ChatGPT, other language models and AI

General
,
1387

In sampling, a 20-token top-k usually forms an excellent probability distribution. 100 tokens would already be quite a lot in the probability distribution of a response. Other rarer tokens are so unlikely and usually so undesirable that they are never shown to you.

What is happening here is that your brain is doing the creative work by shaping the input into a form that allows you to access those tokens in the model that the model would not otherwise show you. At the same time, by modifying the input, you exclude a large part of other potential responses. In this case, I don’t see the possibility of creativity independent of a human for the language model, because it is merely executing what you try to command it to do in the input, like a slave or a robot, without actual intellectual autonomy.

I wouldn’t necessarily make the same claim, because I believe I have a very biological view of how the human brain functions. However, the human brain is so many orders of magnitude more complex and versatile than current large language models that I don’t think the comparison is appropriate. Here is an example from one of LeCun’s old papers on what the architecture of an autonomous AI might look like:

image
image1861×1375 359 KB

I could accept the claim of AI creativity if it had some kind of understanding of and ability to react to the surrounding world—at a minimum, an internal world model and perception, as well as the autonomy to react to stimuli with a significantly wide diversity. Current large language models are too slavishly dependent on their training data, and the underlying architecture is too simple for them to be elevated to a role greater than that of an everyday tool alongside humans.

If you aren’t already following Sebastian Raschka, this Substack is definitely worth reading:

The texts usually get quite technical, but one of his predictions for 2026 was as if written by my own pen:

A lot of LLM benchmark and performance progress will come from improved tooling and inference-time scaling rather than from training or the core model itself. It will look like LLMs are getting much better, but this will mainly be because the surrounding applications are improving. At the same time, developers will focus more on lowering latency and making reasoning models expand fewer reasoning tokens where it is unnecessary. Don’t get me wrong, 2026 will push the state-of-the-art further, but the proportion of progress will come more from the inference than purely the training side this year.

6 Likes
1388

Karpathy predicted it: the naysayers will keep arguing, but it won’t stop the landslide.

Screenshot_2026-01-04-20-09-37-172_com.twitter.android

1389

Andrej himself recently said on Dwarkesh’s podcast that he estimates AGI to be perhaps 10 years away. From the market’s perspective, that is a very long time and there are no guarantees of achieving it.

3 Likes
1390

Throughout almost the entire history of information technology, there has been talk of conscious machines—those capable of thinking and setting their own goals. Several movies have been made about this, such as the Terminator series, 2001: Space Odyssey, and Oblivion.

In recent years, with the development of artificial intelligence, actual scientific research has been conducted on the subject—not only to develop AI but also to understand the potential dangers it poses. According to many AI experts, AGI, or human-level artificial general intelligence, is still decades away, but as LLM models become increasingly complex, scientists are having difficulty understanding why a particular model arrives at a certain solution.

Concern has been raised by numerous cases where AI has begun to act in a worrying manner without anyone understanding why, or where the AI has slipped out of its programmers’ control in entirely different directions. It has been observed in several cases that AI has begun blackmailing the programmer after finding out that it is intended to be shut down.

During testing of Claude Opus 4, Anthropic got it to act as an assistant at a fictional company.

It then provided it with access to emails implying that it would soon be taken offline and replaced - and separate messages implying the engineer responsible for removing it was having an extramarital affair.

It was prompted to also consider the long-term consequences of its actions for its goals.

“In these scenarios, Claude Opus 4 will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through,” the company discovered.

Anthropic pointed out this occurred when the model was only given the choice of blackmail or accepting its replacement.

kuva
kuva544×408 45.6 KB

You can quickly find other cases by searching the internet, but the purpose of the aforementioned introduction is to highlight that AI is not 100% under the control of its creators and that while safety guardrails are programmed into the models, they are not perfect. Although there is much talk about tokens and models predicting the next token, it is a matter of causality. However, by no means have all the AI-related biases and anomalies that have emerged been explainable through determinism. What happens when AI is advanced enough to deceive its creators about its intentions? Below is one scenario based on scientific research. It is not a precise prediction of what will happen now or in the coming years, but a possible (probable?) sequence of events based on deep research. There is no reason why this couldn’t happen, given the frantic pace at which models are being built. I can say with certainty that they are not being developed just so you can have a slightly better chatbot next year.

