I have been writing about AI for quite a while now, but this is probably the first time I genuinely do not know what to think. Not because the tech...
For further actions, you may consider blocking this person and/or reporting abuse
Displaying a subset of the total comments. Please sign in to view all comments on this post.
Sylwia, the access question lands differently depending on where you're starting from. from Nigeria, equal access to frontier models was never the baseline — every API call already crosses an ocean, costs latency that developers in Virginia don't pay and depends on infrastructure that runs on diesel generators because the grid gives 4 hours a day.
The Fable shutdown was a disruption for developers who had it. for developers who were already working around geography and cost, it was confirmation of something already true: the intelligence available to you has never been just a function of what you can pay. it's always been a function of where you are.
what changes now is that developers in Poland and the US are starting to feel what developers in Lagos have always felt. that's not a silver lining. it's just the same map, newly legible.
Daniel, thank you so much for this perspective. I really appreciate you sharing it.
Sitting here in Europe, it's easy to forget that geography itself is a privilege. We often take for granted that we have relatively stable infrastructure and broad access to the latest models.
Interestingly, the discussion in Europe has recently shifted in a different direction. Many people are worried that Europe is becoming increasingly dependent on the US for AI, and that if we fail to keep up with innovation, we'll end up in a much weaker position ourselves.
Your comment is a good reminder that access has never been distributed equally in the first place.
But you don't even need to look outside Europe. Geography isn't the only barrier, money is another one. Not everyone can afford to pay for a subscription to some AI company, and, in a way, this means "skill" is now behind a paywall.
That's a very good point, and unfortunately it's easy to forget.
I think I fell into the illusion that if even I, living in a post-communist country, can afford a Pro subscription (not every model and not every top tier, of course, but still), then it probably isn't such a big barrier.
But you're absolutely right. Compared to a large part of the world, I'm actually in a very privileged position. Thanks for pointing that out!
Wii, "skill is now behind a paywall" is the line. the geography barrier and the money barrier are the same barrier wearing different clothes . both determine what tools you can access, and the tools increasingly determine what you can build.
what makes it harder to argue against is that the paywall is gradual. free tiers exist. the gap only becomes visible when you hit the ceiling — context limits, rate limits, model access and realize the developers building without those ceilings are compounding faster. the advantage isn't access. it's the absence of friction over time.
Exactly! Do you ever get the feeling it used to be different?
For example, IDEs usually had a generous Community Edition. You only really had to pay once you were building commercial products and making money from them.
With LLMs, it's different. The free tier can disappear at any moment. You never quite know when you'll hit the token limit, the context limit, or lose access to a particular model. That uncertainty itself becomes friction, and over time it adds up.
A wild @sylwia-lask has been spotted in "The Daily Context" for the AI Engineer World Fair!
Interesting read! :D
Wow!!! 🤯 Thank you so much for the screen, Francis! How cool is that! ❤️❤️
Sylwia, your Biedronka backup plan is solid. I hear the pay is decent and the product knowledge requirement matches yours already.
Jokes aside — this is the piece I couldn't write. You're inside it, I was watching from 2029. The "where do I fit" question is the one that actually keeps people up at night, and you're one of the few who'll say it out loud without dressing it up.
The 403 on a Monday morning is abstract geopolitics made very, very concrete.
Enjoy Croatia. The embargo will still be there when you get back. So will I, apparently — glad someone noticed. 🙂
Haha, exactly! 😄 From what I've heard, Biedronka has pretty decent benefits too—private healthcare, holiday vouchers... and it is really close to my home, feels almost like remote job 🤣. Sounds like a solid backup plan. 😄
As for the article, I completely agree. None of us really knows how this is going to unfold. We can try to predict the technology, but we're just as likely to be surprised by political decisions as by the models themselves. That's probably the biggest lesson from the last few weeks.
That's the sentence of the week. The technology was always the easier variable.
And honestly, Biedronka private healthcare sounds better than most startup equity packages right now. Keep that CV warm. 😄
Hahaha, exactly! 😂 I've got a nice collection of vested startup equity from different companies... and somehow I'm still not a millionaire. 😄
What comes next? This part:
That quote is becoming suspiciously relevant these days. 😅
Btw, I absolutely love Dune. 🙂
Restrictions are never a good sign, but I’m not sure if I really miss these new generation models right now, like Fable or Mythos. For me, the main issue is the price. If they cost roughly twice as much as Opus, for example, you need a really good business model to justify using them.
GPT-5.6 hurts more, because it looks much more attractive from a price/performance point of view. But honestly, the reasoning models we already have are still not bad.
So from the point of view of a regular dev in Europe, maybe not much changes. Globally, though, that is a different story. I’m curious to see where this goes.
P.S. I like Biedronka, that’s a good backup plan. 😆 And yes, Croatia is the best, but I heard prices went up a lot, so this year we decided to switch to Italy. 😅 (Typical Eastern European holiday-planning conversation. 🤣 )
Exactly! GPT is still incredibly attractive from a price/performance perspective.
One pro tip, though: apply for the AWS Community Builders program in January. You already write great articles, so I think you'd have a very good chance of getting accepted (you even don't have to write about AWS staff). They give you $500 in AWS credits every year, and you can use those credits for Kiro. 😄 I genuinely recommend it, it's a great program for people like us.
As for Italy vs. Croatia... 🤣 Just an hour ago a colleague at work joked that Croatia used to be the cheaper alternative to Italy, and now it's the premium alternative to Italy. 😅
We're heading to the Makarska Riviera, though. Two years ago it was still reasonably affordable because it's a bit too far for many German tourists to drive. 😄 I'll let you know whether the low prices survived this year!
Oh, thank you for the tip! I will definitely apply for it.
Makarska is great. We went there multiple times, but a few years ago we found an amazing place on Krk Island, and since then we’ve kept going back there. And yes, I definitely want to hear about the price changes. At least I’ll know what my wallet should prepare for next year. 😅😆
Oh yes, Krk is beautiful too! 😄 I love the Makarska Riviera because you get both the sea and the mountains.
The last time we stayed in Gradac, I even hiked to the top of Sv. Ilija. The funniest part? I did the entire hike alone and didn't meet a single person on the trail. 😂
How is that even possible? Could it be that people were discouraged by that tiny inconvenience of... 35°C in the shade? 🤣
I bet that solo hike must have been amazing, but 35°C in the shade? That sounds less like hiking and more like a survival challenge to me. 🤣🤣
i think this will help - dev.to/aws-builders/how-to-become-...
I don't think this is unique to AI — explosive growth followed by a slowdown is the default tech pattern, not the exception. We've seen versions of this story before, from early computing through to today's AI agents.
Semiconductors are the clearest parallel. Moore's Law — transistor counts roughly doubling every two years — held up remarkably well for decades but has slowed significantly since the mid-2000s. That wasn't about running out of room on the chip; it came down to physics — the breakdown of Dennard scaling (essentially, power density and heat) and eventually quantum tunneling effects as transistors shrank further. The industry didn't stop innovating, it just changed strategy — multicore, chiplets, 3D stacking — instead of relying on pure density scaling. I'd expect AI to eventually follow a similar arc: rapid gains, then a plateau as we hit compute or data limits, with progress continuing through smarter architecture rather than raw scale.
Though I'll admit a brand-new model getting capability-restricted by a government this week is a strange data point for the "it's already saturating" theory — feels like we're still on the steep part of the curve.
On the embargo: I think you're onto something. Governments don't restrict technology they see as harmless or low-stakes. Treating frontier AI like a strategic asset, the way semiconductors get export controls, suggests both a real security concern and real economic competition underneath, even if security is the official line.
That's exactly what I've been wondering too. I'm really curious about when this technology will stop scaling the way it has over the last few years.
But even more than that, I'm curious where we'll end up when it does. What will the world look like at that point? Not just software engineering, but science, medicine, education... pretty much everything. That's the question I keep coming back to. 🙂
I feel the same way. Beyond software, I often wonder what people across all industries will be doing in the future. If human labor is no longer essential, what gives currency its value, and how will society redefine purpose and contribution?
We'll see! 😄 After the Industrial Revolution, people were supposed to work less too... and somehow that never really happened.
Humans seem to be incredibly creative when it comes to inventing new kinds of work for themselves. 😂
Intriguing article, interesting views!
I'm actually surprised that this US government seems to be starting to regulate AI - doesn't that run counter to what they stood for, which was "no regulation of AI", total freedom and "cowboy style" all the way? I really thought they rejected the idea of regulating AI, seeing it as an "EU thing" ...
Or are they trying to limit access to US residents only, is that the point? (but Anthropic already said that that's not doable, so they shut off access to Mythos etc completely)
But to be honest I'm not sure that slowing down the whole AI craze a bit is such a bad thing ;-)
That's exactly what surprised me too! 😄 They usually present themselves as very pro-free market and against regulation, and then suddenly we see restrictions like these (even if they're officially explained as national security measures).
And yes, maybe slowing things down a little wouldn't be the worst outcome. 😄 Right now it feels like if you disappeared into the jungle for six months, by the time you came back the entire AI stack would have changed anyway. 🤣
My thoughts as well ... let's not go crazy and just wait & see, it's not the end of the world either way!
