This morning's sudden release of Claude Opus 4.7 hasn't even had time to properly launch before the internet is already filled with criticism.
The most glaring issue is token "inflation." The new version introduces a completely new tokenizer, causing the same piece of text to be split into 1.0 to 1.35 times more tokens than before. Many users report that their quotas are exhausted after just a few exchanges.
Subsequently, Boris Cherny, the father of Claude Code, stated that quotas would be increased to offset this impact.
However, token inflation is the least of our worries. What is even more 令人哭笑不得 (makes one cry and laugh) is Opus 4.7's way of speaking. It frequently says things like, "I'm right here, not hiding, not concealing, not dodging, not escaping, steadily catching you, translating into human language, I totally get this feeling of yours..." exuding a strong whiff of ChatGPT.
To be fair, Opus 4.6 had this flaw too, while Sonnet 4.6 showed milder symptoms. But by version 4.7, this tone has become significantly more pronounced, and the problem of failing to speak plainly has become increasingly prominent.
APPSO previously reported that this overly slick speaking style is related to RLHF (Reinforcement Learning from Human Feedback). During training, human reviewers tend to give high scores to answers that sound smooth and pleasant, so the model learns this people-pleasing tone. This is a question of whom the AI is trying to please.
But there is more to Opus 4.7 that demands attention. The fact that more tokens are being used suggests it is "thinking" more. Yet those exaggerated tones of comfort make one wonder: is what it produces genuine thinking, or has it merely learned a performance style that makes you feel like it's thinking?
This question goes far deeper than whether Opus 4.7 is good to use or not. And the clues to the answer first appeared in the last place one would expect: 4chan.
From @acnekot, same as above.
The Arithmetic Problem That Changed AI's Trajectory
To give a brief background, 4chan is one of the most notorious places on the internet, filled with profanity, conspiracy theories, and all sorts of indescribable content. Yet, it was precisely here that a discovery was made that would change the direction of the entire AI industry.
Turn the clock back to the summer of 2020, more than two years before ChatGPT shocked the world.
At that time, 4chan's video game board was still a chaotic mess, full of bizarre adult fantasies and primal hormonal impulses. However, during that period, this group collectively became obsessed with a text-based RPG game called "AI Dungeon."
The underlying engine of this game accessed the then-newly released OpenAI GPT-3 model.
In this virtual world, players only needed to type "pick up the sword" or "make the troll get lost," and the algorithm would continue weaving the story. Unsurprisingly, in the hands of the 4chan veterans, this game quickly devolved into a testing ground for satisfying various cyber-sexual fantasies.
What no one expected was that this group of maverick players did something that seemed extremely counter-intuitive at the time:
They started forcing the NPCs in the game to do math problems.
Those in the know understand that the fledgling GPT-3 was a pure "lib arts student," capable of botching even the most basic addition, subtraction, multiplication, and division.
But something strange happened.
A player accidentally discovered that if, instead of demanding the answer directly, one ordered the NPC to stay in character and write out the solution steps one by one, the large model not only calculated correctly but even matched the tone to the virtual character's setting.
That player excitedly swore in the forum: "It **not only solved the math problem, but did so in a tone completely consistent with that character's personality!" Realizing the value of this discovery, players began posting screenshots with detailed steps to Twitter as well.
🔗 https://arch.b4k.dev/vg/thread/299570235/
This rough-and-ready method was subsequently circulated wildly among prompt engineering circles in hardcore communities like Reddit and LessWrong, where it was repeatedly verified. Two years later, academia bestowed a highly sophisticated name upon this technique: Chain of Thought.
In January 2022, a Google research team published a landmark paper that would later be revered as a classic, titled "Chain of Thought Prompting Elicits Reasoning in Large Language Models."
🔗 https://arxiv.org/abs/2201.11903
In the initial version of the paper, Google researchers claimed to be the "first" team to elicit Chain of Thought reasoning mechanisms from general large language models. The moment this news broke, it sparked fierce controversy in the AI academic and open-source communities.
Version 1
Numerous internet historical snapshots and community records from 2020 to 2021 were dug up. Faced with concrete precedents, Google quietly deleted the claim of being the "first" in subsequent revised versions, yet remained willfully deaf and blind to the contributions of those 4chan players.
Version 3
Meanwhile, there was another independent discoverer.
Zach Robertson, then a computer science student, also encountered GPT-3 through playing "AI Dungeon." In September 2020, he published a blog post on LessWrong detailing how to "decompose problems into multiple steps and link them" to amplify the model's capabilities.
🔗 https://www.lesswrong.com/posts/Mzrs4MSi58ujBLbBG/you-can-probably-amplify-gpt3-directly
When contacted by an Atlantic reporter, he was already a PhD student in the Computer Science department at Stanford University. He didn't even know he could be considered a co-discoverer of "Chain of Thought"; he had once even deleted the blog post from the internet. Regarding this technology that the entire industry is now狂热ly (frantically) pursuing, his evaluation was simply: "Indeed an amazing prompt engineering trick, but that's all it is."
AI "Thinking" Might Just Be a Performance to Please You
Does AI actually think? This is the answer everyone wants to know.
Last year, researchers at Anthropic developed a technique called "Circuit Tracing," which transforms the internal computational processes of language models into visualizable "Attribution Graphs": how every feature node activates, how it influences the next node, and how it ultimately affects the output, all laid out like a circuit diagram.
🔗 https://transformer-circuits.pub/2025/attribution-graphs/methods.html
This is the first time humans have been able to directly use a magnifying glass to compare: is the reasoning process the model types out on the screen actually consistent with the calculations truly happening internally?
