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

📌 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 Eksklusif ai: Railway secures $100 million to challenge AWS with AI-na

Railway, a San Francisco-based cloud platform that has quietly amassed two million developers without spending a dollar on marketing, announced Thursday that it raised $100 million in a Series B funding round, as surging demand for artificial intelligence applications exposes the limitations of legacy cloud infrastructure.

TQ Ventures led the round, with participation from FPV Ventures, Redpoint, and Unusual Ventures. The investment values Railway as one of the most significant infrastructure startups to emerge during the AI boom, capitalizing on developer frustration with the complexity and cost of traditional platforms like Amazon Web Services and Google Cloud.

"As AI models get better at writing code, more and more people are asking the age-old question: where, and how, do I run my applications?" said Jake Cooper, Railway's 28-year-old founder and chief executive, in an exclusive interview with VentureBeat. "The last generation of cloud primitives were slow and outdated, and now with AI moving everything faster, teams simply can't keep up."

The funding is a dramatic acceleration for a company that has charted an unconventional path through the cloud computing industry. Railway raised just $24 million in total before this round, including a $20 million Series A from Redpoint in 2022. The company now processes more than 10 million deployments monthly and handles over one trillion requests through its edge network — metrics that rival far larger and better-funded competitors.

Why three-minute deploy times have become unacceptable in the age of AI coding assistants

Railway's pitch rests on a simple observation: the tools developers use to deploy and manage software were designed for a slower era. A standard build-and-deploy cycle using Terraform, the industry-standard infrastructure tool, takes two to three minutes. That delay, once tolerable, has become a critical bottleneck as AI coding assistants like Claude, ChatGPT, and Cursor can generate working code in seconds.

"When godly intelligence is on tap and can solve any problem in three seconds, those amalgamations of systems become bottlenecks," Cooper told VentureBeat. "What was really cool for humans to deploy in 10 seconds or less is now table stakes for agents."

The company claims its platform delivers deployments in under one second — fast enough to keep pace with AI-generated code. Customers report a tenfold increase in developer velocity and up to 65 percent cost savings compared to traditional cloud providers.

These numbers come directly from enterprise clients, not internal benchmarks. Daniel Lobaton, chief technology officer at G2X, a platform serving 100,000 federal contractors, measured deployment speed improvements of seven times faster and an 87 percent cost reduction after migrating to Railway. His infrastructure bill dropped from $15,000 per month to approximately $1,000.

"The work that used to take me a week on our previous infrastructure, I can do in Railway in like a day," Lobaton said. "If I want to spin up a new service and test different architectures, it would take so long on our old setup. In Railway I can launch six services in two minutes."

Inside the controversial decision to abandon Google Cloud and build data centers from scratch

What distinguishes Railway from competitors like Render and Fly.io is the depth of its vertical integration. In 2024, the company made the unusual decision to abandon Google Cloud entirely and build its own data centers, a move that echoes the famous Alan Kay maxim: "People who are really serious about software should make their own hardware."

"We wanted to design hardware in a way where we could build a differentiated experience," Cooper said. "Having full control over the network, compute, and storage layers lets us do really fast build and deploy loops, the kind that allows us to move at 'agentic speed' while staying 100 percent the smoothest ride in town."

The approach paid dividends during recent widespread outages that affected major cloud providers — Railway remained online throughout.

This soup-to-nuts control enables pricing that undercuts the hyperscalers by roughly 50 percent and newer cloud startups by three to four times. Railway charges by the second for actual compute usage: $0.00000386 per gigabyte-second of memory, $0.00000772 per vCPU-second, and $0.00000006 per gigabyte-second of storage. There are no charges for idle virtual machines — a stark contrast to the traditional cloud model where customers pay for provisioned capacity whether they use it or not.

"The conventional wisdom is that the big guys have economies of scale to offer better pricing," Cooper noted. "But when they're charging for VMs that usually sit idle in the cloud, and we've purpose-built everything to fit much more density on these machines, you have a big opportunity."

How 30 employees built a platform generating tens of millions in annual revenue

Railway has achieved its scale with a team of just 30 employees generating tens of millions in annual revenue — a ratio of revenue per employee that would be exceptional even for established software companies. The company grew revenue 3.5 times last year and continues to expand at 15 percent month-over-month.

Cooper emphasized that the fundraise was strategic rather than necessary. "We're default alive; there's no reason for us to raise money," he said. "We raised because we see a massive opportunity to accelerate, not because we needed to survive."

The company hired its first salesperson only last year and employs just two solutions engineers. Nearly all of Railway's two million users discovered the platform through word of mouth — developers telling other developers about a tool that actually works.

"We basically did the standard engineering thing: if you build it, they will come," Cooper recalled. "And to some degree, they came."

From side projects to Fortune 500 deployments: Railway's unlikely corporate expansion

Despite its grassroots developer community, Railway has made significant inroads into large organizations. The company claims that 31 percent of Fortune 500 companies now use its platform, though deployments range from company-wide infrastructure to individual team projects.

Notable customers include Bilt, the loyalty program company; Intuit's GoCo subsidiary; TripAdvisor's Cruise Critic; and MGM Resorts. Kernel, a Y Combinator-backed startup providing AI infrastructure to over 1,000 companies, runs its entire customer-facing system on Railway for $444 per month.

"At my previous company Clever, which sold …

Konten dipersingkat otomatis.

🔗 Sumber: venturebeat.com