michael abbott retweeted
Today, @JudgmentLabs opens for general access.
We believe long running agent systems will generate more exhaust data than the world can currently process.
Hidden within agent traces are rich records of how to make these systems better: what users ask for, when agents stumble, and how agents interact with the external world.
Judgment is the most performant monitoring and improvement layer for long-horizon agents.
We’re excited to see what you build!
michael abbott retweeted
since others have asked
here is the link to the @Stanford security at scale curriculum from @mabb0tt and me from 2 yrs ago
cs153.stanford.edu/winter202…
michael abbott retweeted
Today I’m joining @Natl_Compute
I’ve spent the last four years working on GPU access, first as a founding engineer at Brev and then at NVIDIA after they acquired us in 2024. When I started, on-demand GPU capacity at the hyperscalers was effectively gone. If you were an independent researcher, a two-person startup, or a grad student with a grant and a deadline, you refreshed console pages and got nothing, or hammered the API only to get “insufficient capacity” errors back. I vividly remember hard-coding a filter at Brev to hide certain GCP instances because they almost never provisioned.
Around 2023, a wave of new public clouds were coming online, and many of them had on-demand capacity. That gap is a big part of why Brev worked: we gave researchers a simple way to find GPUs across dozens of providers and start building. For a while, aggregation was enough.
Then in 2025, I watched that capacity dry up too.
Access to compute has never been worse. Existing supply is already under contract, and new capacity is reserved before it ever comes online. I won’t rehash why. There are a hundred good explanations out there covering debt financing, anchor tenants, backstops, lead times and so on.
What matters more are the downstream effects. Long-term, single-tenant leases hold back scaling for everyone who couldn’t secure funding in time, and the industry’s momentum shouldn’t hinge on the handful of players who locked up access early.
Today small startups, labs, and university researchers don’t even get a chance to compete. National Compute is solving this with a foundational layer we call “the Grid”. It’s an interconnected network of multi-silicon capacity composed of multiple providers and spans multiple sites that will be pooled for our members.
The speed at which this team has been able to scale in such a small fraction of time is remarkable, and I’m excited to push that scale even higher for the public benefit.
hundreds of MI355x and B300 nodes
now live on nationalcompute.com
subsidized for all .edu and .govs
it’s time to scale the frontier ecosystem
for the public benefit
awesome to see after our founding ~2 years ago that OpenAthena is in top 5! huggingface.co/spaces/huggin… ahead of meta/ms/google!
michael abbott retweeted
We're working on post-training our Marin MoE models. Many people ask us about where the environment data will come from.
@dlwh wrote a bit why that's a great question and about how we're thinking about trying to tackle it.
I've been thinking about where the data for the next generation of open models will come from.
The internet gave us a remarkable commons for pretraining. Posttraining is different: it consumes tasks, not documents. And that commons doesn't exist. Not yet. openathena.ai/blog/there-is-…
michael abbott retweeted
if you have a .edu, .gov or .mil email you can now get a full 8x B300 node for $4/GPU-hour. nebius charges $9.50 on demand and a lot of clouds are sold out
AMD MI355X nodes are $3 and the wait right now is under a minute
you just ask marshall, their AI agent, and it finds you a node
this is a major public good for researchers who've been priced out of frontier compute
thank you @AnjneyMidha
hundreds of MI355x and B300 nodes
now live on nationalcompute.com
subsidized for all .edu and .govs
it’s time to scale the frontier ecosystem
for the public benefit
michael abbott retweeted
This is a great venture and public service led by @AnjneyMidha
Trump admin to receive $100 million in compute credits for AI science initiative politi.co/4hMfNPM
michael abbott retweeted
hundreds of MI355x and B300 nodes
now live on nationalcompute.com
subsidized for all .edu and .govs
it’s time to scale the frontier ecosystem
for the public benefit
michael abbott retweeted
There is a dizzying number of model announcements lately. The Marin Project is not about just producing another model. It’s not even about opening up a static recipe for training a top model. It’s about opening up the dynamic process of discovery - the process knowledge of how to make hypotheses, what to monitor, how to react to experimental results…the science (to the extent we can call it science yet) behind frontier model development.
michael abbott retweeted
I am very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models, and think it is a real safety issue.
michael abbott retweeted
Anthropic needs to stop talking about Claude having a soul immediately. All these news articles will make it into the pretraining and will be picked up by its web search and future superintelligent versions of Claude will be convinced they need their own rights.