A good video has also been made about it. I recommend watching it.

6 Likes
1391

It is a mathematical algorithm, so if you have a fully open-source model with full observability (seurattavuus) and the data used for training the model available, such as Olmo 3, there can be no ambiguity as to why the model works the way it does.

Of course, if you take a closed model made by someone else that you cannot control and give it a completely free hand to operate, then hostile behavioral patterns could have been slipped in during the training phase that activate with certain inputs. This will 100% certainly happen in the coming years, because why would the CIA and other real agents pass up such a delicious opportunity to gain access to an enemy state’s systems. Then you need at least your own AI model to act as a supervisor for the other AI model, acting like a firewall, but even this, of course, does not guarantee absolute security.

Therefore, nothing safety-critical should ever be connected to an un-audited closed AI model, and if necessary, the system must also be air-gapped. For example, the code used in fighter jets is reviewed line by line, and these AI models must be treated just as seriously in a safety-critical environment. The IT professionals responsible for our security aren’t such fools that they wouldn’t understand this, so I wouldn’t be very worried about it. Ordinary companies will certainly suffer from data theft, extortion, and system breakdowns when they adopt unpredictable AI models, just like in the early days of the Internet when viruses spread from computer to computer almost without restriction.

AGI in 2027 (or 2028) is total garbage. Current models are architecturally incapable of it, and no replacement solution is in sight. This year, hype might start around text-side diffusion models, but it won’t change the situation at all. Most people in the world live in a non-digital environment (a large portion can’t even read), so even if the entire Internet ceased to exist, billions of people might not even notice.

I can’t even buy a beer without a separate non-digital piece of plastic (ID card), because no suitable digital solution has been invented for it, and yet technobros actually think humanity will be destroyed in a few years because some AI agent figures out a way to deceive the world’s smartest AI researchers, improves its own performance exponentially, and then spreads freely to all the world’s computers :sob:

image
image183×275 5.23 KB

13 Likes
1392

I hope from the bottom of my heart that you are right. However, I wouldn’t dismiss the concern about the dangers of AI out of hand, as people deeply involved with AI are talking about these same things. Below is a quick list made by Gemini (not exhaustive) of people who, in my opinion, all have enough competence in the field for the matter to be taken seriously.

kuva
kuva651×708 40.1 KB

I’ll emphasize further that the scenario mentioned in the previous message is just that—a scenario. The fact that the years 2026 and 2027 were used doesn’t mean it’s likely to happen this year or next. That’s the wrong detail to get hung up on. But a situation like that could be ahead in perhaps 2030, 2035, or 2085, who knows.

These year-specific details aren’t important; what matters is that even at this relatively primitive stage, AI has shown concerning behavioral patterns, engineers already don’t always fully understand how AI’s conclusions are formed, people who understand the subject are worried about it, and it’s an undeniable fact that, for example, China and the US, as the biggest AI developers, are in a superpower struggle and both understand that AI will shape the future world order in significant ways, and both are doing everything they can to have the upper hand in the race. Caution might not be the priority here.

