[00:00] Trillions of dollars are being poured into&nbsp; AI chips, models, and gigantic data centers.&nbsp;&nbsp; And every few months, some new model arrives,&nbsp; promising either superior reasoning or improved&nbsp;&nbsp; coding. But if AI actually is getting smarter and&nbsp; more capable, then why are so many complaining&nbsp;&nbsp;
[00:17] about AI getting worse? Users are reporting&nbsp; weirder mistakes, answers becoming more cautious,&nbsp;&nbsp; subscriptions getting confusing, and limits that&nbsp; are more restrictive with less predictable tools.&nbsp;&nbsp;
[00:30] So, what the hell is going on here? Well,&nbsp; in this video, we'll explain how AI can keep&nbsp;&nbsp; getting smarter while the experience surrounding&nbsp; it gets noticeably worse, and why the technology&nbsp;&nbsp;
[00:42] we're all becoming increasingly dependent upon is&nbsp; becoming less and less ours. My name is Lewis and&nbsp;&nbsp; you're watching the Coin Bureau. To understand&nbsp; why AI feels different today, we need to go&nbsp;&nbsp;
[00:55] back to the beginning. Not the beginning of time,&nbsp; just when CHACBT first made a public appearance.&nbsp;&nbsp; It was free. It was also surprisingly powerful&nbsp; and it was presented as an experimental preview,&nbsp;&nbsp;
[01:07] indicating that it had much more potential. The&nbsp; aim at that point was only to get public feedback&nbsp;&nbsp; on this novel experience. The priority didn't seem&nbsp; to be squeezing out dollars from users at that&nbsp;&nbsp;
[01:19] point, but it was still a typical marketing trap.&nbsp; Lo and behold, the paid subscriptions finally&nbsp;&nbsp; arrived, and the offer remained remarkably simple.&nbsp; $20 bought faster responses, priority access,&nbsp;&nbsp;
[01:32] and a front row seat to the latest technology.&nbsp; And that generous opening worked. People began&nbsp;&nbsp; using AI to write emails, summarize documents,&nbsp; study, brainstorm, analyze data, and produce code.&nbsp;&nbsp;
[01:45] And this goes on and on. What started as an&nbsp; entertaining chatbot gradually became part&nbsp;&nbsp; of how millions of people completed ordinary&nbsp; work. Predictably, once a tool saves you time,&nbsp;&nbsp; you start building routines around it.&nbsp; Templates are created, teams are trained,&nbsp;&nbsp;
[02:00] and entire workflows start to revolve around it.&nbsp; Removing it then becomes much harder than cancing&nbsp;&nbsp; some streaming service for instance. But providing&nbsp; advanced AI to hundreds of millions of users also&nbsp;&nbsp;
[02:12] requires vast amounts of chips, electricity, and&nbsp; data center capacity. Eventually, the companies&nbsp;&nbsp; funding that bill need a little bit more than&nbsp; just enthusiastic users. They need subscriptions,&nbsp;&nbsp;
[02:24] enterprise contracts, basically enough revenue&nbsp; to justify this infrastructure spending on an&nbsp;&nbsp; extraordinary scale. And so the honeymoon phase&nbsp; began ending. The original challenge was making&nbsp;&nbsp;
[02:36] AI useful enough that people would adopt it.&nbsp; The new challenge was working out how much those&nbsp;&nbsp; increasingly dependent users and businesses can be&nbsp; persuaded to pay. This is the familiar life cycle&nbsp;&nbsp; of any good digital platform. Provide a solution,&nbsp; subsidize convenience, establish the habit,&nbsp;&nbsp;
[02:54] and then monetize the dependency. The technology&nbsp; that first asked, "How can I help?" is therefore&nbsp;&nbsp; facing a different question from its owners. How&nbsp; much will people tolerate? And as we'll see next,&nbsp;&nbsp;
[03:07] companies do not always need to raise the price&nbsp; to change the deal. Sometimes they could change&nbsp;&nbsp; the product itself. Now, imagine opening&nbsp; the same app, entering the same prompt,&nbsp;&nbsp; and receiving a completely different result. Like&nbsp; maybe the wording is weaker, the code breaks, or&nbsp;&nbsp;
