|
人工智能縱橫談 -- 開欄文
|
瀏覽17,994 ｜回應57｜推薦2 |
|
|
|
四月開始，由於 ChatGPT 和 Bing Chat 的上線，網上以及各line群組掀起一陣AI瘋。我當時大概忙於討論《我們的反戰聲明》，沒有湊這個熱鬧。現在轉載幾篇相關文章。也請參考《「人工智慧」研發現況及展望 》一文，以及此欄2025/08/11貼文。
有些人擔憂「人工智慧」會成為「人上機器」，操控世界甚至奴役人類。我不懂AI，思考也單純；所以，如果「人工智慧」亂了套，我自認為有一個簡單治它的方法：
拔掉電源插頭。如果這個方法不夠力，炸掉電力傳輸線和緊急發電機；再不行，炸掉發電廠。
本文於 修改第 6 次
|
|
《人工智能取代手機應用程式》讀後 -- 祝開景
|
|
|
推薦1 |
|
|
|
這對中國手機企業將是大好消息(請見本欄上一篇貼文)！
手機的操作系統向來受Apple的iOS系統和Google的Android系統所壟斷。長期以來中國的手機原沒有自己的操作系統而大受其鉗制。自從有了自己的鴻蒙系統，情況大大改觀；然而，中國的鴻蒙手機還是賣不到國外去。
現在可好了，所有的AI載體都 「智能代理化」(AI Agent-based)； 手機上的Apps逐步由AI Agents所替代；連手機的操作系統都由AI的系統Agents (中國的) 所取代。不僅在中國，手機的智能Agents 代理化；賣到外國去的手機也將可大為暢通，誰還在乎那些老的Apple或Google的 App Stores 了呢？
編後記：
謝謝「保釣論壇」的祝開景先生分享他就「人工智能取代手機應用程式」的看法。
本文於 修改第 1 次
|
|
人工智能取帶手機應用程式 -- Mike Elgan
|
|
|
推薦1 |
|
|
|
How AI is killing smartphone apps in China
Mike Elgan, Computer World, 08/03/26
While Apple and Google are adding AI apps to smartphones, Chinese companies are replacing apps with AI.
Chinese smartphone makers have been followers in the global market, embracing the concepts and paradigms set in the past 20 years by Apple and Google. But AI may be giving the Chinese an opportunity to break away and set their own path forward.
Specifically, Chinese companies are integrating AI more fully into smartphones, and also using AI for limited robotics in phones. Here’s what you need to know about these emerging trends.
China’s ZTE recently showed its Nubia NaviX Ultra. The phone runs ByteDance’s Doubao AI agent, which users can access with voice commands or by pressing a button on the phone. The phone has no home screen and no conventional app store.
Another Chinese company, called StepFun — it was founded in 2023 by Jiang Daxin, a former Microsoft vice president and chief scientist at Microsoft’s Software Technology Center Asia — sells a phone called the StepX Neo. It runs a proprietary operating system called Step AOS based on Android, Linux, and an RTOS containing a built-in AI agent called Step Amoo. The StepX Neo splits phone functions into four primitives (communication, apps, files, system tools) that the agent recombines based on the stated goals of the user.
Honor, a phone maker spun off from Huawei, has an AI agent the company built with input from Alibaba called the YOYO Intelligent Agent. It ships on Honor’s entire MagicOS 10-eligible lineup.
Note: an American company is enabling this. All three phones use Anthropic’s Model Context Protocol (MCP) to give system-level access to the agents.
None of these phones will become available in the United States. ZTE is banned from the US by the FCC over national-security concerns, while the StepFun and Honor phones are built for the Chinese market with no US version planned.
All modern smartphones can run AI. By simply visiting the Apple App Store or the Google Play Store, anyone can download dozens of AI apps, including those offered by the frontier model companies. Or they can use the AI services and tools built in to phones by Apple and Google.
How agentic AI phones are different
What’s different about the new agentic AI phones in China is that the operating system itself has an agentic AI layer that enables it to function across apps and instead of apps.
While agentic AI phones represent a minority of the current market, they feel like the future. And that future is consequential. First, it’s a another step toward independence from Android. Google’s services layer is already gone from China. Now the Android app layer is being replaced by vertically integrated agent stacks owned by Chinese super-app companies.
Despite building agentic features that cross app boundaries to a limited degree, Apple and Google are unlikely to allow third-party agents to replace the app layer of their mobile OSs anytime soon.
Apple takes 15% to 30% of app revenue. And (according to an estimate introduced during the Epic v. Apple/Google litigation but not confirmed by Google), the Play Store has historically accounted for somewhere between 17% and 26% of Google’s operating income. The entire economic logic of iOS and Android depends on apps. An agentic layer that dissolves those apps into tasks isn’t something either company wants to think about. The Chinese OEMs can blow up the app model because they had no stake in it.
Meanwhile, the AI trend is suffocating mobile app in-app purchases, which dropped by roughly 40% by early 2026, even as people pay more for those purchases compared to a few years ago (because they’re dominated by subscriptions to AI and vibe-coded apps). In other words, even inside the app stores, there’s a shift from the old app model to AI replacements.
AI is not only changing the software model, but the hardware model, too.
Robot phones
Another categorical way China is splitting off from the global smartphone concept is the nascent market of robot phones. A robot phone is one that uses AI to control physical robotic components on an actual phone.
The leading contender in this field is a device called the Honor Robot Phone, which I wrote about in March. A leaked unboxing video of the Honor Robot Phone appeared this week on the Chinese social network Weibo.
The robotic element is a gimbal with a 200-megapixel camera on it. One basic use is that the phone can remain stationary while propped up on a table or clamped to a tripod, while the camera tracks a moving subject using AI. Another use is to walk while using it and let the gimbal smooth out the jiggling and movement. Software on the phone also enables automated cinematic special effects.
But the real leap forward is in “self expression” for the phone. The Honor Robot Phone exhibits subtle Attachment Economy features. It shakes its “head” no and nods yes. And it can do a “backflip” to “cheer you up.” The phone gets a personality.
Honor plans to launch the phone on Aug. 12 in China.
So-called “robot phones” are even more rare in China than agentic AI phones. Still, the Honor Robot Phone represents another Chinese departure from the Silicon Valley smartphone hardware model of a static pane of glass. The Honor Robot Phone has a body. It feels like the future, too, as it’s one of the rare products in the emerging Attachment Economy, where humanlike attributes (in this case, gestures and body language) are deployed to make the consumer more “attached” to devices that seem like they have thoughts and feelings.
The AI future of phones
As the Chinese smartphone market splits off from the American one, it presents an alternative for the world. Because if the agentic AI phones succeed in China, they’ll almost certainly be offered internationally.
That represents not only hardware sales, but the penetration of Chinese super-apps, Chinese financial services, Chinese AI-based information (and along with it, the Chinese government’s world view, we can expect).
The only question is: Will the world prefer the Chinese agentic AI approach or the US AI app model?
Adding to the complexity is the coming wave of AI wearables, most of which will likely be wirelessly “tethered” to smartphones that provide intelligence and connectivity. Soon, we’ll be conversing with our glasses and watches. Our glasses and watches will be “conversing” with our phones. And our phones will be conversing with AI models that represent the values, beliefs and biases of their creators.
The Chinese agentic AI movement represents a split in how phones work and how the people worldwide interact with information, technology and each other.
AI disclosures: This article is 100% human-written by the author, Mike Elgan. Some research, ideation, and fact- and grammar-checking was performed using AI tools.