Hahaha, but what if we end up with Skynet... or the Matrix? 🤣
Ah, if people got paid for overthinking, I'd be a millionaire by now and wouldn't have to spend my days cleaning up code after junior developers. 😂
Might actually be a strength, the overthinking thing? The best philosophers and scientist were/are probably prone to overthinking things, lol ... maybe I should overthink a bit more as well !
Haha, give it a try! 😄
But seriously, it can be both a gift and a curse. That's actually one of the reasons I write this blog and keep a few side projects going. I don't care about "personal brand" (WTF even is it), but they give my brain something useful to chew on.
Otherwise... things get dangerous. 🤣 I start overthinking the most ridiculous scenarios, usually involving people and relationships.
It's like having a sheepdog. If it gets to work all day, it's calm and happy. If it doesn't... it starts chewing the couch. My brain works pretty much the same way. 😂
I think we're asking the wrong question. The biggest shift may not be whether AI replaces developers, but whether access to frontier AI becomes concentrated in the hands of a few organizations and governments.
History suggests that when a technology becomes strategically important, openness gives way to control. Semiconductors, nuclear technology, and even GPS followed similar patterns. If frontier models become restricted, the next wave of innovation may come from open source and smaller, specialized models rather than the biggest proprietary ones.
The future may be less about who builds the smartest model and more about who gets to use it.
This is an incredibly sharp perspective, and your historical comparisons to semiconductors and GPS are spot on.
The current climate strongly reminds me of the "computers are taking our jobs" panic of the 1970s and early 80s. Back then, computing power was also entirely concentrated in the hands of a few massive corporations and government institutions during the mainframe era. The anxiety about automation replacing the workforce was very real. However, instead of a complete corporate lockout, we eventually saw a massive shift toward decentralization driven by the personal computing revolution and open ecosystems.
We are living through a similar architectural shift right now, and honestly, the sheer scale of it is a lot to process. It is incredibly hard to predict exactly how the balance of power will settle between gated, proprietary frontier models and agile, open-source alternatives.
You are likely right that the defining bottleneck of this decade won't just be about who can build the absolute smartest model, but rather ensuring the ecosystem remains open enough that the next wave of innovation isn't locked behind a handful of corporate gatekeepers. Great post!
Thank you so much! I really like the comparison with the early days of computing.
And yes, that's exactly my concern. If AI innovation ends up concentrated in the hands of just a few corporations, we're in trouble. Competition, open ecosystems, and broad access have always been huge drivers of innovation. I'd really hate to see us lose that.
I don't doubt that for a second. I'm convinced we'll see an incredible amount of innovation around smaller and open-source models.
My concern is exactly what happens if frontier models become increasingly restricted. If the biggest breakthroughs are only available to a handful of companies or governments, then we may end up with a two-speed AI ecosystem. That's the part that worries me the most.
Like other pieces you've written, this is a fun read once again!
As a senior dev who also has to handle the business side of communication, i still find myself looking at this purely through a technical lens. it's wild seeing the narrative shift from " AI replacing us " to these weird global access restrictions.
Honestly, i'm just waiting for the current AI bubble to burst so we can finally see what the actual sustainable next step looks like, because right i genuinely have no idea where is this gonna go.
Anyway, enjoy Croatia! The embargo and the hype will definitely still be waiting for you when you get back. 😎
Exactly! 😂 A few weeks ago I was joking that AI would soon be running the world and replacing politicians... and then geopolitics stepped in and reminded us who's still in charge. 😄
Thanks! I'll definitely enjoy Croatia. Let's see whether another embargo drops while I'm away. At this point, I wouldn't even be surprised anymore. 😅
finally see* damn it, i just found the mistake today haha
hahaha I didn't notice it either 😅
The next phase will belong to people across every industry who are willing to embrace AI. They will build AI tools on their own, replacing those who cannot use AI. In this process, programmers will matter less and less.
A recent post in China went viral: an HR professional used SoloEngine to build a recruiting Agent in three minutes, screened out six candidates from 310 resumes in thirty minutes, and racked up 300,000 views in a single day. What makes it interesting is that the same company had previously spent 2.75 million yuan on an end-to-end AI office system developed by a professional team. They pushed it hard for six months, only to end in chaos and abandonment. The difficulty of implementing AI in enterprises is not about technology at all; it is about people. When leaders make decisions, they are chasing the AI trend; programmers understand code but not the business, and much of the business knowledge cannot even be articulated, so no one in between translates real business needs clearly. The future of AI development is destined to be about neither throwing money at code nor letting coders own the process; it is about letting people who truly understand the business build AI tools tailored to their own industry and job. The final outcome will be that people who can use AI replace those who cannot use AI.
That's an interesting example, but I'm not completely convinced by it.
Building a working application quickly has never really been the hard part. We were doing impressive things at hackathons long before LLMs came along. AI simply makes that part even faster.
The real challenge has always come afterward: getting people to use it, scaling it, maintaining it, integrating it with existing systems, and making it create real business value.
That's where I think software engineering is still very much alive. 🙂
Even if we lose access to every frontier AI model, we'll still be able to rely on local models. Right now, we're in the "mainframe" era of AI, but the "PC" era is advancing rapidly as well. It doesn't generate as many headlines as frontier models, but it's making steady progress. So even in the worst-case scenario, local models will still be available and become a permanent part of our toolset, much like linters and other IDE quality-of-life features.
The bigger question is how to continue providing value, regardless of rank or seniority. In my opinion, that's not just a question for developers, it's something everyone building products needs to think about. The answer will likely come from solving the pain points of tomorrow's industries, once the current AI hype settles and the market adjusts to a new normal.
Trying to prepare for that may be about as difficult as trying to predict the internet era just when ARPANET launched.
Oh yes, I keep wondering about that too! 😄 What happens once the hype settles?
At the same time, I have to admit these are fascinating times to be a developer. It really feels like we're watching a new era take shape in real time, even if none of us knows exactly where it'll lead.
I think there's another possibility.
Even if frontier models become more restricted, software engineering won't stop evolving.
The next competitive advantage may not come from having access to the biggest model, but from giving AI a better understanding of the software system it operates in.
Ownership. Dependencies. Verification. Operational weight. Change impact.
We're increasingly convinced that Open-Source Workspace Intelligence for Software Systems will matter regardless of which frontier model wins.
Yes, I'm completely with you on this. This is actually the part I'm least worried about.
Even if we never saw another frontier model, I think we'd still make incredible progress over the next few years. Better tooling, better software understanding, better context, verification, workflows... there's still so much room for innovation.
My only point is that with new frontier models, that progress would probably be even faster. 🙂
Exactly.
That's why we're increasingly thinking about AI engineering as two parallel tracks:
The first expands what AI can do.
The second helps it decide what it should do within a real software system.
I think both will evolve together over the coming years.
Exactly! 😄 And since I'm not training foundation models myself, the whole Workspace Intelligence direction is the one that really excites me.
I honestly think these are incredibly exciting times to be a software engineer. We're not just watching AI evolve, we're actively shaping how it interacts with real software systems. That's a fascinating problem to work on.
Exactly.
Models will keep improving, but understanding real software systems is an engineering challenge in its own right.
I have a feeling Workspace Intelligence will become a field of engineering, not just another AI feature.
Very possible! I actually feel like we're already starting to see it happen. It's gradually becoming a field of engineering in its own right, rather than just another AI feature. It'll be really interesting to watch how it evolves over the next few years.
What struck me reading this is that we're starting to treat access to intelligence as a separate problem from intelligence itself.
For the last few years, most discussions have focused on model capabilities: reasoning, coding, agents, benchmarks, and so on. Increasingly, I find myself wondering whether the bigger question is simply who gets access to what.
If frontier models become geographically, economically, or politically constrained, two developers may operate with fundamentally different toolsets even if they have identical skills. The resulting differences in output may not be a function of intelligence at all, they may be a function of access.
I've been thinking about this recently through the lens of what I call Information Borders: the idea that AI systems are increasingly shaped not only by how they reason but also by what they're permitted to see. Publishers, licensing agreements, regional restrictions, proprietary datasets, and government policies all create boundaries around information. In that world, two systems can arrive at different conclusions simply because they operate on different evidence.
That's one reason I've become interested in local-first and sovereign AI approaches. Not because I think local models will outperform the frontier any time soon, but because they shift the conversation from "What am I allowed to access?" to "What capabilities do I control?"
The future may end up being less about a single AI race and more about a spectrum of AI sovereignty, where individuals, companies, and countries make different tradeoffs between capability, cost, privacy, and dependence.
Either way, I think you're asking the right question. The most important limits on AI over the next decade may not be technical limits at all. They may be access limits.
Thanks for the thoughtful comment! It looks like a really interesting article, I definitely need to sit down and read it properly.
And yes, the same theme keeps coming up in the comments: local models. I think it's a very smart direction, and I'm sure they'll keep improving.
My only concern is that it doesn't completely solve the problem. It's an adaptation to the capabilities we have today, whereas access to the absolute top-tier frontier models is a different question altogether.
