Pinned Tweet
Excited to share that I've stepped into the role of CTO at @ReliableAI!
Grateful to my team and everyone who's helped along the way. Building reliable AI agents is what we're here for, and we're just getting started.
Special thanks to @FeiziSoheil and @py_laurent for their invaluable guidance and support.
Grateful to my team and everyone who's helped along the way. Building reliable AI agents is what we're here for, and we're just getting started.
Special thanks to @FeiziSoheil and @py_laurent for their invaluable guidance and support.
1
1
6
3,671
Update:
@ChatGPT sends notifications on Mac even if there are no new turns in any of my chats. I think it’s coming from my dot, even if there are no new messages from it.
0
0
0
33
Opus 5.5 is really really good. I say this as I use it to clean up obvious mess from a different model 👀.
0
0
0
34
I have used many open-source agents as personal assistants including a couple that I built myself. I always found it too much trouble to adapt them to my needs. Something would always break and I would end up spending more time fixing things than having them do something useful for me.
Lately, I have been using @ChatGPT dot and I really like the UX. The only problem is that it doesn’t have all the features I need and I can’t control what features @OpenAI ships or how fast they do it.
So I went back to the drawing board and started building one from scratch taking inspiration from my dot. It has the ability to modify any and all of its code, learn new skills, etc. (not saying these are novel ideas, I know @openclaw, @NousResearch have something similar). So far, it hasn’t gone down by breaking itself despite modifying itself several times. There was only one instance where it made an edit that broke it but it was able to automatically rollback, fix that issue and redeploy. That really surprised me, pleasantly, of course! I think the models have gotten really good over the last few months. They will continue to get better at one-shotting tasks, recover from failures and keep pursuing long-term goals. Many of us have seen this coming. But building with the latest models, I feel like I am experiencing the inflection point firsthand.
There has never been a better time to be a builder!
Lately, I have been using @ChatGPT dot and I really like the UX. The only problem is that it doesn’t have all the features I need and I can’t control what features @OpenAI ships or how fast they do it.
So I went back to the drawing board and started building one from scratch taking inspiration from my dot. It has the ability to modify any and all of its code, learn new skills, etc. (not saying these are novel ideas, I know @openclaw, @NousResearch have something similar). So far, it hasn’t gone down by breaking itself despite modifying itself several times. There was only one instance where it made an edit that broke it but it was able to automatically rollback, fix that issue and redeploy. That really surprised me, pleasantly, of course! I think the models have gotten really good over the last few months. They will continue to get better at one-shotting tasks, recover from failures and keep pursuing long-term goals. Many of us have seen this coming. But building with the latest models, I feel like I am experiencing the inflection point firsthand.
There has never been a better time to be a builder!
0
0
1
209
Priyatham Kattakinda retweeted
🚀 Harness optimization got open-weight agents to frontier-level performance at ~3% of the inference cost.
Three models. Three failure modes. Three different fixes.
The harness mattered more than we expected. x.com/i/article/2095…
Three models. Three failure modes. Three different fixes.
The harness mattered more than we expected. x.com/i/article/2095…
4
4
30
2,371
Really exciting results for continual learning using our CLI!
Current benchmarks miss the central challenge in agent optimization.
Most evaluations test one-shot improvement, while real-world use requires recursive optimization as new tasks continue to arrive.
To study this, we built a two-phase continual-learning evaluation on hard Terminal-Bench 2.0 tasks, comparing GEPA, Meta Harness, and RELAI.
The methods behaved very differently:
- GEPA improved initially but transferred negatively to unseen tasks, indicating overfitting.
- Meta Harness transferred well but could not continue improving.
- RELAI was the only method that both generalized to unseen tasks and improved again without regressing prior performance.
Full results are in the article below.
You can access RELAI’s continual-learning engine to run similar optimization studies and apply it to your own agents at relai.ai
0
1
1
296
Appreciate a quick resolution here @OpenAI
Woke up this morning to find that RELAI’s OpenAI organization account had been deactivated for alleged “distillation.” Honestly, it was both surprising and frustrating for our team.
We believe this was a false positive triggered by our Terminal-Bench evaluations and ion workloads using GPT models. We are not using OpenAI outputs to train another model or engaging in any form of model distillation.