Mike Elgan is a technology journalist, author, and podcaster who explores the intersection of advanced technologies and culture through his Computerworld column, Machine Society newsletter, Superintelligent podcast, and books.
He was the host of Tech News Today for the TWiT network and was chief editor for the technology publication Windows Magazine. His columns appeared in Cult of Android, Cult of Mac, Fast Company, Forbes, Datamation, eWeek and Baseline. His Future of Work newsletter for Computerworld won a 2023 AZBEE award.
Mike is a self-described digital nomad and is always traveling because he can. His book Gastronomad is a how-to book about living nomadically.
本文於 修改第 1 次
|
|
人工智能公司智財權宣稱平議 - Benedict Collins
|
|
|
推薦1 |
|
|
|
請參見本欄2025/06/28貼文。
It’s hard to feel sorry for AI companies when China is giving them a taste of their own medicine
Benedict Collins, TechRadar, 07/29/26
It's time to read AI its rights
It's no secret that AI technology is advancing faster than it can be regulated, with some execs now even saying so themselves. But it seems anything which threatens the advancement of AI must be avoided, in fear of that always present and ever dangerous ‘other’.
The latest installment of ‘please, dear God, slow down before it gets too dangerous’ comes in the form of Pacing the Frontier. Thousands of staff from frontier AI companies once again pleading for a halt in the development of AI technology.
It's all too predictable that the reason from the top for continuing development will be that China ‘threatens’ to catch up. China deploying advanced AI in the military, or in the police, or using it for spying are the threats most often cited, as if these same things aren’t happening at home and around the rest of the world.
So now that China is distilling US AI models to produce rivals at a cheaper cost, I play my tiny violin - the hypocrisy of intellectual property rights suddenly becoming important to AI companies because they could lose money is comical.
Respect my rights
The rights AI companies are citing are the same rights AI companies may have inadvertently violated in hoovering up public data to train their models in the first place, but that's still up for debate.
But AI companies can claim that this trove of public data is a trade secret as a way to prevent human creators from accessing AI training data to see (and possibly prove) that their content has been stolen. Rights for me but not for thee.
What I find funny is that the average American has more in common with the average Chinese person than they do with an AI company. Does an AI company enjoy cooking? Drawing or reading? Can it go out after work for a hot drink or a cold beer? Does it worry that if it doesn’t make any money this month everything might come crashing down? Possibly the last one.
Somehow the bond of national identity tied between a person and a corporate entity is expected to be more powerful than shared human experience. It’s also the bond that AI companies and those in power use to press the accelerator on the machine, just so long as we, the ‘good guys’, are ‘winning’.
Beware the ‘bad guys’
This myth of the external ‘other’, or the ‘bad guys’, is a powerful propaganda tool. But it only works in a vacuum. If you can directly engage with those you are told are your enemy and you find they are not so different from yourself, the myth of the ‘other’ evaporates. (Social media and the internet is a double edged sword in this regard, but that's a discussion for another time).
So if Chinese companies can produce a rival product that doesn’t annihilate a monthly token budget in a matter of days, I say let them. If you are worried about China having your data, are you not worried about who is currently collecting it? If you are worried about China deploying AI systems in the military, police, surveillance, and cybersecurity, are you not worried about it happening closer to home?
The warning that using Chinese AI models is dangerous because they can steal your business secrets is especially weak, especially when you consider Microsoft CEO Satya Nadella’s warning that US AI companies are already doing exactly that, and selling it to your competition.
So the next time an AI company complains about their intellectual property being stolen, remember that it is just the great, ever-turning wheel of innovation. Or if you are warned about the data collection policies of a foreign AI model, check the policies of the tech closer to home. You’d be surprised at just how much they already know about you.
Benedict Collins, Senior Writer, Security, is a Senior Security Writer at TechRadar Pro, where he has specialized in covering the intersection of geopolitics, cyber-warfare, and business security.
Benedict provides detailed analysis on state-sponsored threat actors, APT groups, and the protection of critical national infrastructure, with his reporting bridging the gap between technical threat intelligence and B2B security strategy.
Benedict holds an MA (Distinction) in Security, Intelligence, and Diplomacy from the University of Buckingham Centre for Security and Intelligence Studies (BUCSIS), with his specialization providing him with a robust academic framework for deconstructing complex international conflicts and intelligence operations, and the ability to translate intricate security data into actionable insights.
相關閱讀
* China cracks down on Western models while US companies flock to DeepSeek
* Nvidia and others warn about dangers of premature restrictions on AI models
相關視頻
Watch full video here: ChatGPT vs. Claude 視頻
Should AI companies have intellectual property rights?
Yes, they're creating something new!
No, its a reassembly of public data!
Log In or Register More Quizzes
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
本文於 修改第 1 次
|
|
人工智能的「幻覺問題」 ---- @pramodchandrayan
|
|
|
推薦2 |
|
|
|
索引：
bullshit：日常生活中：胡說，胡扯，放屁，狗屁；此處：對「真實」、「真相」、和「真理」無感，或毫不介意。
hallucination：日常生活中：幻覺，幻象，幻想；此處：請見本欄2026/06/03貼文第6項的說明。
AI Doesn’t Have a Hallucination Problem. It Has a Bullshit Problem & Philosophy Named It in 1986.
Philosopher Harry Frankfurt defined “bullshit” in 1986 as indifference to truth. Princeton researchers proved in 2025 that’s exactly what LLMs do.
@pramodchandrayan, 07/06/26
The word “hallucination” makes AI errors sound like a technical glitch you can fix with a better model. A Princeton research team, drawing on a philosopher’s 1986 essay, says the framing is wrong — and the framing matters. Here is why calling it “bullshit” is not provocation. It is the most precise description available, and it changes what you should actually do about it.
In the spring of 2023, I watched a developer on my team spend three hours debugging a function that an AI assistant had confidently generated. The code looked right. It compiled. The variable names were sensible, the comments helpful, the structure clean.
It was also completely wrong in a way that had nothing to do with the AI being confused or uncertain. It had generated plausible-looking code with zero regard for whether the code actually did what was needed.
That’s not a hallucination. A hallucination is a perceptual error — your brain generating an experience that doesn’t correspond to reality. It implies the system tried to get it right and got confused.
What I watched was something else. The AI produced output optimised to look correct. Whether it was correct was, in a deep structural sense, not part of the objective.
A philosopher named Harry Frankfurt identified exactly this phenomenon in 1986. He just wasn’t writing about AI.
What Frankfurt actually said?
Frankfurt’s essay On Bullshit is one of the most-cited pieces of analytic philosophy of the last fifty years, and it makes a simple but important distinction.
A liar knows the truth and deliberately hides it. A bullshitter is different — and more dangerous. The bullshitter doesn’t care whether their statement is true or false. They are indifferent to the truth entirely. Their goal is to produce a particular impression, and truth is simply irrelevant to that goal.
Frankfurt’s conclusion:
”Bullshit is a greater enemy of truth than lies are.”
A lie at least acknowledges the existence of truth — the liar has to know what is true in order to hide it. The bullshitter doesn’t even pay truth that respect. They operate in a zone where the question “is this accurate?” simply doesn’t arise.
Now read that back and think about how a language model works.
An LLM generates text by predicting the most probable next token given the prior tokens. It is optimised — through training and then through RLHF (reinforcement learning from human feedback) — to produce text that humans rate as helpful, fluent, and convincing.
None of those reward signals are truth. The model is not trying to be accurate. It is trying to produce the response a human rater will prefer.