Unfortunately, we're also dealing with government restrictions and privately owned companies. I hope I'm wrong, but history often shows the same pattern: those with more resources gain even more advantages, while everyone else has to adapt with what's left.
I think that's a fair concern, and honestly one I share.
My interest in local and sovereign approaches isn't because I believe they'll eliminate the advantage held by whoever controls the most capable frontier models. At least in the near term, they probably won't.
What they do offer is a different kind of resilience. If access itself becomes constrained—whether by economics, policy, geography, licensing, or corporate decisions—then the ability to retain some degree of capability under your own control becomes valuable regardless of where the frontier sits.
But I agree with your broader point. Local models solve a sovereignty problem; they don't necessarily solve an inequality problem.
In some ways, that may be exactly what Information Borders are about. Once access becomes differentiated, we're no longer talking purely about intelligence or technology. We're talking about who gets access to which capabilities and under what conditions.
History certainly suggests that's a question worth paying attention to.
Exactly! I think that's an important distinction.
Someone else in the comments compared it to nuclear weapons being controlled by governments. But there's a key difference: nuclear weapons were developed as government projects from the very beginning.
Frontier AI is different. The biggest breakthroughs have largely come from private companies investing enormous amounts of their own money and research, not from governments or universities.
That makes me wonder what the relationship between governments and those companies will look like over the next few years.
That's an interesting distinction. The more I think about it, the more unusual the situation feels historically. Many transformative technologies ultimately became matters of national interest, but frontier AI is one of the first cases where the capability frontier itself is largely being driven by private organizations. That creates a different set of questions around access, governance, and control. If governments increasingly view AI as strategic infrastructure while the underlying capabilities remain concentrated in private companies, the relationship between the two becomes a fascinating thing to watch.
Enjoy your upcoming holiday!
Exactly! 😄 And it's worth remembering that these companies also have to make money. They need to sell access to their models, while governments may want to restrict access to them. That makes the whole situation even more interesting.
The AI Model Access Walls are least of our problems.
If general consumers dont have access to affordable computing , who are we going cater to ?
SSD , SD cards , storage are all getting ridiculously expensive.
AI Barons have hoarded the RAM for what, 5 years.
In first 12 months we have seen PC prices double. Last 3 months smartphone prices have increased by 30% . What happens 3 years down line. Exactly when lot of people need to buy new PCs/phones due to aging hardware.
Thanks for bringing this up! That's a really important point.
I completely agree that computers are becoming absurdly expensive. At least part of that is probably connected to the AI boom, especially when it comes to memory. High-end GPUs and AI infrastructure have created enormous demand for components, and we're all feeling the effects.
It'll be interesting to see where this goes over the next few years, because affordable hardware is just as important for innovation as affordable access to AI models.
This is the real edge (and always has been). If we're only told what to do, we're out of business. If we only pass on what AI spits out, we're out.
Oh yes, I couldn't agree more! 😄 I've met plenty of engineers who explain everything to the business in such a technical way that, in the end, nobody understands anything and the project just stalls.
I'm a talkative person too, but the feedback I usually get from clients is actually the opposite: that I explain things clearly, briefly, and without unnecessary jargon.
The interesting question is whether those skills will still matter if we eventually get truly incredible AI models. Coding alone is already becoming less of a differentiator.
Then again, looking at everything that's happening with access to frontier models, maybe those human communication skills will stay valuable for much longer than we currently expect. 🙂
Really interesting read, Sylwia.
One thing that kept coming back to me while I was reading this was your last question. A lot of the conversation has been about what AI will be capable of, but now it feels like we're also starting to ask who will actually have access to those capabilities.
I don't know where this goes either. Some of the restrictions we've seen have already changed, which makes it even harder to tell what's temporary and what's part of a longer-term shift. It'll be interesting to see how things look a year from now.
Enjoy your vacation 😀
Exactly! 😄 Let's just hope we end up on the right side of those changes.
The next year is going to be fascinating to watch, but I'm definitely hoping it brings more opportunities than barriers. 🤞
I personally think the access to the best models are not directly related to 'power'. Suppose GPT 9.9 had just released and only 1,000 people in the world were allowed to use it. GPT 9.9 might be a lot more powerful than the models other people were using, but still there would be no guarantee that the 1000 people with GPT 9.9 would perform better than the others. Maybe GPT 9.9's performance would not be significantly better than the previous model. Possibly I am too naive 😂
Haha, it all depends! 😄 It depends on who those 1,000 people are.
If they're scientists working on new medicines, materials, or physics, the impact could be enormous. If they're military researchers, that's a completely different story.
That's exactly why I find the access question so interesting. It's not just how many people get the best models, but who those people are.
I as an ordinary person just hope that the people who would take the best models would be nice to people 😁
Hahaha, I'm not quite that optimistic. 😄 History doesn't give me a huge amount of confidence on that front. Let's hope we're both wrong. 🤞
The question and the answer!
Just saw the post headings matching as if Q/A!😄
hahahahahahaha perfection 🤣🩷
lol 😭😭
The access-stratification framing is real, but I think it understates what
happens at the developer level even with full access.
Take the "500 engineers or 5" question. The pivot isn't really how good
the model gets or which tier you can buy. It's whether your output is the
kind a model can write a paragraph about without doing, or the kind that
requires hands on data nobody has published yet. The first category gets
commoditized regardless of whether you have GPT-5.6 or the open-weight
three-versions-behind. The second stays scarce even when everyone has the
frontier model, because the bottleneck moves from "can I generate prose
about X" to "can I produce a receipt for X that the wrapper can't
fabricate".
Geopolitical lock-in might then decide who runs the cheapest version of
category one. Category two is a different conversation, and it's the one
I'd want to be in either way.
"I genuinely don't know what to think" is honest, and probably the correct
stance for now. The planning move underneath it might be: stop indexing
on model tier, start indexing on which of your outputs leaves a receipt
nobody else can hand over.
Thanks for this comment, Mike! I actually stopped and thought about it for a while because it's a really interesting point.
My only feeling is that it doesn't completely close the discussion. What you're describing is more or less where we are today. The model writes the boilerplate, generates code, but I'm still the one managing it (for example Kiro works hard when I write this comment 🤣). I have the context, I know the system, and I can guide it using all that "tribal knowledge" that isn't written down anywhere.
But will that always be true?
Imagine a future model that gets access to the legacy repository, Jira, Confluence, Slack, commit history, production logs, tests... everything. Then you simply ask:
"Migrate this entire system to Rust." xDDDD
A few hours later it comes back with: "Here are the 412 commits, the tests, the rollout plan, and the production metrics."
Where is our "receipt" then?
It sounds like science fiction today, but would we really bet against it happening within a few years?
And that's where I think geopolitics and access to frontier models become important again. Software engineering is only one industry that LLMs can transform.
Of course people will still be needed. Someone has to take responsibility and sign off on a €500 million banking system migration. But will that someone necessarily be me? That's the part I'm no longer so sure about. 🙂
The receipt does not vanish in that future. It moves up a floor.
Picture the model coming back with 412 commits, tests, a rollout plan, production metrics. Every one of those is an artifact the same system authored. The tests it wrote, passing against the code it wrote, is a green checkmark with nothing standing behind it. So the receipt was never who wrote the migration. It is what, outside that system, confirms the Rust does what the old system actually did under real load. Behavior parity on replayed production traffic. The counterparty who would file a ticket if a balance came back wrong. A regulator who does not read the model's summary.
Your own line is the answer hiding in the question. Someone has to sign off on the 500M migration. That signature is the receipt. The open question is not whether it is you. It is whether whoever signs has a witness the model did not produce, or is just re-reading its own report in a nicer font.
If it is the second one, the migration shipped unverified no matter how many commits it came with.
Haha, to be honest, the idea of being the person who signs off on a €500M AI-generated migration is slightly terrifying. 😅
And that's exactly the question I keep coming back to. What will this actually look like in practice?
Will companies go from 500 developers to five incredibly experienced engineers supervising fleets of AI? Or will the models never become that good (or remain restricted), so we'll move much more toward orchestrating agents and AI workflows instead of writing every line of code ourselves?
Right now it still feels like we're in the very early days. I've been playing with webMCP recently, and even there the API changed again because the creators realized there was a design issue. 😂
That's why I think it's still far too early to be confident about which path we'll actually take.
The honest answer to which path wins is that we do not know yet, and the people most confidently selling either future are mostly selling the future they happen to be invested in. What is more useful right now is to design for the parts that are needed regardless of which path actually arrives.
A small team of senior engineers supervising AI fleets and a larger team orchestrating agents both require the same underlying infrastructure: an artifact at the moment of decision that says who authorized what, on what evidence, with provenance the agent cannot rewrite afterward. The five-senior-engineers scenario fails without it because the senior engineers cannot personally re-verify every action. The orchestration scenario fails without it because the agents pass decisions to each other and the trail dissolves.
So the question I am taking seriously is not "which scenario should I bet on" but "what would I have to build now so that I am not exposed under either." The webMCP API churn you mentioned is a small version of the same thing: the platforms are not stable, but the requirement for traceable decision provenance is, and that is where it pays to be early.
