2025 - My Look Back

00:00:00: So another year's almost over and you know what that means,

00:00:04: time for me to look back at it, and yeah, obviously that video

00:00:08: will be about AI. Not just about AI,

00:00:11: but a lot of AI because, you know, it's the thing that

00:00:15: influences everything and has an impact on everything,

00:00:19: And believe me, I'm not a fan of having, uh,

00:00:23: AI everywhere. But it is what it is. But yeah, let's take a step back.

00:00:27: You, you might remember this year was meant to be the year of

00:00:31: AI agents. It was the term you heard and read

00:00:35: everywhere, especially at the beginning of the year.

00:00:38: You still, of course, hear and read it everywhere,

00:00:41: year started, everywhere on any- every website you could read,

00:00:45: "2025 is the year of AI agents." And of course, an

00:00:49: AI agent was basically anything that used

00:00:52: AI for anything. You wrote a script that uses

00:00:56: AI to summarize a PDF? Congratulations, you built an AI

00:01:00: agent. Now, of course AI has made

00:01:04: significant advances to some degree when it comes to

00:01:08: agentic workflows. Now, what is an agentic workflow?

00:01:11: Again, depending on how you define it, it's anything that uses AI,

00:01:15: but I would say it is any kind of workflow where AI

00:01:19: makes decisions. And keep in mind, we're still talking about

00:01:22: token generators, but of course those tokens can describe

00:01:26: tools the AI wants to use. And

00:01:30: one big category of agents that you definitely

00:01:34: know is of course all these coding assistants.

00:01:37: I mean, they did not start this year, you know that as well as I

00:01:41: do. But in this year, we saw lots of development on

00:01:44: Cursor, we saw the release of Claude Code, and all these

00:01:48: tools are of course about giving us developers

00:01:52: assistants that help us write code, that can

00:01:56: generate code based on our instructions or, if you're not

00:01:59: that much into writing code, they can of course generate

00:02:03: code based on pretty vague instructions and you can enter the

00:02:07: territory of live coding. And live coding

00:02:11: definitely was another big theme and thing this

00:02:15: year. The term was coined by Andrej Karpathy

00:02:19: in 2025 and as he described it, the idea

00:02:23: really was that you don't really read the code, that you

00:02:27: go with the vibes, you just basically see if it works

00:02:31: or it doesn't and you just tell the AI what to change or make it

00:02:35: work and... Well, that is how you can build software.

00:02:38: Now, to be fair and very clear here, he made it very

00:02:42: clear that he still also writes code himself.

00:02:46: But of course, you could use live coding for pieces of

00:02:50: software that are maybe just a one-time thing you need,

00:02:54: maybe just a utility script you need, or maybe if you're exploring some

00:02:57: idea, building a prototype, you could definitely

00:03:01: try live coding for that. Now, I have tried my fair

00:03:05: share of live coding and I will say, results have been rather mixed

00:03:09: and it's definitely not my thing. Wasn't my thing, still

00:03:13: isn't. I will say though that of course it can be quite

00:03:16: helpful to let AI build a feature where you maybe don't really know

00:03:20: how to tackle it because you need to use some browser

00:03:24: APIs, for example, which you don't know yet,

00:03:28: help of the AI, which is not the same as just letting the AI

00:03:32: build something and then use it and move on if it seems to work.

00:03:35: That's a different thing. But yeah, live coding was a

00:03:39: huge thing and still is a huge thing.

00:03:41: We have platforms like Lovable, v0, and many others

00:03:45: which are, in the end, live coding platforms. Yeah.

00:03:49: It's here, it will probably stay here.

00:03:52: We saw lots of people experiment with it, we definitely

00:03:56: have live coders out there that do live coding

00:04:00: for a living, more or less I guess. But yeah, let

00:04:04: me be very clear here. Live coding is not a

00:04:08: good idea. Being an expert, knowing how to write code, and

00:04:11: using that expertise and that knowledge in combination with AI is

00:04:15: so much better than live coding. That's...

00:04:18: I can only recommend taking this path because your knowledge

00:04:22: plus AI can get you really good results.

