Just yesterday, the global AI race experienced its most violent shock since the birth of ChatGPT.
All along, on the road to the holy grail of ASI, the four knights—Anthropic, OpenAI, Google, and xAI—had been racing neck and neck, maintaining a delicate and brutal balance of power.
Several Silicon Valley powers were contesting for supremacy, evenly matched.
However, on May 7th, the landscape collapsed.
Elon Musk personally overturned xAI's chessboard.
Following xAI's dissolution, the iron army of computing power that once made the industry tremble, composed of 220,000 top-tier GPUs, was directly integrated and leased entirely to Claude.
This is a structural collapse comparable to a 'blitzkrieg': from this point on, the four-power structure has completely disintegrated, causing what was originally a long semi-final to leap directly into the final round in a nearly violent manner before people could even react.
Now, only two solitary peaks remain in Silicon Valley: on one side is Anthropic, bolstered by Google's computational power and now having swallowed the torrent of Musk's 220,000 GPUs, making rivals shiver with its unprecedented evolutionary speed.
On the other side, deeply surrounded yet still holding the hegemony of the first-mover advantage, OpenAI stands alone defending the last pass leading to ultimate intelligence.
From a chaotic battle of four armies to a death struggle between two titans, AI's endgame has entered the final five minutes of 'bayonet combat'.
This is not just the exit of one company, but the end of an era. With every passing second from now on, we are witnessing a miracle.
OSS Capital founding partner Joseph Jacks predicts that if Anthropic maintains its current momentum, its revenue will surpass Google's parent company Alphabet by mid-2028!
Even more, if the current exponential growth continues, Anthropic's annualized revenue could reach 100% of global GDP by early 2028.
Just today, the Anthropic Institute officially made its debut.
This time, Factory A rarely disclosed its internal changes: software engineering is undergoing a dramatic shift, and AI is accelerating the R&D of AI itself.
This time, they posed 53 ultimate AI questions: When a 3-person team replaces 300 people, who will train the next generation of experts? When an intelligence explosion occurs, businesses and professions reshape, and billions share the same AI models, what will happen to this world?
Meanwhile, at a recent opening keynote, the three major directions for Anthropic's next-generation model were also revealed!
53 Questions, Each Asking the Same Thing
Just today, Anthropic established TAI (The Anthropic Institute) and released a complete research agenda.
53 research questions covering four major directions: Economic Diffusion, Threats & Resilience, AI Societal Impact, and AI-Driven R&D Acceleration.
What really hits a nerve here is a detail disclosed by Anthropic itself—
'If a 3-person team can now accomplish tasks that previously required 300 people, what happens to industrial organization?'
This is the reality that Anthropic has internally observed, rapidly approaching. Three people replacing 300 signifies the collapse of the entire human resource structure.
Another set of questions in TAI's research agenda is even more thought-provoking: When AI drives AI research, could recursive self-improvement emerge? If it does, will humans still be able to supervise this process?
Anthropic's answer is: We don't know. But we must research it.
They believe that we should now establish a 'hotline' infrastructure for AI crises, ensuring at least one communication channel remains in the worst-case scenario.
Relying solely on safety teams is no longer sufficient; an independent institution is needed to ask the questions that companies dare not ask internally.
TAI is not an ivory tower. Its research findings will directly feed into Anthropic's Long-Term Benefit Trust (LTBT), influencing the company's core decisions.
The LTBT has a single mission: to ensure that Anthropic always optimizes its actions for the long-term interests of humanity.
Simultaneously, Anthropic also launched a Fellowship program: open for applications from researchers worldwide, inviting experts to join the Claude ecosystem and TAI's frontier exploration.
Anthropic is putting real money on the table to recruit people to answer these questions.
Anthropic's Major Preview: Claude's 'Infinite' Memory Revolution Begins!
As an extremely powerful side in the global AI duel, how is Anthropic laying out its upcoming model roadmap? This question captivates countless AI researchers.
During the Code with Claude opening keynote, Anthropic generously disclosed three major directions for its next-generation models: higher judgment, near-infinite context, and multi-agent collaboration.
Anthropic's Head of Product R&D, Dianne Penn, systematically elaborated on the three core upgrade directions for the next generation of Claude models.
The first is 'higher judgment and code taste', aiming to enable the model to handle complex, autonomous engineering tasks—not just writing code, but making high-quality technical decisions.
The second is an '''infinite' context window', which, by combining high-quality memory mechanisms, allows the model to maintain contextual coherence and deep understanding over long-running tasks, thereby enhancing overall output quality.
The third is 'multi-agent coordination capabilities', supporting multiple Claude instances to work together, forming an 'agent team' to achieve grand, complex goals that a single model would struggle with.
These three directions point to the same thing: extending the task horizon from hours to days, and then to 'always online'.
Anthropic is attempting, through a 'memory revolution' and 'engineering judgment', to transform Claude from a 'single-response dialog box' into a 'long-term autonomous engineering partner'.
The core logic is: When the task time domain leaps from minutes to hours, even days, the essence of AI shifts from a 'tool' to 'labor'.
The Developer's Battlefield is Here
Research PM lead Dianne Penn threw out the core judgment of the entire conference:
Model capabilities are growing exponentially, but most organizations are still adopting AI on a linear path.
The gap between these two curves is the developer's battlefield.
