Learning Library

← Back to Library

2025 AI Breakthroughs: Code and Images Unlock

Key Points

  • 2025 didn’t bring sensational sci‑fi AI, but it clarified where real value lies in the AI revolution and exposed critical gaps that are now visible.
  • The breakthrough that most exceeded expectations was allowing LLMs to use code as a tool, unlocking agentic workflows and making AI accessible to non‑technical users through plain‑English computer interaction.
  • A suite of emerging technologies—cloud‑code integration, model‑context protocols, skills, Codeex, and tools like Cursor—combined to let anyone manipulate files and automate tasks simply by describing what they want.
  • Advances in image generation proved equally transformative, enabling detailed infographics, layouts, slides, generative UI for games, interior redesign, fashion, and more, effectively turning visual creation into a new, powerful AI capability.

Full Transcript

# 2025 AI Breakthroughs: Code and Images Unlock **Source:** [https://www.youtube.com/watch?v=uVZMc-i5EEs](https://www.youtube.com/watch?v=uVZMc-i5EEs) **Duration:** 00:13:52 ## Summary - 2025 didn’t bring sensational sci‑fi AI, but it clarified where real value lies in the AI revolution and exposed critical gaps that are now visible. - The breakthrough that most exceeded expectations was allowing LLMs to use code as a tool, unlocking agentic workflows and making AI accessible to non‑technical users through plain‑English computer interaction. - A suite of emerging technologies—cloud‑code integration, model‑context protocols, skills, Codeex, and tools like Cursor—combined to let anyone manipulate files and automate tasks simply by describing what they want. - Advances in image generation proved equally transformative, enabling detailed infographics, layouts, slides, generative UI for games, interior redesign, fashion, and more, effectively turning visual creation into a new, powerful AI capability. ## Sections - [00:00:00](https://www.youtube.com/watch?v=uVZMc-i5EEs&t=0s) **Code‑Enabled LLMs Unlock 2025** - In a 2025 recap, the speaker emphasizes that the most surprising breakthrough was allowing large language models to use code, a development that dramatically expanded agentic workflows and made advanced AI capabilities accessible to non‑technical users. - [00:03:31](https://www.youtube.com/watch?v=uVZMc-i5EEs&t=211s) **Generative Interfaces and System Design** - The speaker argues that while generative AI can create adaptive web interfaces, practical value lies in balancing novelty with familiar habits, and success often stems from system design skills rather than specialist AI developers. - [00:07:18](https://www.youtube.com/watch?v=uVZMc-i5EEs&t=438s) **The Undervalued AI Middleware Layer** - The speaker argues that the middle tier—transforming raw model outputs into structured, domain‑specific workflows—remains underbuilt yet holds immense value, and startups are now confidently building in this space despite concerns about hyperscaler dominance. - [00:11:45](https://www.youtube.com/watch?v=uVZMc-i5EEs&t=705s) **AI Shifts Toward Creativity and Quality** - The speaker highlights a recent pivot from viewing AI solely as a cost‑cutting tool to embracing its role in enabling creative expression and enhancing human expertise, emphasizing new opportunities for interdisciplinary collaboration and a focus on delivering higher‑quality outcomes. ## Full Transcript
0:00You know, 2025 didn't deliver the 0:02science fiction version of AI that gets 0:04lots of clicks, but it exceeded my 0:06expectation in ways that matter more. It 0:09clarified where value is actually coming 0:11from in the AI revolution, and it made 0:14the gaps that we still have visible in a 0:16way that I think is useful. The number 0:17one thing that I want to call out as we 0:19sort of look back at the year 2025 0:22is I think that we all or almost all of 0:26us underestimated how powerful it is 0:31when you allow an LLM to use code as a 0:34tool. That turns out to be an absolutely 0:37massive unlock. It's at the heart of a 0:39lot of our agentic workflows these days 0:41and it is on the verge in 2026 of being 0:45one of the biggest places where non-code 0:48non-technical folks can lean in and 0:51start to figure out how agents can work 0:53for them in ways that are far removed 0:56from writing code. And I think that that 0:59that core unlock of wow an LLM can work 1:01with code therefore it can work with any 1:03part of the computer. That was one of 1:05those pieces that you could see that 1:08vision as something that the model 1:10makers had in their heads at the 1:12beginning of 2025 and they did talk 1:14about it but it wasn't realized. It was 1:16very much a someday we will see this 1:19come true and then over the course of 1:21the year you know we get cloud code we 1:23start to get model context protocol 1:25starts to dominate we start to get 1:26skills we start to get codeex and you 1:29start to gradually see these pieces come 1:30together you see the development of 1:31cursor etc. And what you start to see is 1:34that these pieces allow everyone, not 1:39just technical users, to start to use 1:42the tool because now plain English 1:45allows you to talk with your computer. 