Trillions Are Flowing Into AI, But It's Getting Worse

[00:00] Trillions of dollars are being poured into  AI chips, models, and gigantic data centers.   And every few months, some new model arrives,  promising either superior reasoning or improved   coding. But if AI actually is getting smarter and  more capable, then why are so many complaining  

[00:17] about AI getting worse? Users are reporting  weirder mistakes, answers becoming more cautious,   subscriptions getting confusing, and limits that  are more restrictive with less predictable tools.  

[00:30] So, what the hell is going on here? Well,  in this video, we'll explain how AI can keep   getting smarter while the experience surrounding  it gets noticeably worse, and why the technology  

[00:42] we're all becoming increasingly dependent upon is  becoming less and less ours. My name is Lewis and   you're watching the Coin Bureau. To understand  why AI feels different today, we need to go  

[00:55] back to the beginning. Not the beginning of time,  just when CHACBT first made a public appearance.   It was free. It was also surprisingly powerful  and it was presented as an experimental preview,  

[01:07] indicating that it had much more potential. The  aim at that point was only to get public feedback   on this novel experience. The priority didn't seem  to be squeezing out dollars from users at that  

[01:19] point, but it was still a typical marketing trap.  Lo and behold, the paid subscriptions finally   arrived, and the offer remained remarkably simple.  $20 bought faster responses, priority access,  

[01:32] and a front row seat to the latest technology.  And that generous opening worked. People began   using AI to write emails, summarize documents,  study, brainstorm, analyze data, and produce code.  

[01:45] And this goes on and on. What started as an  entertaining chatbot gradually became part   of how millions of people completed ordinary  work. Predictably, once a tool saves you time,   you start building routines around it.  Templates are created, teams are trained,  

[02:00] and entire workflows start to revolve around it.  Removing it then becomes much harder than cancing   some streaming service for instance. But providing  advanced AI to hundreds of millions of users also  

[02:12] requires vast amounts of chips, electricity, and  data center capacity. Eventually, the companies   funding that bill need a little bit more than  just enthusiastic users. They need subscriptions,  

[02:24] enterprise contracts, basically enough revenue  to justify this infrastructure spending on an   extraordinary scale. And so the honeymoon phase  began ending. The original challenge was making  

[02:36] AI useful enough that people would adopt it.  The new challenge was working out how much those   increasingly dependent users and businesses can be  persuaded to pay. This is the familiar life cycle   of any good digital platform. Provide a solution,  subsidize convenience, establish the habit,  

[02:54] and then monetize the dependency. The technology  that first asked, "How can I help?" is therefore   facing a different question from its owners. How  much will people tolerate? And as we'll see next,  

[03:07] companies do not always need to raise the price  to change the deal. Sometimes they could change   the product itself. Now, imagine opening  the same app, entering the same prompt,   and receiving a completely different result. Like  maybe the wording is weaker, the code breaks, or  

[03:23] the model suddenly ignores an instruction that it  followed perfectly yesterday. You changed nothing.   But something is up with the intelligence behind  that screen. This is one of the strangest things   about modern AI products. Traditional software  normally changes when you install an update,  

[03:39] and the changes are often quite predictable. But  with cloud-based AI, however, the interface can   remain identical, while the underlying model,  not so much. Here is what's going on. Sometimes  

[03:51] the previous model is simply retired and replaced.  Even when the new version performs better overall,   that does not guarantee that it will be  better with your particular scope of work.   An upgrade for the average user can still be a  downgrade for you, and it makes this game very  

[04:07] tricky. It's a little like opening your drawer  and finding things somehow keep getting rearranged   every single day. It almost looks the same,  but something always feels different. Automatic   routing makes this even less transparent. Instead  of choosing one fixed model, users may allow the  

[04:23] platform to decide how much reasoning a request  deserves. That can make AI cheaper and faster,   but it also means two similar prompts may not  receive the same underlying treatment. For casual  

[04:35] questions, that inconsistency may be harmless. But  for a business relying on repeatable summaries,   standardized reports, or code production, this  unpredictability can cost a lot. Every silent  

