Kyle Corbitt
@corbtt
Followers
16K
Following
5K
Media
209
Statuses
2K
Currently building @OpenPipeAI. Formerly @ycombinator, @google. I am always down to go on a quest.
Seattle, SF
Joined September 2012
đ Meet ART¡Eâour open-source RL-trained email research agent that searches your inbox and answers questions more accurately, faster, and cheaper than o3. Let's go deeper on how we built it. đ§ľ
40
123
997
AI-as-arbitrator feels like a very obvious niche and I'm surprised I haven't seen it productized.
Never thought I'd become a 'take your relationship problems to ChatGPT' person but when the 8yo and I have an argument it actually works really well to mutually agree on an account of events for Claude and then ask for its opinion.
2
0
20
Every consumer company gets RLHF'd, but some get more RLHF'd than others.
Weâre making GPT-5 warmer and friendlier based on feedback that it felt too formal before. Changes are subtle, but ChatGPT should feel more approachable now. You'll notice small, genuine touches like âGood questionâ or âGreat start,â not flattery. Internal tests show no rise in.
3
0
27
RT @jamievoynow: Agent Design Pattern: Parallel Rollouts. Inspired by Tree-of-Thought [1] and @corbtt's Universal Reward Function [2], lateâŚ.
0
50
0
RT @mattshumer_: One super important point: for companies using LLMs at scale, the new open source OpenAI models, combined with RL from folâŚ.
0
4
0
RT @mattshumer_: The @OpenPipeAI team continues to push towards making RL easy. You can now make agents that are superhuman at using speciâŚ.
0
6
0
RT @mattshumer_: @heyjchu I think you're describing @OpenPipeAI. Check out their work on RULER (, it's essentiallyâŚ.
0
1
0
0
1
0
RT @mattshumer_: @iScienceLuvr ART by @OpenPipeAI. Makes it so easy to do multi-turn agentic rollouts.
0
1
0
I generally don't post about politics; it isn't my area of expertise and isn't what my followers are here for. But if, like me, you've had a hard time understanding the situation in Gaza given the hyper-partisan reporting on both sides, this story may be interesting. I find it.
âWhen I entered Gaza the Israeli military had a rule: I was only allowed to bring in three kilos of food. As I was weighing out protein bars, trying to get under the limit, I said to my husband: âHow sinister is this?â Iâm a humanitarian aid worker. Why would there even be a
3
9
90
Killer feature for the ChatGPT UI would be a "flag this answer as terrible" button. I want to see a filtered list of all my failed chats so I can immediately retry them with GPT-5.
2
0
17
Agent Reinforcement Trainer has taken off like a rocket since we launched RULER a couple weeks ago. Today, we passed 5,000 stars on GitHub!. The community is super friendly and active and it has never been easier to get started with RL. Come join us on GitHub/Discord!
16
24
420
@dvdcrbt And if you find this useful (or just think it's cool), please âď¸ the Agent Reinforcement Trainer repo on GitHub! That'll help more people find us!.
github.com
Agent Reinforcement Trainer: train multi-step agents for real-world tasks using GRPO. Give your agents on-the-job training. Reinforcement learning for Qwen2.5, Qwen3, Llama, and more! - OpenPipe/ART
0
3
40
MCPâ˘RL is fully open source and is released as part of the Agent Reinforcement Trainer (ART) project. We have an example notebook training Qwen2.5 to use an MCP server here!
4
7
118
How does it work? When you connect a server, MCPâ˘RL:. 1. Queries the server to get a list of tools.2. Uses a strong model to brainstorm tasks that the tools might be useful for.3. Tries to complete the task using the tools.4. Improves using RULER. In practice, it trains great!
2
7
115
Announcing MCPâ˘RL: teach your model how to use any MCP server automatically using reinforcement learning!. Just connect any MCP server, and your model will start playing with it and (using RL) "learn from experience" how to use its tools most effectively!
55
198
2K
Interesting take from the HF comments. Would make sense that it's pretrained primarily on synthetic data vs internet text -- reduces the risk of jailbreaks, accidental harmful content, copyright etc. (I still think it's a useful model though!)
2
0
20
symptom:
gpt-oss 120B is very blatantly incapable of producing linguistically correct german text. đ§ľ.
1
1
20