1:48Now plain English allows you to 1:49manipulate the files on your computer 1:51any way you want. Absolutely massive 1:53unlock. And I think it is hard to 1:56correctly estimate going in how big that 1:59turned out to be. The second one I think 2:02that turned out to be absolutely massive 2:04is images. And people have talked about 2:06this, but we have lived in a world for 2:09most of 2025 where images were getting 2:11better, but text remained the most 2:13accurate way to work with LLMs. And in 2:16that world, code is just a subset of 2:18text, right? Accurate code is just 2:20another language that the LLM has to 2:22learn. But images are how humans process 2:24information quickly. And when we finally 2:26got to the point where we had images 2:28solved, where you could do detailed 2:30infographics, you could do detailed text 2:32and image and it wouldn't look weird. 2:34You could do full maps, you could do 2:36layouts, you could do slides. I don't 2:38think we realized how big of an unlock 2:40getting that right was. And that doesn't 2:42just mean we got powerpoints, right? 2:43Like we did solve powerpoints, but think 2:45of it more broadly than that. Think of 2:47it as generative UI for gaming. Think of 2:50it as you have a chance to redecorate 2:53the rooms in your house now in ways that 2:55you never did before. Uh there's all 2:58kinds of personal applications that 3:00people are building on with fashion that 3:02come with this. Basically, getting 3:04images right enables us to realize a 3:08vision of graphical user interfaces that 3:13has always been out of reach. the idea 3:17that the graphical user interface isn't 3:19locked to your screen and locked to what 3:20the developer says. That the graphical 3:22user interface becomes something that 3:24evolves with you. Maybe you end up with 3:25it as a wearable. Maybe you end up with 3:28it as a combination of phone and laptop. 3:30Maybe you end up with it being 3:31generative in the side like it is for 3:33Comet. Those are all variants of this 3:36idea that with the right graphics solve 3:40from an AI perspective, that whole 3:43surface becomes continuous. and you can 3:46evolve the surface of engagement 3:48digitally with where an individual is at 3:51and what they're looking for in that 3:52moment. Now, we're not all the way there 3:54yet. I'm not here to tell you that the 3:56future of the web is everybody has a 3:58generative interface that changes for 4:01everything that they need. There is a 4:03value in habit. There's a value in 4:05steadiness for common use cases. We 4:07don't need to reinvent the wheel where 4:09it's incredibly obvious. But I do think 4:12solving images is one of those things 4:14that we probably underestimated how big 4:16a deal that was going in and like 4:17looking at it on the other side a month 4:19or two in it's absolutely massive. Like 4:22it's it's huge. What are some other 4:24ones? I think those ones are pretty 4:25common reflections. Let's go a little 4:26bit deeper. I think one of the things 4:28that surprised me is how far you can get 4:30without AI developers. So we we were 4:33told at the beginning of the year AI 4:35developers are everything. Like you need 4:37to have a developer that knows AI. I 4:39actually think you need someone who can 4:41design systems. I have watched 4:43individuals outexecute entire 4:45development teams because they treated 4:48engineering as a workflow they could 4:50design and they didn't worship at the 4:52altar of a particular model. And so they 4:55were okay building low tech things like 4:57templates, like validators, like 4:59retries. And they iterated really 5:01quickly and aggressively. And they 5:03didn't confuse Agentic with good. And 5:05before you know it, those folks are 5:07understanding the new principles for 5:10evolving agentic systems. And then they 5:12become more and more valuable. And and I 5:14think we are going to have to throw away 5:15the idea that there are technical and 5:17non-technical people. I think a more 5:19accurate description is that everyone 5:21picks up the degree of technical skills 5:24they need to solve the problems they're 5:26interested in. And increasingly the 5:28question is going to be, are you curious 5:31about the problems that are relevant in 5:33your domain? And are you willing to dive 5:35in and pick up the AI skills you need to 5:37solve those problems, including 5:39technical skills, because those are 5:40increasingly approachable. You can get a 5:42scheduled task with a nice learning 5:45composed just for you in chat GPT every 5:48morning if you want. And I do I get 5:50them. It's really nice. I get nice 5:51little coding reviews every morning. You 5:54can get it for whatever