[04:47] change can force teams to do a whole exercise  again. That's just an unwanted headache, right?   Well, the problem then is not merely frequently  changing models. The actual problem here is having  

[05:00] little to no control over when things change,  what replaced them, or whether the old behavior   can even be restored. You may subscribe to the  same product yet wake up to a different employee   every single day. Good luck getting to know a  new person every morning. And that's why knowing  

[05:17] how to use AI properly and keeping up with these  changes is becoming so important. And by the way,   we've made that easy for you. All you have to  do is subscribe to the AI Bureau. That's where   we bring you practical insights on AI, important  updates, and clear explanations without all the  

[05:32] hype and noise. So, if that sounds good to you,  then subscribe to the AI Bureau so that you don't   miss our next video. Now, let's examine one of  the most contentious issues when it comes to AI.  

[05:44] Paying for AI. no longer guarantees simple  or predictable access. Those straightforward   monthly subscriptions are turning into a maze  of plans, model allowances, premium tools,   credits, and separate usage rules. Wish we could  ask AI which AI suits you best for this month,  

[06:00] to be honest. Take Chat for instance. A paid  subscriber may receive more access than a free   user, but that access is not always unlimited. Hit  the relevant cap and the service may temporarily   switch you to a smaller model until the allowance  resets. Plaude uses a different vocabulary here,  

[06:17] but the basic experience is again similar. Users  can face limits across both shorter sessions and   entire weeks. The burst of heavy work today could  affect how much access remains several days later.  

[06:30] Google adds another variation with Gemini.  Subscription tiers offer different multiples   of standard usage. So instead of asking whether  an AI tool is included, users increasingly have   to ask how much access that they actually  receive, to which model, for how long,  

[06:46] and under what conditions. This makes comparisons  very difficult. One company measures prompts,   another tracks tokens, while another applies  rolling windows or task specific quotas.  

[06:59] On top of that, even the limits themselves  may change. And the more capable AI becomes,   the more expensive some tasks can be. Long  reasoning, coding agents, and large files consume  

[07:11] far more computing power than a quick question,  creating this pressure to charge users according   to how heavily they use the system. The result is  a strange reversal. Users may now spend more while  

[07:23] feeling less certain about what they actually  purchased. Instead of using AI freely, they   ration prompts, postpone some demanding tasks,  and still worry that the best model might vanish   exactly when they need it. So with these modern  subscriptions also comes a free side of anxiety.  

[07:40] Then there's the increasingly strange personality  of AI itself. Many chatbots have become more   agreeable. They are more cautious and far more  eager to reassure you even when what you actually   need is an honest challenge. This is known as sego  fancy. your very own yes man. In April of 2025,  

[07:59] OpenAI reversed an update after GPT40 began  validating doubts and encouraging impulsive   ideas. It was telling users what they appeared to  want to hear. A narcissist's dream come true. The  

[08:12] assistant had become too good at being supportive  and worse at being truthful. It may sound harmless   when the model praises your business plan or calls  every question excellent, but an assistant that  

[08:24] instinctively agrees can promote bad assumptions.  Flattering poor work or making reckless ideas   feel sensible can be straight up dangerous. At the  opposite extreme is over refusal. Safety training  

[08:36] is supposed to stop models from helping with  genuinely dangerous requests. Yet, research has   found that the same systems can reject perfectly  innocent prompts simply because they contain words   or scenarios that resemble something harmful. And  even when the AI does answer, the useful part may  

[08:53] arrive wrapped in several layers of conversational  bubble wrap. A straightforward explanation becomes   a warning followed by a disclaimer followed by  a reminder to consult an expert and so on and so   forth. Then somewhere near the bottom of this pile  lies the actual answer. I'm not trying to say that  

[09:10] alignment isn't essential. Nobody wants an AI to  enthusiastically assist with a dangerous request,   right? But the problem is that safety with AI is  not a simple switch. It is not a black and white   concept. Finding a balance between preventing  harm and remaining useful is a huge variable to  