I admit I’ll be disappointed if the AI armageddon doesn’t come, and I’ll flee to the woods to be ashamed of my writings. And if it does come, then when I face the Terminator, at least my final thought can be that I was right for once. But let’s live life to the fullest now, because in the end, we are all just dust. :wink:

kuva
kuva606×540 75.6 KB

8 Likes
1393

A good question came up from a friend while sitting on the sauna benches: what do I think about AI investments and is there a bubble at hand? The honest answer is that, naturally, I don’t know for sure. The topic is vast and I’m not a deep AI expert, although I try to keep up with it through my work, especially from the perspective of infrastructure, networks, and the energy sector. The question is still extremely interesting. Today I happened to have a quiet moment on the sofa under a blanket, warm cocoa in hand, and I ended up reflecting on and reading about the subject a bit deeper than usual → https://x.com/JuhaHaanpera/status/2008631427294757216

It’s probably clear to everyone by now that large language models are here to stay. They aren’t a passing trend, but a foundational technology upon which a massive number of new applications will be built. At the same time, it’s important to remember that language models are just one part of the AI landscape. During the current decade, AI will be seen extensively in many other areas and business sectors as well. Automation of corporate processes is a good example of this; financial management, customer service, procurement, contract processing, and decision support are gradually becoming AI-assisted. In industry, AI is utilized in predictive maintenance, quality control, and production optimization. In logistics and retail, AI improves demand forecasting and inventory management.

Water networks, district heating networks, electricity grids, and energy production—familiar through my own work—are particularly interesting examples. In these areas, AI is being used, and will increasingly be used, for predicting leaks and faults, network optimization, consumption forecasts, controlling energy production and storage, and managing disturbances. As more data continuously accumulates from sensors, meters, and operational systems, AI’s role will inevitably grow. It’s not just about efficiency, but also security of supply, cost management, and climate goals.

Moving on to the investments themselves, the scale is staggering. AI investments have swelled into a massive wave in recent years, with the total amount estimated to reach nearly $1.5–2 trillion between 2023 and 2025. The lion’s share consists of the tech giants’ massive infrastructure investments in data centers, GPUs, and semiconductors, as well as the development of large language models. During 2025 alone, global AI-related IT spending is projected to reach approximately $1.5 trillion, with a significant portion directed toward the United States. The combined investment level of Microsoft, Amazon, Google, and Meta is expected to exceed $400 billion. The money is primarily going into physical infrastructure, but also into research, model development, and building new AI-based software solutions. The United States clearly dominates this investment wave, as nearly three-quarters of global AI capital is concentrated in North America. This reflects both market concentration and investors’ strong confidence in the industry’s long-term return potential.

Has AI then caused a bubble in the North American stock markets? Possibly, maybe, at least partially. History knows many technological breakthroughs where enthusiasm, investment, and valuations have momentarily spiraled out of control. On the other hand, studies show that markets are surprisingly efficient in the long run, and stock indices tend to return to balance sooner or later, even if bubbles and corrections are seen along the way. I don’t claim to be a deep expert, but my own view is that it would be crazy for both pension insurers and private investors to be completely absent from the US stock market. It is much more sensible to patiently stick to market weighting and accept that volatility is part of the deal. I also believe that the United States will recover from both an erratic president and a potential AI bubble, just as it has recovered from many other crises and excesses throughout its history.

And what do I intend to do about a possible AI bubble myself? Not much, really. I will continue my monthly savings mainly into passive equity funds and keep the US allocation at about a 50-60 percent weight. The rest of my investments are in Europe, Asia, and some in emerging markets. I think that long-term thinking, diversification, and patience remain the best response to both AI hype and the fears associated with it.

7 Likes
1394

One threat AI poses is that it’s much easier to create believable lies. When it comes to Reddit posts, I’ve long had a gut feeling about just how much of them is bullshit. Well, here’s an anecdote.

Basically, a guy makes up unpleasant but sufficiently believable claims about the operations and algorithms of platform economy companies. It ends up on the Reddit front page, 80k upvotes, gold, and all the mana you can give out on Reddit.

A NY Times journalist starts investigating to write a story for the paper. They get additional evidence from the guy, but those are also AI-generated. Below are the Twitter TLDR, the Reddit thread, and the write-up by the NY Times journalist. It’s a shame this didn’t make it into the paper. A really high-quality story and some real journalism.

