MAROKO133 Eksklusif ai: MIT Researchers Unveil “SEAL”: A New Step Towards Self-Improving A

📌 MAROKO133 Update ai: MIT Researchers Unveil “SEAL”: A New Step Towards Self-Impr

The concept of AI self-improvement has been a hot topic in recent research circles, with a flurry of papers emerging and prominent figures like OpenAI CEO Sam Altman weighing in on the future of self-evolving intelligent systems. Now, a new paper from MIT, titled “Self-Adapting Language Models,” introduces SEAL (Self-Adapting LLMs), a novel framework that allows large language models (LLMs) to update their own weights. This development is seen as another significant step towards the realization of truly self-evolving AI.

The research paper, published yesterday, has already ignited considerable discussion, including on Hacker News. SEAL proposes a method where an LLM can generate its own training data through “self-editing” and subsequently update its weights based on new inputs. Crucially, this self-editing process is learned via reinforcement learning, with the reward mechanism tied to the updated model’s downstream performance.

The timing of this paper is particularly notable given the recent surge in interest surrounding AI self-evolution. Earlier this month, several other research efforts garnered attention, including Sakana AI and the University of British Columbia’s “Darwin-Gödel Machine (DGM),” CMU’s “Self-Rewarding Training (SRT),” Shanghai Jiao Tong University’s “MM-UPT” framework for continuous self-improvement in multimodal large models, and the “UI-Genie” self-improvement framework from The Chinese University of Hong Kong in collaboration with vivo.

Adding to the buzz, OpenAI CEO Sam Altman recently shared his vision of a future with self-improving AI and robots in his blog post, “The Gentle Singularity.” He posited that while the initial millions of humanoid robots would need traditional manufacturing, they would then be able to “operate the entire supply chain to build more robots, which can in turn build more chip fabrication facilities, data centers, and so on.” This was quickly followed by a tweet from @VraserX, claiming an OpenAI insider revealed the company was already running recursively self-improving AI internally, a claim that sparked widespread debate about its veracity.

Regardless of the specifics of internal OpenAI developments, the MIT paper on SEAL provides concrete evidence of AI’s progression towards self-evolution.

Understanding SEAL: Self-Adapting Language Models

The core idea behind SEAL is to enable language models to improve themselves when encountering new data by generating their own synthetic data and optimizing their parameters through self-editing. The model’s training objective is to directly generate these self-edits (SEs) using data provided within the model’s context.

The generation of these self-edits is learned through reinforcement learning. The model is rewarded when the generated self-edits, once applied, lead to improved performance on the target task. Therefore, SEAL can be conceptualized as an algorithm with two nested loops: an outer reinforcement learning (RL) loop that optimizes the generation of self-edits, and an inner update loop that uses the generated self-edits to update the model via gradient descent.

This method can be viewed as an instance of meta-learning, where the focus is on how to generate effective self-edits in a meta-learning fashion.

A General Framework

SEAL operates on a single task instance (C,τ), where C is context information relevant to the task, and τ defines the downstream evaluation for assessing the model’s adaptation. For example, in a knowledge integration task, C might be a passage to be integrated into the model’s internal knowledge, and τ a set of questions about that passage.

Given C, the model generates a self-edit SE, which then updates its parameters through supervised fine-tuning: θ′←SFT(θ,SE). Reinforcement learning is used to optimize this self-edit generation: the model performs an action (generates SE), receives a reward r based on LMθ′’s performance on τ, and updates its policy to maximize the expected reward.

The researchers found that traditional online policy methods like GRPO and PPO led to unstable training. They ultimately opted for ReST^EM, a simpler, filtering-based behavioral cloning approach from a DeepMind paper. This method can be viewed as an Expectation-Maximization (EM) process, where the E-step samples candidate outputs from the current model policy, and the M-step reinforces only those samples that yield a positive reward through supervised fine-tuning.

The paper also notes that while the current implementation uses a single model to generate and learn from self-edits, these roles could be separated in a “teacher-student” setup.

Instantiating SEAL in Specific Domains

The MIT team instantiated SEAL in two specific domains: knowledge integration and few-shot learning.

• Knowledge Integration: The goal here is to effectively integrate information from articles into the model’s weights.
• Few-Shot Learning: This involves the model adapting to new tasks with very few examples.

Experimental Results

The experimental results for both few-shot learning and knowledge integration demonstrate the effectiveness of the SEAL framework.

In few-shot learning, using a Llama-3.2-1B-Instruct model, SEAL significantly improved adaptation success rates, achieving 72.5% compared to 20% for models using basic self-edits without RL training, and 0% without adaptation. While still below “Oracle TTT” (an idealized baseline), this indicates substantial progress.

For knowledge integration, using a larger Qwen2.5-7B model to integrate new facts from SQuAD articles, SEAL consistently outperformed baseline methods. Training with synthetically generated data from the base Qwen-2.5-7B model already showed notable improvements, and subsequent reinforcement learning further boosted performance. The accuracy also showed rapid improvement over external RL iterations, often surpassing setups using GPT-4.1 generated data within just two iterations.

Qualitative examples from the paper illustrate how reinforcement learning leads to the generation of more detailed self-edits, resulting in improved performance.

While promising, the researchers also acknowledge some limitations of the SEAL framework, including aspects related to catastrophic forgetting, computational overhead, and context-dependent evaluation. These are discussed in detail in the original paper.

Original Paper: https://arxiv.org/pdf/2506.10943

Project Site: https://jyopari.github.io/posts/seal

Github Repo: https://github.com/Continual-Intelligence/SEAL

The post MIT Researchers Unveil “SEAL”: A New Step Towards Self-Improving AI first appeared on Synced.