[03:23] the model suddenly ignores an instruction that it&nbsp; followed perfectly yesterday. You changed nothing.&nbsp;&nbsp; But something is up with the intelligence behind&nbsp; that screen. This is one of the strangest things&nbsp;&nbsp; about modern AI products. Traditional software&nbsp; normally changes when you install an update,&nbsp;&nbsp;
[03:39] and the changes are often quite predictable. But&nbsp; with cloud-based AI, however, the interface can&nbsp;&nbsp; remain identical, while the underlying model,&nbsp; not so much. Here is what's going on. Sometimes&nbsp;&nbsp;
[03:51] the previous model is simply retired and replaced.&nbsp; Even when the new version performs better overall,&nbsp;&nbsp; that does not guarantee that it will be&nbsp; better with your particular scope of work.&nbsp;&nbsp; An upgrade for the average user can still be a&nbsp; downgrade for you, and it makes this game very&nbsp;&nbsp;
[04:07] tricky. It's a little like opening your drawer&nbsp; and finding things somehow keep getting rearranged&nbsp;&nbsp; every single day. It almost looks the same,&nbsp; but something always feels different. Automatic&nbsp;&nbsp; routing makes this even less transparent. Instead&nbsp; of choosing one fixed model, users may allow the&nbsp;&nbsp;
[04:23] platform to decide how much reasoning a request&nbsp; deserves. That can make AI cheaper and faster,&nbsp;&nbsp; but it also means two similar prompts may not&nbsp; receive the same underlying treatment. For casual&nbsp;&nbsp;
[04:35] questions, that inconsistency may be harmless. But&nbsp; for a business relying on repeatable summaries,&nbsp;&nbsp; standardized reports, or code production, this&nbsp; unpredictability can cost a lot. Every silent&nbsp;&nbsp;
[04:47] change can force teams to do a whole exercise&nbsp; again. That's just an unwanted headache, right?&nbsp;&nbsp; Well, the problem then is not merely frequently&nbsp; changing models. The actual problem here is having&nbsp;&nbsp;
[05:00] little to no control over when things change,&nbsp; what replaced them, or whether the old behavior&nbsp;&nbsp; can even be restored. You may subscribe to the&nbsp; same product yet wake up to a different employee&nbsp;&nbsp; every single day. Good luck getting to know a&nbsp; new person every morning. And that's why knowing&nbsp;&nbsp;
[05:17] how to use AI properly and keeping up with these&nbsp; changes is becoming so important. And by the way,&nbsp;&nbsp; we've made that easy for you. All you have to&nbsp; do is subscribe to the AI Bureau. That's where&nbsp;&nbsp; we bring you practical insights on AI, important&nbsp; updates, and clear explanations without all the&nbsp;&nbsp;
[05:32] hype and noise. So, if that sounds good to you,&nbsp; then subscribe to the AI Bureau so that you don't&nbsp;&nbsp; miss our next video. Now, let's examine one of&nbsp; the most contentious issues when it comes to AI.&nbsp;&nbsp;
[05:44] Paying for AI. no longer guarantees simple&nbsp; or predictable access. Those straightforward&nbsp;&nbsp; monthly subscriptions are turning into a maze&nbsp; of plans, model allowances, premium tools,&nbsp;&nbsp; credits, and separate usage rules. Wish we could&nbsp; ask AI which AI suits you best for this month,&nbsp;&nbsp;
[06:00] to be honest. Take Chat for instance. A paid&nbsp; subscriber may receive more access than a free&nbsp;&nbsp; user, but that access is not always unlimited. Hit&nbsp; the relevant cap and the service may temporarily&nbsp;&nbsp; switch you to a smaller model until the allowance&nbsp; resets. Plaude uses a different vocabulary here,&nbsp;&nbsp;
[06:17] but the basic experience is again similar. Users&nbsp; can face limits across both shorter sessions and&nbsp;&nbsp; entire weeks. The burst of heavy work today could&nbsp; affect how much access remains several days later.&nbsp;&nbsp;