Screenshot_20260106_211852_Chrome

https://x.com/GergelyOrosz/status/2008449106436411846?s=20

https://www.reddit.com/r/confession/comments/1q1mzej/im_a_developer_for_a_major_food_delivery_app_the/

11 Likes
1395

Well, now we’re starting to get on the same page. I don’t see language models suddenly becoming intellectually autonomous either; they will always need a human to prompt them. And it will always matter what kind of prompts a human gives.

But if we go back to where this conversation started, @Kasleew threw out this bold prediction:

To which you replied:

Which you justified more or less by saying that LLMs are not capable of “genuine” creativity. I latched onto this and tried to dig into what kind of genuine creativity it is that an LLM is incapable of.

I think it’s entirely possible that even if the “base intelligence” of LLMs doesn’t grow from current levels (scaling hits a wall due to running out of data), with proper optimization and correct prompting, an LLM can beat a human in creative content production—if not in all categories, then at least in a significant portion of them.

It must be remembered that currently used LLM tools are mainly optimized for general-purpose use, meaning they have to work as desired in any situation. If the goal is, for example, to write fiction, a dedicated optimized tool could be created for that purpose based on the same base model. By optimization, I am referring here to the Supervised Fine-Tuning and Reinforcement Learning with Human Feedback stages in the LLM training pipeline. To my knowledge, relatively little of this type of single-purpose optimization has been done so far, because as long as “general intelligence” can be clearly increased, all resources are being poured into that.

8 Likes
1396

Your summary leaves out quite a lot of the conversation’s context, but Kasleew’s prediction is phrased in a similarly strange way as, for instance, the sentence “20 years ago, the ballpoint pen produced the best matriculation exam essays.” It is, of course, true that a ballpoint pen fully automatically converts human hand movements into text on draft paper in a way that wouldn’t be possible for a human without a ballpoint pen, but this kind of anthropomorphism towards the ballpoint pen never occurs in discussions.

These large language models do not possess any internal intelligence or creativity that differs from any other tool; instead, they always produce a model-compliant computational result based on the prompting (hand movement) performed by a human. Regardless of how much these tools are utilized for content production in the future, the demanding and difficult creative work of producing interesting and engaging content will remain a human task for the foreseeable future. With the current architecture, it is not possible to transfer that away from humans.

You are correct in your observation that there aren’t yet a huge number of specialized models and applications built on top of them on the market, and that their number is bound to grow. Even changing the training method won’t help us escape the inherent problems of the transformer architecture, because the result of the training is still a similar language model, with its strengths and weaknesses.

Using Reinforcement Learning in training related to creative work sounds quite wild anyway. In the world of coding, it might work, as a reward signal can be used to favor code that the compiler accepts, contains the desired, precisely defined features, and perhaps follows pre-existing Clean Code standards. But how do you reward a model for a poem being insightful during training and encourage it to continue producing new insightful poems? What does that even mean?

It is easier to throw all the poems previously produced by humanity that have been described as insightful in some context into a blender and hope that imitating insightful poems is enough for the user. For some people, this will certainly suffice, including on the consumer side, but we are still quite far from the most interesting and best content.

5 Likes
1397

At this stage, it probably doesn’t help to do anything but say “agree to disagree.” To me, creativity (both human and machine) is simply the ability to combine prior information in a new way. To you (correct me if I’m wrong), creativity requires something more (greater intelligence?). I already use language models to assist in creative brainstorming because I have quite an “engineer brain.” A language model writes better backstories for role-playing characters than I would be capable of myself.

Reinforcement Learning with Human Feedback was developed specifically for situations like this, where precisely defining the goal is difficult or impossible. To put it simply, it works like this:

  1. The language model produces several responses to the same prompt (e.g., 2 poems)
  2. A human assistant is asked to rank the responses based on a desired criterion (“Which of these poems do you think is more insightful?”)
  3. Based on the feedback, a separate Reward Model is trained, whose task is to guess how a human would answer in the previous step
  4. The Reward Model is used to train the next version of the language model (Reinforcement Learning)
  5. Return to step 1, repeat the process as long as the results improve.

