Skyfall AI

2

Jaideep Dhanoa@jdhanoa·58m

@skyfallai Congrats @skyfallai, enterprise super intelligence is a completely new ballgame. Excited to see what comes next.

English

0

1

5

Skyfall AI@skyfallai·8h

Yesterday, we launched Morpheus, the first Continual Reinforcement Learning Environment grounded in the Big-World hypothesis to evaluate and train RL agents. ✨ We have released five enterprise environments for researchers to start training and evaluating their agents in the real world. No, they are not toy problems. These are complex operational processes that run global businesses. 1. Process Outbound - Outbound Logistics 2. Process Inbound - Receiving & Putaway 3. Inventory Optimisation & Policy Simulator 4. Order-to-Cash - Portfolio Credit & Collections 5. Production Planning Start training now at hub.morpheus.skyfall.ai We're also on HF: huggingface.co/spaces/Skyfall… PS: Share your feedback and any environments you would like us to add.

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

English

3

0

19

1.7K

Skyfall AI@skyfallai·7h

Blog: skyfall.ai/blog/llms-are-…

English

122

Skyfall AI@skyfallai·7h

Why do we need a Big World Environment for Continual Learning? Simply because currently RL benchmarks such as Atari, OpenAI Gym, MuJoCo, and Procgen share the below three properties that have nothing to do with the real world and will never help reach Continual Learning. ▪️Episodic: every task ends and the world resets to a clean state ▪️Stationary: the rules and dynamic of the environment stay the same ▪️Non-persistent: past decisions don't affect future ones. A failed episode carries no consequence into the next one. In the real world, the environment never resets 🧬and it constantly evolves causing humans to continually adjust and adapt to new environments. This is why we need a Big World environment like Morpheus. Learn in our blog. Link below 📌

English

0

11

818

Skyfall AI@skyfallai·1d

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

English

57

109

561

235.1K

Skyfall AI@skyfallai·1d

The "never" applies to LLMs as evaluated here, i.e., inference-only, fixed-weight agents with no access to self-modification or fine-tuning loops, and the unattributable failure signature problem holds in that setting regardless of context window size. However, if an LLM could dynamically fine-tune itself from reward signals or delegate to specialized subagents, it would effectively be a different system, and the attributability argument would need to be re-examined against that architecture. The more precise version of our claim is that LLMs as typically deployed cannot be continual learners, and that the opacity of their failure modes is a fundamental obstacle to making them so.

English

6

565

Srinath Mahankali@srinathm1359·1d

@skyfallai Never being continual learners feels a bit strong and it probably also depends on what tools they have access to. Were they able to spin up subagents, modify their own context, or finetune themselves for instance?

English

0

3

773

Skyfall AI@skyfallai·1d

@kjaved_ @RichardSSutton Congratulations @kjaved_ @RichardSSutton Today we launched Morpheus, a continual reinforcement learning environment built on the Big World Hypothesis. Thank you for paving the way -- looking forward to seeing what comes next from @oaklab_ai x.com/skyfallai/stat…

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

English

4

274

Khurram Javed@kjaved_·1d

Last month, @RichardSSutton and I left Keen to do our own thing. I had ~two absolutely wonderful years at Keen, and I learned a lot working with John, Gloria, Joseph, and the rest of the team. If you want to work on some of the foundational unsolved problems in AI, such as continual learning, then I would strongly recommend applying to Keen. Going forward, Rich and I have founded a small company called Oak Lab. We have a fairly complete roadmap to building animal-like intelligence that learns purely from its own experience (the OaK architecture), and Oak Lab is going to follow this roadmap aggressively with a small, focused team. We will be sharing our progress often and aim to build a prototype of the complete OaK architecture in the next few years. A successful prototype will be closer to a baby learning in its first year than it will be to any of the current AI systems. Our strategy is to demonstrate the limitations of current methods in simple settings, and then work out algorithms that overcome these limitations in a domain-independent way. Only after we have made sufficient progress on the core algorithms will we build large-scale artifacts. If you are interested in learning about some of the specifics of our approach, then follow @oaklab_ai. We will be sharing more details in the coming weeks and months.

English

36

51

790

60.1K

Skyfall AI@skyfallai·1d

@rohanpaul_ai You get it Rohan. The models were struggling, we explain it in details in our blog. Check it out!

English

3

591

Rohan Paul@rohanpaul_ai·1d

@skyfallai Super interesting findings. The attribution problem might be the hardest part here. Do the models notice the shift before their actions change?

English

2

1.5K

Skyfall AI@skyfallai·1d

@henrythe9ths Thank you!

English

640

Henry Shi@henrythe9ths·1d

@skyfallai Congrats and excited to see what this unlocks!

San Francisco, CA 🇺🇸 English

0

2

892

Skyfall AI@skyfallai·1d

@JozsefSzalma @Scobleizer This is the exact reason why we built Morpheus, the industry need a real-world environment. Have you tried Morpheus yet? hub.morpheus.skyfall.ai

English

12

Jay@JozsefSzalma·1d

@Scobleizer Some of us have been yelling from the rooftops about this for a while.

