Pinned Tweet
“don’t train your own model” is common ai advice. it's wrong. your token bill's the proof.
today, we’re excited to launch castform into open preview. castform is the easiest way for you to train your own model, on your own data.
open-weights models are performant and much cheaper. when trained on your task & proprietary data, they beat closed models. the thing standing between you and that was weeks of plumbing & years of ml expertise.
with castform, model training is as simple as prompt engineering. @castformai
bring your agent traces or raw corpora. castform turns it into training data, picks the right algorithmic recipes, manages gpus, and gives you an ide to watch and chat with your model as it learns.
see what you can build with castform👇
girish retweeted
the hardest part of enterprise AI is no longer just building the agent – it's getting it adopted, changing how work happens, and capturing the value created.
we think owning the business changes the equation. here's how we think about buying, building, and deploying at @llmh
girish retweeted
Excited to launch Before They Break Out, where I’ll spotlight promising startups before the rest of the world catches on.
Kicking off with Vals AI, a startup building critical evaluation infrastructure for the future of AI. @ValsAI @RayanKrishnan
Introducing Before They Break Out, where we explore promising startups and why they’re worth joining early. First up: Vals AI, a company building the measurement layer for artificial intelligence. @ValsAI @RayanKrishnan
beforetheybreakout.substack.…
girish retweeted
We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.
It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.
girish retweeted
we’re releasing REMAG, a derivative GPU first implementation of SIESTA that is purpose built for magnetic materials. we’ve seen ~111x speed ups for select compounds, and we hope that this inspires folks to bring other computing packages into the GPU era project-14.us/remag/
girish retweeted
Narrower surfaces in gaming are unusually hard for models. We're excited to share how we approached post training in areas where users have very peculiar taste.
Great partnering with @castformai on this.
coding models have gotten much better at ui generation, but still struggle with stylistic and domain-specific requirements.
we worked with @gg_lemonade to post-train hy3 to more consistently build roblox uis that follow their visual styles.
here’s how: castform.com/blog/lemonade
girish retweeted
coding models have gotten much better at ui generation, but still struggle with stylistic and domain-specific requirements.
we worked with @gg_lemonade to post-train hy3 to more consistently build roblox uis that follow their visual styles.
here’s how: castform.com/blog/lemonade
girish retweeted
Today, @usefootprint has raised a $25M Series B.
In the last few years, Footprint has seen incredible growth. We went from inventing something meaningful once a year to once a week.
And it’s made me think deeply about “timing.” That we’re always limited by the technology of our time.
Some of the most brilliant people who’ve ever lived, with even a fraction of what we have access to today, would have built things that would stop you in your tracks.
That’s felt especially true this past year, so raising right into this moment has got the whole team running on a whole new tank of gasoline.
We’re tackling the most interesting problems I’ve worked on in my life, and I can’t wait to continue investing in the future we want to build: real judgment applied to the systems that every bank, fintech, and business relies on to keep global financial crime out.
A huge thank you to our lead investor @QEDInvestors, our new investors @mufgcr_official , @CommerceVC , @lightbank, @alumniventures ; and returning investors @IndexVentures, @LererHippeau, @BoxGroup, Operator Partners, Definition, and Animal Capital.
Today, we're announcing our Series B fundraise. As the volume of financial crime in the global economy expands, we're building the agentic defense system that to meet and defeat it. onefootprint.com/learn/serie…
girish retweeted
Many folks professing support for “pacing” view it as the path to shirk real regulation.
I don’t think we should pace. But if we do, and we believe extinction is possible, we shouldn’t use a fig leaf like METR to reinforce the cartel.
To the DoE or to everyone.
girish retweeted
I live in SF in 2026, so naturally my landlord decided to jack up my rent by 50%.
I used Supergood to build a bunch of integrations, then automatically email my landlord a report every morning to show her she's out of her fucking mind.
it:
1. Generated APIs to pull specs on a specified apartment (did this so my friends could use this)
2. Generated APIs to use those specs to find similar listings within a 1-2 mile radius of the specified apartment
3. Compiled a report and emailed the landlord every morning.
this mostly worked. she decided to only jack our rent up by 25%.
DM if your landlord is also out of their fucking mind
girish retweeted
What if every judgment call became institutional knowledge?
𝘠𝘰𝘶𝘳 compliance program run through 𝘰𝘶𝘳 agentic system — that remembers and compounds.
Today we're (re)introducing Footprint: The AI operating system for risk. onefootprint.com
example run: dominant-color
the task involves showing the model 3 images, each dominated one color. it only gets the first one - it has to call a tool to see images 2 and 3.
then it names all three colors in order. get one wrong, get nothing.
app.castform.com/train/11a57…
github.com/castform-ai/bench…
@castformai now supports post-training vision models 👁️
you can train on images, diagrams, and screenshots as input, beyond just text.
quick thread on what this unlocks 🧵
the environment contract doesn't change. two ways to feed images:
in the prompt - content can be a mix of text and prompt
from a tool call - run_tool can return content parts
multi-frame / video reasoning
sports mechanics, assembly line steps, security footage. judging what happened across frames is something general models are still weak at.
technical diagrams → structured output
circuit diagrams, floor plans → json. these are highly domain-specific with conventions general models haven't seen, so post-training pays off fast.
5/ TITO (token-in-token-out) = the trainer trains on the exact tokens the model generated in multi-turn rollouts.
decode → re-encode breaks this, because encode(decode(x)) ≠ x.
a perfect match matters for training stability.