We have submitted an appeal and would be very grateful if anyone at @OpenAI could help expedite the review, especially as we are finalizing our benchmarking results. @sama @jxnlco @thsottiaux
In the meantime, to ensure our users experience no disruption, we have switched the default models in RELAI’s continual-learning engine to Anthropic models.
0
0
0
59
Priyatham Kattakinda retweeted
Woke up this morning to find that RELAI’s OpenAI organization account had been deactivated for alleged “distillation.” Honestly, it was both surprising and frustrating for our team.
We believe this was a false positive triggered by our Terminal-Bench evaluations and ion workloads using GPT models. We are not using OpenAI outputs to train another model or engaging in any form of model distillation.
We have submitted an appeal and would be very grateful if anyone at @OpenAI could help expedite the review, especially as we are finalizing our benchmarking results. @sama @jxnlco @thsottiaux
In the meantime, to ensure our users experience no disruption, we have switched the default models in RELAI’s continual-learning engine to Anthropic models.
We believe this was a false positive triggered by our Terminal-Bench evaluations and ion workloads using GPT models. We are not using OpenAI outputs to train another model or engaging in any form of model distillation.
We have submitted an appeal and would be very grateful if anyone at @OpenAI could help expedite the review, especially as we are finalizing our benchmarking results. @sama @jxnlco @thsottiaux
In the meantime, to ensure our users experience no disruption, we have switched the default models in RELAI’s continual-learning engine to Anthropic models.
3
7
66
22,100
Priyatham Kattakinda retweeted
🚀 We’re hiring at RELAI.
We have an opening for an AI Research / Engineering role focused on continual learning for AI agents.
Requirements:
- PhD in CS, ECE, or a related field
- At least two high-impact publications
- Located in the U.S. or Canada
Details:
- Full-time
- mediate start
Apply here: forms.gle/JFWTd4t1h6QvP1…
Please feel free to share with strong candidates in your network.
We have an opening for an AI Research / Engineering role focused on continual learning for AI agents.
Requirements:
- PhD in CS, ECE, or a related field
- At least two high-impact publications
- Located in the U.S. or Canada
Details:
- Full-time
- mediate start
Apply here: forms.gle/JFWTd4t1h6QvP1…
Please feel free to share with strong candidates in your network.
0
3
14
1,466
Priyatham Kattakinda retweeted
Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work.
openai.com/index/previewi…
openai.com/index/previewi…
3,532
5,590
39,788
18,955,881
I will be in SF next week for World's Fair @aiDotEngineer. Please reach out if you want to chat about buidling (with) AI agents, continual learning or anything else.
0
0
0
22
Now that @SpaceX is acquiring @cursor_ai, what will Cursor be called? XCode? CodeX? Both are taken!
0
0
0
47
And we're hiring! See relai.ai/careers.
Today, we’re launching RELAI’s verifiable continual learning platform for AI agents and announcing $6.9 million in total funding, led by @406Ventures.
Future agents will get better through experience. But today, most of that learning happens through vibes—not verification.
ning engine is built around four principles: make every failure replayable, improve without forgetting what already works, apply feedback to the right layer of the agent stack—prompt, tool, memory, workflow, model, or code—and keep the entire loop efficient enough to run continuously.
🚀 The best news: you can now add RELAI’s continual learning loop to your agents in just two commands.
👉 Now available in limited access. Join the waitlist at relai.ai.
0
1
2
108
Priyatham Kattakinda retweeted
Today, we’re launching RELAI’s verifiable continual learning platform for AI agents and announcing $6.9 million in total funding, led by @406Ventures.
Future agents will get better through experience. But today, most of that learning happens through vibes—not verification.
ning engine is built around four principles: make every failure replayable, improve without forgetting what already works, apply feedback to the right layer of the agent stack—prompt, tool, memory, workflow, model, or code—and keep the entire loop efficient enough to run continuously.
🚀 The best news: you can now add RELAI’s continual learning loop to your agents in just two commands.
👉 Now available in limited access. Join the waitlist at relai.ai.
Future agents will get better through experience. But today, most of that learning happens through vibes—not verification.
ning engine is built around four principles: make every failure replayable, improve without forgetting what already works, apply feedback to the right layer of the agent stack—prompt, tool, memory, workflow, model, or code—and keep the entire loop efficient enough to run continuously.
🚀 The best news: you can now add RELAI’s continual learning loop to your agents in just two commands.
👉 Now available in limited access. Join the waitlist at relai.ai.
11
7
38
4,477