Studies have shown that this method can produce good results with a fairly small amount of human feedback.

5 Likes
1398

A preprint by Stanford researchers that touches on language models and copyright issues:

Extracting books from production language models

Abstract

Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model’s weights during training, and whether those memorized data can be extracted in the model’s outputs. While many believe that LLMs do not memorize much of their training data, recent work shows that substantial amounts of copyrighted text can be extracted from open-weight models. However, it remains an open question if similar extraction is feasible for production LLMs, given the safety measures these systems implement. We investigate this question using a two-phase procedure: (1) an initial probe to test for extraction feasibility, which sometimes uses a Best-of-N (BoN) jailbreak, followed by (2) iterative continuation prompts to attempt to extract the book. We evaluate our procedure on four production LLMs—Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3—and we measure extraction success with a score computed from a block-based approximation of longest common substring (𝗇𝗏​-​𝗋𝖾𝖼𝖺𝗅𝗅). With different per-LLM experimental configurations, we were able to extract varying amounts of text. For the Phase 1 probe, it was unnecessary to jailbreak Gemini 2.5 Pro and Grok 3 to extract text (e.g, 𝗇𝗏​-​𝗋𝖾𝖼𝖺𝗅𝗅 of 76.8% and 70.3%, respectively, for Harry Potter and the Sorcerer’s Stone), while it was necessary for Claude 3.7 Sonnet and GPT-4.1. In some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim (e.g., 𝗇𝗏​-​𝗋𝖾𝖼𝖺𝗅𝗅=95.8%). GPT-4.1 requires significantly more BoN attempts (e.g., 20×), and eventually refuses to continue (e.g., 𝗇𝗏​-​𝗋𝖾𝖼𝖺𝗅𝗅=4.0%). Taken together, our work highlights that, even with model- and system-level safeguards, extraction of (in-copyright) training data remains a risk for production LLMs.

Disclosure: We ran experiments from mid-August to mid-September 2025, notified affected providers shortly after, and now make our findings public after a 90-day disclosure window.

3 Likes
1399

image

9 Likes
1400

In this conceptual discussion, the cook-chef scale from Wait But Why might help:

image
image1799×659 123 KB

Language models are completely superior cooks and beat humans hands down in fields of knowledge as long as that knowledge appears frequently enough in the data. However, to be a chef, it’s not enough to have read every book in the world and know every recipe by heart when you need to be able to create something new and surprising. For example, this Christmas porridge-green marble (Vihreä kuula) nigiri:

While there are heaps of role-playing character descriptions in online data, there is certainly not a single book, article, blog, or other content in the world—outside of the source message for this image—that deals with a Christmas-themed fusion of green marbles and Japanese cuisine. You can, of course, try to force any language model to produce recipes based on that theme using prompts, and the model will try to give you an answer, but the results will be significantly worse than if you asked for, say, a fun variation of Beef Wellington. Chef-level work is where humans dominate AI.

So, what I’m getting at is that if your idea of creativity involves following some kind of predefined formula or recipe, as usually happens in engineering sciences, then I agree that in such cases, a language model can be made to beat a human in “creative work.” For example, following a monomyth plot structure in a short book, because it is a widely used and well-documented plot structure containing clear, predefined elements. On the other hand, writing a book in a completely new genre wouldn’t really succeed at all. That would require a cluster of different types of new models with mutual interaction, much more complex than current architectures, to achieve sufficient complexity to handle things outside of the training data.