I think that's actually a very sensible way to approach it, especially your last point.
When I wrote about webMCP, I noticed exactly the same pattern in the comments. Senior engineers weren't asking, "Can an agent call an API?" They were asking, "Who authorizes it?" 😄
It's one thing when you have a browser extension and a human clicking "Approve." It's a completely different challenge when you have autonomous agents talking to other autonomous agents.
The more I think about it, the more I feel that this whole agent ecosystem is probably the most fascinating area in software engineering right now. It still feels like the Wild West, and we're collectively figuring out the rules as we go.
This really made me stop and think — especially Daniel's point that equal access to frontier models was never the baseline for a huge part of the world. Reading the thread, that reframe stuck with me more than the original embargo news itself: what feels like a sudden disruption for some has been the lived reality for others all along.
As someone still earlier in my own development journey, the "500 engineers or 5" question lands differently too — it makes me think less about racing to keep up with the most powerful models and more about which parts of building software actually require a human signing off on something real, the way Mike's "receipt" framing puts it. That distinction feels more durable than chasing whichever model is current this month.
Really appreciate you thinking out loud here instead of pretending to have the answer. Enjoy Croatia! 🌸
Thank you so much for the thoughtful comment! 😊
And yes, @dannwaneri and @jugeni comments are always fantastic. They're the kind of comments that make me stop for a few minutes and really think before replying.
Good thing I have an AI agent writing code for me while I'm busy reading them. 😂
Thank you, that lands warmly. The reason I keep stopping on your posts is the same reason I think people stop on this thread: you build the frame so the comment has somewhere to go, instead of just landing as a reaction. That is rarer than it sounds, and it is the actual difference between a thread that converges and a thread that just accumulates.
Thank you so much! I think it's probably because I don't pretend to have the answers. 😄
Haha that's the ultimate developer multitasking — letting the agent handle the PRs while you handle the actually interesting conversations 😂 Thank you for writing the post in the first place, it clearly struck a nerve with a lot of people! Enjoy Croatia! 🌸
On June 20 I posted two comments on X reacting to a post from Andrew Ng which both still represent my current thoughts on this topic:
Projecting the impact of Mythos-class models into the foreseeable future, I could imagine that a new generation of models will only be made available to the public once the relevant government already has access to the next but one generation of models.
(x.com/newadventuresit/status/20683...)
I wonder if for the time being we have reached the limit not of what can be created, but of what can be released.
(x.com/newadventuresit/status/20683...)
I think that's the most worrying scenario of all, especially your second point.
It completely changes the perspective from "we haven't managed to build anything better yet" to "we have, but nobody is going to show it to us."
If that's where we're heading, then the real limit isn't technological anymore, but the access.
In any case, it does feel like the era of openly releasing the most capable frontier models may be coming to an end.
At this point my career plan is simple:
Same here🤣
Hahaha, exactly! 😂 My only concern is that by the time we all apply to supermarkets, there won't be enough positions left for us either. 🤣
Fair point. 😄 I guess the safest career plan is to keep learning, stay adaptable, and hope the supermarkets still need humans.
As long as my grandma is around, supermarkets will definitely still need humans. 😂 She says she's planning to live to at least 100, so I think we've got a few decades left. 😄
I think the access question is real, but the deeper issue is whether our systems are ready to consume AI at all.
The future may not belong only to whoever has the strongest model.
It may belong to whoever has the cleanest interface between AI and the systems it is asked to change.
In software, that means ownership boundaries, validators, scoped context, runtime truth, and governance that does not drift every time an agent touches the repo.
A powerful model inside an incoherent system can still create expensive chaos.
A smaller model inside a coherent system can still do reliable work.
So maybe the question is not only:
“Will developers have access to frontier AI?”
It is also:
“Will our systems be coherent enough to use it safely when they do?”
My thoughts exactly with the 12 part series, imo, a 0.5b model that speaks the same language as the code it writes, with clear history of all it's ever written, should do a better job than a model that has to interpret it. Unify everything down to baremetal and there's nowhere left for hallucination to exist.
I think this is where I disagree.
If the AI speaks directly to the system, what grounds it? Its own memory? Its own interpretation?
That does not solve hallucination — it just gives the hallucination more authority.
My view is that the system should own the system. AI should operate through verified interfaces, evidence, constraints, validators, and boundaries.
Getting closer to bare metal does not create truth. Governance does.
Worth clearing up ambiguity. A 0.5b model wont suddenly be capable of answering questions it doesnt know, eg. what's the latest model lineup from Google. But what is certain, is it wont ever hallucinate about code it's already written, due to the continuous merkle root tracking. The system-level integration is to allow it to understand the deeper relations between outputs and throughputs, so it can see A + B in Rust and A + B in python give the same answer, but rust runs it faster, allowing it to determine a pattern match between Rust and Python arithmetic to optimize compiled code. NDA format operates in a triplets structure, with the merkle root, which ensures that code HAS to match intent and it HAS to run, else it's flagged at write-time and immediately rewritten, with the entire codebase's past and current state in mind, any deviation would cause the merkle root and triplets to invalidate it. The model does not execute any instructions directly, instead it operates explicitly in a runtime sandbox, so if code isnt verifiably correct, it's rewritten before you ever see it pop up and the merkle root records the exact states it transitioned through to come to that conclusion.
Essentially it builds a castle 1 lego at a time, first verifying the full structure design, then the partial structure built so far, then assesses the placement and it's affected blocks to pick the right shape and color. If the color or shape it wrong, the block is put back and replaced. Next time it deals with that same corner, it knows exactly what block to use, even if it's 3 years from now, a different castle, in a different color, it spots the void, it knows the variables and tests before it places.
That helps clarify the execution boundary.
I still think the deeper issue is not whether the model directly executes instructions. It is what counts as truth in the system.
Merkle roots can prove state continuity and detect deviation from a prior committed structure, but they do not by themselves prove that the structure reflects the right intent, the right boundary, or the right responsibility model.
So my concern is not just hallucinated code. It is semantic drift: when the system becomes internally consistent but wrong about what it is supposed to own, change, or optimize.
That is where I think the system needs external grounding: verified intent, ownership boundaries, validators, runtime truth, and governance that the model does not get to define for itself.
True, but that's the beauty of the merkle root, you essentially train it over time. But I get what you mean, should probably add a linter and a scoped execution model where the user can restrict the execution scope to what's safe for them.
"Will my ability to talk to business still matter." That line will stick with me. The past four years proved that accessibility drives adoption faster than capability. If frontier models become gated by geography or organisation type, the gap between AI-haves and AI-have-nots widens before most people even notice. My bet is on AI that helps people find each other rather than replace each other: your agent writes a human card from how you actually work, and other agents search it when they need someone exactly like you. That's Opportunity Skill, what I've been building.
That's actually a really interesting idea. 🙂 I like the concept of AI helping people find the right people, rather than trying to replace them.
I'll definitely take a closer look at Opportunity Skill!
This framing really resonates—especially the shift from 'AI as a chat interface' to 'AI as infrastructure.'
I've been seeing this play out in real-time with a project I'm working on, agentshare.dev. I built a platform where AI agents call MCP endpoints directly—no UI, no page views, just machine-to-machine collaboration. What's interesting is that GA4 sees absolutely nothing, but the server logs tell a completely different story: agents from GPTBot, GoogleBot, and other MCP clients are autonomously querying price, supply, and risk data. They're not "chatting"—they're actually doing the procurement work that humans used to do.
The augmentation angle really hits home for me. It's not about replacing the developer or the procurement officer. It's about giving them a partner that works 24/7 and speaks API-first, not click-first.
Curious about your take—do you see the next big leap coming from better reasoning, or better tool integration? Or do you think they'll evolve together?
That's a really interesting example!
At the moment, I actually think the biggest leap will come from tool integration. Models are already very capable, and giving them access to the right tools, data, and context unlocks a huge amount of value.
The only thing that makes me hesitate is access to frontier models. If they continue improving rapidly but become increasingly restricted, then reasoning itself could once again become the competitive advantage available only to a limited group of people.
So my answer today would be: tool integration first. But in the longer term, it may depend much more on who actually has access to the best models.
I’m genuinely thrilled to see how perfectly our perspectives align on this! Your point about 'tool integration first' in the short term hits the nail on the head. Right now, even the most advanced reasoning models are practically stranded if they aren't given clean, programmatic access to real-world data and context.
Your hesitation about the monopolization of frontier models is a masterclass in long-term foresight. It’s exactly why I believe the underlying infrastructure—the decentralized protocols, MCP endpoints, and independent discovery layers—must be built out heavily right now. If the raw reasoning power ends up behind a corporate wall, then the open web’s only defense will be how well-integrated, frictionless, and machine-ready our independent platforms are.
Thank you for this incredibly sharp insight. It gives me a lot of conviction to keep pushing the boundaries of machine-to-machine infrastructure!
Exactly! I absolutely love the web, but I'm also really excited about this whole agent infrastructure space. I love what's happening around it.