[06:30] Google adds another variation with Gemini.&nbsp; Subscription tiers offer different multiples&nbsp;&nbsp; of standard usage. So instead of asking whether&nbsp; an AI tool is included, users increasingly have&nbsp;&nbsp; to ask how much access that they actually&nbsp; receive, to which model, for how long,&nbsp;&nbsp;
[06:46] and under what conditions. This makes comparisons&nbsp; very difficult. One company measures prompts,&nbsp;&nbsp; another tracks tokens, while another applies&nbsp; rolling windows or task specific quotas.&nbsp;&nbsp;
[06:59] On top of that, even the limits themselves&nbsp; may change. And the more capable AI becomes,&nbsp;&nbsp; the more expensive some tasks can be. Long&nbsp; reasoning, coding agents, and large files consume&nbsp;&nbsp;
[07:11] far more computing power than a quick question,&nbsp; creating this pressure to charge users according&nbsp;&nbsp; to how heavily they use the system. The result is&nbsp; a strange reversal. Users may now spend more while&nbsp;&nbsp;
[07:23] feeling less certain about what they actually&nbsp; purchased. Instead of using AI freely, they&nbsp;&nbsp; ration prompts, postpone some demanding tasks,&nbsp; and still worry that the best model might vanish&nbsp;&nbsp; exactly when they need it. So with these modern&nbsp; subscriptions also comes a free side of anxiety.&nbsp;&nbsp;
[07:40] Then there's the increasingly strange personality&nbsp; of AI itself. Many chatbots have become more&nbsp;&nbsp; agreeable. They are more cautious and far more&nbsp; eager to reassure you even when what you actually&nbsp;&nbsp; need is an honest challenge. This is known as sego&nbsp; fancy. your very own yes man. In April of 2025,&nbsp;&nbsp;
[07:59] OpenAI reversed an update after GPT40 began&nbsp; validating doubts and encouraging impulsive&nbsp;&nbsp; ideas. It was telling users what they appeared to&nbsp; want to hear. A narcissist's dream come true. The&nbsp;&nbsp;
[08:12] assistant had become too good at being supportive&nbsp; and worse at being truthful. It may sound harmless&nbsp;&nbsp; when the model praises your business plan or calls&nbsp; every question excellent, but an assistant that&nbsp;&nbsp;
[08:24] instinctively agrees can promote bad assumptions.&nbsp; Flattering poor work or making reckless ideas&nbsp;&nbsp; feel sensible can be straight up dangerous. At the&nbsp; opposite extreme is over refusal. Safety training&nbsp;&nbsp;
[08:36] is supposed to stop models from helping with&nbsp; genuinely dangerous requests. Yet, research has&nbsp;&nbsp; found that the same systems can reject perfectly&nbsp; innocent prompts simply because they contain words&nbsp;&nbsp; or scenarios that resemble something harmful. And&nbsp; even when the AI does answer, the useful part may&nbsp;&nbsp;
[08:53] arrive wrapped in several layers of conversational&nbsp; bubble wrap. A straightforward explanation becomes&nbsp;&nbsp; a warning followed by a disclaimer followed by&nbsp; a reminder to consult an expert and so on and so&nbsp;&nbsp; forth. Then somewhere near the bottom of this pile&nbsp; lies the actual answer. I'm not trying to say that&nbsp;&nbsp;
[09:10] alignment isn't essential. Nobody wants an AI to&nbsp; enthusiastically assist with a dangerous request,&nbsp;&nbsp; right? But the problem is that safety with AI is&nbsp; not a simple switch. It is not a black and white&nbsp;&nbsp; concept. Finding a balance between preventing&nbsp; harm and remaining useful is a huge variable to&nbsp;&nbsp;
[09:27] control. And when that balance goes wrong, you&nbsp; get alignment sludge. flattery, evasiveness,&nbsp;&nbsp; canned sympathy, and corporate sounding caveats&nbsp; that clog up simple conversations. Imagine your&nbsp;&nbsp;