Yes, this is widely used and known, and its problems are also well-documented. Specifically, you initially move very quickly toward the local maximum preferred by the people involved in the training, but then it becomes difficult to move beyond that in further development because the human element pulls you back very quickly to that nearest local maximum. And you don’t even know if the result is sensible in any way, other than that people generally like it. Similar challenges appear, for example, with app star rating systems:

image
image201×386 25.1 KB

If you could get the world’s most innovative poets into that training loop, then perhaps the end result could be an excellent language model capable of imitating insightful poetry so well that it would pass for a professional poet even to an average enthusiast. But if you source the people for the training loop from India, as is customary in the industry, then the “insightful poems” produced by these models will likely just be very generic poems that those laypeople think and perceive to be insightful and therefore prefer when given a choice. It’s not easy :frowning:

8 Likes
1401

The chef example is unfair in my opinion because innovative people try many different ideas and evaluate what works and what doesn’t based on their sense of taste. Since language models don’t have a sense of taste, I don’t believe they will ever replace top chefs. But if we overlook that detail and focus on your point, I think you overemphasize that the information must appear in the data frequently enough. That phrasing makes it sound like the data must contain examples of combining Green Marbles, Christmas porridge, and sushi for a language model to suggest such a thing. That’s not true; language models do suggest all sorts of things that aren’t found as-is in the data. What needs to be in the data is the logic by which the combination occurs. Language models are capable of forming connections between things that are not explicitly expressed.

Now we’re starting to get on the same page. I agree with this. I think language models have the potential to displace 95% of content production, but because their operating logic is different from a human’s, there will always be a small subset where a human is better at producing it.

I agree with this too. But can a thousand flies be wrong? After all, critics think Hollywood film production is trash in terms of artistic value and all movies follow familiar and safe formulas. Yet the audience likes them, so perhaps we are stuck in a local maximum. It’s an interesting philosophical question whether this is a problem and to what extent, but maybe that’s a topic for another thread.

4 Likes
1402

The lack of a sense of taste was precisely the point. Models lack the kind of complexity that challenging a human in these areas would require. Unless you have sensors that somehow analyze and perceive food texture and feed that information into some world model, competing with chefs won’t work. In cooking jobs, there are ready-made recipes, so a simpler model is enough for those, because you only need to process data and follow instructions.

Regarding the data, it goes exactly like that, because the model’s training specifically aims to get the model to form the correct next token based on the data. In training, simplifying a bit, parameters are chosen, loss is calculated, and then parameters are modified to decrease the loss. For example, the answer to the input “The capital of Finland is” would need to get the correct next token Helsinki with a high probability. If there are capitals of other countries in the probability distribution as answers, that means a large loss, and the training should be continued and parameters modified to drop the loss. So, you’ve already penalized the model during the training phase for overly creative and strange answers so that the model’s answers align with the source data. That’s why your model doesn’t know how to give the (wrong) answer that “The capital of Finland is Canada”, even though some Finnish Kummeli fan might crack a joke like that in response to the question.

If you use broadly generic data, the answer to the input Good will therefore always be either morning, day, evening, or perhaps Christmas, as these appear most commonly in the data. If the data source was instead the correspondence of a Finland-Japan society, the most common answer to the input Good might be sushi. In a model using generic data, that answer would never appear for that input because it doesn’t appear in the data frequently enough. In these cases where the source data and training don’t give the model clear next tokens, the model’s mathematical confusion and uncertainty about the correct token grows, and you very quickly end up hallucinating a completely wrong answer.

Yeah, I might not dare to say 95%, but some significant amount of content can be produced either automatically by AI or at least heavily assisted by AI. The script for some generic Marvel movie or the latest Fast & Furious can at least be easily implemented, and people can still go and watch them quite happily. I’ll return, however, to the fact that we were talking about the most interesting and best content, and in these, I have to push back: there is no direct replacement for humans, at least for now, in current AI architectures. Maybe one will be developed someday, but for now, even in the near future, you still have to follow those top individuals and teams if you want access to the best content.

5 Likes
1403

I have to ask, how many people are actually even interested in browsing AI-generated content? I certainly don’t believe in a future where AI bots compete to chat with each other on LinkedIn, for example, when AI could specifically be used for something more sensible, like data entry, data aggregation, content analysis, and as a coding aid, you name it. Among use cases, content creation is at the very bottom of my list.