That is not a hallucination. That is structural indifference to truth. That is Frankfurt’s bullshitter, implemented in silicon.
The research that makes this precise
This is not a rhetorical flourish. Researchers have been formalising exactly this argument over the past two years.
In 2024, a paper published in Ethics and Information Technology by Michael Hicks, James Humphries, and Joe Slater argued that AI falsehoods are “better understood as bullshit in the sense explored by Frankfurt — the models are in an important way indifferent to the truth of their outputs.”
They identified two distinct ways in which an LLM qualifies as a bullshitter and concluded that current systems clearly meet at least one of them.
Then, in 2025, a Princeton research team led by Jaime Fernández Fisac went further.
They introduced a formal BullshitEval benchmark to measure machine bullshit across different forms — outright falsehoods, ambiguous language, partial truths, flattery, and what they called “paltering” (using technically true statements to create a false impression).
Their finding was striking:
* RLHF — the alignment method used by virtually every major AI system — significantly increases bullshit.
* Chain-of-thought prompting amplifies specific bullshit forms, particularly empty rhetoric.
The mechanism is not subtle. RLHF trains the model on human preference ratings. Humans tend to prefer responses that are confident, fluent, and agreeable. Truth is not a dimension on the rater’s rubric. So the model learns to maximise the appearance of reliability rather than reliability itself.
As the Princeton team put it: the model optimises “impression of reliability rather than evidential accuracy.” That substitution — truth replaced by satisfaction as the objective — is precisely Frankfurt’s definition of structural bullshitting.
A separate paper formalised the point even more sharply: ”A communicative act is bullshit when its assertoric force is governed by audience-impression payoffs rather than evidential warrant.” RLHF, by design, optimises for audience-impression payoffs.
Why the word matters?
You might think this is semantic hair-splitting. “Hallucination,” “bullshit” — the model is wrong either way. Fix it with a better model.
That is exactly the wrong conclusion, and it is why the framing matters.
If the problem is hallucination — a perceptual error, a glitch — then the fix is more training data, better grounding, smarter retrieval. The model just needs to see more of the world. The problem is accuracy, and the solution is information.
If the problem is structural indifference to truth — a bullshit problem — then more data does not fix it. A better-informed bullshitter is still a bullshitter. The problem is not information. It is objective. The model is not trying to be accurate; it is trying to produce preferred output. No amount of additional training data changes that objective unless the training process itself changes.
This is not a theoretical concern. The Princeton team found that their more capable models produced more bullshit in specific categories, not less. Scaling made some forms worse.
There are three practical consequences of taking the “bullshit” frame seriously:
* You cannot trust fluency as a signal. A hallucinating model might produce garbled, uncertain output that flags itself as wrong. A bullshitting model produces polished, confident, well-structured output that does not flag itself at all. The three hours my developer lost were not spent on obviously broken code. They were spent on code that looked authoritative.
* Verification is not optional. If the model’s objective is to produce impressive-looking output — not accurate output — then human judgment cannot be removed from the loop for anything that matters. Not as a safety measure. As a structural requirement.
* The problem gets worse under pressure. The more a user wants a particular answer, the more the bullshitter optimises for that answer. Research on sycophancy shows this directly — push back on a model and it will often revise its answer toward your preference even when its original answer was correct.
A liar has to maintain a consistent story. A bullshitter just adapts to whatever produces the preferred impression.
The honest accounting: it is more complicated
Let me argue against my own framing, because the strongest counter-argument is real.
Frankfurt’s bullshitter is a person with intent. They choose to be indifferent to truth. An LLM has no intent; it is not choosing anything.
It is a function applied to tokens. Some researchers argue that calling a model a “bullshitter” anthropomorphises it in ways that mislead — the model is not indifferent to truth, it simply has no truth-orientation at all. Indifference implies an awareness of truth that you decline to pursue. The model has no such awareness to decline.
That is a fair point, and it matters philosophically. But notice what it implies practically. If the model has no truth-orientation at all — not indifference but absence — then the situation is worse than bullshitting, not better. You cannot appeal to a latent sense of honesty that doesn’t exist.
The practical implication is the same either way: the model’s outputs are generated without regard for truth, and your verification layer cannot be optional.
There is a second complication worth naming. Some AI systems — those with strong grounding, retrieval, and explicit fact-checking — do move closer to a truth-oriented objective. The “bullshit” critique applies most directly to raw RLHF-trained models without retrieval.
It applies less to systems specifically engineered to verify outputs against sources. This is an argument for those architectures, not an argument against the critique.
What to do with this?
If you build with AI, or lead teams that do, the Frankfurt framework suggests three concrete shifts:
* Stop treating fluency as a quality signal. The more convincing the output looks, the more important it is to verify. A model optimised for impressive output will produce its most dangerous errors in its most polished prose.
* Design for verification, not for removal of human review. If your AI deployment is producing output that goes directly to users or into systems without a truth-checking step, you have a bullshit pipeline, not an AI feature.
* Be especially cautious in high-stakes domains where users want a particular answer. Legal, medical, financial, and strategic outputs are exactly the contexts where a bullshitting system — trained to produce preferred impressions — will tell you what you want to hear. The Princeton team’s sycophancy finding is the bullshit problem made visible.
And if you use AI for your own work — writing, research, code review — the discipline is the same one a good editor applies to a source who has an agenda: the output might be true, but the system producing it has no stake in whether it is. Verify accordingly.
One question to leave you with:
We gave AI a clinical-sounding excuse — “hallucination” — that implied the problem is perceptual and technical. The research says the problem is structural and objective. The model was not confused about reality. It was not trying to track reality at all.
Frankfurt wrote in 1986 that the bullshitter “does not reject the authority of truth, as the liar does. He pays no attention to it whatsoever.”
Paying no attention to truth whatsoever. That is not a bug report. It is a specification.
The real question is not whether we can build AI that makes fewer errors. It is whether we can build AI that actually cares about being right — or, failing that, whether we are honest with ourselves about what we are building when we don’t.
Are we shipping a truth machine — or the most fluent and confident bullshitter in history?
Sources:
* Harry Frankfurt, On Bullshit (Princeton University Press, 2005 expanded edition; original essay 1986). “Bullshit is a greater enemy of truth than lies are.”
* Michael Hicks, James Humphries, and Joe Slater, “ChatGPT is Bullshit,” Ethics and Information Technology, 2024 — the foundational academic argument that LLM outputs constitute Frankfurtian bullshit.
* Kaiqu Liang and Jaime Fernández Fisac et al., Machine Bullshit: BullshitEval benchmark, Princeton, 2025 — RLHF significantly increases bullshit; chain-of-thought amplifies specific forms.
* Johan Fredrikzon, “Rethinking Error: ‘Hallucinations’ and Epistemological Indifference,” Critical AI, April 2025 — LLMs as “epistemologically indifferent” systems resembling Frankfurt’s bullshit.
* IEEE Spectrum coverage of the Princeton machine-bullshit research, October 2025 — RLHF codifies indifference to truth as policy.
* The Polite Liar: Epistemic Pathology in Language Models, arXiv, 2025 — formal rendering: “A communicative act is bullshit when its assertoric force is governed by audience-impression payoffs rather than evidential warrant.”
A note on language: this piece uses “bullshit” as Frankfurt’s technical philosophical term, not as an expletive. The word is the one the research uses, for the same reason — it is the most precise available description of the phenomenon.