She used a concept called 'task horizon'. It measures how long Claude can work independently while the output quality continues to improve.
This time last year, models could run independently for minutes. Now, agents operate continuously for hours. Next step? 24/7 online, proactively perceiving tasks without needing human commands.
To make this exponential curve feel tangible, Dianne traced back a timeline:
A few years ago, Claude's greatest skill was writing a decent email, and everyone was quite pleased.
A year ago, with the release of Opus 4, an agent running independently for an hour was considered remarkable.
Six months ago, agents started running tasks overnight, with work completed by the next morning.
Then last month, Mythos read the entire OpenBSD source tree. It found a 27-year-old vulnerability. Something that all human reviewers, all fuzzers, all static analyzers had missed for three decades, was dug up by an AI.
Recently, the Mozilla Firefox team achieved a feat worthy of recording in security history:
In April alone, they successfully fixed 423 security vulnerabilities using the Anthropic Claude Mythos preview.
This number directly crushed the record of 31 total fixes over the previous 15 months.
Even more incredible, this AI demonstrated extremely high precision in detecting vulnerabilities, with zero false positives.
The leaps are getting larger, and the intervals shorter.
Minutes → One hour → Overnight, each step a leap of an order of magnitude, not a linear increment.
Dianne mentioned a number that is easily overlooked: she has been involved in 18 versions of Claude. Haiku, Sonnet, Opus, Mythos—four product lines iterating in parallel.
Eight frontier models were shipped in the last 12 months, each version tackling a core challenge.
Format reliability → Long-range stability → Behavioral appropriateness → Controllable reasoning, from 'can output' → 'can write long' → 'understands social norms' → 'can think'.
Each step smoothed over the weaknesses exposed by the previous generation. By Opus 4.7, the nature of the change was different.
The coding agent Amp switches its entire intelligence mode to Opus 4.7.
The reason is singular: It scores the highest, and it no longer needs the assistance of complex scaffolding. The model is sufficient on its own.
An internet service company ran it for internal testing and solved three times the number of production engineering tasks as before.
More intriguing is another piece of feedback: Opus 4.7 can discover logical errors on its own during the planning phase, backtrack, and correct itself, resulting in faster and cleaner execution in the end.
The model has begun to self-correct.
This is a qualitative change signal.
The day after the Opus 4.7 release, Anthropic Labs launched Claude Design.
Design and code running on parallel tracks; some are already using this combination to build production-ready interfaces directly.
The AI Trend: Those Who Embrace It Will Thrive
As developers, how should we face all of this?
Here are a few thoughts from within Anthropic.
Exponential progress will not stop. You must build for capabilities that do not yet exist, not just today's Claude. The new models will be vastly more powerful than they are now.
In the past, scaffolding was meant to compensate for the model's shortcomings; now, its role is to amplify the model's inherent intelligence.
In the past, you had to design complex iterative loops and figure out retry mechanisms; now, these can be handed back to the model itself—letting it think things through and do them correctly.
The trend is already very clear. The next exponential leap is approaching, and it's not a small step.
The way humans collaborate with Claude must change accordingly.
First, design for the next version of Claude, not the current one.
The developers who ultimately win are those who optimize their architecture to accommodate the next leap in intelligence, not those who merely chase today's incremental accuracy.
This means: maintaining harder evaluation sets, building ambitious prototypes that you think won't work right now. When a previously impossible thing suddenly works, that's the signal—exponential progress is happening under your feet, and some magical experience previously unattainable is about to become possible.
The teams best at leveraging Claude have already discovered: model upgrades are a business opportunity.
Through automated evaluations, streamlined scaffolding, and bold use cases others haven't yet imagined, they make upgrades cheap and efficient.
That exponential curve will maintain its slope.
AI 2027: The Prophecies Are Being Fulfilled One by One
A year ago, a report titled 'AI 2027' sparked huge controversy in the tech world.
The report warned that humanity could face an existential-level threat from AI by 2030. Many people's initial reaction was: this is an alarmist doomsday forecast.
However, looking back and comparing now, everything that happened to Anthropic on this day validates that report's logical chain.
Today, Google Docs co-founder Steve Newman published a lengthy blog post titled 'Is AI 2027 Coming True?'
Model capabilities unexpectedly spilling over into security offense and defense—the report predicted it.
AI-driven acceleration of AI R&D, the emergence of recursive self-improvement—the report predicted it.
The industry shock of 3 people replacing 300 is already happening—the report predicted it.
Anthropic's own revenue data also corroborate the acceleration: 2025 revenue has nearly increased tenfold, annualized to $30-40 billion, having already surpassed McDonald's and Mastercard.
If OpenAI's growth rate is incredible, then Anthropic's growth rate is simply unimaginable. Its revenue nearly increased tenfold in 2025, and then accelerated further, tripling in just three months to a $30 billion run rate.
This is real-time evidence of the AI industry entering an exponential curve.
Fortunately, the report's authors left one loophole: AI still has obvious shortcomings in generalizing to real-world scenarios, enterprise adoption, and self-correction.
This chart from METR is one of the most famous charts in the AI field.
The only certainty is the report's final sentence: The current pace of change might already be the slowest we will experience for the rest of our lives.
References:
https://www.youtube.com/watch?v=GMIWm5y90xA
https://www.anthropic.com/research/anthropic-institute-agenda