00:04:26: Just relying on AI, probably not so much.

00:04:29: Of course, coding assistants like Cursor or Claude

00:04:33: Code are not the only agentic systems we

00:04:37: saw develop this year. OpenAI, for example, also

00:04:40: integrated an agent into ChatGPT that can do stuff

00:04:44: for you, that can click around on websites

00:04:48: and do work for you. Now,

00:04:51: maybe it's just my bubble, but it doesn't really seem like a feature

00:04:55: that everybody's using all the time.

00:04:58: But yeah, it's here, and combined with deep research

00:05:02: agents, which I think we also had last year

00:05:06: which definitely became more of a thing this year, you can

00:05:09: offload some tasks, especially some research tasks, to

00:05:13: AI. Though as everything and always, it's

00:05:17: typically not a super good idea to just take those

00:05:21: AI results and use them as they are, but instead

00:05:25: it's much better to use them as a starting point and build

00:05:29: up on them with your own knowledge, maybe your own additional research,

00:05:33: and then AI, in my experience, can save you some

00:05:37: time in certain situations. And I definitely, for

00:05:40: example, like using-... Gemini Deep Research

00:05:44: for doing research for a video like this.

00:05:47: I still bring in my own expertise. There are many points I

00:05:51: cover in this episode which were not brought up by

00:05:54: Gemini, and also not by ChatGPT, but yeah, it's a nice

00:05:58: additional help. From that angle, we could call this

00:06:02: year the year of agents, but to be very honest,

00:06:05: not in the sense all these newspapers meant

00:06:09: at the beginning of the year. It doesn't really seem like all those

00:06:13: companies out there are using agents for everything.

00:06:17: Quite the opposite. There recently was an

00:06:20: article about, uh, Copilot by Microsoft not

00:06:24: really delivering on its promise or not being as

00:06:27: popular as Microsoft would hope. It really seems

00:06:31: to be the case that companies want to use AI, they

00:06:35: all feel like they have to, of course, and therefore, they are really

00:06:39: searching for use cases for AI and it's hard to find them.

00:06:43: Coding and, um, AI-assisted coding is definitely one of

00:06:47: them, but many others might not be there yet and

00:06:51: we don't have all these agents that do all the work humans used to

00:06:55: do. And by the way, when we talk about coding, I am well

00:06:59: aware that in many companies out there, there are executives that

00:07:02: basically tell their teams, "You either get double the work done

00:07:06: more, or we'll have to let some people go

00:07:10: use AI." And that, of course, is a horrible situation

00:07:14: because whilst AI can be a very good

00:07:17: assistant for coding, and I've said that more than once, I say

00:07:21: it in almost a- all my videos, whilst it can be a good assistant,

00:07:25: it's not a good replacement. And yeah, it can make you more

00:07:29: productive, but twice or three times the output?

00:07:33: I'm not so sure about that. But yeah, it is

00:07:37: what it is. That is something which will probably get better as

00:07:41: companies realize that even when it comes to coding, AI is not the

00:07:45: magic bullet, but just another helpful tool.

00:07:48: Or, of course, all these AI models get much, much

00:07:52: better, but until that point, it's a

00:07:55: really, really bad situations. But yeah, agents

00:07:59: everywhere? I'm not so sure I don't see

00:08:02: them. Now all that, of course, does not mean that

00:08:06: AI is not getting better and that there is no chance

00:08:10: of more agents doing useful stuff in the

00:08:13: future. But when we look at this year, we

00:08:17: definitely saw the release of many, many AI models.

00:08:21: (laughs) It, it feels like a new model was released basically every

00:08:25: week, right? Especially now in November, we got like Gemini

00:08:29: 3, we got new models by OpenAI, by xAI.

00:08:32: We get so many new models, and not just in November.

00:08:35: It's really model inflation out there.

00:08:38: If you're a web developer like I am, it feels like the

00:08:41: JavaScript framework wars we had in 2019

00:08:45: around that time. Uh, but even worse because we didn't have as

00:08:49: many frameworks as we have models now.

00:08:51: Obviously, that will only continue, but the question is

00:08:55: will those models get better and better?