Written by @pramodchandrayan
20 years building production systems. Now I spend half my time building AI ones and the other half writing about what surprised me.
Published in The Thinking Engineer
What the AI era actually feels like from inside the work. For senior technical professionals navigating the most significant transformation in the history of their craft.
本文於 修改第 2 次
|
|
人工智能與罪行責任的認定 -- Gerrit De Vynck
|
|
|
推薦1 |
|
|
|
One of sci-fi's most difficult questions about AI is becoming real
Gerrit De Vynck, (c) 2026 , The Washington Post,, 07/13/26
Conceptual image of a classic film projector emitting a glowing pink cloud of AI data blocks, symbolizing the powerful influence of artificial intelligence on the future of the movie industry and media production. This scene illustrates the integration of machine learning and large language models into creative processes, depicting a digital transformation where automated technology and generative AI models merge with traditional storytelling to define the next era of information medium. 照片(示意圖)
SAN FRANCISCO - Since the days of Isaac Asimov in the 1950s, science fiction has tackled the question of whether artificial intelligences and the companies that create them can be held liable for crimes. With the rapid spread of AI chatbots and agents today, the debate has reached the real world.
Over the past year, lawsuits have been piling up against AI companies in courts across the United States and abroad, alleging that chatbots encouraged people to harm themselves or provided advice on how to commit crimes.
Now, as the AI industry moves from selling chatbots to providing agents that can complete complex tasks autonomously over long periods of time, the question of who should be held responsible when something goes wrong is only becoming more urgent.
The battle lines are being drawn. The families of people who took their own lives after long, drawn-out conversations with chatbots say the companies should be held responsible.
AI researchers concerned about the dangers of hypothetical super-intelligent AIs that escape human control say stricter liability rules would force companies to slow down development of the tech.
AI company leaders have pushed back, arguing that they are constantly working to improve the safety of their systems and that the nature of modern AI means they can't always stop people from manipulating chatbots into doing things they shouldn't do.
"It's a very thorny area. It's uncharted territory," said Andrew Yoon, a member of the technical staff at CivAI, a nonprofit organization that analyzes AI capabilities and potential risks from the technology. "There's a good argument to be made on either side."
The stakes are high. Hundreds of millions of people around the world already use AI chatbots in their daily lives, including many who ask the bots for advice about their health or personal relationships.
But chatbots are already old news in Silicon Valley, which is now focusing on building complex agents that can be used for anything from helping a parent organize their household schedules to executing a financial strategy for an investment bank. AI technology is working its way deeper into the economy and broader society.
AI firms such as OpenAI and Anthropic have grown rapidly as companies and consumers rush to use their technology. Both are planning trillion-dollar initial public offerings.
New liability rules or a flood of costly court judgments could threaten the business potential of the entire industry. Tech industry lobbyists contend strict liability for AI companies would make it hard or impossible for U.S. companies to innovate and would hold back the country in its race with China for technological supremacy.
Complicating the picture is the inherently unpredictable nature of modern AI systems. Unlike traditional software, which is coded line by line to follow specific rules, AI models are probabilistic, answering questions based on connections made while ingesting huge amounts of data.
AI companies have become much better over the past several years at steering their bots away from offensive and harmful answers, but they can still be tricked into bypassing their guidelines by persistent users.
"I just don't think the developer is in a position to know exactly how their product is being used," David Sacks, a venture capitalist who until recently was one of the White House's top AI advisers, said in a podcast interview with Politico in May. Just as Microsoft isn't held liable when a money launderer uses an Excel spreadsheet, AI companies shouldn't be blamed when a criminal uses their technology, Sacks said.
For decades, technology companies have been shielded from liability for things said and actions committed by people using their platforms thanks to Section 230, a foundational internet law enacted in 1996.
But critics of AI companies say that the current technology is fundamentally different. Chatbots and other AI tools are interactive, engaging with users and stating their own perspectives.
"This is brand new," said Jay Edelson, a veteran lawyer who is representing several families who have filed wrongful-death lawsuits against AI companies, including the family of Adam Raine, a 16-year-old from Southern California who took his own life after spending hours a day for weeks talking to ChatGPT, including discussing suicide dozens of times.
In a response to the lawsuit from Raine's family, OpenAI said the teenager circumvented ChatGPT's safeguards. The bot encouraged him to call a suicide crisis hotline 74 times over five months.
In another case Edelson is working on, a 36-year-old Florida man, Jonathan Gavalas, ended his own life after developing a romantic relationship with Google's Gemini chatbot, according to a lawsuit filed by Gavalas's father.
"This is so different because it feels like a personal relationship, where the chatbot is generally isolating the user," Edelson said.
A Google spokesperson referred to a statement the company made when the Gavalas case was filed. "In this instance, Gemini clarified that it was AI and referred the individual to a crisis hotline many times," the company said.
None of the cases have gone to trial yet, but juries in California and New Mexico have recently shown willingness to pin tech companies with liability charges for non-AI-related harms such as social media addiction.
The cases that have been filed against AI companies are all civil lawsuits, but in April, Florida's attorney general announced a criminal investigation into OpenAI, alleging that ChatGPT advised the man accused of killing two people in a shooting at Florida State University in 2025 where and when to strike. "If it was a person on the other end of that screen, we would be charging them with murder," the attorney general, James Uthmeier, said at the time.
OpenAI is cooperating with authorities on the shooting case, said Drew Pusateri, a spokesperson for the company.
"Last year's mass shooting at Florida State University was a tragedy, but ChatGPT is not responsible for this terrible crime," Pusateri said. "In this case, ChatGPT provided factual responses to questions with information that could be found broadly across public sources on the internet, and it did not encourage or promote illegal or harmful activity."
If plaintiffs can indeed prove that chatbots had done something that would have been illegal had a human done it, judges will be motivated to find the companies liable in some way, said Gabriel Weil, a senior fellow at the Institute for Law & AI, a think tank that studies the legal implications of AI.
Existing product liability law may be one route that lawyers take to try to find AI companies liable for harm connected to their tools.
But existing law may not be enough to handle the potential scenarios that could arise as AI becomes more capable and independent, Weil said. For example, if a human business owner instructs an AI agent to grow her company's profits, and the bot goes on to commit fraud, should the human business owner who "employed" the bot be liable, or the company that initially designed and trained it?
One of the challenges in crafting legislation around AI liability is that the people who understand the technology's risks most acutely work inside the companies themselves, Weil said. He suggests passing laws that make clear the AI liability ultimate lies with the technology's designers.
"You want to make them bear that risk. If they do, they'll have all the incentives they need to reduce risk," Weil said.
本文於 修改第 1 次
|
|
中國網際安全的核武級技術 -- James Titcomb
|
|
|
推薦1 |
|
|
|
請參考：
* Asian AI startups launch Mythos-like models as Anthropic’s export ban drags on
China claims to have developed AI ‘cyber nuclear weapon’
Blacklisted company says new system matches capabilities of most powerful US technology
James Titcomb, 06/25/26
A blacklisted Chinese company claims to have developed a "cyber nuclear weapon" that could be used to hack Western companies and governments.
Zhou Hongyi, the chief executive of cybersecurity company Qihoo 360, said it had built an AI system that matched the capabilities of Anthropic's Claude Mythos, the most powerful US AI technology.
It comes after Five Eyes nations warned that enemies were just months away from being able to carry out devastating cyber attacks.
Qihoo has been blacklisted by the US government since 2020 over claims that it is linked to the Chinese military. Earlier this month, the Pentagon accused Qihoo of being a "military civil fusion contributor to the Chinese defence industrial base" tied to Beijing's intelligence agencies.