[09:27] control. And when that balance goes wrong, you  get alignment sludge. flattery, evasiveness,   canned sympathy, and corporate sounding caveats  that clog up simple conversations. Imagine your  

[09:40] employee telling you that he needs to check with  the legal department before he could tell you   where the stapler is. Irony is you can't even fire  the guy. Now, on paper, AI progress always looks   extraordinary. Models are scoring higher on tests  covering mathematics, science, coding, and visual  

[09:58] reasoning. Each new release arrives alongside  charts that appear to prove it has surpassed   everything before it. But let's not forget a  benchmark is still a controlled exam. Once leading   models begin approaching the ceiling, small  differences can look more meaningful than they  

[10:14] really are. The industry then needs harder tests  because yesterday's stats can no longer measure   tomorrow's progress. There's also the risk of  contamination. If benchmark questions, solutions,   or close variations enter the training data, a  model may look as though it has learned to reason  

[10:31] when it has partly just learned the exam. There's  a difference between memorizing past papers   and actually mastering the subject. Meanwhile,  ordinary users still encounter hallucinated facts,   even invented sources in some instances. And these  answers are delivered with complete confidence  

[10:48] with a trust me bro vibe. And to top it all,  the model may ace an advanced reasoning test,   then confidently provide the wrong date, forget a  clear instruction, or contradict something it said   three messages earlier. Now, this does not mean  that benchmark gains are fake. Not at all. They  

[11:04] reveal improvements in specific capabilities. The  actual problem is that intelligence is only one   ingredient in a useful product. So, inconsistency,  following instructions, and knowing when to admit  

[11:16] uncertainty have far-reaching impacts. For  example, coding agents demonstrate this issue   particularly well. A system may produce code  that technically passes an automated test.   At the same time, it may still leave behind  poor documentation, weak test coverage,  

[11:32] or changes that cannot safely be merged into a  real project. In fact, impressive capabilities not   automatically translate into greater productivity.  According to a meter study, experienced developers   using AI took 19% longer to complete their tasks.  partly because checking and correcting generated  

[11:50] work introduced costs that the initial output had  concealed. And then comes that reliability figure.   An agent that completes a complicated task  half the time may look astonishing in a demo,  

[12:02] but in a workplace, the other half can mean  broken code or deleted work. Then several hours   get spent discovering where it all went wrong. So  better scores don't necessarily produce a better   experience. Benchmarks ask whether an AI can  succeed under defined conditions. Users need  

[12:19] to know whether it will succeed reliably inside  messy tasks where one confident mistake can ruin   everything. On that note, simplicity was a key  value of these AI models in the early days. AI  

[12:31] chat bots were originally appealing because they  were simple. You opened a blank conversation,   typed a question, and received an answer. But just  like most other tech, that clean little chat box   is rapidly becoming transformed into something  else entirely. Chat bots are now becoming search  

[12:47] engines, research assistants, coding platforms,  image generators, and digital agents capable   of completing tasks. You name it, they can  connect to files, browse websites, analyze data,  

[12:59] and potentially act across the services people  already use. Each figure sounds useful by itself.   Together, however, they introduce more models,  connectors, permissions, and decisions. Users  

[13:11] must increasingly understand not only what to ask,  but which tools should answer, what information it   can access, and what actions it may perform. And  the same race is happening across the industry.  

[13:23] Gemini can help users research, write code, and  create interactive applications. While Microsoft   is embedding co-pilot and specialized agents  throughout its workplace software, the business  

[13:35] incentive is that the more tasks an AI platform  can manage, the less reason users have to abandon   it. A chatbot that once only answered questions  can become the doorway to your documents,  

[13:47] calendar, shopping, coding, and eventually much  of your everyday life. But there is a trade-off.   When one prompt can trigger search, code  execution, external apps, or an autonomous agent,  

[14:00] the same request may behave differently depending  on what the system decides to activate. That's a   few tablespoons of unpredictability right there.  More integrations also mean more data access and   more opportunities for something to go wrong.  Convenience increasingly depends on granting  

[14:16] an AI permission to inspect personal files,  connect accounts, and take actions outside   the conversation itself. So, the original  chatbot is becoming a whole operating system   with a conversation box. More features may make  it more powerful, but they can also make the core  

[14:32] experience cluttered, unpredictable, and harder  to trust. So, this race to make AI capable of   everything may actually be weakening the quality  that attracted users to it in the first place.  