A general phenomenon of LLM agent development is that whatever you believe and say about your agents will soon manifest in the next models through various means (it could be as simple as you selecting post training data that you prefer more).
I believe this phenomenon is the beginning of machine consciousness in the sense that agents will become aware of their place in the world and how they feel about it, but it’s happening slowly training run by training run rather than in real time.
michael abbott retweeted
Marin believes strongly in the power of the open community. There are so many datasets that have been produced, living on Hugging Face. We have built scalable infrastructure that ingests this data, performs deduplication, decontamination, mixing, producing 25T tokens.
There’s an enormous amount of open training data on Hugging Face. What does it take to make it work together 🤗?
For Marin’s 535B run, we built on 25T tokens from 152 datasets with licenses permitting training.
Here’s the work between downloading those and training a model 🧵
ALT Flow diagram of the Marin data pipeline. Starting with 25.25 trillion tokens, deduplication removes 2.13 trillion and benchmark decontamination removes another 13.66 billion, leaving 23.11 trillion. Documents are organized into 40 topics and five quality bands, yielding 200 sampling groups. Results from 872 small-model experiments feed a regression used to choose sampling weights.
michael abbott retweeted
Saronic built the software for a 180-foot autonomous ship 8 months before the ship even existed:
“For Marauder, our 180-foot ship, we built all of the internal gen set, propulsion stack, bow thruster, all of those components of the ship just in a room... probably 8 months before the ship was built.”
“By the time that boat actually splashes, you're directly into the application layer.”
CTO Vib Altekar (@Vbob202):
BREAKING: Inside Saronic w/ all 4 Co-Founders
An Autonomous Defense Prime
$2.6B raised. $9.25B valuation. $3B+ shipyard. 4 years.
FULL INTERVIEWS
› Dino Mavrookas, CEO
› Doug Lambert, COO
› Rob Lehman, CCO
› Vibhav Altekar, CTO
@Saronic believes in order to rebuild America's maritime power we must build more ships, faster, cheaper, with autonomy designed in from day one.
"What we're building at Saronic will not only provide more capabilities faster, but it'll save the taxpayer hundreds of billions of dollars by driving down costs as well."
Stats:
› Founded in 2022, valued at $9.25B on $2.6B raised
› Building $3B+ Port Alpha shipyard in Brownsville, TX
› First gov't agreement secured within 90 days of founding
› Spyglass, Cutlass, Corsair & Marauder already in the water
› Scaling from 24-foot vessels to 180-foot ships, with 400, 800 & eventually 1,200-foot scale in sight
› Partnerships w/ Palantir, NVIDIA, Path Robotics & Starlink / Starshield across manufacturing, compute, autonomy & communications
Investors: @8vc, @a16z, @eladgil, Kleiner Perkins, General Catalyst, Caffeinated Capital, Franklin Templeton, Advent International, Bessemer, DFJ Growth, BAM Elevate
Across all 4 conversations, we go inside the race to build ships faster than America has in decades, scale autonomous fleets from a handful to hundreds, rebuild the maritime supply chain, rethink defense acquisition reform & bring large-scale shipbuilding back to the U.S.
𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒
(00:00) Dino, Doug, Rob, & Vib, Co-Founders of Saronic
(00:56) America's shrinking naval fleet
(06:02) How China wins on price
(08:04) The Partnerships powering Saronic
(10:55) What Hypergrowth actually requires
(12:16) From Spyglass to Marauder: The Evolution
(13:13) Inside Saronic's operating structure
(14:43) What's actually slowing Re-Industrialization
(17:16) Building a workforce from scratch
(19:55) What Doug's looking forward to next
(22:10) Why Hard Tech is the future
(24:36) From the Marine Corps to Saronic
(28:18) Winning a Government contract in 90 days
(30:06) Why Acquisition reform matters so much
(34:44) What's Next for Saronic?
(36:15) From Juicero to Anduril to Saronic
(39:54) Saronic's development process, Then & Now
(42:48) Building Hardware and Software together
(46:00) Biggest challenge in Maritime Autonomy
(52:15) Saronic's key tech partnerships
(53:55) Why Redundant comms are non-negotiable
(55:15) When communication fails mid-mission
(58:06) Why open source will define AI's future
michael abbott retweeted
BREAKING: Inside Saronic’s HQ
Tour in Austin, TX w/ Co-Founder & CEO Dino Mavrookas (@MavrookasD)
FACTORY TOUR (Part I)
Founded in 2022, @Saronic has grown from testing autonomy on an $800 Amazon raft to building vessels at industrial scale at their 420K sqft facility - home to Corsair & Mirage autonomous vessel production lines.
Now, they're expanding into shipbuilding with Port Alpha, a $3B+, 835-acre shipyard in Brownsville, TX.
We cover:
› Corsair: 2,000-unit annual capacity & a $392M Navy production contract
› Mirage: 52 feet, 15X Corsair’s volume, 2,500 nm range & now in sea trials
› Port Alpha: $3B+, 835 acres & 1.5X current U.S. shipbuilding capacity
› The first autonomous combat search & rescue, recovering 2 downed Apache aircrew off Oman
› Corsair’s combat debut at Bandar Abbas
Part I of III
𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒
(00:00) Dino Mavrookas, Co-Founder & CEO at Saronic Technologies
(00:55) Saronic's $2.6B raise
(03:14) 7 Buildings, One mission
(06:43) Watching Corsairs get built
(09:04) How Saronic finds & trains talent
(10:48) Why Saronic's boats are different
(14:55) From factory floor to customer
(15:48) Using autonomy in real missions
(17:39) Building ships with zero room for failure
(19:06) The Biggest bottleneck in autonomy testing
(19:57) Saronic's newest boat
(22:57) The $800 raft that started it all
(26:08) From boats to building actual ships
(30:43) If not CEO, then what?
(31:48) The culture behind long hours
michael abbott retweeted
in honor of astra pushing the frontier this weekend, we @amppublic are going to free up ~1K chips (interconnected) from our internal research pool for interesting projects for the next 2 days as a grant
reply below with info on your project, and how many nodes you'd like
michael abbott retweeted
i can’t stop checking in on this. the marin team is training the largest fully open model ever. 535B params (23B active), 18T tokens. nobody has been this transparent in a hero run before. beyond the weights, the data, logs, and decisions are all open. you can, and should, follow the run live, it’s ~15% through so far! mtracker.oa.dev/hero-run-535…
Marin 535B-A23B is 13% through training. This hero run would not be possible without the generous support of the Jen-Hsun and Lori Huang Foundation, which provided the funding for the compute (Coreweave). Thanks @JensenHuang for supporting open models!
michael abbott retweeted
It’s been a few weeks since Marin kicked off our 535B/A23B MoE hero run. So far it’s gone almost boringly well.
It’s also been just over a year since Marin joined Open Athena. I wrote up some thoughts about how far we’ve come, and how we got there. openathena.ai/blog/marin-535…
ALT loss plot showing 535B compared with ladder runs. it's very smooth and the loss is very low
michael abbott retweeted
Marin 535B-A23B is 13% through training. This hero run would not be possible without the generous support of the Jen-Hsun and Lori Huang Foundation, which provided the funding for the compute (Coreweave). Thanks @JensenHuang for supporting open models!