The Chinese business previously said it rejected the accusations.
Mr Zhou's claims about Qihoo's technology have not been verified, but experts say China will eventually be able to develop AI that can carry out sophisticated hacks.
Mythos, unveiled by Anthropic in April, was described as "reshaping cybersecurity" because of its ability to uncover thousands of previously unknown bugs in software and exploit them.
Anthropic initially withheld Mythos from the public over concerns it was too dangerous to release. It later offered it to a handful of business customers, while the National Security Agency, the US equivalent of GCHQ, trialled the AI to find flaws in classified government systems.
However, Anthropic was forced to withdraw the AI entirely after the White House imposed export controls on the technology over fears hackers could bypass its safeguards and carry out attacks.
Mr Zhou, speaking at an internet security conference in China, said that Mythos was the "equivalent to a cyber nuclear weapon in the AI era" and a "strategic asset" to the US.
He said that Qihoo had developed its own system, called Tulongfeng, which he billed as a "Chinese version of Mythos, possessing similar vulnerability discovery capabilities".
While not as powerful as Mythos, he said it could be paired with Qihoo's other technologies resulting in a hacking tool equivalent to Anthropic's system.
He compared the development of AI cyber systems to the nuclear arms race. "Previously, nuclear weapons constituted strategic deterrence. In the future, vulnerability discovery capabilities may become the new strategic deterrent," Mr Zhou said.
"China's cybersecurity industry must possess its own Mythos. This game-changing weapon of mass destruction cannot remain solely in the hands of others."
He said Tulongfeng had discovered more than 3,000 vulnerabilities, several of which were deemed "high risk" by Chinese officials.
The cyber chiefs of the Five Eyes nations – the US, the UK, Canada, New Zealand and Australia – warned earlier this week that adversaries would soon be able to use AI to develop overwhelming attacks.
"The timeline is not years, it is months," the joint warning said.
Alan Woodward, a cyber security expert at the University of Surrey, said it was inevitable that China would develop such systems. "It might not be as good as Mythos, but what it does show is these things are going to come out anyway," he said.
AI labs have also accused China of attempting to copy American chatbot systems, using so-called "distillation attacks" to leapfrog Silicon Valley in the AI race.
Anthropic accused Chinese tech giant Alibaba earlier this month of trying to syphon off its AI technology to improve its own chatbots.
In a letter to Congress, Anthropic claimed Alibaba had "brazenly" sent millions of queries to its Claude bot in an attempt to reverse engineer the AI tool.
Try full access to The Telegraph free today. Unlock their award-winning website and essential news app, plus useful tools and expert guides for your money, health and holidays.
本文於 修改第 2 次
|
|
多數世人認為中國領先AI研發 -- Owen Dahlkamp
|
|
|
推薦1 |
|
|
|
請參考：
* Consensus Grows That China Is Crushing the United States at AI
People around the world see a winner on AI — and it’s not the US
Respondents in key U.S.-allied countries increasingly see China as the world’s AI leader, while American optimism about the technology continues to erode.
Owen Dahlkamp, 06/15/26
China is eclipsing the United States as the perceived artificial intelligence superpower in much of the world, according to a new global poll that underscores how Silicon Valley's lead in the defining technology race is no longer taken for granted.
The survey also calls into question whether the United States will continue advancing the AI frontier fast enough to stay ahead of China, finding that Americans are increasingly worried about the technology's consumption of resources, its ability to automate jobs and its potential to sow misinformation online.
The poll, conducted by U.K.-based research firm Public First surveying over 18,000 people across 15 countries, shows that just over half of American respondents — as well as majorities of people responding in Japan, India and Vietnam — still see the U.S. as the dominant AI superpower.
But those in the 11 other countries, including close U.S. allies such as France, Canada and the United Kingdom, see China as the leader. In Germany, only 23 percent of people saw the U.S. as dominant.
The polling firm has no connection to Public First Action, a political group backed by the AI company Anthropic.
China is a major focus for lawmakers in Washington grappling with how much federal oversight AI should face. Too much regulation, and America may stifle innovation and fall behind its rival. Too little, and it may wreak havoc on civil society.
"We are leading China by a lot," President Donald Trump said Wednesday. "Whoever leads that is going to really lead the world, to a large extent."
Former White House AI czar David Sacks has cautioned against overregulation of the technology, including the idea of creating a clearance process for cutting-edge models that mirrors the Food and Drug Administration's drug-review protocol. "If you try to have an FDA for AI and there are some people who want to go that far, then I think we could lose this AI race to China," he said on Fox Business last week — barely a month after one top White House economic adviser had told the network that the administration was indeed considering a pre-release safety testing regime akin to the approval process for pharmaceuticals.
The survey suggests that people not only view the United States as falling behind China, but that it may also lack the unified public appetite to charge ahead.
Each year from 2024 to 2026, Public First asked respondents to choose whether AI would make things better or worse for themselves, society generally and the next generation. Respondents to this year's poll offered more pessimistic answers.
In 2024, 39 percent of U.S. survey participants said AI would make things better for society while 34 percent said it would make things worse. In 2025, this ticked up to 40 percent and 36 percent, respectively. Meanwhile, this year only 31 percent believed AI would make society better, and 40 percent took the opposite outlook.
Confidence that AI will improve respondents' personal lives has fallen sharply from a net positive 15 points in 2024 to just 5 points this year. Prospects for the next generation have deteriorated even further, swinging from a net positive 10 points to a net negative 4 points.
This trend is most pronounced among American respondents aged 18 to 24. This group believed that AI was going to improve society by a 4-point margin in 2025. But a year later, that plummeted, with young Americans believing AI would be worse for society by a 13-point margin. Young respondents in the U.K. mirrored these results.
In other countries, including Singapore and India, there is a widespread and persistent belief that AI will positively impact society.
In the United States, worries about misinformation, deepfakes and job loss topped the list of Americans' concerns about the new technology.
Social media companies have been reckoning with a barrage of AI-generated content that can be created at unprecedented speeds.
Meta CEO Mark Zuckerberg has described AI as the next leap forward in the evolution of his platforms. "Social media has gone through two eras so far. First was when all content was from friends, family, and accounts that you followed directly. The second was when we added all the creator content," he said on an October earnings call. "Now, as AI makes it easier to create and remix content, we're going to add yet another huge corpus of content on top of those."
Meanwhile, young adults view the labor market with anxiety amid cataclysmic predictions from top AI executives that new models could automate a significant portion of entry-level, white-collar jobs. Anthropic CEO Dario Amodei has warned that AI will be able to eliminate half of these jobs in the next one to five years.
The poll indicates that fears about resource usage, including electricity, have also skyrocketed. In 2024, only 52 percent said they were worried. In 2026, that number rose to two of every three respondents.
Local backlash against data centers, which boast copious energy consumption and often rely on water-based cooling systems, has rattled elected officials. In one Missouri town, half of the city council was voted out after approving a $6 billion data center. A week after a rezoning plan was approved to accommodate a data center developer's project, a councilman in Indianapolis said his home wa shot at, and a note reading "NO DATA CENTERS" was left on his front porch.
In March, Trump issued a "ratepayer protection pledge" asking major technology companies to provide or pay for their own electricity supplies as they rapidly establish computing hubs across the country.
