TrainLoop
@TrainLoop_ai
Followers
387
Following
40
Media
13
Statuses
32
Reasoning fine-tuning.
san francisco, ca
Joined January 2025
🧵 Labs and VC's are throwing cash at RL environments, especially for computer and browser use. Yet, with just 4 customers and over 30+ vendors, is cloning every website in the world really the path to scale? of course not. Introducing TRACE: Trajectory Recording and Capture of
6
10
72
Only the best companies use @superset_sh had a lot of fun onboarding @TrainLoop_ai. If you are a cracked MLE and want to jump on a rocket ship with crazy high talent density hit them up 🚀
@TrainLoop_ai is one of those companies in our group office hour that has something special. They're hitting an exponential curve and are now hiring more cracked engineer! Had a great time onboarding @jackson_stokes and @mlpierce22 to Superset yesterday with @avimakesrobots
2
2
11
New on the TrainLoop blog: MAE, MSE & R² — Making Sense of Model Errors We break down Mean Absolute Error (MAE), Mean Square Error (MSE), and R-Squared -- three core metrics that shape how we judge model performance. Link to blog -- https://t.co/mBcaqf903X
0
0
2
A bite-sized explainer on how LLMs learn - end to end. Core ideas without the math. Follow @TrainLoop_ai for our plain-language blog series that dives deeper into each concept.
0
0
3
With @TrainLoop_ai , AI model training outcomes are predictable, repeatable, and sustained --- not blind trial-and-error.
0
0
2
Precision vs. Recall 🤔 Always confused between the two? You’re not alone. We broke it down in plain English + a 2-min “Wild Fire” game 🔥 After this, you’ll never mix them up again : 👉 https://t.co/GVAs7e6DAw
0
0
5
AI model training isn’t about getting lucky -- your competitive advantage depends on it. With TrainLoop fine-tuning, you trade chance for control.
0
0
5
Remember the last time someone brought up "Mean Squared Error" in a conversation, and you nodded your head like you knew what it meant? Learn the concept in depth -- https://t.co/H0kZEAt4je
#learn_the_basics_with_trainloop
0
0
7
New office at @trainloop_ai Comes with cardboard and bubble wrap
So many good-looking things in one frame. Also, here’s our new office in North Beach. If you’re around, come say hello.
0
1
9
Have you ever spotted a four-leaf clover? 🍀 We’d love to hear your story if you have!
0
1
4
How accurate is your model? Before you think about it, Here's another one - what is 'accuracy', really?
0
0
3
The coffee machine is in the house! And we have started running some experiments. We are a Research Lab after all!
1
0
6
Model evaluation is the broccoli of Machine Learning. 🥦🥦 At least we made it simpler -- our evals framework is now open-source → https://t.co/R6LQZ2KVE6
0
1
2
So many good-looking things in one frame. Also, here’s our new office in North Beach. If you’re around, come say hello.
4
1
16
The intern now runs one mile every day at 7 am with @mlpierce22 and comes to the office by noon. We'll update when he hits the 1.1-mile target.
“I’m excited to tell my girlfriend that I run now. She’s been trying to get me to do this for the last year.”
0
0
2
When you fine-tune your AI model without a “map”, it will wander. Random tweaks = random outcomes. TrainLoop gives your fine-tuning process structure, direction, and control - so you land where you intend, not where the currents take you
0
0
6
ODML is tricky, MLX-> CUDA is a great first step.
MLX supporting CUDA might be key to Apple's Siri strategy, here's why...
0
0
0
Love this explanation on why AI models always generate outputs that seem good but don’t get anybody excited. This is the main reason why custom models are the future. A one-size-fits-all solution actually fits nobody
Every single RL paper for image-gen I've seen honestly the outputs look like SLOP. my gut is that the mean aesthetic preference, whats optimized for, is not a good preference, its kitsch. Is there a way we can sample individual aesthetic targets?
1
2
9
Cannot say this enough Working in robotics right now is probably one of the highest ROI bets in history It’s like web in the 90s, mobile in 2007…
*really* glad i decided to do robotics instead of generally swe otoh i do think if you’re a good SWE, there’s never been a better time for you
7
7
151