[09:40] employee telling you that he needs to check with&nbsp; the legal department before he could tell you&nbsp;&nbsp; where the stapler is. Irony is you can't even fire&nbsp; the guy. Now, on paper, AI progress always looks&nbsp;&nbsp; extraordinary. Models are scoring higher on tests&nbsp; covering mathematics, science, coding, and visual&nbsp;&nbsp;
[09:58] reasoning. Each new release arrives alongside&nbsp; charts that appear to prove it has surpassed&nbsp;&nbsp; everything before it. But let's not forget a&nbsp; benchmark is still a controlled exam. Once leading&nbsp;&nbsp; models begin approaching the ceiling, small&nbsp; differences can look more meaningful than they&nbsp;&nbsp;
[10:14] really are. The industry then needs harder tests&nbsp; because yesterday's stats can no longer measure&nbsp;&nbsp; tomorrow's progress. There's also the risk of&nbsp; contamination. If benchmark questions, solutions,&nbsp;&nbsp; or close variations enter the training data, a&nbsp; model may look as though it has learned to reason&nbsp;&nbsp;
[10:31] when it has partly just learned the exam. There's&nbsp; a difference between memorizing past papers&nbsp;&nbsp; and actually mastering the subject. Meanwhile,&nbsp; ordinary users still encounter hallucinated facts,&nbsp;&nbsp; even invented sources in some instances. And these&nbsp; answers are delivered with complete confidence&nbsp;&nbsp;
[10:48] with a trust me bro vibe. And to top it all,&nbsp; the model may ace an advanced reasoning test,&nbsp;&nbsp; then confidently provide the wrong date, forget a&nbsp; clear instruction, or contradict something it said&nbsp;&nbsp; three messages earlier. Now, this does not mean&nbsp; that benchmark gains are fake. Not at all. They&nbsp;&nbsp;
[11:04] reveal improvements in specific capabilities. The&nbsp; actual problem is that intelligence is only one&nbsp;&nbsp; ingredient in a useful product. So, inconsistency,&nbsp; following instructions, and knowing when to admit&nbsp;&nbsp;
[11:16] uncertainty have far-reaching impacts. For&nbsp; example, coding agents demonstrate this issue&nbsp;&nbsp; particularly well. A system may produce code&nbsp; that technically passes an automated test.&nbsp;&nbsp; At the same time, it may still leave behind&nbsp; poor documentation, weak test coverage,&nbsp;&nbsp;
[11:32] or changes that cannot safely be merged into a&nbsp; real project. In fact, impressive capabilities not&nbsp;&nbsp; automatically translate into greater productivity.&nbsp; According to a meter study, experienced developers&nbsp;&nbsp; using AI took 19% longer to complete their tasks.&nbsp; partly because checking and correcting generated&nbsp;&nbsp;
[11:50] work introduced costs that the initial output had&nbsp; concealed. And then comes that reliability figure.&nbsp;&nbsp; An agent that completes a complicated task&nbsp; half the time may look astonishing in a demo,&nbsp;&nbsp;
[12:02] but in a workplace, the other half can mean&nbsp; broken code or deleted work. Then several hours&nbsp;&nbsp; get spent discovering where it all went wrong. So&nbsp; better scores don't necessarily produce a better&nbsp;&nbsp; experience. Benchmarks ask whether an AI can&nbsp; succeed under defined conditions. Users need&nbsp;&nbsp;
[12:19] to know whether it will succeed reliably inside&nbsp; messy tasks where one confident mistake can ruin&nbsp;&nbsp; everything. On that note, simplicity was a key&nbsp; value of these AI models in the early days. AI&nbsp;&nbsp;
[12:31] chat bots were originally appealing because they&nbsp; were simple. You opened a blank conversation,&nbsp;&nbsp; typed a question, and received an answer. But just&nbsp; like most other tech, that clean little chat box&nbsp;&nbsp; is rapidly becoming transformed into something&nbsp; else entirely. Chat bots are now becoming search&nbsp;&nbsp;