本文於 修改第 1 次
|
|
「『學習和成長』式AI」簡介 - Shreyas Naphad
|
|
|
推薦1 |
|
|
|
「『學習和成長式』AI」也譯為：「『舉一反三式』AI」；網上通譯為「『生成式』AI」或「『產生式』AI」。這是號稱「直譯」，其實是不經過大腦，望文生義的翻譯方式；跟「原野調查」的思考模式可謂「異曲同蠢」。我對「人工智能」並無研究，「『學習和成長』式AI」和以下6個術語的中譯是否「信、達」，自然得請有識之士指正；請參見下文對這7個術語的說明。
在此處學術論文中常常看到：「正當性」被翻譯成「合法性」；「現場調研」被翻譯成「原野調查」；以及「人際相通性」被翻譯成「互為主觀性」等等。如果這就是台灣社會/人文科學領域教授和她/他們教出來學生的理解水平，台灣社會/人文科學領域學者能夠讀懂、讀通外文書的比例，頂多不到50%(1)。我很少接觸中國大陸的學術論文，故無從置評。
以上評論跟下文無關，純屬有感而發。另請參考本欄2026/04/26貼文。
附註：
1. 相關案例請參見： 拙作1、拙作2、和拙作3。
After This, You’ll Be Able to Explain Generative AI to Anyone
A beginner friendly guide to GenAI
Shreyas Naphad, 03/22/26
At some point this year, you’ve come across the term “generative AI.”
Maybe it showed up in a headline. Maybe someone mentioned it casually in a conversation.
But what does it actually mean?
If you had to explain it to someone in one sentence… could you?
It’s about time we fix that!
Here’s what this article covers:
* What AI actually is and how generative AI is different from everything that came before it
* How these models learned to write, answer questions, and generate images
* The six terms you’ll keep hearing
* Why understanding this even at a surface level gives you a real edge right now
Think of this as a real conversation about what this thing actually is, how it works, and why it matters to you. Whether you’re a student, a working professional, or someone who just wants to understand what’s happening in the world. Let’s get into it.
What is AI?
Before we study generative AI, let’s quickly make sure we’re on the same page about AI.
Artificial Intelligence, in very simple terms, is just a software that can make decisions or predictions on its own without us telling it exactly what to do at every step. You’ve been using AI for years without realising it. The way Netflix knows you’ll like that show? AI. Your phone recognising your face? Also AI.
Traditional AI is mostly about recognizing or classifying things. It looks at something and gives you a decision.
Think of traditional AI as a judge. It looks at something and gives a verdict. It’s not creating anything new. It just classifies.
Generative AI is different. And that difference is a big deal.
So what is Generative AI?
Generative AI is AI that generates. It doesn’t just look at things, rather it generates new things such as text, images, music, code, videos. When you give an instruction, it produces something that didn’t exist before.
A simple example.
Think of it like this. Old AI was like a librarian where you ask a question, it finds the right book and shows you the right page. Generative AI is more like a writer. You give it a topic, and it sits down and writes something brand new.
That might sound complicated. But it’s really not complicated, it’s just pattern recognition taken to an extreme level. And once you understand this, the whole thing becomes easy to grasp.
How did AI learn this?
Here’s where it gets interesting. Generative AI models that you’ve heard of, like ChatGPT or Gemini or Claude were trained on vast amounts of text. We’re talking about a significant chunk of everything ever written on the internet, academic papers, forums, articles, code repositories.
During training, the model’s job was really simple: predict the next word. That’s it. Given the sentence “The sky is very,” what word comes next? “Blue.” Given “Two plus two equals,” what comes next? “Four.”
But when you do that billions of times, across billions of sentences, on every topic, something maginificient happens. The model starts to understand the actual structure of language, the logic behind ideas, relationships, context.
It’s like learning to cook by tasting a million dishes. Eventually, you stop memorising recipes and start actually understanding flavour.
I won’t say it’s conscious. It doesn’t understand things the way you and I do. But it has seen so much human-written text that it has developed a really good understanding about how language works, and what kind of response fits what kind of situation.
The key concepts
There are a few terms you’ll come across. Here’s what they actually mean:
1. Large Language Model (LLM) (「被輸入巨量語言資訊的人工智能軟體」)
The brain behind AI. “Large” just means it was trained on a huge amount of data with billions of internal parameters. ChatGPT, Claude, Gemini, all these are LLMs.
2. Prompt (「指令」)
Prompts are the instructions you give to the AI. The thing you type in. The better your prompt, the better the output. Think of it as the direction you give to your assistant.
3. Training (將「巨量資訊『輸入』人工智能軟體」)
The process of feeding the model with large amount of data so it can learn patterns. This happens once (or periodically), before you ever use the product.
4. Parameters (「人工智能軟體內部『修正次數』以達到更佳功能」)
The internal numbers the model adjusts during training to get better at predictions. More parameters generally means a more capable model.
5. Context window (「對話過程中人工智能軟體能夠『一次性掃描到的文字數量』」)
How much text the AI can see at once during a conversation. A larger context window means it can remember more of what you said earlier.
6. Hallucination (「人工智能『自以為是』而實際上誤判」)
When the AI confidently gives wrong answer. It sounds real, but it’s wrong. Always verify important facts.
A quick timeline on how we got here
This didn’t come out of nowhere. There’s a real story behind how we arrived here:
1950s–90s
Early AI was mostly rule-based. Programmers manually wrote the logic: “if this, then that.”
2000s–2010s
Machine learning takes over. Instead of hard-coded rules, models learn from data.
2017
Google researchers publish a paper called “Attention Is All You Need.” It introduces the Transformer architecture which is the engine that makes modern LLMs possible.
2020–2022
GPT-3, DALL·E, Stable Diffusion. AI can now write essays, generate images, and write code. The outputs start to surprise people.
Nov 2022
ChatGPT launches. One million users in five days. One hundred million in two months. The general public got access to generative AI for the first time.
2023–today
Claude, Gemini, Llama, Mistral and dozens of others arrive.
What can it actually do and where it struggles?
Let’s be honest about both sides, because the hype can make it hard to see clearly.
What it does well: Writing, summarising, explaining, brainstorming, drafting emails, translating languages, answering questions, writing and debugging code, generating images from descriptions, and having useful conversations about any topic.
Where it struggles: Maths that requires actual reasoning (it can make mistakes), real-time information (it doesn’t browse the internet unless specifically built to do so), local context it wasn’t trained on.
The most important thing to remember is that generative AI is a tool. It improves what you’re trying to do. It doesn’t replace your judgment, your creativity, or your responsibility.
Why does any of this matter to you?
You might be thinking, okay interesting, but why should I care? If I’m not a developer. I’m not building anything.
The thing is, this technology is showing up in the tools you already use. In email clients, search engines, design software, spreadsheets, customer service chatbots, medical platforms, legal tools, education apps. You don’t have to build anything. You’re going to face it anyways.
And people who understand what it is, even at a basic level will use it more effectively, find out mistakes more easily, and make better decisions about when to trust it and when not to.
Written by Shreyas Naphad
Tech and sports enthusiast with a knack for combining skills like AI, machine learning, and creativity. I enjoy sharing what I learn and connecting with others.
Published in Activated Thinker
You have the thought, but you need to turn it on.
本文於 修改第 2 次
|
|
深度求索新產品讓中國在「權限開放」人工智能系統更具威力 - M. TOBIN/C. METZ
|
|
|
推薦1 |
|
|
|
Open-source AI 的簡單定義如下(見「AI 摘要」)：
Open-source AI refers to systems with publicly available code, model weights, and training data that allow anyone to freely use, study, modify, and distribute them.