00:08:58: To some extent, to me it feels like they're not really

00:09:03: I know they are continuously getting better in all these

00:09:06: benchmarks, but we've seen more than once this year

00:09:10: new model is released, does very well in a benchmark, and then

00:09:14: kind of falls flat in the popular

00:09:18: opinion. Isn't really living up to its promise and

00:09:21: people are not really, uh, happy about the model.

00:09:25: We've seen that more than once. One prominent example would be LLaMA

00:09:29: 4, which was released by Meta, uh, in, in the first half of the

00:09:33: year. LLaMA 4 did amazing in all these

00:09:37: course, especially interesting because it's an

00:09:40: But people didn't really like it. It really felt like

00:09:44: it was fine-tuned for these benchmarks as it often feels these

00:09:48: days with model releases and, well, I'm not sure if

00:09:52: anybody's still using LLaMA 4. Uh, the result, of course

00:09:55: (laughs) , of this, uh, failure, if you wanna call it like this,

00:09:59: was that Mark Zuckerberg started hiring

00:10:03: all the talents in the AI space by paying

00:10:07: astronomic amounts of money for the top talent

00:10:11: and we now have all that talent war going on between OpenAI

00:10:15: and Meta and Google and bunch of other companies.

00:10:19: But yeah, regarding the raw performance of the

00:10:23: we have yet to see if we'll see a steep increase

00:10:26: in the next year, 2026, or if the curve kind of

00:10:30: flattens out. Benchmarks alone certainly do not tell

00:10:34: the full story. Now, of course, there's a lot of money being

00:10:38: thrown on the development of all these models and therefore, the year actually

00:10:42: started with a shock when DeepSeek R1 was released

00:10:46: because that was some Chinese underground model

00:10:50: which nobody expected, which at the time

00:10:53: seemed to be trained very cheaply and seemed to be able to

00:10:57: compete with top-tier models back then.

00:11:00: And of course now, at the end of the year, we can say that all these

00:11:04: Chinese models, not just the DeepSeek models

00:11:08: related to video or image generation, play an important role

00:11:12: in the AI landscape and China, despite all

00:11:16: the problems they should in theory have with getting graphic

00:11:19: cards, GPUs, plays a super important role and

00:11:23: will be an important competitor in the year to come.

00:11:26: And on the less, as mentioned, we'll have to see if AI

00:11:30: models improve like this or maybe a bit flatter

00:11:34: like this. Of course, large language models aren't everything,

00:11:38: text generation isn't everything.

00:11:41: We saw pretty amazing improvements this year when it

00:11:44: comes to image and video generation.

00:11:48: I mean, if you compare the kind of videos you can

00:11:52: View 3 by Google or of Sora 2 by OpenAI

00:11:56: with what was possible a year ago, that's mind-blowing.

00:11:59: It's so much better, so much, uh, better quality, higher

00:12:03: character consistency, audio is better, and that's just the

00:12:07: US models. If you take a look at some Chinese models

00:12:11: like the Kling models-Um, you can see some

00:12:14: amazing development and results there too.

00:12:17: And it's not just the case for video.

00:12:19: Image generation was revolutionized first by

00:12:23: OpenAI's GPT image model, but then that model

00:12:27: now looks quite old compared to Google's NanoBanana,

00:12:31: and now its NanoBanana Pro model, which excels at

00:12:35: generating all kinds of images, including, uh,

00:12:39: images that include a lot of text like this beautiful infographic

00:12:43: here, which was created with, uh, my site, buildmygraphic.com,

00:12:47: which gives you a nice feature-rich editor and beautiful

00:12:50: style templates to build graphics like this.

00:12:54: So there is lots of development in the AI space when it comes

00:12:57: to other modalities other than text, and of

00:13:01: course, I'm also convinced that we'll see more

00:13:05: useful tools being developed on top of AI models.

00:13:08: Even if those AI models shouldn't become much smarter, we will

00:13:12: definitely see better and better tools

00:13:15: models and therefore deliver more value.

