MAROKO133 Breaking ai: Sakana AI's CTO says he's 'absolutely sick' of

📌 MAROKO133 Update ai: Sakana AI's CTO says he's 'absolutely sick&#

In a striking act of self-critique, one of the architects of the transformer technology that powers ChatGPT, Claude, and virtually every major AI system told an audience of industry leaders this week that artificial intelligence research has become dangerously narrow — and that he's moving on from his own creation.

Llion Jones, who co-authored the seminal 2017 paper "Attention Is All You Need" and even coined the name "transformer," delivered an unusually candid assessment at the TED AI conference in San Francisco on Tuesday: Despite unprecedented investment and talent flooding into AI, the field has calcified around a single architectural approach, potentially blinding researchers to the next major breakthrough.

"Despite the fact that there's never been so much interest and resources and money and talent, this has somehow caused the narrowing of the research that we're doing," Jones told the audience. The culprit, he argued, is the "immense amount of pressure" from investors demanding returns and researchers scrambling to stand out in an overcrowded field.

The warning carries particular weight given Jones's role in AI history. The transformer architecture he helped develop at Google has become the foundation of the generative AI boom, enabling systems that can write essays, generate images, and engage in human-like conversation. His paper has been cited more than 100,000 times, making it one of the most influential computer science publications of the century.

Now, as CTO and co-founder of Tokyo-based Sakana AI, Jones is explicitly abandoning his own creation. "I personally made a decision in the beginning of this year that I'm going to drastically reduce the amount of time that I spend on transformers," he said. "I'm explicitly now exploring and looking for the next big thing."

Why more AI funding has led to less creative research, according to a transformer pioneer

Jones painted a picture of an AI research community suffering from what he called a paradox: More resources have led to less creativity. He described researchers constantly checking whether they've been "scooped" by competitors working on identical ideas, and academics choosing safe, publishable projects over risky, potentially transformative ones.

"If you're doing standard AI research right now, you kind of have to assume that there's maybe three or four other groups doing something very similar, or maybe exactly the same," Jones said, describing an environment where "unfortunately, this pressure damages the science, because people are rushing their papers, and it's reducing the amount of creativity."

He drew an analogy from AI itself — the "exploration versus exploitation" trade-off that governs how algorithms search for solutions. When a system exploits too much and explores too little, it finds mediocre local solutions while missing superior alternatives. "We are almost certainly in that situation right now in the AI industry," Jones argued.

The implications are sobering. Jones recalled the period just before transformers emerged, when researchers were endlessly tweaking recurrent neural networks — the previous dominant architecture — for incremental gains. Once transformers arrived, all that work suddenly seemed irrelevant. "How much time do you think those researchers would have spent trying to improve the recurrent neural network if they knew something like transformers was around the corner?" he asked.

He worries the field is repeating that pattern. "I'm worried that we're in that situation right now where we're just concentrating on one architecture and just permuting it and trying different things, where there might be a breakthrough just around the corner."

How the 'Attention is all you need' paper was born from freedom, not pressure

To underscore his point, Jones described the conditions that allowed transformers to emerge in the first place — a stark contrast to today's environment. The project, he said, was "very organic, bottom up," born from "talking over lunch or scrawling randomly on the whiteboard in the office."

Critically, "we didn't actually have a good idea, we had the freedom to actually spend time and go and work on it, and even more importantly, we didn't have any pressure that was coming down from management," Jones recounted. "No pressure to work on any particular project, publish a number of papers to push a certain metric up."

That freedom, Jones suggested, is largely absent today. Even researchers recruited for astronomical salaries — "literally a million dollars a year, in some cases" — may not feel empowered to take risks. "Do you think that when they start their new position they feel empowered to try their wild ideas and more speculative ideas, or do they feel immense pressure to prove their worth and once again, go for the low hanging fruit?" he asked.

Why one AI lab is betting that research freedom beats million-dollar salaries

Jones's proposed solution is deliberately provocative: Turn up the "explore dial" and openly share findings, even at competitive cost. He acknowledged the irony of his position. "It may sound a little controversial to hear one of the Transformers authors stand on stage and tell you that he's absolutely sick of them, but it's kind of fair enough, right? I've been working on them longer than anyone, with the possible exception of seven people."

At Sakana AI, Jones said he's attempting to recreate that pre-transformer environment, with nature-inspired research and minimal pressure to chase publications or compete directly with rivals. He offered researchers a mantra from engineer Brian Cheung: "You should only do the research that wouldn't happen if you weren't doing it."

One example is Sakana's "continuous thought machine," which incorporates brain-like synchronization into neural networks. An employee who pitched the idea told Jones he would have faced skepticism and pressure not to waste time at previous employers or academic positions. At Sakana, Jones gave him a week to explore. The project became successful enough to be spotlighted at NeurIPS, a major AI conference.

Jones even suggested that freedom beats compensation in recruiting. "It's a really, really good way of getting talent," he said of the exploratory environment. "Think about it, talented, intelligent people, ambitious people, will naturally seek out this kind of environment."

The transformer's success may be blocking AI's next breakthrough

Perhaps most provocatively, Jones suggested transformers may be victims of their own success. "The fact that the current technology is so powerful and flexible… stopped us from looking for better," he said. "It makes sense that if the current technology wa…

Konten dipersingkat otomatis.

🔗 Sumber: venturebeat.com

📌 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