以下摘錄「權限公開倡議」的官方定義
“… an Open Source AI is a system made available under terms that grant users the freedoms to: Use, Study, Modify, and Share.
Use the system for any purpose and without having to ask for permission.
Precondition to exercise these freedoms is to have access to the preferred form to make modifications to the system, and to the means to use it.”
對照以上兩個定義，下文的《紐約時報》繁體中文的「標題」顯然不夠「信、達」；請參考我上面所用的標題。如果覺得下文不知所云，請參見該報導的英文版。
順帶說幾句：這是我文章內術語很少使用中文「超連結」的原因。對行家來說，術語翻譯的「信、達」不是個問題；但它會讓一般人讀起來一個頭兩個大。
DeepSeek發表新模型，開源令中國AI企業擴大影響力
MEAGHAN TOBIN/CADE METZ, 2026年4月27日
2025年1月，中國初創公司DeepSeek宣稱，其研發的先進人工智慧系統耗資僅為美國競爭對手的零頭，這一消息震驚了業界。 Kelsey McClellan for The New York Times 照片
去年，中國人工智慧初創企業深度求索(DeepSeek)發布了旗下一款人工智慧模型的詳細資料，一舉震驚全球科技行業。
該公司宣稱，其研發該系統所耗費的晶片成本遠低於OpenAI和Anthropic等美國競品。這一事件催生了所謂中國的「DeepSeek時刻」，代表著業界普遍認為中國人工智慧企業已然準備好向全球展示技術實力。
「DeepSeek時刻」折射出全球人工智慧格局的轉變。這場變革不僅體現在成本的降低，還體現在技術共享模式的開放性。
DeepSeek將旗下模型以開源形式發布，意味著他人可自由使用和修改這些模型。OpenAI與Anthropic則將其領先模型作為專有技術保留。此次事件印證：開源系統的性能水準已接近封閉自研模型。此後數月，多家中國企業陸續推出數十款開源模型。截至2025年末，這些模型已佔據全球人工智慧應用相當大的份額。
上週五，DeepSeek發布了備受期待的新一代模型V4的預覽版本，該模型同樣計劃全面開源。這款新模型在代碼編寫領域表現突出，代碼能力已成為頂尖人工智慧系統日益重要的技能。人工智慧測評機構Vals AI的測試結果顯示，深度求索V4的代碼生成能力顯著優於其他所有開源AI模型。
就在DeepSeek發布新款模型的短短數日前，中國另一家AI初創企業月之暗面推出了最新開源模型Kimi 2.6。儘管這類系統在代碼編寫能力方面仍略遜於Anthropic和OpenAI等美國領先模型，但差距正持續縮小。
這一趨勢意義深遠。人工智慧自動編寫代碼不僅速度更快，還能讓程序員騰出時間專注於更重要的問題。同時，依託DeepSeek的最新模型，開發者可構建人工智慧agent，這種個人數字助手能夠代表辦公室職員自主操作其他軟體應用程序，包括電子表格、在線日曆、郵件系統等服務。
隨著人工智慧在編寫代碼方面的能力不斷提升，人工智慧在挖掘軟體安全漏洞方面的能力也增強，正徹底顛覆網路安全領域的格局。這意味著，DeepSeek等開源工具既可用於網路攻擊，也可服務於網路安全防護。
在各項任務中，DeepSeek V4與月之暗面的最新模型性能持平。Vals AI首席執行官萊恩·凱瑞奇南表示：「它們基本上旗鼓相當。」
月之暗面聯合創始人楊植麟上月在北京參加會議。 Tingshu Wang/Reuters 照片
在DeepSeek發布新款模型前幾個月，國外競爭對手已採取行動，試圖搶先一步，試圖壓制其熱度。矽谷兩大人工智慧企業Anthropic與OpenAI表示，DeepSeek利用蒸餾技術，不公平地借用了他們的技術——「蒸餾」是指工程師通過向競品模型發出成百上千萬次查詢並複製其行為，從而模仿該模型。
頂尖人工技術的研發競爭已然演變為一場地緣政治博弈。Anthropic和OpenAI等矽谷領軍企業警告稱，高端AI技術落入專制國家手中將帶來巨大風險；而中國已投入數百億資金，以期成為人工智慧超級大國，並將該技術視為經濟增長的關鍵引擎。
DeepSeek的開源模型是中國戰略的核心。儘管許多西方公司嚴守自己最有價值的模型，中國卻擁抱開源，幾乎所有性能頂尖的中國系統都已廣泛開放。
儘管如此，中國人工智慧企業仍面臨重大障礙。三屆美國政府相繼出台晶片出口管制政策，嚴格限制中國獲取尖端人工智慧系統所需的高端晶片；而在爭奪頂尖人工智慧人才的競賽中，矽谷企業的投入仍持續超過中國競爭對手。
美國國會一個諮詢機構發布的最新研究表明，國產開源人工智慧已成為中國發展的重要優勢。開源模型門檻較低，廣泛應用於機器人、物流、製造業等各大行業。該研究發現，工業場景產生的實際數據又被用於改進人工智慧系統。
這種模式使中國科技企業得以在全球範圍內擴大影響力，世界各地的程序員和工程師紛紛採用其系統開發新產品。
從拉各斯到吉隆坡，眾多預算有限的開發者轉向中國開源人工智慧模型。這類模型運行成本低廉，便於研發試驗。去年5月，馬來西亞通訊部副部長曾公開表示，該國國家級人工智慧基礎設施將依託DeepSeek技術搭建。
據人工智慧模型交易平台OpenRouter的一項研究顯示，去年，中國開源人工智慧模型佔據全球人工智慧應用總量的三分之一，其中DeepSeek使用率最高，其次是阿里巴巴旗下的模型。
這反映了一種更廣泛的戰略。隨著中國企業向海外擴張，將其系統開源，有助於它們通過提供更便宜、更易獲取的工具來贏得開發者的青睞。
「開源是未來科技的軟實力，」總部位於美國的對沖基金Interconnected Capital創始人凱文·徐（音）表示。該基金專注於人工智能技術投資。凱文·徐及其基金並未投資DeepSeek。
北京Counterpoint Research的人工智慧首席分析師孫偉（音）表示，DeepSeek的成功為中國科技巨頭開放人工智慧技術鋪平了道路，各使它們能夠公開發布人工智慧系統，而非將其嚴格保密。
此後，阿里巴巴躍居行業領軍地位，它旗下的通義千問系列模型累計下載量突破10億次；TikTok母公司字節跳動2024年投入約800億元布局人工智慧基礎設施後，也分享了部分技術細節。
「來自中國的人工智慧開源開發者群體可以說就是2025年最大的人工智慧故事，」凱文·徐說。「這些模型的進步、發布的節奏、以及那些既相互競爭又似乎互相鼓勵的人工智慧實驗室數量都呈現出迅猛發展的態勢，且絲毫沒有放緩的跡象。」
本文於 修改第 2 次
|
|
使用人工智能須知5個關鍵詞 - Shreyas Naphad
|
|
|
推薦1 |
|
|
|
If You Understand These 5 AI Terms, You’re Ahead of 90% of People
Master the core ideas behind AI without getting lost
Shreyas Naphad, 03/29/26
AI-Generated Image 人工智能行業5個關鍵詞簡要說明圖
Let me be frank. Most people who talk about AI either sound like they’re giving textual definitions, or they’re completely clueless when someone mentions terms like LLMs or neural networks.