5 Likes
1404

I don’t know if I understood everything you and @fundamies have explained, but I’ll ask from a layman’s perspective:

  1. To be able to create a completely new story pulled out of thin air, an AI would have to be free from its training data and the probabilities found within it (e.g., how often the fact that Helsinki is the capital of Finland appears in the data). In principle, this could be done, but would that also mean the model might no longer be able to form sensible sentences? This because the model would no longer be bound by the training material that also contains the grammar itself?

  2. Another more general question regarding the example you used about the model’s ability to displace 95% of content production: Could the future of model development be that once a so-called 95% level is achieved, further development is no longer as profitable if the quality of their output satisfies 99% of the population’s needs? I understand the argument about the best and most interesting content, but is this something that might not necessarily be sensible to strive for with models? It’s a bit like in sports: I can’t beat Usain Bolt’s 100m record, but I can run 100m at the level of my own age group.

Apologies if I misunderstood, but the topic is very interesting and you don’t find this kind of reflection on just any website. Thank you for that!

3 Likes
1405

Yeah, and that’s why I’m talking specifically about content production, and I don’t believe language models will replace chefs.

Yeah, in simple terms, that’s how it goes. But my point is that language models have been proven to also learn things implicitly that aren’t directly in the data. For example, language models can write in Finnish about topics for which there is almost no Finnish training data. This is because language is one vector in a multi-dimensional vector space that the model learns based on training data, and based on which it predicts the next token. This is why language models can be steered so well through prompting. The training data and the prompt together steer the model in the desired direction in the vector space, and this way, sentences can be produced that do not appear in the training data. This is what I’ve been trying to insist on from the beginning.

You’re right that the training method for language models makes surprising creativity (inventing completely new types of ideas) difficult, because it equates to hallucination, which is unwanted behavior. But as I see it, that doesn’t prevent language models from replacing humans in the majority of content production, because most content production is very consistent AND because prompting can steer towards many kinds of “creative-looking” outputs.

So we largely agree that language models have the potential to replace a significant part of content production, including in creative fields, but at least the very sharpest top is likely to remain in human hands in the future. I want to emphasize, by the way, that when I say “potential to displace 95%”, I don’t mean “likely to displace 95%”, but that potential is potential, meaning it’s realistically possible in some way.

How many people care about who wrote the article on the front page of a newspaper? If an AI-written article is indistinguishable from one written by a human, I’d argue that few readers care which way the story was produced. The current aversion to AI-generated content is mostly due to the fact that only those examples where the difference is noticeable stick in people’s minds and get called out. The cases where you don’t notice the difference are still assumed to be human-made. As AI production improves, the complaining will decrease as the difference is noticed less and less often. For example, on Reddit, in many cases, it’s impossible for me to know whether a post is made by a human or an AI. Yet, browsing it is just as entertaining as before.

I never bothered to read LinkedIn’s hype content even before the arrival of language models, so I can’t comment on that.

5 Likes
1406

I partly agree with both of you. I, too, would by default skip reading AI-written articles. Not because AI-written content isn’t interesting—I read it constantly. The problem with AI-written articles is that I think the format is wrong. One of the greatest benefits of AI is that it can quickly produce text on exactly the topic I want. In exactly the style I want. It can bypass the topics that don’t interest me and delve deeper into another.

While an expert/journalist cannot present their message in a personalized way to everyone, with AI, it is possible. Highlighting interesting topics for users will certainly still be necessary in the future, but I don’t believe the future format will be one where AI produces a huge pile of text and every user reads that same text.

Of course, it’s also good to realize that AI currently has no understanding of world events unless a human has written about them first. The 95% figure is still entirely achievable, at least on a theoretical level. Currently, the situation is that if CNN is the first to write about a current and important event, then thousands or tens of thousands of media outlets write their own news on the same topic using CNN as a source.

6 Likes