00:13:19: Just like Cursor is not just some AI model being

00:13:22: used in its raw form, but instead an AI model that is

00:13:26: heavily adjusted by engineering the context in a

00:13:30: clever way, having good system prompts and all that stuff that

00:13:34: leads to better results than if you were just to use

00:13:38: the raw model through the OpenAI API or whichever

00:13:42: provider you are using. So that was a

00:13:45: lot of talk about AI. I'm aware of that, as I mentioned.

00:13:49: I'm sorry. But it is obviously an important part of

00:13:53: we're living in if we like it or not.

00:13:55: And believe me, I would be very happy if we were not

00:13:59: living in a world that's full of AI slop and all

00:14:03: that AI hype. But it is what it is. And of course,

00:14:07: I'm fully aware that this also was and is

00:14:11: another very difficult year for software developers,

00:14:15: especially for web developers, especially also for junior

00:14:18: developers because we're coming still from

00:14:22: that all-time high during the pandemic where all the

00:14:25: companies couldn't get enough developers.

00:14:29: Well, turns out they had too many. So you all probably

00:14:33: know that chart about the software developer jobs.

00:14:36: And, well, at least that seems to finally have reached the

00:14:40: bottom, but now of course there is AI.

00:14:42: And as I mentioned before, there are these discussions

00:14:46: or the pressure inside of companies about

00:14:49: using AI instead of developers for certain tasks.

00:14:53: And especially junior developers, therefore, are not in high demand because,

00:14:56: hey, AI can do it, right? Wrong, totally wrong.

00:15:00: And who's going to be the senior of tomorrow if there is no junior being

00:15:04: hired today? But yeah, it is what it is, and therefore,

00:15:08: unfortunately, this was another difficult year, and next year

00:15:11: is probably going to be diff- difficult too,

00:15:14: which is really bad. There is no other

00:15:18: way of putting that. Now despite AI, thankfully,

00:15:22: there are also some other interesting topics to talk

00:15:26: about when looking back at this year.

00:15:28: For example, we, of course, saw new versions being released for

00:15:32: Next.js, React, Angular, and

00:15:36: all those other frameworks and library stats are important for

00:15:40: us developers. It's also fair to say that it

00:15:44: looks like React and TypeScript and Next.js won

00:15:48: the AI race. I mean, take a look at the React

00:15:51: downloads. AI clearly has a favorite.

00:15:55: And I've talked about this in other episodes before,

00:15:59: this is a huge problem in my opinion because no matter if

00:16:02: you like, um, the fact that we had times where a

00:16:06: new framework seemed to come out every day or not, it's

00:16:10: definitely not good to have one clear winning stack,

00:16:14: which is React, TypeScript, Next.js, ChatGPT, and Tailwind.

00:16:17: That's it. That is definitely not good, and I

00:16:21: don't see that changing, unfortunately.

00:16:24: And that is not to say that any of these technologies would be

00:16:28: They're not. They're great. Competition and

00:16:31: innovation just also is good, and it looks like that's going to get

00:16:35: harder with AI. But hey, okay, th- there was AI again.

00:16:38: Let's, let's move away from AI. So yeah, we saw these new versions bringing

00:16:42: us new features, but especially React

00:16:46: and to some degree Next.js, but it was React, had

00:16:50: some problems, some security problems.

00:16:53: It was just a few weeks ago that a

00:16:57: couple of vulnerabilities were reported related to React

00:17:01: and React server components. I did create a separate

00:17:05: episode about that, so take a look at that if you wanna learn more.

00:17:08: That, unfortunately, wasn't the only security issue

00:17:12: related to web development we saw this year.

00:17:15: Instead, we also had multiple NPM-related

00:17:19: attacks, node modules-related attacks, where

00:17:23: some packages were infiltrated with

00:17:27: malicious code in a way that that code would

00:17:31: spread automatically to other packages and

00:17:34: ultimately potentially reach your machine to then

00:17:38: steal crypto or crypto

00:17:41: credentials and stuff like that from your machine.

00:17:45: And of course, that all is probably

00:17:49: just the beginning. I mean, vulnerabilities and

00:17:53: attacks have always been a thing, but especially with

00:17:56: AI, I see a huge danger of that trend

00:18:00: accelerating for two main reasons.

