Unfolded AI

AI Engineer exploring world of AI.
Large Language Models | Reinforcement Learning | Deep Learning | Computer Vision | Data Science
Let's Unfold AI!

[Unfolded AI @TheUnfolded_AI]

#AI #TechTwitter
Looking to connect with people passionate about Artificial Intelligence, Machine Learning, LLMs, &AI research.
Whether u're building AI applications, exploring research or experimenting with AI, let's connect, share ideas, and learn 4m each other!
Say hello. 🤝

[Unfolded AI @TheUnfolded_AI]

NEW: Brute force computation is not enough for intelligence. Clever new methods are needed. The science of Artificial Intelligence ( AI ) is far from finished.
~ AI Pioneer Richard Sutton
#ArtificialIntelligence #AI
So, it's "Bitter Lesson 2.0" !
RIP Scalable Computation.

[Unfolded AI @TheUnfolded_AI]

#AI model alone doesn't make an agent.
The harness is what makes it operational: tools, memory, context management, execution loops, permissions, error handling & feedback. The model reasons. Harness enables it 2 act, observe, recover & complete tasks.
Agent = Model + #Harness.

[Unfolded AI @TheUnfolded_AI]

NEW: Brute force computation is not enough for intelligence. Clever new methods are needed. The science of Artificial Intelligence ( AI ) is far from finished.
~ AI Pioneer Richard Sutton
#ArtificialIntelligence #AI
So, it's "Bitter Lesson 2.0" !
RIP Scalable Computation.

[Unfolded AI @TheUnfolded_AI]

#AI model alone doesn't make an agent.
The harness is what makes it operational: tools, memory, context management, execution loops, permissions, error handling & feedback. The model reasons. Harness enables it 2 act, observe, recover & complete tasks.
Agent = Model + #Harness.

[Unfolded AI @TheUnfolded_AI]

@omarsar0 V should stop confusing scaling architecture with understanding intelligence. Transformers have shown what's possible with scale, but continual learning, reasoning, planning & knowledge acquisition remain open challenges. A successful architecture is not finished theory of intel.

[Unfolded AI @TheUnfolded_AI]

1/ On-policy RL can't teach an agent a behavior it never tries.
"Stabilizing On-Policy Targeted Guidance in Agentic Reinforcement Learning" turns LLM-judge text feedback into a token-level signal 🧵
#ReinforcementLearning #LLM #artificialIntelligence #AIResearch #Research

[Unfolded AI @TheUnfolded_AI]

#AI #TechTwitter
Looking to connect with people passionate about Artificial Intelligence, Machine Learning, LLMs, &AI research.
Whether u're building AI applications, exploring research or experimenting with AI, let's connect, share ideas, and learn 4m each other!
Say hello. 🤝

[Unfolded AI @TheUnfolded_AI]

Agentic AI has an energy problem caching doesn't fix.
New law for #LLM inference energy (13 dense LLMs, H100):
• Agentic requests: 13.0× the energy of non-agentic
• Prefix caching cuts prefill cost 93%
• Cached agentic requests still cost 4.8× more
The KV cache is the bill.

[Unfolded AI @TheUnfolded_AI]

Kimi K3.1 Pro open weights are out. The best part isn't a benchmark win.

It's an ablation: Moonshot AI trained the same 2.8T model with and without cyber-related data and published both scores.

What it says about open-model AI safety (1/9)

#OpenWeights #AISafety #LLM #AI

[Unfolded AI @TheUnfolded_AI]

Hello everyone! I’m looking to connect with people interested in:
- Large Language Models LLMs
- Agentic AI
- #Automation
- Reinforcement Learning
- #deeplearning
- Data Science
- #computervision
Let’s connect and build our own little network, lets grow togather #AIAgents

[Unfolded AI @TheUnfolded_AI]

Paper: The SIMPLE Law for Energy Cost of LLM Inference https://t.co/EKzsfB3ZMQ

Also: Intelligence per Joule peaks at a middle reasoning budget. Accuracy rises with more reasoning tokens, but energy rises faster.
#ArtificialIntelligence #LLM #AgenticAI #AI #Research

[Unfolded AI @TheUnfolded_AI]

Agentic AI has an energy problem caching doesn't fix.
New law for #LLM inference energy (13 dense LLMs, H100):
• Agentic requests: 13.0× the energy of non-agentic
• Prefix caching cuts prefill cost 93%
• Cached agentic requests still cost 4.8× more
The KV cache is the bill.

[Unfolded AI @TheUnfolded_AI]

@kloss_xyz= learn systems architecture
@emollick= learn AI science
@Hesamation= learn AI/ML@claudeskills101

[Unfolded AI @TheUnfolded_AI]

The best 20 accounts to follow in AI:
@karpathy= learn LLMs
@steipete= learn openclaw
@gregisenberg= learn startup ideas
@rileybrown= learn vibecoding
@corbin_braun= learn cursor
@jackfriks= learn solo apps
@levelsio= learn solo startups
#ArtificialIntelligence #AI

[Unfolded AI @TheUnfolded_AI]

@marclou= learn solo startups
@EXM7777= learn AI ops + systems
@eptwts= learn AI money
@ai_explorer25= learn about ai tools
@mardehaym= learn enterprise AI
@vasuman= learn AI agents
@AmirMushich= learn AI ads
@0xROAS= learn AI UGC
@AndrewYNg= learn AI engineering

[Unfolded AI @TheUnfolded_AI]

15/ Takeaway: rubric judges already explain failures in text. Distilling it only where the agent failed, alongside RL, let a coding agent learn behaviors outside its support without losing task performance.

Paper: https://t.co/DPM32YVkqe

#MachineLearning #AIResearch

[Unfolded AI @TheUnfolded_AI]

1/ On-policy RL can't teach an agent a behavior it never tries.
"Stabilizing On-Policy Targeted Guidance in Agentic Reinforcement Learning" turns LLM-judge text feedback into a token-level signal 🧵
#ReinforcementLearning #LLM #artificialIntelligence #AIResearch #Research

[Unfolded AI @TheUnfolded_AI]

14/ Stated limitations:

• Each reward needs a hand-written feedback string
• Only multi-turn SWE tested
• Targets mostly style/format, not core capability
• Many OPSD variants weren't compared head-to-head
Experiments use public Qwen3; Cursor used a version for Composer 2.5.