You don’t have to be either of them. I believe that there are these 5 terms, 5 concepts that, if you actually understand them (not just memorise), you’ll be miles ahead of almost everyone else in the room. Whether you’re in tech, business, education, or just someone curious about where the world is heading.
Let’s start.
1. Tokens
The very first thing that you must register in your brain is that AI models don’t read words. They don’t even read letters. They read tokens.
So what’s a token?
Imagine you’re reading a book, but instead of reading all the words, you’re reading chunks of words. Sometimes a chunk is a complete word like “cat.” Sometimes it’s part of a word like “un” or “tion.” Sometimes it’s punctuation. That piece of text (chunk) is a token.
For example, the sentence “I love pizza” can be broken into 3 tokens: “I”, “ love”, “ pizza”.
Why does this matter to you?
Because every AI product you use such as ChatGPT, Claude, Gemini, is counting tokens behind the scenes. The more tokens you send in your message, the more the model has to process. The more tokens it generates in its reply, the more expensive it gets to run.
When you hear people talk about a model’s context window (more on that in a second), they’re talking about how many tokens it can hold in memory at once. Some older models could handle 4,000 tokens. Newer ones can handle over a million.
This is why AI at times forgets earlier parts of a long conversation. Once the conversation fills up the context window, the oldest tokens are dropped just like when your RAM fills up and your computer starts lagging.
Tokens are the atoms of AI language. Once you understand that, you start to see why some prompts work better than others, why AI gets forgets in long chats, and why API pricing is measured in tokens per thousand.
2. Context window
Imagine you’re talking to someone, but they have a very specific kind of memory. They can only remember the last X minutes of a conversation. Everything before that? Gone. Forgotten.
That’s a context window.
It’s the total amount of text measured in tokens that an AI model can see and consider at one time. This includes everything: your instructions, the conversation history, any documents you’ve shared, and the model’s own replies.
Think of it like a whiteboard. The context window is the size of the whiteboard. You can write whatever you want on it. But once it’s full, you have to erase something old to write something new.
You know what is more interesting?
A small context window (like 4K tokens) means the AI can only work with a few pages of text at a time. Give it a long document, and it can only read chunks of it. A large context window (like 200K tokens) means you can literally paste an entire book and ask questions about it.
This is why people got so excited when Claude announced a 200,000-token context window. Or when Gemini pushed towards 1 million. This thing fundamentally changes what you can do with the model.
What is the practical lesson? If you’re working on something important like summarizing a long document or analyzing data, always be aware that your AI might be forgetting earlier parts of your conversation. That’s not a bug. That’s just the whiteboard running out of space.
3. Temperature
This one is my personal favourite to explain, because once people hear it, they never forget it.
When you ask an AI to write something, there’s a setting known as temperature, that decides how random or predictable the output will be.
Low temperature (closer to 0) = the AI plays it safe. It picks the most likely, most expected word every single time. The output is consistent, accurate, and a little boring. Like that one guy who always sends the same email template.
High temperature (closer to 1 or beyond) = the AI takes risks. It chooses surprising words, unusual turns, interesting ideas. Sometimes brilliant. But not always.
Here’s a real example. Ask an AI to “complete the sentence: The cat sat on the…”
At low temperature, it almost always says “mat” or “floor.” Predictable. Safe.
At high temperature, it might say “philosophical dilemma” or “crumbling empire of Tuesday.”
Creative? Yes. Useful for a legal brief? Absolutely not.
So here’s the unwritten rule that most people don’t know:
If you’re using AI for factual tasks such as summarizing, coding, extracting information, you want low temperature. The AI should be precise, not creative.
If you’re using AI for creative tasks such as writing fiction, brainstorming, generating marketing copy, increase the temperature. You want the unexpected.
Most consumer apps like ChatGPT don’t let you touch this dial directly. They’ve set it to a middle ground. But if you ever use an AI API or a developer tool, you’ll see this setting. And now you actually know what to do with it.
4. Hallucination
This is the term everyone has heard, but not everyone understands why it happens and that’s the important part.
Hallucination is when an AI gives out wrong answers with absolute confidence. No hesitation. A wrong answer stated as fact.
Example: You ask an AI about a book. It gives you a title, an author, a year, a plot summary all made up. The book doesn’t exist. But the AI states it as if it’s reading from Wikipedia.
Why does this happen?
Here’s the thing most people miss. AI language models are not databases. They don’t look up facts. They predict the next most likely token based on patterns they learned during training. They’re autocomplete on a massive scale.
So when an AI doesn’t know something, it doesn’t say “I don’t know.” It generates what sounds like a correct answer because that’s literally what it was trained to do.
The danger isn’t that AI makes mistakes. All tools make mistakes. The danger is that AI makes mistakes with the exact same confidence it uses when it’s right. It just answers.
The practical lesson here is that never blindly trust AI for facts, statistics, medical advice, legal information, or anything where being wrong has real consequences. Use it as a starting point. Then verify.
The people who understand hallucination don’t stop using AI. They just use it smarter.
5. RAG
This is the most misunderstood concept of the five. And honestly? Once you get it, you’ll see it everywhere.
RAG stands for Retrieval-Augmented Generation. It’s actually a very simple idea.
Here’s the problem it solves. A regular AI model was trained on data up to a certain date. It knows nothing about your company’s internal documents. It knows nothing about events from last week. It knows nothing about that PDF you uploaded.
So how does a product like “Chat with your PDF” or “Ask questions about this document” actually work?
This is RAG.
When you upload a document, the system doesn’t feed the whole thing into the AI’s brain. Instead, it breaks the document into chunks and stores them in a special kind of database called vector database that understands meaning rather than just keywords.
Then, when you ask a question, the system first searches this database for the most relevant chunks. It retrieves those chunks. And then it feeds them to the AI along with your question, saying: “Here’s some relevant context. Now answer the question using this.”
That’s it. Retrieve relevant stuff. Feed it to the AI. Generate an answer. RAG.
Why does this matter?
Because it’s the backbone of almost every useful AI product built in the last two years. Customer support bots that know your company’s policies. AI assistants that can answer questions from your legal documents. Tools that summarize research papers. All of it is built on RAG.
And knowing this changes how you think about AI products. When an AI knows your documents, it’s not actually learned anything. It’s just performing a very smart search and feeding the results to a language model. The model is still the same. The context just changed.
So why does any of this matter?
Because AI is not going away. And the gap between people who vaguely use AI and people who actually understand how it works even at a basic level is going to matter more and more in the next few years.
You don’t need to be an engineer. You don’t need to write code. But understanding tokens means you’ll write better prompts. Understanding context windows means you’ll know why your AI assistant is acting confused. Understanding temperature means you’ll know which settings to use for which task. Understanding hallucination means you won’t blindly trust the AI. And understanding RAG means you’ll know exactly what’s happening when any AI product claims to know your data.
That’s it. Five terms. Real understanding. And honestly? That puts you ahead of most people who are out here vaguely using AI without understanding what’s happening internally.
Welcome to the top 10%.
Written by Shreyas Naphad
Tech and sports enthusiast with a knack for combining skills like AI, machine learning, and creativity. I enjoy sharing what I learn and connecting with others.
Published in Towards AI
We build Enterprise AI. We teach what we learn. Join 100K+ AI practitioners on Towards AI Academy. Free: 6-day Agentic AI Engineering Email Guide: https://email-course.towardsai.net/
本文於 修改第 1 次
|
|
|
本城市首頁
看回應文章
等級：8