you want. 5:55Another one that I think positively 5:56surprised me about 2025, verification 5:58loops in Agentic systems turn out to be 6:02incredibly powerful. Uh the idea that 6:05you can measure correctness in different 6:07dimensions turns out to have incredibly 6:09wide ranging implications for good 6:11system design. And that's not super 6:13surprising if you know how software is 6:15designed. But hooking that up to an 6:18agent that iterates is it's like hooking 6:21up a jet engine to an airplane. Like 6:23it's it's amazing how fast you can go 6:25when you stick an agent against a 6:28verification loop that is hard to game 6:30and you say go get it done. And I think 6:31that we figured that out as a community 6:35partway through the year and started 6:36really really practicing it and that has 6:39supercharged our progress. And the nice 6:40thing is that's one of those primitives 6:42that we can really build on heading into 6:442026. Like getting more into 6:46verification loops is something that I 6:48think we'll see from more and more 6:50teams. And I think we're going to start 6:51to see some consistency around those. So 6:53there's things like accessibility where 6:55you just want to have a standard set of 6:56eval verification loops across the 6:58industry and people just need to get 7:01their agents to pass them when they're 7:02building software and it should not be 7:03something you have to reinvent every 7:05time. So I think we're going to start to 7:06see a really interesting ecosystem build 7:07up around verifications. That's that's 7:09compelling. Another one that has been 7:11really good to see is the messy middle 7:14turned out to be the entire game. 7:16Everyone wants to talk about the idea of 7:18the front end and there was a lot of 7:20talk during the middle of the year about 7:22super model makers or hyperscalers 7:24owning the entire stack like Chad GPT 7:26owning the stack. Cloud code launches 7:28does cloud code own the entire stack. Is 7:30there a competitive advantage to being 7:31in the middle? Is is cursor game over 7:34etc. It turns out that there is so much 7:36value in transforming messy inputs into 7:39structured representations, in routing 7:41intent, in orchestrating calls, in 7:43handling exceptions, in providing useful 7:46user interfaces for specific things. 7:48That the middle layer is still feeling 7:50underbuilt relative to how much value we 7:53can unlock. I think most of us got that 7:55one wrong. And I think that was one of 7:56the pleasant upsides of the year is that 7:58the messy middle, yes, maybe it's 8:01vulnerable. Maybe you worry about sort 8:02of the launch of a particular 8:04hyperscaler's product idea, but there is 8:06so much value in taking raw AI model 8:10outputs and getting them into a 8:11particular domain that we are I feel 8:14like we're really underbuilt. I feel 8:15like uh probably the most prominent 8:17example of that is cursor which everyone 8:19refers to all the time but there's a 8:21whole host of other startups in the in 8:23the space not just in coding and non- 8:25tech spaces too that are are building 8:28aggressively now into the middle because 8:30we've we've realized we don't have to be 8:32afraid of it. We've realized that the 8:34model makers are essentially forming a 8:36very competitive substrate of 8:37intelligence that we can build over the 8:39top of to deliver outputs to users that 8:42are much more valuable than they're 8:44really going to be able to get from 8:45models alone. That's a really 8:47interesting one. Another one that 8:49surprised me or at least it's a positive 8:51reflection from the year. I think we are 8:53realizing now how much value there is in 8:56effectively scoping our workflows. I 8:58think agents were oversold. That was 9:00that was something that a lot of people 9:01were disappointed by because they were 9:02sold as magic buttons. But the flip side 9:05is when you put an agent in a good 9:06workflow, that's a really pleasant 9:08surprise because instead of promising 9:10that it can do everything, it turns out 9:12that when you can use it as a tool, it 9:14can do a tremendous amount reliably and 9:16you can start to really move volume over 9:18to it. And so I think the pleasant 9:20surprise is how much you can accomplish 9:23when you properly harness your agents 9:25and how big companies are leaning in and 9:29able to actually get volume done on that 9:31basis. You know, another positive for 9:33the year, I know that we had a lot of 9:35conversation around AI slop this year, 9:37but I think one of the things that I 9:38learned is that AI slop is a symptom of 9:43unconstrained and unmanaged artificial 9:45intelligence. And that companies that 9:47start to get into marketing, start to 9:49get into producing content at scale for 9:51AI, if you build the right systems, you 9:54can produce really compelling, very 9:56performant ad flows, very performant 9:58email marketing, very performant content 10:00marketing