidation get only the key they need.
1
4
219
Timmy Ghiurau retweeted
Most failures in our DolphinBench run came after the facts reached the model.
89.1%: right tool, facts in context, response omitted a requirement.
7.9%: retrieval misses.
3.0%: expected tool never called.
Timmy Ghiurau retweeted
had to make the national anthem of europe with this
Timmy Ghiurau retweeted
We wanted to do more for the people who’ve been with us the longest. So with Fable 5.2, the longer you’ve been subscribed, the more Fable resets you get.
An agent can have the right facts in front of it and still miss a requirement. That’s what stood out to me in our DolphinBench failure analysis.
It raises the next question for us: how do we make remembered context translate into reliable action? At @MidbrainAI , our longer-term goal is agents that apply what they remember, adapt when instructions change, and learn from corrections. Understanding where they fail today helps us decide what to test next.
Most failures in our DolphinBench run came after the facts reached the model.
89.1%: right tool, facts in context, response omitted a requirement.
7.9%: retrieval misses.
3.0%: expected tool never called.
4
171
Timmy Ghiurau retweeted
this METR report on AI R&D acceleration at Anthropic is very inward looking imo. it counts the lab's compute, hires and internal data, then credits ~80% of the speed-up to AI
what I see from the outside: labs are becoming schelling points for money, redistributing it to an ecosystem they created to accelerate their own development
it's a symbiosis. the best data and RL environment companies now know where models will fail before the labs do, and the labs just buy what they propose. that feeds back into every lab at once
a lot of the "AI accelerating AI R&D" happens outside the lab
Timmy Ghiurau retweeted
The post: midbrain.ai/blog/dolphinbenc…
The benchmark: dolphinbench.ai/
Timmy Ghiurau retweeted
83.2% on DolphinBench, across 600 tests.
Same agent and model as the leaderboard: Hermes with GPT 5.6 Luna. $64 in harness cost, plus 52k memory credits.
Production MidBrain, with two experimental prefetch features on.
Timmy Ghiurau retweeted
Canva is laying off 22% of its workforce which is 1271 employees, and this is what their CEO said.
We need more founders like thissss.
83.2% on @mem0ai DolphinBench: Mapping the Pareto frontier of agent memory, across 600 tests.
Same agent and model as the leaderboard: @NousResearch Hermes with GPT 5.6 Luna. $64 in harness cost.
Production @MidbrainAI , with two experimental prefetch features on.
The benchmark: dolphinbench.ai/
The post: midbrain.ai/blog/dolphinbenc…
1
1
72




















