Your LLM’s personal trainer. Get with the program, nerds.
Joined March 2024
- Tweets232
- Following11
- Followers1.5K
- Likes43
For interested readers and learners:
youtube.com/playlist?list=PL…
Multi-Agent Reinforcement learning (MARL) is the Future of training the next generation of AI applications
👀
today we are announcing reinforcement finetuning, which makes it really easy to create expert models in specific domains with very little training data.
livestream going now: openai.com/12-days/
alpha program starting now, launching publicly in q1
This is where Haptic can provide a valuable service to upcoming AI agents. To avoid malicious feedback this ability can be token gated to top holders and aligned users
Feedback is something that can be crowdsourced and becomes the easiest way to level up these agents
What if users and token holders can signal to the AI agent which of their responses are appreciated by the users and which responses seem like low quality responses
Programmatic Upgrades - Controlled by either the deployer of the AI agent or platform wide upgrades introduced by the Agentic platform team
Data Quality - Work intensive process of acquiring relevant and non repetitive data to improve the agent through retraining
How will the winning agents be differentiated from every other AI agent out there replying to twitter posts?
Through constant upgrades, training data richness and Feedback on their response generation
There is dilution of AI agents category through rapid launches via platforms like @virtuals_io @vvaifudotfun @ai16zdao @SimulacrumIO
HapticAI retweeted
Tulu 2.5 work is still underrated RLHF paper. Lots of industry interest, not that much immediate academic uptake. Really great empirical study on how DPO and PPO work across datasets and implementation.
Core to a lot of our efforts on Tulu 3.
If you are a team or developer looking to build and deploy your AI agents - feel free to DM us and hop on a call
RLHF isn't just about making AI agents better - it's about making them better partners for humans. It's a bridge between raw capability and practical utility.
Another Future development we are looking to include in our product suite that will unlock major upgrades for partners:
Multi-agent RLHF - Agents training each other further
Automated feedback loops
Real-world improvements from RLHF that Agent creators can expect by integrating Haptic's RLHF:
1- More natural dialogue
2- Better task comprehension
3- Improved safety boundaries
4- Enhanced reasoning
Solution approaches our team has been exploring in the last month:
1- Diverse evaluator pools
2- Structured feedback frameworks
3- Iterative refinement loops
4- Robust validation methods
Common challenges in RLHF that become more challenging with improving Agentic behavior
1- Reward hacking
2- Preference inconsistency
3- Scaling human feedback quickly to keep up with the Agents outputs
4- Maintaining model stability
This balance is crucial: we want the agent to improve without losing the beneficial capabilities it learned during pre-training.
Technically building for these AI agents involves PPO (Proximal Policy Optimization) - an algorithm that helps the model improve while staying close to its original behavior.
Think of RLHF as teaching AI agents not just what to do, but how to do it in a way that's most helpful and aligned with a set of predetermined values.
RLHF can be the factor that decides which agents win in the long term. Core operations of Agents that benefit from RLHF:
1- Task planning
2- Decision making
3- Response generation
4- Safety constraints
5- Goal alignment