Anthropic Warns Globally, OpenAI Has Crossed the 'Reliability Threshold': AI Self-Acceleration Activated

marsbitPublished on 2026-06-06Last updated on 2026-06-06

Abstract

Anthropic issues a global warning, urging a pause in AI research due to fears of recursive self-improvement where AI accelerates its own development, nearing a "self-building" tipping point. Simultaneously, OpenAI's Yann Dubois provides a key insight: AI's perceived "jump" in usefulness stems from crossing a "reliability threshold." Before this point, AI is an unreliable toy; after (around Dec 2023 for OpenAI), it becomes a dependable tool, triggering self-acceleration. This is evident as AI now assists in its own R&D, boosting researcher productivity. Dubois argues AI development is more "craft" than pure science, relying heavily on intuition. He highlights "the last-mile AI dividend": if current models were frozen, focused development on vertical application harnesses (orchestration systems connecting AI to real-world data and permissions) could deliver AGI-like performance in many domains. The main bottleneck isn't model intelligence, but integration—granting access, connecting data, and embedding into workflows. However, major challenges remain, like enabling continual learning so AI improves with experience rather than plateauing. For startups, the opportunity lies in solving these intricate, real-world integration problems—the hard work of bringing powerful models down to ground level.

The AI world was jolted by a sudden thunderclap!

Anthropic has issued a warning to all humanity: Stop researching AI!

Internal data from Anthropic indicates that AI is accelerating the development of AI, and a path towards recursive self-improvement may have emerged.

In other words, AI is approaching the tipping point of "building itself."

This process is faster than Anthropic anticipated, prompting them to call for a slowdown or pause in AI research.

Meanwhile, Yann Dubois, leader of OpenAI's post-training team, offered a more micro, yet equally thought-provoking perspective in a recent interview:

AI Evolution Isn't About Suddenly Cheating, It Just Passed the Pass Mark!

In his latest interview, he revealed several internal perspectives:

The growth of AI capabilities is linear and continuous, but the "usefulness" experienced by users is discrete and jumpy.

Because before reaching a certain "Reliability Threshold," AI is just a clever trick; once it crosses that point, it becomes a reliable worker you can delegate tasks to, initiating self-acceleration.

This threshold, OpenAI crossed around last December.

Furthermore, Yann Dubois made a counter-intuitive assertion: AI development resembles a "Craft" more than a "Science."

This insight holds immense tension: in this field emphasizing raw computing power, what ultimately triumphs is something akin to an alchemist's "flare (intuition/inspiration)."

He also introduced the concept of the "Last-Mile AI Dividend."

If we froze all current models and focused solely on developing vertical applications (Harnessing), we could already achieve AGI.

The bottleneck isn't the model's brain, but in "permissions, connectivity, and data." This directly pours cold water on hesitant developers while simultaneously pointing to where the gold lies.

Reliability Threshold Crossed, AI Self-Accelerates

The past few weeks have been lively in the AI world: GPT-5.5 was released, Claude Mythos also emerged.

Especially in areas like cybersecurity and AI agents writing code, it feels like things are changing daily, and AI progress feels like it suddenly "leaped a grade."

Dubois puts it rather bluntly: Capability improvement has actually been quite continuous. The feeling of being on a rocket stems from a "reliability gate" standing in the middle.

Before crossing that gate, AI is like a smart but unreliable intern: it can write, calculate, and offer ideas, but you wouldn't dare hand over real work to it.

After crossing it, you dare let it "actually get to work."

He estimates OpenAI crossed this line around "last December," leading to the externally perceived "step-change leap."

More stimulating is the second reason: when models become good enough, they accelerate R&D itself.

This is exactly what Anthropic is most worried about.

Dubois mentions, particularly in programming scenarios, researchers code daily. When the model gets stronger, it's like the entire team gains a partner that doesn't sleep — helping researchers build their toolchains and "feeding AI with AI" when training the next generation of models.

Once this acceleration loop starts spinning, it spins faster and faster. It's no surprise recent months have felt "increasingly intense."

This is also happening inside Anthropic. By Q2 2026, the code contributed per person per quarter was already 8 times that of Q1 2024.

The third driving force comes from the "transformation and upgrade" of Reinforcement Learning (RL).

Early reasoning models like o1 mainly focused on scoring high on tasks with "verifiable rewards" — math problems, programming contests, because right/wrong was clear, and rewards easy to define.

But over the past year, they've migrated the tools honed in competitions to more real-world, ambiguous work scenarios: no longer just optimizing for "problems with standard answers," but optimizing for "things users find genuinely useful."

In short: evolving from test-takers to workplace professionals.

AI Engineers Aren't Scientists, AI is 'Cultivated'

But once you step into the real world, trouble arises: How do you improve reliability?

Dubois offers a straightforward "probability model":

Given many systems are now AI-agentic, you can roughly think of it as "a certain probability of error every two minutes"; the longer it runs, the higher the chance the final answer goes off the rails.

So-called "improving reliability" essentially means continuously pushing down this "error rate per two minutes."

This is an inherent, tough challenge for AI agents.

