Pinned Tweet
Late night rant: Spark is an awesome piece of software. But a horrible developer experience.
What happened to OSS that was simply `apt install` and 🚀? Why should software be excused for slow local performance because it was built for "production scale"?
So much of "big data" JVM-based tooling was hacked together on the giant datacenters of tech giants. The world has changed, and so too must our big data tooling.
⭐️ Rust: self-contained compiled native binaries that have no dependencies. Hello, clean installs, my old friend.
🐍 Python: the undeniable winner of iterative plumbing for data/ML. Build with a Python API in mind. Using the JVM through a Py4J gateway should be an automatic disqualification.
☁️ Cloud: Build cloud-first, lightweight, ephemeral software. Cattle vs Pets. S3, not NFS/HDFS. Spot instances, not machines on a rack.
🤓 Dev UX: build for the single developer, on their laptop, then think about scaling. A docker-compose local dev story is lazy bundling of overly complicated software.
☀️ Open Formats: let software TALK to each other, so devs can choose the right tool for the right job, and so devs can keep building better tooling. This is why JSON is awesome. Arrow is awesome. Iceberg is awesome. Parquet and CSV are (I begrudgingly admit) somewhat awesome. And please build flexible SDKs for these formats, in C++ or Rust, not just for the JVM.
And its a wrap! Had such an amazing time at #CoRL2026.
TBH I started 2026 as a skeptic of "Physical AI" but its undeniable now that this technology is going to have such immense potential in the next few months 🧵
We're seeing an incredible explosion of interest: capital, talent and GPU-hours.
Frontier labs are seeing massive leaps in intelligence by training on perception data (see: Fable/Astra).
It's clear that this is where the frontier of AI is today. And the ceiling keeps going up.
1
39
Had a ton of fun with @CMUDB and @andy_pavlo talking about databases AND AI/robots! SPOILER: we’re giving away our simple recipe for how to work with millions of video hours with just good ol’ parquet, iceberg and mp4.
youtu.be/S6_HFMpz_Yk?si=RoME…
Now go oneshot this w Fable 🤣
Jay Chia - eventual.ai retweeted
I'm hiring a Chief of Staff at Eventual.
We build data curation for Physical AI. Looking for someone with high agency to work with me on GTM, customer support, and ops.
$5,000 referral bonus if your introduction leads to a hire.
Interested or know someone? DM me.
Jay Chia - eventual.ai retweeted
Gemini 4 Argon dropped today, so we ran a quick bakeoff: Astra vs Argon vs Opus on action segmentation in egocentric video, scored against Eventual's golden set.
Accuracy: Astra 59.4 | Argon 55.3 | Opus 48.7
Astra's still king of the frontier models. But all three are a long way from annotation-grade.
Jay Chia - eventual.ai retweeted
RoboJEPA builds on V-JEPA 2.1 and links better predictions to better robot planning.
My bet: choosing which interactions to train on gets more valuable as we scale.
At Eventual, we're building signals to help labs make that choice.
We ingested >100TB of data over the last 3 days
Omg big data is not dead. Not when it comes to video.
2
185
Mental health has significantly improved reading Opus 5.5 outputs v.s. 5.1
But sorry I already churned to Codex over the last 2 weeks 🤣
2
14
1,175
For robotics researchers, evaluation can be the slowest part of their research loop. Manifold makes it the fastest.
Introducing Manifold. The platform for accelerating robotics research with failure analysis that explains itself.
A month of experiments in a week.
Simulate a thousand scenarios in 30 minutes.
Dense labels on every atomic sub-goal and action.
Agents that cluster failure modes across every rollout.
The robotics research loop that moves at the speed of compute.
One eval harness for every benchmark, simulator, or world model. Native support for @NVIDIA’s Isaac Lab Arena.
VLAs ground language unevenly, and apparently in a consistent order...
We keep seeing VLMs fumble scene understanding in robotics.
According to the paper the order goes: color first, then object, spatial (left vs right), verb, size
The verbs come last.
arxiv.org/abs/2607.21582
This is measurable.
And the fix is simple - reallocate data mixtures towards under-represented scenarios.
2
2
75
This is consistent with our experience pointing VLMs at robot logs.
They're great at describing what things look like, much weaker on describing what actually happened...
arxiv.org/abs/2607.21582
43