00:18:02: For one, AI can, of course, help with developing

00:18:06: malicious code. It can help run such attacks.

00:18:09: It can help with social engineering.

00:18:11: I mean, you can create realistic voices

00:18:15: days. So that will be a huge problem going

00:18:18: think.But of course, the other problem potentially

00:18:22: is that AI can also lead to gaping

00:18:26: security holes in software. And just to be

00:18:28: clear, the React vulnerability

00:18:32: and the NPM attacks are very

00:18:36: likely not linked to AI in any way. I still

00:18:40: see these problems coming up in the future for the two

00:18:44: reasons I just described. And we saw with the React

00:18:48: attacks and with the NPM attacks how

00:18:51: vulnerable this web development, this front-end web development

00:18:55: ecosystem really is. And especially in a world where more

00:18:59: and more applications are powered by React, well, it's pretty

00:19:03: clear that that is a quite interesting

00:19:06: target for an attack and for looking for

00:19:10: vulnerabilities. And the same, of course, is true for NPM and Node modules

00:19:14: because even we experienced

00:19:17: developers often just install packages.

00:19:21: And of course, we trust many of those established

00:19:24: Why wouldn't we? So even we are very

00:19:28: vulnerable to these attacks. Now take it a step further

00:19:32: and think about people that don't know anything about coding

00:19:36: too much about it and just live code their way through the

00:19:39: world. Yeah, these are quite

00:19:43: interesting targets, I'd say. So yeah, that was a thing, that will

00:19:47: be a thing, I think. Speaking of problems and

00:19:51: attacks, not really an attack but still kind of worrying

00:19:55: was that we had two big outages this

00:19:59: fall, AWS and Cloudflare. AWS

00:20:03: took significant chunks of the internet down this October for

00:20:07: multiple hours, almost a day overall.

00:20:10: Cloudflare kind of did the same on a slightly smaller

00:20:14: scale. Again, is AI a problem here?

00:20:17: Were those engineers using AI for these configurations

00:20:21: that blew, uh, things up? I don't know.

00:20:24: Probably not, but we don't know. But it's

00:20:28: kind of worrying how vulnerable the

00:20:31: overall internet is and how dependent it is on a

00:20:36: few big providers. And we knew that before.

00:20:39: It's not new and it's not the first AWS outage.

00:20:42: But these two big incidents which happened only a few

00:20:46: weeks apart, yeah, they really show us

00:20:51: what the internet has become. It's been that place of

00:20:55: decentralization and it kind of still is,

00:20:58: but on the other hand, it kind of still isn't.

00:21:01: If AWS, if one big region of AWS

00:21:06: gets taken out, boom. Here go

00:21:09: 30% of all websites as it seems. So yeah, that is

00:21:14: definitely also something we saw this year, that happened this year

00:21:17: implications in one way or the other.

00:21:20: And that is kind of, well, showing us

00:21:23: what the internet has become, I guess.

00:21:26: Speaking of what the internet has become, and here I have to go back to AI

00:21:30: briefly. I apologize. Significant portions of all traffic

00:21:34: on the internet and of all content on the internet are

00:21:38: already generated by AI. So we got bots reading

00:21:42: content generated by other bots generated by AI.

00:21:45: Great new world, I guess. So that, of

00:21:48: course, is also something that accelerated this year.

00:21:52: I mean, just two or three years ago, we had almost no

00:21:56: AI-generated content on the internet because it

00:21:59: sucked. But now it's, it's everywhere.

00:22:03: And guess what? That will also only

00:22:06: accelerate. I mean, if you're on the same bubble as I am on

00:22:10: X, everybody's just building AI slop

00:22:14: there. Great. That is, uh, a lovely new future.

00:22:17: And don't get me wrong, you can generate very useful stuff with

00:22:21: AI. I love that I'm able to generate images or

00:22:25: icons with AI. Of course, I am aware that we

00:22:29: also have some problems here if we think about copyright and so

00:22:33: on. But it is useful, I will admit

00:22:36: that. And the same is true for some text, being able to

00:22:40: summarize a PDF document. There are many use cases where AI is

00:22:44: very useful. I wanna be very clear about that.