This also explains why Dubois says building AI resembles "craftsmanship" more than textbook "scientific experiments."

The realistic process is often: first, use experience, intuition, trial-and-error to build something, even with a touch of "alchemy"; then, once it actually works and is useful, go back and supplement it with more scientific explanations and methodology.

He also mentions a rather ironic anecdote —

When ChatGPT first went public and mentioned using RL, his initial reaction was, "Isn't that too complex? Supervised Fine-Tuning (SFT) should be enough," which was precisely the approach he aimed to validate while working on Alpaca at Stanford.

But later facts showed that once model scale crosses a certain level, RL really does "suddenly start working well," albeit at a high cost — sampling many answers, judging which are right/wrong — requiring significant computing power and systems engineering.

Vertical Harnessing Has Reached AGI

When it comes to "bringing AI into reality," you can't avoid the term entrepreneurs love these days: Harness (orchestration system).

Some see it as the "external skeleton" for AI agents, while others suspect models will eventually "consume" it.

Dubois's stance is pragmatic:

In the short term, vertical-scenario Harnesses are valuable, able to push reliability from 80% to 85%.

But the premise is accepting that models are continuously improving, and Harnesses need constant re-tuning.

Attempting to create a long-term stable, universally applicable "General Harness" is, in his view, essentially a dead end.

He even throws out a provocative judgment: If we "froze" current models today and focused solely on meticulously polishing Harnesses and training around them, many in specific fields might "clearly sense the flavor of Artificial General Intelligence (AGI)."

The Last Mile

But what truly excites and worries Dubois is the perennial challenge of "continual learning."

Three years ago, when ChatGPT first exploded in popularity, he and a friend seriously discussed starting a venture for personalized memory and continual learning.

At the time, they thought, "OpenAI will solve this within 6 months," so they didn't proceed. Three years later, he's now at OpenAI, yet this issue remains unresolved.

The current model's awkwardness lies in this: on day one at a company, it might be more capable than most new hires (high starting point); but afterwards, it largely "stays the same," because it doesn't learn more about you or become more efficient within a specific environment.

The human learning curve climbs upward; AI's curve tends to flatten.

Bending AI's curve from "flat" to "continuously rising" is, in Dubois's view, one of the most important problems ahead.

So, is there still room for startups to build vertical applications?

Dubois's answer is clear: Not only is there room, but it's significant.

Because the real bottleneck often isn't "whether the model is smart enough," but the last mile — how to grant permissions, how to connect data, how to build connectors, how to integrate into specific business workflows.

No matter how high foundation models fly, if they don't land, they're just fireworks. Pulling them down to earth, giving them the right keys, and opening the right doors is the most valuable, grunt work.

References:

https://x.com/Potatoloogs/status/2062494654885749126

https://www.youtube.com/watch?v=DhD1zZ8w8Mw&t=3s

This article is from the WeChat public account "XinZhiYuan" (New AI Era), Author: ASI Apocalypse

Related Questions

QAccording to the article, what significant threshold did OpenAI reportedly cross around December of last year, and what was the perceived impact?

AAccording to the article, OpenAI reportedly crossed a 'reliability threshold' around December of last year. The impact was that AI transitioned from being seen as an unreliable 'toy' or 'intern' to becoming a dependable 'employee' that users felt confident delegating real work to, leading to a perceived discontinuous leap in usefulness despite continuous underlying capability improvements.

QWhat are the two main factors, as described in the article, that have contributed to the recent acceleration in AI development?

AThe two main factors contributing to the recent acceleration are: 1) AI models themselves have become good enough to assist in the research and development of newer AI models, creating a self-accelerating feedback loop, particularly in areas like coding. 2) The application of reinforcement learning (RL) techniques has shifted from optimizing for tasks with clear, verifiable answers (like math problems) to optimizing for more subjective, real-world usefulness.

QHow does Yann Dubois characterize the process of building advanced AI, and what analogy is used to describe the initial phase?

AYann Dubois characterizes the process of building advanced AI as more of a 'craft' or 'handiwork' rather than a strict 'science.' The analogy used for the initial phase is 'alchemy,' suggesting it relies heavily on intuition, experience, and iterative trial-and-error before a more scientific methodology can be applied retroactively.

QWhat controversial claim does Dubois make about achieving AGI (Artificial General Intelligence) with current models?

ADubois claims that if all current AI models were 'frozen' and developers focused solely on building and refining vertical application harnesses (orchestration systems) around them, people in many domains would 'noticeably feel the taste of AGI.' He argues the bottleneck is not the model's intelligence, but the 'last mile' issues of permissions, data connectivity, and integration into specific workflows.

QWhat major unsolved challenge in AI does Dubois identify as critical for the future, relating to how AI systems perform over time in a specific environment?

ADubois identifies 'continual learning' as a major unsolved challenge. He points out that while an AI might be highly capable on its first day in a specific environment (like a company), it typically 'stays the same' afterward because it doesn't learn and improve from ongoing experience within that context, unlike a human employee whose performance curve rises over time. Making AI's performance curve 'continuously upward' is deemed a crucial upcoming problem.