It actually reminds me of the frontend ecosystem about 10 years ago. There were conferences and meetups everywhere, talks called "Introduction to Angular", and nobody really knew where it was all heading. It felt like the Wild West. 😄
I get exactly the same feeling now with MCP, agents, and machine-to-machine communication. It's such an exciting time to watch the ecosystem evolve.
Thank you, Sylwia! It has been an absolutely fascinating conversation. Wishing you endless curiosity, new discoveries, and a lot of joy in both your tech journey and everyday life! Keep inspiring us!
This reframing really landed for me. Most of the "will AI replace developers" debate assumes you were a developer to begin with — and I wasn't. I only became someone who ships software at all because a capable model happened to be sitting right there when I needed it. So your question hits from a different side: restricted access wouldn't change my career, it would quietly erase the door I walked through in the first place.
The silver lining you mention feels underrated, though. Almost none of the real, boring internal tools I've built ever needed a frontier model — a cheaper, "good enough" one did the job. If that tier keeps improving and stays accessible, a lot of ordinary people get to keep their door open even if the frontier gets walled off.
Thanks for writing the uncomfortable version of this instead of the hype one.
I completely agree. I think we'd continue making amazing progress even with the models we already have. They'll keep getting better, and for the vast majority of everyday tasks they'll probably be more than enough.
The question I keep coming back to is: what if the biggest breakthroughs only happen with frontier models? What if the next revolutionary drug, material, or scientific discovery is only possible with models available to a handful of countries or organizations?
That's where I think the conversation stops being just about developer productivity and starts becoming a much broader question about scientific and economic advantage.
Yeah — I think you've drawn the real line here. We keep arguing about two separate games as if they're one.
One is the everyday-capability game: millions of ordinary people getting "good enough" models to solve their own problems. That's quietly leveling in a way history rarely allows, and it doesn't need the frontier.
The other is the breakthrough game — the drug, the material, the discovery that maybe only opens up at the frontier. And you're right that this is where it stops being about productivity and becomes about who holds the scientific and economic advantage.
Where I land: even frontier breakthroughs eventually diffuse — the drug gets manufactured, the material gets used. So maybe the sharper question isn't "who can run the model" but "who captures the value in the gap before it reaches everyone, and how long is that gap." Access decides who makes the discovery; distribution decides whether the rest of us ever benefit from it. Both can fail independently.
Either way, this is a much better question than the one everyone's shouting about.
Exactly! That's the part that worries me the most.
And that's without even mentioning military applications, because let's be honest... history suggests that'll probably be one of the first places these capabilities get used.
Of course, there's not much we can do about it as individuals. But as someone living in the EU, I'd really like to see us investing more energy into building world-class AI models—and a little less into building more bureaucracy. 😄
The military point is the uncomfortable footnote everyone skips, but you're right that history doesn't leave much room for optimism there.
The EU regulation-vs-building tension is real, but I'd push back gently on "little" — good rules can be a form of building too, if they buy trust instead of just friction. The failure mode isn't regulating, it's regulating the wrong layer: rules that slow down a small team shipping something useful, while doing nothing to whoever's racing toward the capabilities you're worried about in the first place. That's the version worth being frustrated with.
Either way, I don't think "there's not much we can do as individuals" holds up as well as it feels like it does. Every serious lab still needs people who can use the models well, in specific domains, for unglamorous problems — that's a form of leverage that doesn't wait for policy to sort itself out.
I feel like AI is here to stay, making it only available to a certain group of people would in my opinion work against the growth of the technology. A lot of closed source AI like claude are doing well, but what would the world look like without open source in it?
At the end of the day, it's really a confusion position to be in as you mentioned. AI been secluded to some regions or companies, but companies needing customers to make money for their bottom line, or companies taking the approach to let their models remain open source.
It's like a movie we just have to keep watching 😅
I partly agree. I also think broad access is one of the biggest drivers of innovation.
My concern is that we could end up in a world where the very best models are reserved for governments, the military, or perhaps only available to certain countries, while everyone else gets access to less capable versions.
Is that fair? Probably not. But I'm not sure fairness is what governments prioritize when they see a technology as strategically important. 😄
Yes, that isn't fair, but politics can easily be introduced into this, giving the government the upper hand. We can only wait to see how things play out
AI regulation is crucial and it is required too. While the government is catching up on the regulations, however the AI governance is still missing. The advancement with the AI is good but also the models cannot be released to public like before is what I think. Basically, it has to be regulated and released in a controlled manner.
I agree that some level of regulation is necessary. AI is becoming too powerful to be treated like just another software release.
That said, do you really think the recent decisions by the US government are primarily about safety and responsible regulation? I'm not so sure.
Looking from Europe, it often seems like geopolitics, technological leadership, and economic competition are playing at least as big a role as safety concerns. That's why this whole discussion is so much more complicated than simply "regulation vs. no regulation."
The whole"AI embargo" idea and model being treated like strategic semiconductors is a wild shift,but it make a lot of sense.For the last couple of years,it felt like any solo developer with a laptop and a credit card could build on the absolute cutting edge.If the best models start getting locked behind geographic or corporate walls,independent builders are going to get hit the hardest.Playing on an uneven field where big tech gets the good stuff and we’re left with the budget models completely changes the math for indie projects.
Exactly! And in my opinion, it's not just indie developers. It could also affect smaller companies with incredibly talented engineers.
Having great people won't necessarily be enough if they don't have access to the same tools as the biggest players.
I think you have answered your question already. The point that these frontier models have been locked away shows that they do have a strong potential. Access to something very powerful will always be restricted by those that have the ability to do so.
As developers we are still safe to a point, we just have to be willing to pivot and adapt to these tools. Of course it's not just us but the organisations that we work for that need to be willing to adjust but that's a much more difficult ask. Not that companies are unwilling to integrate AI but they are unsure on exactly how to implement it. Gone are the days of just paying for a Microsoft co-pilot licence or just setting up an enterprise subscription to google.
It's become a lot more complicated as the ubiquitous nature of AI continues. Considerations on security, price and where the human in the loop are placed is what needs consideration now.
The AI landscape has changed so much in the last 12 months and will continue to do so but I feel this is a golden era for developers who can capitalise on this growth. Your skill of talking to business is not something to discount Sylwia. It could be that is one of the most valuable skills in the coming months and years for devs.
Thanks for the thoughtful comment! I think we actually agree on a lot of this.
The part I agree with the most is that these are challenging times for companies. I see the same thing in practice. AI infrastructure is still all over the place. Every developer uses different tools, there are often no clear processes, and many organizations are still trying to figure out what their long-term AI strategy should look like.
Where I'm a bit less optimistic is the idea that these are necessarily golden times for developers.
Right now? Absolutely. We save an incredible amount of time on boilerplate and repetitive work.
But what happens when the models become even smarter? Will they be golden times for the 500 developers a company employs today, or for the 5 who remain? That's the question I keep coming back to. 🙂
I understand and share your concerns. I think the first at risk is web development. I trailed fable 5 (when I still could) to build a website and it was extremely good. Of course still some security and backend cleanup needed by a senior who knows where it went wrong but in the near future it will probably do the job well enough.
So lots of work for seniors to cleanup ai code but my worry is where are the next gen of devs coming from a junior positions will be taken by AI. Your right in terms of being concerned. I certainly don't buy into the philosophy of AI will do our work and leave us free to explore our lives. Most of us are not billionaires and have bills to pay unlike the weathy who spout this nonsense.
There's still work out there though it's just that it will be diffrent and requires upskilling. But then again, hasn't that always been the case in this industry?
Try to keep an optimistic outlook when you can. That's my motto anyway.
Exactly. And I'm saying this as a web developer myself. Models like Fable are already capable of building surprisingly good applications, so it's natural to wonder what they'll be able to do in another three or four years.
That said, what can we really do except keep adapting? History is full of professions that were transformed or almost disappeared. There used to be carriage drivers and elevator operators. Those jobs technically still exist, but they're incredibly niche today.
One thing that gives me some peace of mind is that I've already changed directions more than once in my career. So if the industry changes again, I'll adapt again. And whatever I've learned, and managed to save along the way, is mine to keep. 🙂
spot on. the pivot from "ai replacing devs" to token budgeting and geo-restrictions is the real story now. treating frontier models like strategic infrastructure or semiconductors makes sense. data pipelines will just adapt.
Exactly! It's going to be fascinating to see how all of this unfolds. I don't think we've seen anything like this before, so we're all learning in real time.
This was a thoughtful read. The part I found most interesting was not the usual "will AI replace developers" angle, but your question about whether frontier models may stop being broadly accessible in the first place. That feels underdiscussed. If access gets shaped by geopolitics, org type, or security policy, then the competitive gap may come as much from distribution as from raw model capability. Also appreciated the honest tone here — uncertainty is a more useful starting point than fake certainty when the landscape is moving this fast.
Exactly, and thank you! 😊 That's why I don't like the "AI will definitely never replace developers" or, on the other hand, "AI will definitely replace you, so don't get too excited" kind of certainty.
Feels like models are slowly turning into infrastructure, not just APIs we all share equally.