🔗 Sumber: syncedreview.com

📌 MAROKO133 Breaking ai: OpenCV founders launch AI video startup to take on OpenAI

A new artificial intelligence startup founded by the creators of the world's most widely used computer vision library has emerged from stealth with technology that generates realistic human-centric videos up to five minutes long — a dramatic leap beyond the capabilities of rivals including OpenAI's Sora and Google's Veo.

CraftStory, which launched Tuesday with $2 million in funding, is introducing Model 2.0, a video generation system that addresses one of the most significant limitations plaguing the nascent AI video industry: duration. While OpenAI's Sora 2 tops out at 25 seconds and most competing models generate clips of 10 seconds or less, CraftStory's system can produce continuous, coherent video performances that run as long as a typical YouTube tutorial or product demonstration.

The breakthrough could unlock substantial commercial value for enterprises struggling to scale video production for training, marketing, and customer education — markets where brief AI-generated clips have proven inadequate despite their visual polish.

"If you really try to create a video with one of these video generation systems, you find that a lot of the times you want to implement a certain creative vision, and regardless of how detailed the instructions are, the systems basically ignore a part of your instructions," said Victor Erukhimov, CraftStory's founder and CEO, in an exclusive interview with VentureBeat. "We developed a system that can generate videos basically as long as you need them."

How parallel processing solves the long-form video problem

CraftStory's advance rests on what the company describes as a parallelized diffusion architecture — a fundamentally different approach to how AI models generate video compared to the sequential methods employed by most competitors.

Traditional video generation models work by running diffusion algorithms on increasingly large three-dimensional volumes where time represents the third axis. To generate a longer video, these models require proportionally larger networks, more training data, and significantly more computational resources.

CraftStory instead runs multiple smaller diffusion algorithms simultaneously across the entire duration of the video, with bidirectional constraints connecting them. "The latter part of the video can influence the former part of the video too," Erukhimov explained. "And this is pretty important, because if you do it one by one, then an artifact that appears in the first part propagates to the second one, and then it accumulates."

Rather than generating eight seconds and then stitching on additional segments, CraftStory's system processes all five minutes concurrently through interconnected diffusion processes.

Crucially, CraftStory trained its model on proprietary footage rather than relying solely on internet-scraped videos. The company hired studios to shoot actors using high-frame-rate camera systems that capture crisp detail even in fast-moving elements like fingers — avoiding the motion blur inherent in standard 30-frames-per-second YouTube clips.

"What we showed is that you don't need a lot of data and you don't need a lot of training budget to create high quality videos," Erukhimov said. "You just need high quality data."

Model 2.0 currently operates as a video-to-video system: users upload a still image to animate and a "driving video" containing a person whose movements the AI will replicate. CraftStory provides preset driving videos shot with professional actors, who receive revenue shares when their motion data is used, or users can upload their own footage.

The system generates 30-second clips at low resolution in approximately 15 minutes. An advanced lip-sync system synchronizes mouth movements to scripts or audio tracks, while gesture alignment algorithms ensure body language matches speech rhythm and emotional tone.

Fighting a war chest battle with $2 million against billions

CraftStory's funding comes almost entirely from Andrew Filev, who sold his project management software company Wrike to Citrix for $2.25 billion in 2021 and now runs Zencoder, an AI coding company. The modest raise stands in stark contrast to the billions flowing into competing efforts — OpenAI has raised over $6 billion in its latest funding round alone.

Erukhimov pushed back on the notion that massive capital is prerequisite for success. "I don't necessarily buy the thesis that compute is the path to success," he said. "It definitely helps if you have compute. But if you raise a billion dollars on a PowerPoint, in the end, no one is happy, neither the founders nor the investors."

Filev defended the David-versus-Goliath approach. "When you invest in startups, you're fundamentally betting on people," he said in an interview with VentureBeat. "To paraphrase Margaret Mead: never underestimate what a small group of thoughtful, committed engineers and scientists can build."

He argued that CraftStory benefits from a focused strategy. "The big labs are in an arms race to build general-purpose video foundation models," Filev said. "CraftStory is riding that wave and going very deep into a specific format: long-form, engaging, human-centric video."

Why computer vision expertise matters in generative AI video

Erukhimov's credibility stems from his deep roots in computer vision rather than the transformer architectures that have dominated recent AI advances. He was an early contributor to OpenCV — the Open Source Computer Vision Library that has become the de facto standard for computer vision applications, with over 84,000 stars on GitHub.

When Intel reduced its support for OpenCV in the mid-2000s, Erukhimov co-founded Itseez with the explicit goal of maintaining and advancing the library. The company expanded OpenCV significantly and pivoted toward automotive safety systems before Intel acquired it in 2016.

Filev said this background is precisely what makes Erukhimov well-positioned for video generation. "What people sometimes miss is that generative AI video isn't just about the generative part. It's about understanding motion, facial dynamics, temporal coherence, and how humans actually move," Filev said. "Victor has spent his career mastering exactly those problems."

Enterprise focus targets training videos and product demos

While much of the public excitement around AI video generation has centered on creative tools for consumers, CraftStory is pursuing a decidedly enterprise-focused strategy.

"We are definitely thinking about B2B more than consumer," Erukhimov said. "We're thinking about companies, specifically software companies, being able to make cool training videos and product videos and launch videos."

The logic is straightforward: corporate training, product tutorials, and customer edu…

Konten dipersingkat otomatis.

🔗 Sumber: venturebeat.com