Related Reads

US Stocks Suffer Worst Plunge Since 2025: Three Triggers Ignite Tech Stock Valuation Reset

The US stock market experienced its most severe sell-off since the 2025 tariff crisis on June 5th, 2025. The Nasdaq Composite plummeted 4.18%, the S&P 500 fell 2.64%, and the Dow Jones dropped 695 points. The panic stemmed from three converging factors. First, Broadcom's earnings report ignited fears of a slowdown in AI growth. While its AI chip revenue surged 143% YoY to $10.8B, its Q3 AI revenue guidance of $16B fell short of the $17.2B consensus. This triggered a massive sector-wide sell-off, with the Philadelphia Semiconductor Index crashing 10.26% and semiconductor stocks losing roughly $1.3 trillion in market value in a single day. Second, a shockingly strong May jobs report crushed hopes for Federal Reserve rate cuts. Non-farm payrolls added 172,000 jobs, doubling expectations. This robust data, combined with persistently high oil prices above $92/barrel due to the ongoing Iran war and blockade of the Strait of Hormuz, drastically increased market expectations for a potential Fed rate hike instead of a cut. Higher interest rates compress the valuations of growth-heavy tech stocks. Third, the prolonged Iran conflict continues to fuel inflationary pressures, complicating the Fed's policy decisions and undermining the "inflation is tamed" narrative. Together, these events challenged the twin pillars of the market rally: the "limitless AI growth" story and expectations for imminent monetary easing. The sell-off spread globally, impacting Asian and European markets and cryptocurrencies. The article posits this is likely a severe "valuation repricing" rather than the end of the AI story. The underlying demand for AI remains strong, but investor expectations for growth speed and the prices they are willing to pay are being recalibrated. Key upcoming factors include the June FOMC meeting, future AI company earnings, and developments in the Iran conflict.

marsbit2h ago

US Stocks Suffer Worst Plunge Since 2025: Three Triggers Ignite Tech Stock Valuation Reset

marsbit2h ago

From Madison Square Garden to Kalshi: Prediction Markets Break into the NBA Finals

From Madison Square Garden to Kalshi: Prediction Markets Break into the NBA Finals Prediction markets are playing a significant role in the 2026 NBA Finals, particularly around the New York Knicks' unexpected 2-0 series lead. Platforms like Kalshi and Polymarket have seen massive trading volumes, exceeding hundreds of millions of dollars on championship and related markets. Their influence extends beyond online trading. Kalshi's official partnership with Madison Square Garden has given it prominent physical branding at the arena. Furthermore, local businesses like The Jeffrey bar are using prediction market contracts to hedge the risk of game-result-based promotions, turning potential losses into manageable costs—a concept similar to the famous "Mattress Mack" strategy from traditional sports betting. These markets differentiate themselves by offering a wider, more entertainment-focused range of "event contracts" beyond typical game outcomes, such as predicting celebrity attendance. They also have broader accessibility across the U.S. compared to age- and location-restricted traditional sportsbooks. However, their rapid integration into sports raises regulatory and ethical questions. The NBA is cautiously engaging, discussing integrity frameworks with regulators like the CFTC. While the league permits minor investments like Giannis Antetokounmpo's stake in Kalshi, it advocates for strict rules to prevent insider trading. Many fans express concern on platforms like Reddit, fearing that the close ties between prediction markets, the league, and players could compromise the game's integrity. The NBA Finals has thus become a high-stakes testing ground, showcasing prediction markets' commercial potential while challenging traditional boundaries between financial trading, entertainment, and gambling.

marsbit4h ago

From Madison Square Garden to Kalshi: Prediction Markets Break into the NBA Finals

marsbit4h ago

Recursive Self-Improvement AI Gains Traction, Google Pours Cold Water, While DeepSeek and Others Approach the Fringes

The term "recursive self-improvement" (RSI), where AI improves itself autonomously, is gaining momentum in the AI industry. Startups like Recursive Superintelligence and projects such as Andrej Karpathy's Auto-Research aim to create systems where AI designs, implements, and validates its own research, moving toward superintelligence. While Google CEO Sundar Pichai cautions that such exponential acceleration is not yet a reality, progress is evident. For instance, Anthropic reported its Claude Code writes nearly 100% of the team's code, though it still lacks true self-direction. Analysts frame RSI development in stages: "adequacy" (systems functioning without humans), "parity" (matching human research quality), and "supremacy" (exceeding human-AI collaboration). Reaching parity could trigger rapid, unpredictable advancement due to AI's continuous operation. In China, companies like DeepSeek and Baidu incorporate self-optimization techniques without explicitly branding them as RSI, focusing on algorithmic efficiency and reinforcement learning. However, challenges remain, including "model collapse" from training on AI-generated data and the immense computational and open-collaboration requirements. Ultimately, RSI represents a trend of increasing automation in AI development, potentially reducing human oversight in the creation process itself.

marsbit4h ago

Recursive Self-Improvement AI Gains Traction, Google Pours Cold Water, While DeepSeek and Others Approach the Fringes

marsbit4h ago

Trading

Spot
Futures

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片