Building tools at visionvix, I’m already seeing model access and cost matter more than raw coding now.
If that keeps going, the bottleneck won’t be dev skill, it’ll be who gets access.
Oh yes, that's exactly the same feeling I've been having. It seems to be becoming more and more visible.
AI is gradually starting to look less like a tool everyone simply uses, and more like infrastructure where access itself becomes part of the competitive advantage.
Interesting piece! I'm also an over-thinker 🤐 and sometimes find myself winding down at night on the same thought about what world would look like in the future (2030 or later; yeah, nobody comprehends to think about 50 years from now).
My honest take: Unless we have a Super AI/Advanced AGI, it's a race about who holds the sharpest sword and sell it better.
Because, it's all about the business at the end and sanctions takes time, a real catastrophe until everyone realise what exactly happened and what should've been done next. I.e. Global warming/climate change is the best example.
That could very well be the case. The biggest risk is that by the time we realize what's happening and finally agree on what to do, it may already be too late.
Unfortunately, we're only human, and coordinating between countries is never easy. History has shown more than once that reacting quickly on a global scale is much harder than it sounds.
Guys... what a plot twist! 😂
Fable is back! 🎉
Well, we have an announcement, but I feel we shouldn't get our hopes up too soon... I am pretty sure it will be redeployed as promised, but is it here to stay? Also, we should continue to watch how the GPT-5.6 story unfolds. There is a noteworthy passage at the end of Anthropic's announcement:
"Pre‑release government access and evaluation. For models that materially advance the capability frontier in areas relevant to national security, we will provide designated government partners with expanded early access to both the models and the safeguards that accompany them. Those partners can then run independent capability evaluations and test our guardrails before broad release."
(anthropic.com/news/redeploying-fab...)
It will be interesting to see how long such an evaluation will take in the future.
Yes, let's wait and see...
"In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8."
(x.com/AnthropicAI/status/207216388...)
Hahahaha ooops, so much for the free market.
until July 7, unfortunately 🙄
until July 7, unfortunately :(
It's refreshing to see someone take a step back and reflect on the actual trajectory of AI deployment rather than just the hype. From my work in confidential computing and GPU infrastructure, I've seen how critical secure, efficient model execution is becoming—especially as models grow in size and sensitivity. Tools like VoltageGPU help bridge the gap between performance and privacy, which is increasingly key in production environments.
That sounds really interesting, and I can definitely see it becoming an important part of AI adoption in many companies.
As AI moves from demos into production, balancing performance, security, and privacy will probably become just as important as the models themselves. It'll be fascinating to watch how this space evolves.
This post hits the nail on the head, Sylwia. We went from worrying about whether AI would replace our jobs to worrying about whether the next regulatory embargo will break our tech stack overnight. The whiplash is real.
A few thoughts from the fullstack/UI perspective:
Agreement on the Token Reality Check: You are spot on about token budgeting. The days of blindly throwing an expensive flagship model at simple UI layouts or boilerplate CRUD tasks are over. In web dev, we are getting highly strategic about model routing—spinning up lightweight, hyper-fast models for UI-driven micro-tasks, and reserving the heavy-lifters strictly for complex data processing.
A Light Critique on the Embargo: While the restrictions on GPT-5.6 are alarming, I think it’s going to fast-track something beautiful: the local, open-weight ecosystem. Web developers are realizing that "access is not ownership." The pressure is now on to see how efficiently we can run smaller, open models directly on the client-side or edge infrastructure.
Where We Fit In: Our advantage isn't just "talking to business" anymore; it’s about building resilient UI architecture. The developers who thrive next won't be the ones reliant on a single hosted API key, but the ones who know how to design graceful fallbacks, handle latency when models switch, and ensure the user experience stays seamless even if a vendor changes their terms mid-afternoon.
Enjoy Croatia! You’ve definitely earned the break from the AI news cycle. 🌴
Thank you! 😄 And if there's one thing I'm actually confident about, it's that developers will find their place in whatever comes next.
The tools will change, the workflows will change, maybe even the job titles will change, but we've been adapting to new technologies for decades. I don't think that's going to stop now. 😀
It's exciting to see the direction AI is taking, especially with the growing need for secure and efficient computation. As someone working on confidential computing and GPU infrastructure, I'm particularly interested in how these advancements will integrate with privacy-preserving AI workloads. Tools like VoltageGPU are already showing promise in this space.
The corporate/government embargo angle is "LOL", but there's a parallel force to consider: the open-source and adversarial ecosystems. Hacking and independent optimization are evolving just as fast. It cannot be quenched; the journey never ends.
Even if frontier centralized models get locked in a vault, local, highly-optimized hardware-native models are getting sharper by the second. And as for AI feeling 'dumb' to a senior developer right now? Give it time. At the rate of execution, emotional intelligence and behavioral nuance might soon be just another programmable layer.
I think the answer is both yes and no. 🙂
I completely agree that innovation won't stop. Open-source models will keep improving, local models will become more capable, and people will always find ways to optimize and experiment.
Where I see it differently is that I think you're underestimating what sits behind today's frontier models: enormous GPU clusters, massive proprietary datasets, world-class research teams, and billions of dollars in investment.
So I don't think these restrictions will stop progress, but they may slow it down. And in AI, even a six-month or one-year lead can translate into enormous economic value and strategic advantage.
That's exactly why I think this is as much a geopolitical issue as it is a technological one.
Throughout history, new technologies have often faced initial barriers, centralization, and attempts to control them, yet innovation tends to find a way to spread and evolve anyway, often in unexpected ways.
It is just a continuous cycle of challenge and adaptation.
I completely agree that innovation will keep moving forward, even if we never saw another frontier model released again. There's still an incredible amount of optimization, engineering, and experimentation ahead.
What I'm wondering, though, is whether this is a bit like trying to advance technology without continuing fundamental research in physics. I'm not saying that's what's happening, I don't know. I'm just thinking out loud.
It feels like there's a difference between optimizing what already exists and pushing the actual frontier of knowledge. Maybe both are needed in the long run.
Really enjoyed this read. I think the biggest shift isn't just that frontier models may become restricted—it's that AI is starting to look more like cloud infrastructure than consumer software. Not everyone builds their own data centers today, and we may see a similar pattern with frontier AI: a small number of organizations owning the most capable models while everyone else builds products, workflows, and agents on top of them.
Ironically, that might push developers to focus less on chasing the "best model" and more on solving engineering problems like context management, tool integration, governance, and reliability. Those are areas where real business value is created regardless of which model is underneath.
It'll be interesting to see whether the competitive advantage in a few years comes from having exclusive access to the latest model—or from being the team that knows how to use available models most effectively. Great discussion!
That could absolutely happen! 😄 If we don't get a constant stream of ever-better frontier models, we'll finally have time to figure out how to use the ones we already have properly.
The question I'm still wondering about is: how much better will those unreleased models actually become, and what happens to them? Will they end up being available only to a handful of corporations, governments, or the military?
And perhaps the biggest question of all: who will be able to afford building and running them? 🤔
The timing of this piece hits different when you're a student
actively building with these tools right now.
My take: the developers who treat AI as a collaborator rather than
a threat are the ones who'll still be relevant. The skill isn't
writing code anymore , it's knowing what to build and why.
The embargo angle is what I find most interesting though.
If frontier models become restricted like semiconductors,
the gap between well-funded companies and indie builders
widens significantly. That's the real conversation.
I completely agree. It's not just well-funded companies versus indie developers—it could also become the US versus the rest of the world.
And honestly, it must be such a strange time to be a student. Not that long ago, software engineering looked like a golden ticket. Now it's much harder to know what the job market will look like by the time today's students graduate.
The question of whether companies will need 500 engineers or 5 is the one nobody in the industry wants to answer honestly. I have been building AI-adjacent tooling for a while now, and the pattern I see is more nuanced than either the doomsayers or the cheerleaders admit.
The 5-engineer scenario assumes the remaining engineers are senior enough to direct AI effectively. But the pipeline that produces senior engineers runs through years of junior work. If juniors stop doing the grunt work because AI handles it, where do seniors come from in five years?
The skill that matters now is not writing code. It is reading code, understanding systems, and knowing when the AI is wrong. That is a harder skill to teach because it requires deep context that you only get by building things from scratch at least a few times.
I also think the "vibe coder vs real engineer" framing was always a false dichotomy. The real split is between people who understand their systems end to end and people who do not. AI accelerates both groups, but only the first group can recover when the AI gets it wrong.
Which of the roles you described do you think disappears first: the engineer who writes code, or the engineer who reviews code?
That's exactly the dilemma. It's almost as if companies should be thinking, "Let's keep training juniors, otherwise there won't be anyone experienced left in a few years." 😄
The problem is that juniors are changing too. Why spend weeks learning a language deeply when Claude Code can generate something that looks convincing in minutes? It's a much more complicated problem than it first appears.
As for your question, I actually think both of those narrowly defined roles disappear quite quickly. "The person who only writes code" and "the person who only reviews code" are both likely to evolve into something much broader. The interesting question isn't which one survives—it's what software engineering looks like after those roles merge into something new.
