There's this weird dynamic at VC events where it feels like everyone is competing to see who can care the least. The more indifferent you act, the cooler you seem. And somehow, the more arrogant or dismissive you are, the more important people perceive you to be. I've always found that a bit strange.
What's interesting is that when you live in a society with very strict social norms - which often reflect economic stagnation, where there isn't enough room for everyone to grow, and society develops increasingly elaborate mechanisms of competition and control that eventually shape even the most subtle behaviors - you quickly grow to resent those constraints and seek greater freedom.
But when you live in a society where social norms have largely disappeared or are no longer enforced, you begin to realize that shared expectations and even social pressure can be healthy and necessary - not only for society to function, but also for individuals to grow and evolve.
People moved to cities not because they were pleasant, but because of job opportunities, networking, and often because they had no other choice. Living in a big city sucks - crime, pollution, lack of space, and high prices
This is wild...
There’s a farmhouse on top of a building in the West Village in NYC, and it just listed for $10.5 million.
Porch. Garden beds. A vegetable patch. City skyline instead of a backyard.
The penthouse at 719 Greenwich Street.
If you care about Europe's future and robotics, and genuinely enjoy building things and solving problems that haven't been solved yet, you should try to join UMA.
I had a chance to visit the team recently, and I think what they're building is simply mind-blowing.
What are the most difficult tasks for robots today? I talk about that with JB, Steve, Medhi, and about all kind of stuff related to Humanoid robots, dexterity, fleet learning. Check this out 😇
youtube.com/watch?v=Mw2q8jZm…
IMHO, training is much more interesting than inference, although inference has a lower barrier to entry and require good old-fashioned computer science problem sets
My girlfriend spends more time with my parents than I do, and they seem to like her much more. I’m not sure how I should feel about that.
People are going to realize that LLMs are much better at organizing large groups of people, we'll see the first projects, events, parties and revolts organized and coordinated by agents.
As a kid, I loved watching Japanese anime so much. I remember how my brain just couldn’t process all the concepts and ideas. One of the craziest ones was Ergo Proxy. I think that even today, it’s one of the few industries where artists prioritize quality, education, and freedom of thought over profits.
Interestingly enough, ScaleResearch/HypothesisLoop - an OSS harness for running experiments - follows a very similar pattern. Initially, it defines an experiment in terms of its goal and the metrics used to benchmark and compare trials. After that, there is a layer that controls the experiment, while agents generate hypotheses and iterate on them.
We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next.
Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon.
This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials.
Read our blog posts below.
People usually pick one domain and live within it, and by doing so, they don't realize how interconnected things are. This applies to hardware and general engineering, but also to serving, training, and infrastructure more broadly. The same data problems keep repeating again and again, regardless of whether it's robotics, bio, LLMs, or materials synthesis.
Everyone needs to store data, query it, clear it, evaluate it, clean it, and use it effectively for training; have good integrations with data partners; and evaluate data quality.
There's also a lack of good LLM-powered interfaces across teams, including for research engineers - the ability to run experiments at scale still isn't really solved.
And the cultural problems are the same too: building coalitions, sometimes much worse than kids at school.
I'm thinking of shutting downsizing.dev down.
After all, I don't think an LLM router outside of the providers makes that much sense, for a few reasons.
The biggest one is that cached context has become much cheaper. Also, for most customers I've talked to, stability matters much more than pricing. They'd often rather pay more for guaranteed quality and consistent responses than save money on their coding-agent budget.
I don't want to pivot into LLM inference, and I haven't found enough customers for LLM observability or context compaction.
I do regret not building the "hacker LLM" I wanted to work on instead.
On the positive side: downsizing works. I use it almost every day, it saves me a lot of my Claude and codex quota, it has some traction, a few piloting customers and the name brings in a surprising amount of free traffic.
People cooperate and build strong social norms => trust enables complex, high-order systems => shared culture, purpose, and trust erode => bad actors exploit norms that cannot be fully enforced => institutions weaken and order declines => people turn to stronger authority to restore stability => the cycle repeats
1377 year!!!! Ibn Khaldun's asabiyyah
The hype around AI safety is already too much. Sixteen-year-old kids hack websites all the time. We're not giving LLMs control over our nuclear weapons. Chatting with an LLM isn't any worse for my mental health than chatting with a random person on Chatroulette. Just releasing a new LLM version into production requires hundreds of people to coordinate.
What would be interesting to know is what roles these agents played and what the justification was for having so many of them - what drops off when you reduce the number, say, to just a hundred?
This model represents a step-function improvement on many benchmarks, and its training is ongoing.
Our internal model group arrived at the Navier–Stokes solution in 88 hours, using around 10,000 coordinating AI agents.
Throughout the effort, we maintained the strict safeguards—including monitoring and isolation—that we apply to all our frontier evaluations.
You can just do things
A bit of an overlooked paper. It introduces a simple trick to stop an RL agent from getting stuck repeating the same behavior over and over - using 3 models:
1. A randomly initialized, frozen model
2. A model trained to mimic model 1's outputs
3. An actual agent (the one being trained to act)
Every time the agent takes a step, we also run that state through models 1 and 2 and measure the difference between their outputs. Small difference = this area is already well-explored. Large difference = novel territory.
That difference gets added into the reward signal, pushing the agent to seek out unfamiliar states.
arxiv.org/abs/1810.12894
