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

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

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 Hot ai: Anthropic launches Cowork, a Claude Desktop agent that works i

Anthropic released Cowork on Monday, a new AI agent capability that extends the power of its wildly successful Claude Code tool to non-technical users — and according to company insiders, the team built the entire feature in approximately a week and a half, largely using Claude Code itself.

The launch marks a major inflection point in the race to deliver practical AI agents to mainstream users, positioning Anthropic to compete not just with OpenAI and Google in conversational AI, but with Microsoft's Copilot in the burgeoning market for AI-powered productivity tools.

"Cowork lets you complete non-technical tasks much like how developers use Claude Code," the company announced via its official Claude account on X. The feature arrives as a research preview available exclusively to Claude Max subscribers — Anthropic's power-user tier priced between $100 and $200 per month — through the macOS desktop application.

For the past year, the industry narrative has focused on large language models that can write poetry or debug code. With Cowork, Anthropic is betting that the real enterprise value lies in an AI that can open a folder, read a messy pile of receipts, and generate a structured expense report without human hand-holding.

How developers using a coding tool for vacation research inspired Anthropic's latest product

The genesis of Cowork lies in Anthropic's recent success with the developer community. In late 2024, the company released Claude Code, a terminal-based tool that allowed software engineers to automate rote programming tasks. The tool was a hit, but Anthropic noticed a peculiar trend: users were forcing the coding tool to perform non-coding labor.

According to Boris Cherny, an engineer at Anthropic, the company observed users deploying the developer tool for an unexpectedly diverse array of tasks.

"Since we launched Claude Code, we saw people using it for all sorts of non-coding work: doing vacation research, building slide decks, cleaning up your email, cancelling subscriptions, recovering wedding photos from a hard drive, monitoring plant growth, controlling your oven," Cherny wrote on X. "These use cases are diverse and surprising — the reason is that the underlying Claude Agent is the best agent, and Opus 4.5 is the best model."

Recognizing this shadow usage, Anthropic effectively stripped the command-line complexity from their developer tool to create a consumer-friendly interface. In its blog post announcing the feature, Anthropic explained that developers "quickly began using it for almost everything else," which "prompted us to build Cowork: a simpler way for anyone — not just developers — to work with Claude in the very same way."

Inside the folder-based architecture that lets Claude read, edit, and create files on your computer

Unlike a standard chat interface where a user pastes text for analysis, Cowork requires a different level of trust and access. Users designate a specific folder on their local machine that Claude can access. Within that sandbox, the AI agent can read existing files, modify them, or create entirely new ones.

Anthropic offers several illustrative examples: reorganizing a cluttered downloads folder by sorting and intelligently renaming each file, generating a spreadsheet of expenses from a collection of receipt screenshots, or drafting a report from scattered notes across multiple documents.

"In Cowork, you give Claude access to a folder on your computer. Claude can then read, edit, or create files in that folder," the company explained on X. "Try it to create a spreadsheet from a pile of screenshots, or produce a first draft from scattered notes."

The architecture relies on what is known as an "agentic loop." When a user assigns a task, the AI does not merely generate a text response. Instead, it formulates a plan, executes steps in parallel, checks its own work, and asks for clarification if it hits a roadblock. Users can queue multiple tasks and let Claude process them simultaneously — a workflow Anthropic describes as feeling "much less like a back-and-forth and much more like leaving messages for a coworker."

The system is built on Anthropic's Claude Agent SDK, meaning it shares the same underlying architecture as Claude Code. Anthropic notes that Cowork "can take on many of the same tasks that Claude Code can handle, but in a more approachable form for non-coding tasks."

The recursive loop where AI builds AI: Claude Code reportedly wrote much of Claude Cowork

Perhaps the most remarkable detail surrounding Cowork's launch is the speed at which the tool was reportedly built — highlighting a recursive feedback loop where AI tools are being used to build better AI tools.

During a livestream hosted by Dan Shipper, Felix Rieseberg, an Anthropic employee, confirmed that the team built Cowork in approximately a week and a half.

Alex Volkov, who covers AI developments, expressed surprise at the timeline: "Holy shit Anthropic built 'Cowork' in the last… week and a half?!"

This prompted immediate speculation about how much of Cowork was itself built by Claude Code. Simon Smith, EVP of Generative AI at Klick Health, put it bluntly on X: "Claude Code wrote all of Claude Cowork. Can we all agree that we're in at least somewhat of a recursive improvement loop here?"

The implication is profound: Anthropic's AI coding agent may have substantially contributed to building its own non-technical sibling product. If true, this is one of the most visible examples yet of AI systems being used to accelerate their own development and expansion — a strategy that could widen the gap between AI labs that successfully deploy their own agents internally and those that do not.

Connectors, browser automation, and skills extend Cowork's reach beyond the local file system

Cowork doesn't operate in isolation. The feature integrates with Anthropic's existing ecosystem of connectors — tools that link Claude to external information sources and services such as Asana, Notion, PayPal, and other supported partners. Users who have configured these connections in the standard Claude interface can leverage them within Cowork sessions.

Additionally, Cowork can pair with Claude in Chrome, Anthropic's browser…

Konten dipersingkat otomatis.

🔗 Sumber: venturebeat.com