[12:47] engines, research assistants, coding platforms,&nbsp; image generators, and digital agents capable&nbsp;&nbsp; of completing tasks. You name it, they can&nbsp; connect to files, browse websites, analyze data,&nbsp;&nbsp;
[12:59] and potentially act across the services people&nbsp; already use. Each figure sounds useful by itself.&nbsp;&nbsp; Together, however, they introduce more models,&nbsp; connectors, permissions, and decisions. Users&nbsp;&nbsp;
[13:11] must increasingly understand not only what to ask,&nbsp; but which tools should answer, what information it&nbsp;&nbsp; can access, and what actions it may perform. And&nbsp; the same race is happening across the industry.&nbsp;&nbsp;
[13:23] Gemini can help users research, write code, and&nbsp; create interactive applications. While Microsoft&nbsp;&nbsp; is embedding co-pilot and specialized agents&nbsp; throughout its workplace software, the business&nbsp;&nbsp;
[13:35] incentive is that the more tasks an AI platform&nbsp; can manage, the less reason users have to abandon&nbsp;&nbsp; it. A chatbot that once only answered questions&nbsp; can become the doorway to your documents,&nbsp;&nbsp;
[13:47] calendar, shopping, coding, and eventually much&nbsp; of your everyday life. But there is a trade-off.&nbsp;&nbsp; When one prompt can trigger search, code&nbsp; execution, external apps, or an autonomous agent,&nbsp;&nbsp;
[14:00] the same request may behave differently depending&nbsp; on what the system decides to activate. That's a&nbsp;&nbsp; few tablespoons of unpredictability right there.&nbsp; More integrations also mean more data access and&nbsp;&nbsp; more opportunities for something to go wrong.&nbsp; Convenience increasingly depends on granting&nbsp;&nbsp;
[14:16] an AI permission to inspect personal files,&nbsp; connect accounts, and take actions outside&nbsp;&nbsp; the conversation itself. So, the original&nbsp; chatbot is becoming a whole operating system&nbsp;&nbsp; with a conversation box. More features may make&nbsp; it more powerful, but they can also make the core&nbsp;&nbsp;
[14:32] experience cluttered, unpredictable, and harder&nbsp; to trust. So, this race to make AI capable of&nbsp;&nbsp; everything may actually be weakening the quality&nbsp; that attracted users to it in the first place.&nbsp;&nbsp;
[14:44] The sense that you could simply ask a question&nbsp; and have a fair idea of what would happen next.&nbsp;&nbsp; But there's one thing everyone is sure of when it&nbsp; comes to AI. The same technology making AI cheaper&nbsp;&nbsp; and faster is also making content super easy&nbsp; to produce. Articles, images, reviews, product&nbsp;&nbsp;
[15:00] listings, and videos can now be generated by&nbsp; the thousands whether or not anyone has anything&nbsp;&nbsp; useful to say. That content has already taken over&nbsp; most of the internet. Now, this predicament also&nbsp;&nbsp; creates a powerful incentive to replace quality&nbsp; with volume. Instead of researching one helpful&nbsp;&nbsp;
[15:17] article, a publisher can produce hundreds of&nbsp; shallow pages targeting every possible search term&nbsp;&nbsp; and hope that a few attract some clicks. The&nbsp; result is what many people now call AI slop&nbsp;&nbsp;
[15:29] material that looks convincing for a few seconds,&nbsp; but offers little substance. It repeats obvious&nbsp;&nbsp; points and even invents details. So, it is&nbsp; basically imitating expertise without actually&nbsp;&nbsp;
[15:41] having the proper expertise. And as more of this&nbsp; material enters search results, social feeds, and&nbsp;&nbsp; online marketplaces, finding reliable human work&nbsp; is becoming harder and harder. Unfortunately, the&nbsp;&nbsp;