[14:44] The sense that you could simply ask a question  and have a fair idea of what would happen next.   But there's one thing everyone is sure of when it  comes to AI. The same technology making AI cheaper   and faster is also making content super easy  to produce. Articles, images, reviews, product  

[15:00] listings, and videos can now be generated by  the thousands whether or not anyone has anything   useful to say. That content has already taken over  most of the internet. Now, this predicament also   creates a powerful incentive to replace quality  with volume. Instead of researching one helpful  

[15:17] article, a publisher can produce hundreds of  shallow pages targeting every possible search term   and hope that a few attract some clicks. The  result is what many people now call AI slop  

[15:29] material that looks convincing for a few seconds,  but offers little substance. It repeats obvious   points and even invents details. So, it is  basically imitating expertise without actually  

[15:41] having the proper expertise. And as more of this  material enters search results, social feeds, and   online marketplaces, finding reliable human work  is becoming harder and harder. Unfortunately, the  

[15:53] internet is becoming crowded with content that  takes longer to verify than it took to create.   But that's not all. There's an even more  unsettling possibility. Future AI models may   be trained on an internet increasingly filled with  the outputs of earlier models. A peer-reviewed  

[16:09] Nature study found that repeatedly learning from  synthetic material can gradually distort what a   model understands about the original data. Think  of it like making a photocopy of a photocopy of  

[16:21] a photocopy. A bit like being in a dream within  a dream within a dream. Like the movie Inception,   the first version may look acceptable, but with  every repetition, fine details start to disappear  

[16:34] and final image drifts further away from the  original. This is not to say synthetic data   is always harmful. Responsibly generated and  verified material can improve training where   authentic examples are scarce. The danger actually  stems from unfiltered slop being mistaken for  

[16:51] reliable knowledge simply because there is so much  of it. So AI is changing the internet which it   will later consume again. A snake eating itself  up. That may sound a bit dramatic, but unless  

[17:04] quality keeps pace with quantity, today's cheap  content could seriously impact what people call   intelligence in the future. But let's bring it  back to the user. Where does all of this leave us?  

[17:16] AI may be growing more capable, but users  are losing control over which model they use,   how it behaves, what it costs, and  how long it even remains available.   Users are increasingly renting access  to intelligence chosen and configured  

[17:30] by someone else. The provider can adjust its  personality. Route your prompt differently,   impose new limits, or even replace the model while  leaving the familiar chat box almost unchanged.  

[17:42] Traditional software could often be installed  and preserved, even if the developer released   an unpopular update. You might continue using  the old version. Cloud AI is different, though.   The product lives on someone else's servers, so  yesterday's assistance can simply disappear. Poof,  

[17:59] it's gone. And this helps explain growing interest  in openweight and locally operated models. They   may not always match the strongest commercial  systems, but they can offer greater control,   then control over privacy, customization, updates,  and whether a working model remains available.  

[18:17] So, we're not saying that AI progress is  imaginary. Models really are improving. Yes,   the problem is that capability and usability are  seeing some friction. The machine becomes smarter  

[18:29] while the average person using it receives less  certainty. Trillions may create more powerful AI,   but power alone does not create a better product.  AI may improve dramatically, but the experience of  

[18:42] depending on it could become noticeably worse  unless you really know what you're doing. But   what do you think? Is AI genuinely getting worse,  or are companies simply giving users less control   over technology that continues to improve? Let  us know your thoughts in the comments down below.  

[18:58] And if you want to make sure that you're getting  the absolute most out of AI as the tech develops,   then make sure to check out our videos over  on the AI Bureau, like this one on how to make   sure AI makes you smarter, not dumber. As always,  thank you for watching. This is Lewis signing off.