English

0

2

108

Robert Scoble@Scobleizer·1d

A new benchmark just called the industry's bluff. Everyone assumes frontier models get better with experience. Skyfall AI published data showing the opposite. On Morpheus, their persistent enterprise sim, reward on the harder task collapsed 95% when conditions shifted. GPT-5.5 hit zero repeatedly and never recovered. The weakest RL baseline closed 99.96% of the same gap. LLMs are fixed policies, not continual learners. Full findings: skyfall.ai/blog/llms-are-…

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

English

19

9

104

12.9K

Skyfall AI@skyfallai·1d

The LLM models are going to keep getting better, but they still have fundamental limitations when it comes to adaptation and continual learning. We recently built Morpheus, a continual reinforcement learning platform to test frontier models like Gemini 3.1 Pro. Our findings showed clear limitations across all of them. For those building in the agentic space and want to train the next generation of agents, check us out here: x.com/skyfallai/stat…

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

English

3

219

swyx@swyx·1d

By the end of the year we should have: GPT 6 Fable 5.5 Gemini 3.5 Pro Grok 5 Spark 2 Kimi 3 Minimax M3.5 GLM 6 DeepSeek v4.5 Mistral 4 Qwen 4 MiMo 3 Never in the history of LLMs has the frontier been so multipolar. The benefits to agent labs and agent orchestration / LLM council judges/sidekicking are ramping up. invest accordingly

English

87

15

432

59K

Skyfall AI@skyfallai·1d

Great research! We just launched Morpheus, which validates this exact thesis. Continual learning is about becoming more capable as the world changes. Morpheus is grounded in the Big World Hypothesis: the environment has to be bigger than the agent for it to continually evolve and learn. For anyone building in the agentic space, try Morpheus and benchmark your methods. Check it out here: x.com/skyfallai/stat…

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

English

0

4

861

Yutong Bai@YutongBAI1002·1d

x.com/i/article/2076…

ZXX

6

31

280

28.6K

Skyfall AI@skyfallai·1d

7/n) Frontier LLMs are fixed policies masquerading as competence with no continual learning skills. Real continual learning is compound intelligence not stable performance. 🧬 The path forward requires algorithms capable of three things: learning from their experience rather than executing pre-trained policies, understanding causality across time horizons that exceed their immediate perception, and failing intelligibly in ways that reveal their limits and point toward improvement. This is what Morpheus is designed to benchmark. It is now open to the research community here: hub.morpheus.skyfall.ai Morpheus website: morpheus.skyfall.ai Blog: skyfall.ai/blog/llms-are-… Catch us on hugging face 🤗here: huggingface.co/spaces/Skyfall… PS: We have five environments on Morpheus, let us know which environment we should add next?

English

27

1.8K

Skyfall AI@skyfallai·1d

6/n) When an RL agent fails, we can identify the causes and hereby find a solution. However, when an LLM fails, it is impossible to diagnose the cause as context overflow, pre-training gap, or reward misalignment all produce identical performance signatures. Due to not producing attributable failure signatures, LLMs can never be continual learners.

English

19

1.6K

Skyfall AI@skyfallai·1d

5/n) LLMs Apply Pre-Trained Heuristics, Not Reward-Optimal Strategies Both Gemini 3.1 Pro and GPT-5.5 pursue different allocation strategies but their scores remain persistent rather than aiming to close the gaps. Simply put, the LLMs perform consistently because they are not learning from the reward signal at all.

English

20

1.7K

Skyfall AI@skyfallai·1d

4/n) Longer detection lags reveal the context window bottleneck GPT-5.5 in the Process-Inbound environment exhibited detection lag of greater than 25 at the second configuration shift and never recovered. This failure is consistent with a context window bottleneck where the delayed effects of configuration changes cause the diagnostic evidence needed for adaptation to fall outside the model's effective context. The outbound task does not encounter this limitation because its consequences unfold over a shorter horizon.

English

22

1.9K

Skyfall AI@skyfallai·1d

3/n) The experimental results revealed that intelligence trained in small closed environments doesn’t transfer to large open, evolving ones. Here’s why: Stability in the reward signal = Lack of Adaptation Both of the models GPT-5.5 and Gemini 3.1 Pro achieved near-identical raw allocation reward across all three configuration intervals in Task 1. But, it shows a clear flaw: the current configuration sequence does not yet push either model outside its pre-training distribution.

English

2

34

2.7K

Skyfall AI@skyfallai·1d

2/n) The evaluation matrix to properly test continual learning abilities were: - Per-Configuration Reward: How well the agent performs within each interval separately? - Adaptation Speed: How fast (based on number of decision steps) can the agent detect and respond to a configuration shift? - Forgetting: What’s the immediate performance drop / increase when a configuration shift happens? Does the agent retain any of its competence from before? - Recovery time: After detecting a shift, how long until the agent's behavior stabilizes in the new configuration? - Stability: Does the agent's policy oscillate wildly, or does it settle into consistent behavior? - Performance gap relative to a configuration: What’s the difference between what an agent actually achieved and the best it theoretically could have achieved in that specific configuration?

English

31

3.1K

Skyfall AI@skyfallai·1d

1/n) Firstly, Morpheus is grounded in the Big World Hypothesis: intelligence learned in small, closed environments does not transfer to large, open, and evolving worlds. Enterprise operations are canonical Big World settings. So, we deployed the below frontier models GPT-5.5 and Gemini 3.1 Pro in two complex environments on Morpheus - Process-Outbound (test delayed reward) and Process-Inbound (test sequential process accuracy) The first task was Dynamic Resource Allocation Under Structured Drift where the agent had to allocate limited resources across competing priorities in real time. The second task was Scheduling Under Drift with Delayed Effects where the agent decides the sequence and timing of jobs. The main trap here was delayed reward signals. For instance, the agent wouldn’t know whether it made the right decision or action until after a week.

English