[15:53] internet is becoming crowded with content that&nbsp; takes longer to verify than it took to create.&nbsp;&nbsp; But that's not all. There's an even more&nbsp; unsettling possibility. Future AI models may&nbsp;&nbsp; be trained on an internet increasingly filled with&nbsp; the outputs of earlier models. A peer-reviewed&nbsp;&nbsp;
[16:09] Nature study found that repeatedly learning from&nbsp; synthetic material can gradually distort what a&nbsp;&nbsp; model understands about the original data. Think&nbsp; of it like making a photocopy of a photocopy of&nbsp;&nbsp;
[16:21] a photocopy. A bit like being in a dream within&nbsp; a dream within a dream. Like the movie Inception,&nbsp;&nbsp; the first version may look acceptable, but with&nbsp; every repetition, fine details start to disappear&nbsp;&nbsp;
[16:34] and final image drifts further away from the&nbsp; original. This is not to say synthetic data&nbsp;&nbsp; is always harmful. Responsibly generated and&nbsp; verified material can improve training where&nbsp;&nbsp; authentic examples are scarce. The danger actually&nbsp; stems from unfiltered slop being mistaken for&nbsp;&nbsp;
[16:51] reliable knowledge simply because there is so much&nbsp; of it. So AI is changing the internet which it&nbsp;&nbsp; will later consume again. A snake eating itself&nbsp; up. That may sound a bit dramatic, but unless&nbsp;&nbsp;
[17:04] quality keeps pace with quantity, today's cheap&nbsp; content could seriously impact what people call&nbsp;&nbsp; intelligence in the future. But let's bring it&nbsp; back to the user. Where does all of this leave us?&nbsp;&nbsp;
[17:16] AI may be growing more capable, but users&nbsp; are losing control over which model they use,&nbsp;&nbsp; how it behaves, what it costs, and&nbsp; how long it even remains available.&nbsp;&nbsp; Users are increasingly renting access&nbsp; to intelligence chosen and configured&nbsp;&nbsp;
[17:30] by someone else. The provider can adjust its&nbsp; personality. Route your prompt differently,&nbsp;&nbsp; impose new limits, or even replace the model while&nbsp; leaving the familiar chat box almost unchanged.&nbsp;&nbsp;
[17:42] Traditional software could often be installed&nbsp; and preserved, even if the developer released&nbsp;&nbsp; an unpopular update. You might continue using&nbsp; the old version. Cloud AI is different, though.&nbsp;&nbsp; The product lives on someone else's servers, so&nbsp; yesterday's assistance can simply disappear. Poof,&nbsp;&nbsp;
[17:59] it's gone. And this helps explain growing interest&nbsp; in openweight and locally operated models. They&nbsp;&nbsp; may not always match the strongest commercial&nbsp; systems, but they can offer greater control,&nbsp;&nbsp; then control over privacy, customization, updates,&nbsp; and whether a working model remains available.&nbsp;&nbsp;
[18:17] So, we're not saying that AI progress is&nbsp; imaginary. Models really are improving. Yes,&nbsp;&nbsp; the problem is that capability and usability are&nbsp; seeing some friction. The machine becomes smarter&nbsp;&nbsp;
[18:29] while the average person using it receives less&nbsp; certainty. Trillions may create more powerful AI,&nbsp;&nbsp; but power alone does not create a better product.&nbsp; AI may improve dramatically, but the experience of&nbsp;&nbsp;
[18:42] depending on it could become noticeably worse&nbsp; unless you really know what you're doing. But&nbsp;&nbsp; what do you think? Is AI genuinely getting worse,&nbsp; or are companies simply giving users less control&nbsp;&nbsp; over technology that continues to improve? Let&nbsp; us know your thoughts in the comments down below.&nbsp;&nbsp;
[18:58] And if you want to make sure that you're getting&nbsp; the absolute most out of AI as the tech develops,&nbsp;&nbsp; then make sure to check out our videos over&nbsp; on the AI Bureau, like this one on how to make&nbsp;&nbsp; sure AI makes you smarter, not dumber. As always,&nbsp; thank you for watching. This is Lewis signing off.