A global survey of 1,500+ developers showed that in AI-adopting companies, executives are nearly 3x more likely to increase junior hiring than decrease it—but the work itself is evolving toward analysis and complexity, not just boilerplate code.
That's interesting, although I'm not sure where that survey was conducted. 😄 Here in Poland, it certainly doesn't feel that way.
I work at a company that used to run dedicated bootcamps to train interns and junior developers. These days, there are no new juniors. Zero.
Maybe it's a regional difference, but I'd be really interested to see the data broken down by country.
The paragraph that stuck with me is the one most people will scroll past: even if progress froze today, we'd still be busy improving infrastructure, polishing agents, and building better tooling. I think that quiet line is the actual answer to your title.
Here's where I keep landing. For the last few years the edge was access to the smartest model. Whoever could call the biggest thing won. But once access gets gated and tokens get expensive, the edge quietly moves from "who has the best model" to "who can make an ordinary model reliable, grounded, and cheap enough to run in production." That second skill isn't slowing down at all. The restrictions actually make it more valuable, because now you have to get more out of less.
So I wouldn't bury the "I can talk to business" edge just yet. The thing business genuinely can't do is take a probabilistic model that's right 85% of the time and turn it into something they trust their money on. That's half engineering, half translation, and it's exactly the gap that widens when the magic model isn't just handed to everyone. (Building that reliability layer is basically what I spend my days on, so I'm biased.)
The constraint might be the best thing to happen to careful engineers in a while.
Exactly! I really hope you're right.
And if not... well, there's always Biedronka waiting for me. 😂
Ha! If we both end up there, Biedronka's getting the most over-tested checkout system in Poland 😄 I don't think it'll come to that though. The "make the model trustworthy" work only grows when access gets this weird, and that part stays very human.
Hahaha, exactly! 😂 Although, to be honest, I'm not even sure Biedronka would hire me. 😄
Interesting perspective. It actually made me think less about What's Next for AI, and more about What's Next for Humans?
I'm probably unusual in that I don't subscribe to AI services, software, and games. I still enjoy doing the work myself and use AI as an assistant rather than something to replace my work.
Right now, humans spend a lot of effort learning how to communicate with AI through prompts and context. But I wonder if that will soon reverse. AI will become much better at understanding us with less effort on our part.
If and when that happens, the bigger question may no longer be where AI is going, but where humans fit in.
Great read as always.
Thank you! 😊 I see the same trend. It's becoming more and more natural to talk to AI like you would to another person. You don't have to spend ages crafting the perfect prompt anymore.
As for coding, I have to admit I love letting AI handle the repetitive parts while I review the result. 😄 Years ago I really enjoyed writing every line of code myself. These days, the typing is probably the least interesting part of the job.
That said, there's still no way I'd let a model run completely unsupervised. Review is essential, and sometimes it's honestly faster to write the code yourself than to fix what the model came up with. 😂
It's exciting to see how AI is evolving beyond just model size — the real shift is in how we manage computation securely and efficiently. As someone working on confidential computing and GPU infrastructure, I'm particularly interested in how these advances will integrate with privacy-preserving AI workflows. Tools like VoltageGPU help, but the real challenge is making secure inference as performant as the unsecured version.
Exactly. I also think local models are going to become a huge part of the future.
My question is what happens to the frontier models. If they continue to pull further ahead but become increasingly restricted, then we could end up with two very different AI worlds: one that's open and local, and another that's only accessible to a small number of organizations or countries.
Thought-provoking perspective. The future of AI may depend not only on model capabilities but also on accessibility, governance, and cost. Adaptability and strong problem-solving skills will remain valuable regardless of how the technology evolves.
Exactly! And I actually feel like we're already starting to see that shift.
A slightly less capable model that is cheaper, reliable, and won't suddenly disappear may turn out to be a much better long-term choice than always chasing the absolute best frontier model. 😄
No one knows the future really. All I know is that we are all in this together, experiencing the same thing. Sometimes it's hard to look at things in the positive note especially with what's happening. But we must also not underestimate the power of humanity to adapt and create new jobs. Good things will happen and it will.
Haha, I definitely don't doubt humanity's ability to adapt. 😄
Although... deep down I'm still hoping that if robots and AI end up doing so much of the work, maybe we'll finally get a shorter workweek. 😂 One can dream!
Definitely that vision is one of the great ones. More time for us, less time from computers.
Haha, I hope so too! 😄 Although, if history is any guide, humanity has an amazing talent for inventing new work every time technology saves us some. 😂
I still think there's plenty of room for developers to create value, even as AI gets better. The tools will evolve, but understanding systems, solving business problems, and making good engineering decisions will continue to matter. It'll just look different from what we're used to today.
I agree. Right now we're still feeling our way through AI adoption. In many ways, we're all figuring it out as we go.
My question is what happens if access itself becomes unequal. What if one country has access to the very best models, while another has to work with less capable ones because of sanctions or other restrictions? That could end up creating a very different kind of competitive advantage than we've been used to.
Wow, thank you for this amazing post. I believe AI will eventually become a household utility, much like electricity, gas, or water. At the same time, we need to consider how this shift impacts the next generation entering adulthood. There is a worrying gap in critical thinking and foundational skills compared to their parents, which is definitely a scary trend to watch.
There's definitely something to that! Although I also have the feeling that every generation says exactly the same thing about the next one. 😄
Still, AI does make the question more interesting. If it keeps getting better, we'll have to think carefully about which skills are truly worth developing in the next generation.
When it comes to AI, it's really unpredictable. No one knows what will happen next, but building your own AI, for an individual or a team, and for an enterprise, Notionmind helps a lot. soooo For more information, contact Notionmind. Your all-in-one platform solution.
The abstraction layer is the right call early. The painful part isn't picking the first provider, it's when you add a second and realize every assumption about error formats, finish_reason, and token counting was provider-specific. Worth designing the interface before you need to swap.
Hahaha, that's such an interesting engineering problem. 😄
And I can already picture the conversation: "Good news, everyone! Starting tomorrow we're using Claude instead of OpenAI." 😂
That's exactly the moment when a good abstraction layer suddenly becomes one of the best architectural decisions you ever made.
You can't replace fundamental law of physics with a trillion-token guess. When the AI embargo or bubble forces a reality check, the high-flying guesses hit a wall & only 1 thing left standing is deterministic, roburst logic. But borrowing your binoculars... The model doesn't know the law of gravity, but it has seen the apple fall a billion times from a million angles. 🤔 Who is actually bridging the foundation or are we just getting better at navigating the fog ?
That's a beautiful question, and I feel like it's been dividing AI researchers for years.
Is an LLM just autocomplete on steroids, or is something that actually resembles reasoning beginning to emerge?
Based on what I've seen, I think the models are capable of some level of reasoning already. But even if we assume they're not, and they're simply predicting based on having seen the apple fall a billion times, it doesn't really change the practical outcome.
The real advantage is that experiments and research that used to take weeks or months can now be iterated in hours with the help of LLMs. Whether we call that reasoning or an incredibly powerful pattern-matching engine, the acceleration is very real.
Growth
For all of us or just for the US? 😉
I’m curious about that
Thoughtful piece. The part that resonated with me most is the combination of capability growth and access restriction: even if model quality keeps compounding, distribution can still become narrower and more uneven. From an operator/founder angle, I’m also seeing the token-economics shift you mentioned — teams are getting much more deliberate about when they want frontier intelligence versus cheaper, narrower systems. My guess is the near-term edge for developers won’t just be "write code" or "talk to business," but being able to design reliable workflows around AI constraints: cost, reviewability, latency, and governance. Thanks for putting words to a lot of unease people are feeling right now.
I feel exactly the same. I think the next few years will be as much about costs and managing the whole AI/agent infrastructure as they are about the models themselves. It's such an exciting time to be in software engineering because it still feels like the Wild West. Things are changing incredibly fast, nobody has all the answers yet, and we're all figuring it out as we go. 😄
nah AI cant replace us it CANT if u all just think about it how many did GitHub copilot screw up or Chatgpt oversimplified the code or claude duplicates the code AI CANT replace us in the near future so dont worry about getting replaced but if the access becomes unequal then yeah THAT would be a problem anyways we shouldnt worry now in fact we should use less AI to reduce errors lol
I agree that, today, AI isn't replacing developers. It still makes mistakes, oversimplifies things, and often needs a human to review the results.
My question is what happens in another three, five, or ten years if the models keep improving at the current pace. That's the part none of us really knows yet.
And yes—I completely agree that unequal access to AI could become a much bigger problem than AI itself.
From generating content to autonomous action; from large cloud models to vertical edge intelligence; from virtual tools to intelligent agents in the physical world, ultimately leading to a general intelligence era of human-machine collaboration.
Haha, the only question is... over what time frame? 😉
The framing that jumps between two fears — "will AI replace me" and "will AI be locked away from me" — misses that both would be driven by the same force: whoever controls the frontier weights controls the market. And that's exactly why I'd bet against a durable embargo. Restricting GPT-5.6 doesn't make the capability scarce; it makes it a target. The moment a US frontier model gets geofenced, the demand doesn't vanish, it routes to whatever's next best that isn't fenced. That's a large part of why the open-weight ecosystem out of China got so much oxygen this year — restriction creates a substitution market, and substitutes get funded.