00:22:46: But the AI slop and the internet becoming more of an AI

00:22:50: machine full of bots and AI-generated content,

00:22:54: yeah, I've said it before, it's, it is what

00:22:58: it is. I'm not super pumped about it though, to

00:23:02: put it like this. Now, I could go on for much longer.

00:23:05: I could talk about TanStack, which has a little

00:23:08: wave of popularity, at least in my bubble now at the end of

00:23:12: I could talk about the TypeScript port to Go

00:23:16: instead of Rust, which was announced at the beginning of the year

00:23:20: soon going to be complete. And you can already try the Go version of

00:23:24: TypeScript, it looks like. It's awesome. It's much faster.

00:23:27: I could talk about the, the trend back to

00:23:30: monoliths and VPS instead of

00:23:34: microservices or managed services.

00:23:36: And all these are just small trends, small building blocks, of

00:23:40: course. In large enterprises, we're still working with

00:23:43: microservices probably for good reasons.

00:23:46: And as I mentioned before, React and Next is still king.

00:23:49: But-These are all nice little things which

00:23:53: still make this overall web development world really

00:23:57: nice to be in, to put it like this. I'm super excited about all these

00:24:01: things. I'm excited about the fact that we got BetterAuth, an

00:24:05: amazing authentication library which became super popular this year,

00:24:09: which I'm using in all my projects, which makes authentication such

00:24:13: a breeze. I enjoy working with Bun and Hono.

00:24:17: I really love what you can do with PostgreSQL, that you can

00:24:21: use extensions like pgvector to build a vector

00:24:24: database and build RAG applications, because again, there are

00:24:28: useful things you can do with AI. I also very much like the

00:24:32: fact that you can use Cursor and Cloud Kode or OpenCode

00:24:35: or all these other AI coding assistants to

00:24:39: be more productive and to outsource some tasks to AI.

00:24:43: I'm absolutely loving the fact that we got a bunch of

00:24:47: amazing new features built into browsers, like the

00:24:50: popover API and many, many other cool

00:24:54: APIs, and also CSS properties and values that you can

00:24:58: use today in all major browsers to make your life

00:25:02: easier. The problem there just is that AI

00:25:06: probably doesn't know about these features yet, and maybe never if it

00:25:09: just is that machine that builds React and Next

00:25:13: applications, and then learns based on what it built

00:25:16: But we'll see how that turns out. I just wanna

00:25:19: say, web development is, in

00:25:22: theory, in great shape. We got all these

00:25:26: amazing features that we can use in browsers today.

00:25:28: Many things are much easier to build than in the past.

00:25:32: We got mature frameworks. We got innovation through

00:25:35: frameworks like TanStack Start or the BetterAuth

00:25:38: library. We got so many amazing stuff going on

00:25:42: can be such a useful tool. I

00:25:46: really, really hope that we'll be able to

00:25:50: make good stuff out of all these building blocks, that we'll be able

00:25:54: to build up on this great foundation in the

00:25:57: future, and that we'll not get consumed by all the

00:26:01: AI slop and all the BS around AI, and

00:26:05: all the problems that come from AI and

00:26:09: companies trying to force AI on us in every aspect of our

00:26:13: life. I really hope that executives

00:26:17: will learn that AI is a great assistant, but

00:26:21: not a great replacement for developers, and that

00:26:25: we can leverage the potential of all these amazing

00:26:29: things we got today, because I truly believe that this is the

00:26:33: best time ever to be a developer and a web

00:26:36: developer. But it's also one of the most difficult

00:26:40: times, unfortunately. So yeah, that was my look back.

00:26:44: I might have missed some things you found important too.

00:26:47: Please let me know. Please let me know.

00:26:49: I try to be comprehensive, and this video is already super long.

00:26:52: I will also create a separate video where I take a look at the next year, and

00:26:57: I create a bunch of videos where I share opinions and thoughts about everything all

00:27:00: the time. So yeah, if you like that, follow me and

00:27:04: have a great end of the year and a great start into the new

00:27:07: year.