that outperforms what humans 10:03can do. And so I guess the the hopeful 10:06thing there is really we don't have to 10:09surrender to and live in a world with 10:11non-performance slop. we can actually 10:13construct these systems so that they're 10:15beautiful, so they sound good, so that 10:18people want to click on them. And I'm 10:21fully aware that right now that if you 10:23announce that your ad is AI, there's 10:25generally a backlash. I think we're just 10:27going to get to the point where we don't 10:28announce it anymore and we just do the 10:29ad that feels right, however we do it, 10:31whether that's manually, whether that's 10:32manually plus AI, whether that's AI, and 10:34we just get it done and we move on. And 10:36I think that's going to be the case with 10:37a lot of content. And the key and the 10:40measure that we should hold is, is the 10:42content useful? Is the content genuinely 10:44helpful? Is the content information 10:46dense? Is it something that I can come 10:48back to again and again and say, there's 10:50something here that I can dig my teeth 10:51into. And what we're discovering is that 10:54AI can actually be really helpful on 10:55that because it can ground you with 10:56research. It can increasingly do 10:58factchecking. It can help you think 11:00through the structure of a piece. It can 11:02help you with uh generating those ads 11:04now that we've sort of solved the images 11:06piece. And I think that gives me hope 11:08because I don't want to live in a world 11:09with slap and I don't think most people 11:10do. And I think that's going to end up 11:12being a phase and we're going to end up 11:13getting through it because there's a lot 11:14of selection pressure for better 11:16content. You know, another one that 11:17surprised me positively in 2025 is how 11:20quickly the market selected for people 11:23with very strong creative problem 11:25solving instincts. I I think we saw a 11:28quicker response from the labor market 11:30than I anticipated. Not on the dramatic 11:32headline stuff of firing and all of 11:34that. There's actually still not a ton 11:36of evidence that AI is driving overall 11:38job market declines as much as there's 11:40newspaper headlines about it everywhere. 11:42But there is a lot of anecdotal evidence 11:45around the degree to which AI is a 11:50creative or liberal arts endeavor. And I 11:52think that signal has swung really 11:53sharply. And I think that's been really 11:55positive to see technical people that 11:57wanted to express their creative side 11:59and creative people that never felt like 12:01they could be technical finally have a 12:03chance to get the AI they're looking for 12:05and get a chance to build and get a 12:07chance to share their talents and get a 12:09chance to stretch their wings in ways 12:10that they couldn't before. And so I I 12:12actually think that one of the great 12:13opportunities of 2026 is people who want 12:16to grow the edges of their domain 12:18expertise and get smarter and do more. 12:21The world is your oyster. We've never 12:23been able to lean in more that way and I 12:25think that's been really fun. And the 12:26last one I I want to call out is that we 12:28did start to see a shift from a lot of 12:31the cost cutting mentality to quality 12:33lift. It's not universal. There are 12:35absolutely still people who will see AI 12:37primarily as cost cutting. But more and 12:40more and more as I sit and talk with 12:42leaders, they're sort of recognizing 12:43after the first wave of vendor purchases 12:46and getting AI installed quickly and 12:48thinking of AI as a magic button, they 12:50still need their people. their people 12:52can't go anywhere because they still 12:53need their people to deliver the kind of 12:56value that only people can deliver in a 12:58customer-f facing organization. And so 13:00what they want is they want to have a 13:02conversation about quality. How can we 13:04level up the quality of the experience 13:05we provide to customers in ways that 13:07were unimaginable because of AI? How can 13:09we lever up the volume of customers 13:11served? How can we make the price more 13:13competitive because we're able to scale 13:15the unit economics of the business? 13:16Those are so much more interesting 13:18questions and so much more compelling 13:20questions than just saying, can we cut 13:22costs and dump AI? There will still be 13:24folks that have that sort of brutalist 13:26mindset. But I think increasingly people 13:28are starting to recognize how powerful 13:29these systems are. And they're starting 13:31to recognize that the firms that win are 13:33firms that regard their people and their 13:35people's attention as a precious asset. 13:37And they're designing AI systems around 13:39them in ways that allow people to put 13:42their expertise to work where it matters 13:44most. And so my question for you is what 13:46did I not mention? What exceeded your 13:48expectations in 2025?