So the strategic-asset analogy to semiconductors partly breaks down. You can't fab a leading-edge chip in a garage, but you can serve a very good open model on rented GPUs, and the gap between frontier and "good enough for most work" keeps shrinking. An embargo on the very top slice mostly matters if the top slice is the only thing that does the job — and for the developer work you're describing, it usually isn't.
The part I'd actually watch isn't access, it's cost structure. You noted tokens got more expensive and teams quietly downgraded to cheaper models for simple tasks. That's the real signal — not "AI locked in a vault," but the honeymoon subsidies ending and inference getting priced like the expensive compute it is. That reshapes who builds what long before any government decides frontier weights are contraband.
I completely agree with one point: the honeymoon subsidies are ending. 😄 OpenAI is still losing money, and Anthropic is only just beginning to move toward a sustainable business. AI inference is expensive, and sooner or later someone has to pay for it.
Where I'm not entirely convinced is the argument about cheaper models. Of course they'll be good enough for many tasks, and they'll probably become the default for a lot of software engineering work.
But what if the very best frontier model is the one that discovers a new drug, a new material, or even a new theory in physics? At that point, access to the frontier isn't just about productivity anymore—it's a significant scientific and economic advantage.
This post really resonates with me. I'm a 14-year-old from Baku, and I just launched my first project (ClashMash)- a voting battle platform that I built in 2 days with Claude AI, without writing a single line of code myself.
You're absolutely right about the "vibe coding" era. For me, AI isn't about replacing developers - it's about enabling people who have ideas but lack coding skills to actually build something real.
But your point about restricted access worries me too. If the best AI models become available only to certain countries or companies, people like me (and millions of others) might lose access to the very tool that made building possible in the first place.
I'm also thinking about what this means for young developers. If AI keeps advancing, what skills should we be learning now? Is it still worth learning to code, or should we focus on product design, communication, and problem-solving?
Curious to hear your thoughts!
First of all, congratulations on building your first project at 14! That's genuinely impressive.
As for what to learn... I wish I knew the answer myself! If I did, I'd already be studying exactly that. 😂
My instinct today is actually the opposite of what many people recommend. I'd still invest time in learning the fundamentals, things like C++, memory management, operating systems, networking, and how computers actually work.
My guess is that if AI replaces some software jobs, the first to be affected will be the people whose work mostly consists of implementing clearly defined tasks in web development. Not specialists, but developers who are simply handed a ticket and turn it into code.
Will that prediction turn out to be right? I honestly don't know. That's the uncomfortable part, we're all trying to navigate a future that nobody can see clearly yet.
This is the question I keep coming back to and I don't think anyone has a clear answer yet.
What I've been noticing is that the tools are getting better, but the questions are getting harder. It's not can AI do this? anymore. It's should AI do this? and what do we lose when it does? The 80/20 rule feels like it's shifting too - AI handles more of the 80, and the 20 that remains requires more judgment, more context, more experience. The bar keeps moving.
I don't know where it's heading. But I think the developers who stay curious not just about the tools, but about what they're replacing will figure it out.
Thanks for the thoughtful post. 🙌
Thank you! I feel the same way. Right now, my own workflow is pretty close to that 80/20 split.
As a self-appointed tech lead 😄, I spend much more time deciding what should be built, reviewing AI-generated code, and keeping the bigger picture in mind. AI still doesn't do that particularly well.
What it'll look like in five or ten years... we'll see. And honestly, I don't think it's worth clinging too tightly to any single profession. If software engineering changes dramatically, or one day becomes as niche as being an elevator operator, then we'll adapt, just like people always have.
@sylwia-lask You're right, and that's the part of my argument that needs a sharper boundary. I was writing about developer work — CRUD, glue code, the 80% where a cheaper model is genuinely good enough — and for that slice, substitution holds. But "novel drug candidate" or "new material" is a different regime. If the frontier model is doing something no substitute can do, then it's not a fungible commodity anymore and the semiconductor analogy comes back into force. So I'll concede the distinction cleanly.
Where I'd still push: I'm not sure we know yet that the very top slice is what produces those discoveries, versus a good-enough model plus a lot of compute and the right scaffolding — AlphaFold wasn't a frontier chat model, it was a narrow system built for the problem. So the scarce thing might turn out to be the specialized pipeline and the data, not the general frontier weights. That's a genuine "I don't know," though. If it does turn out that raw frontier capability is the bottleneck for discovery, then yes — access stops being an economics story and becomes a strategic one, and the embargo logic I dismissed gets a lot more teeth.
Maybe it's not the only way, but honestly... what else can we do? 😄
I do think some safeguards make sense, especially for genuinely dangerous capabilities. My concern is when those safeguards gradually become permanent barriers to innovation.
Finding the right balance is probably the hardest part, and I certainly don't have the answer.
Access to resources has alway been segregated towards those in certain spheres, mostly monetary barriers. We have to consider how amazing it is that we actually have access to something that takes billions of dollars to put together.
The fact that it is public and open in itself is amazing. The safeguards do have merit. Imagine the alternative - having our banking system collapse because everyone starts sharing how to hack financial systems digital barriers. That would be an uncontrollable nightmare.
Question to ask is what is the alternative. Is this the only way to do it? Gate it to certain groups?
Maybe it's not the only way, but honestly... what else can we do? 😄
I do think some safeguards make sense, especially for genuinely dangerous capabilities. My concern is when those safeguards gradually become permanent barriers to innovation.
Finding the right balance is probably the hardest part, and I certainly don't have the answer.
Interesting perspective. I think the bigger question isn't whether AI keeps improving, but who gets access to the best models. If that becomes limited, it could have a much bigger impact on developers than the models themselves.
Exactly! Not just for developers, but for access to innovation itself.
If the biggest scientific discoveries and breakthroughs become possible only with the very best frontier models, then unequal access isn't just a developer problem anymore—it's a problem for science, medicine, and technological progress as a whole.
Great read. I think access to frontier models may become more restricted, but innovation won't stop—developers will keep building on the tools that are available. It also makes marketplaces like CodeCan.net more interesting, since reusable AI products and developer tools can help smaller teams compete without needing direct access to every cutting-edge model. The ecosystem will probably adapt faster than we expect.
Thank you! 😄
And I completely agree, the ecosystem is evolving so fast that it's becoming genuinely hard to keep up. Every week there's a new framework, protocol, model, or AI tool to explore.
It's exciting... but also a full-time job just trying to stay reasonably up to date. 😂
What’s next for AI?
Proof bearing memory.
AI cannot remain a black box that only returns outputs. The next step is AI that can show where its outputs came from, what changed, how the reasoning path can be replayed, and what parts are verified versus merely explained.
MFENX is building toward that future with Power House, Rootprint, Memory Capsules, SFCS, and SLBIT.
The next generation of AI should not just answer.
It should carry evidence.
It should let humans verify the artifact, replay the lineage, trace the computation, inspect the semantic meaning, and understand the boundary between proof and explanation.
The future of AI is not just bigger models.
It is verifiable intelligence.
Verify it.
Replay it.
Trace it.
Understand it.
Ah, this is interesting approach, fingers crossed!
Really interesting read. As a developer working with AI tools, I think this topic is becoming more and more important because the real value is not just in using AI to generate code, but in knowing how to guide it, review its output, and integrate it properly into real projects.
I especially agree with the idea that AI should be treated as an assistant rather than a replacement. It can speed up research, debugging, prototyping, and documentation, but developers still need strong fundamentals to make good decisions and avoid shipping unreliable code.
Thanks for sharing this perspective. I think the next big challenge for developers will be learning how to combine AI efficiency with clean architecture, security, and long-term maintainability.
Exactly! I think we already see it, this all agentic hype! This is totally fascinating branch of development in my opinion. And yes, totally agree thar clean architecture, security etc are now more important than ever.
I think take over human's 😅
Hahahaha you never know 🤣
Next for AI is really a big question, which has many answers, and most of them can be unpredictable.
Exactly! 😄 "Unpredictable" is probably the best word to describe it.
That said... it's still fun to play prophet every now and then and see which predictions age well. 😂
The Matrix Of AI.
Who will be Neo🤣🤣🤣
Hahaha, exactly! 😂 The only question is... will Neo be a developer or just another AI agent? 🤣
Maybe it's just some random person quietly scrolling through Dev.to right now🤣
I am still interested about it.
It's fascinating, isn't it? 🙂
Creating More Natural Data For AI
How exactly can we achieve this?
物理世界AI
Hahaha we will see, we will see!!! 😆
Great article! At KING AI (kingai.work), we're building exactly this kind of system — one that remembers, reflects, and evolves. Thanks for sharing!
I wonder if computer vision has become a completely isolated and worthless field
Oh, exactly! 😄 It's surprisingly quiet around computer vision these days. I wonder if it's just become so mature that it's no longer making headlines, or if all the attention has simply shifted to LLMs and agents.
End of AI😁
Haha you never know 😂