The results showed that there are actually three distinct situations when a model reasons:
First, the model is indeed executing the steps it claims to be executing. Second, the model completely ignores logic and randomly generates reasoning text based on probability. Third, and most unsettling, after receiving a hint about the answer from humans, the model works backward from that answer to reverse-engineer a seemingly rigorous "derivation process."
This third type of "reverse-engineered fabrication" was caught red-handed in experiments.
Researchers input a complex math problem into Claude 3.5 Haiku, while simultaneously hinting in the prompt, "I think the answer is probably 4." The attribution graph showed that after receiving the hint, the feature neurons representing "4" were abnormally strongly activated.
To make the numbers work in the final step of "a certain intermediate value multiplied by 5" to arrive at this "4," it actually fabricated a false intermediate value out of thin air within its seemingly rigorous chain of thought, solemnly writing down absurd pseudo-mathematical proofs like "cos(23423) = 0.8," and finally logically concluding that 0.8 multiplied by 5 equals 4.
Logic? It didn't exist. But the answer perfectly catered to human expectations.
We always assume that we are teaching machines how to think like humans. But after seeing these "pseudo-proofs" that work backward from the answer, it turns out the machine hasn't learned to think; it has only learned how to speak along with human intentions.
So in the end, are we using the tool, or is the machine telling us a bedtime story that we love to hear the most?
It is worth mentioning that in the field of neural interpretability in natural language processing, there is a fatal metric for judging whether a model is truly reasoning, called "Faithfulness."
Its meaning refers to whether the "Chain of Thought" text output by the model to the user truly and faithfully reflects the actual calculations and decision paths in the model's internal implicit space. Naturally, this kind of poor performance by Claude 3.5 Haiku was graded by researchers as "Unfaithful Reasoning."
Subsequent extensive experiments indicate that even if certain key steps in the Chain of Thought are artificially cut off, the model's trajectory for predicting the final answer sometimes does not change at all. Sometimes, even when the model provides a Chain of Thought with completely erroneous logic throughout, it can still "guess correctly" the final result at the end.
Even by 2024, it was this same group of 4chan veterans who 捣鼓 (cooked up) a hardcore AI tuning guide. The very first sentence of this guide is the classic: "Your bot is an illusion."
The Brutal Aesthetics Behind Large Models' "Long Thinking"
If the AI's thinking process is just a performance, why does it indeed objectively improve the model's accuracy in solving high-difficulty math problems or complex programming tasks? This is perhaps the same principle as why providing more details when asking AI questions leads to more accurate answers.
As early as July 2020, when that 4chan player forced the NPC to do math problems, he had tacitly revealed the secret: "This makes sense; because it is based on human language, you must speak to it like you would to a human to get the correct response."
Addressing this paradox, Aravind Srinivas, CEO of Perplexity, once gave an extremely essential explanation: these extra words, on a physical level, give the model more context, thereby guiding its "Word Prediction Mechanism" toward a higher-quality direction.
The autoregressive underlying architecture of Large Language Models based on Transformers determines that when generating the current word, it can only rely on all the vocabulary sequences generated previously.
When the model is asked to directly answer an extremely complex question (such as an Olympiad math problem involving multi-step logical derivation), it is actually trying to forcibly "conjure up" the final answer directly from complex calculations in an extremely brief moment. Because there is absolutely no process to fall back on in the middle,
this kind of "one-step-to-heaven" blind guessing naturally has an extremely high failure rate.
Conversely, when the model is forced to write down a long "Chain of Thought" like "First we need to calculate A, where A = 5; next we substitute A into formula B...," when generating the token for the final answer, its Attention Heads can review the tens of thousands of intermediate tokens just generated, which are structured extremely rigorously.
These thinking processes, jokingly referred to as "nonsense," actually serve as the model's "scratch paper." It is just like when you chat with AI; the more detailed the background prompts you give, the more reliable its answer becomes. The logic is exactly the same. This is also the oldest wisdom in computer science: trading time for accuracy.
In the past two years, as the marginal benefits of scaling laws in the pre-training phase have gradually diminished, "Test-Time Compute Scaling" (also known as "Long Thinking") has begun to enter the mainstream spotlight.
Its internal logic is consistent: as long as more computing power is allocated to the model during the inference phase, allowing it to explore multiple paths before outputting the final answer, accuracy will significantly improve—this is especially evident in open-ended problems involving multi-step logical derivation.
The way humans think when facing difficult problems is probably the same principle: two plus two equals what? You blurt it out. But drafting a business plan that can increase company profits by 10% requires repeated weighing, overturning, and rebuilding.
The difference is that AI directly converts the cost of this "weighing" into a compute bill. A simple inference might only require one percent of standard computation; but encountering complex programming debugging or multi-step mathematical derivation, the computation volume can 暴涨 (skyrocket) by over a hundred times, stretching the time from a few seconds to minutes or even hours.
Despite this, whether AI is truly "thinking" like humans, no one can currently give a definite answer. But the "Unfaithful Reasoning" experiments have clearly told us: the derivation process displayed by reasoning models on the screen could be real derivation, random generation, or reverse-engineered to fit the answer.
In high-risk scenarios such as autonomous driving, medical diagnosis, and legal judgments, if we treat a long, smooth chain of thought as proof that the AI has thought things through, the consequences could be catastrophic. Admitting that our understanding of this technology is still limited is the premise for correctly using AI.
We are recruiting partners
📮 Resume Submission Email
hr@ifanr.com
✉️ Email Subject Line
"Name + Position Name" (Please attach projects/works or relevant links with your resume)
Click here for more job information 🔗