Transcription
We all use our chatbots too much. We do. We all use our chatbots because that is the default thing to do. And I think it doesn't help that there's a hundred,000 other tools. And I'm not actually making that number up. That's roughly what the number of total AI tools out there are. It's just too many. How do we figure out which one to use? And so we end up defaulting back into the chatbot space. And the model makers know that our mind share is allocated there. And so they continue to invest in making those experiences more sticky. That's why ChatGPT has memory, for instance. It's a stickiness feature.
When you think about it that way, it makes sense that we would periodically poke our noses out beyond chatbot land and actually look around the landscape and say, what other tools fill gaps that LLMs are inherently struggling to close? So, first in this video, I'm going to lay out some of those gaps that LLMs may not be the best in the world at. Not because it's impossible for the model makers to close the gaps, but because the model makers are preoccupied with larger what I would call generic problem sets like frankly finding enough GPUs to serve their model to all the people who want it. And that is also not something I made up. That's something very well documented as a prime concern for both Anthropic and OpenAI right now.
So what are the six structural limitations of LLMs that are being partially compensated for inside chatbots right now? But maybe there are specialized tools that would help us get farther and do our work more effectively. And then from there we'll get into 12 tools, two for each of the the structural gaps that you can survey. My goal here isn't to convince you to use these tools. It's to help you get a sense of how to think about structural gaps in the chatbot experience and then to understand what tools might be useful for closing the structural gaps that matter to you.
So gap number one, spatial reasoning. Yes, LLMs are absolutely getting better at this. I still am impressed that GPT-4 can produce 3D graphs, but fundamentally if you are trying to get to design, LLMs are not phenomenal designers. I have yet to see an LLM do a great job at that. I think my favorite anecdote here is agent mode where with great effort ChatGPT taught an AI agent to make a PowerPoint. The results are less than stellar. Uh to put it very kindly, the text tends to run over. It tends to run off the slide. It tends to be poorly organized on the slide. It doesn't work well with the visuals. The visuals feel slapped on. I know interns that could do a vastly better job.
Gap number two, spreadsheet context. We have an issue with spreadsheets because spreadsheets have orthogonal meaning. In other words, they have relationships horizontally, relationships orthogonally, and in complex spreadsheets, there's relationships between tabs. There's relationships between special columns and rows and the regular columns and rows of data. There's formulas. It is really, really challenging for LLMs that are designed for next token prediction to master spreadsheets. Again, we see some progress. I'll go back to agent mode. It can make a spreadsheet. It can make a spreadsheet with a simple formula. Now, it cannot process your existing spreadsheet. Well, it can't build a fully complex spreadsheet yet. I have tried it. Eh, it's okay. When you ask other LLMs like uh Claude or Opus uh Claude Opus 4 or ChatGPT-4 or Gemini 2.5 Pro, they range from insisting on CSVs, which are comma delimited and therefore more friendly to tokens to trying to ingest and process Excels and still struggling. Still struggling if they're large, still struggling to read all the detail. I've uploaded 40 or 50 row Excel spreadsheets and have found that even at that scale, which anyone who's using an Excel sheet will know is tiny, they can still sometimes struggle to list every row. They just can't seem to read all of the data. So, spreadsheets are a problem.
Code execution also remains a challenge. Fundamentally, none of the LLMs were constructed with the idea of being code execution environments. And I don't anticipate them becoming code execution environments anytime soon. And for those of you who are not coders, that means running the code. The fact that Claude can spin up a little React component and you can kind of run a little applet inside a preview window is about the best it gets right now. And that's still very very minor, right? It's not really a full code execution environment. Certainly not something that you would want to put into production. And that may seem obvious and you may think that isn't even an AI related thing. But increasingly because prompts and because AI generated code and because LLMs themselves are integrated into our production pipelines for software, we do need software that has code execution and AI capabilities.
Another gap, operational visibility. Again, why would you expect this? But LLMs are not built to give you any kind of operational visibility on your AI software in production. They're just not. No big surprise there.
And last but not least, narrative structure is a huge problem for AI. And this is one that I don't think gets talked about a lot. Text versus experience is very difficult for LLMs to convey. They often will respond with various versions of text because they can output text, but they can't think through the visual hierarchy. They have trouble sometimes thinking through the structure of the story in a way that's accessible. This is an area where I would expect breakthroughs like ChatGPT-5 to be helpful, but I still think that there's going to be a complex interplay between the structure of a narrative and the way a narrative is visually presented that is going to be hard for traditional LLMs to master. And I think it's it's just not something that is easy to do unless it's your sole focus. And even then, it's quite difficult.
One more, last but not least, voice processing. ChatGPT famously launched meeting notes recently. I have used them. They are only okay. They don't give you live transcriptions. They give you only one generic summary. You can't really access the transcript. It's very much a bolt-on feature. And that is exactly what I would expect from a team that is fundamentally resource constrained and trying to ship a lot of things to an 800 million or more user base. Now, they cannot do everything perfectly. And therein lies the opportunity for builders like the 12 tools that we're going to outline here. Again, these are not the best 12 tools ever. I think they are great answers for these six gaps that I've called out. The six gaps being uh voice processing, narrative structure, operational visibility, code execution, spreadsheet context, and spatial reasoning. Those are not the only gaps, but I thought they were really illustrative of the kinds of gaps that LLMs have. And these 12 tools do a good job hitting those gaps. So look at these. Think about the strengths. Think about the weaknesses. I'll call out and think about where your workflow doesn't work well with a chat.
Tool number one. This is in interface builders. Magic Patterns. Magic Patterns has just I've had people coming to me with Magic Patterns. I've not been the one sort of sharing it out, but people have come to me and showed me Magic Patterns because they like them so much. Fundamentally, it makes it extremely extremely easy to extract a design out of a screenshot or something else, turn it into working components and get something back that is a working piece of compliant stylewise front-end code that illustrates a vision for a new interface, which is a complicated way of saying it is really easy now to copy the style off the website and change it and show your engineers. And that is something every single marketer and PM and program manager and anyone else CS who has an idea for something that should be different about the tool or the app or the website. We have all wished for this. We have wished for it to be magically easy to say here's my sketch. Here's my concept. But magically it's in the right style. Now that's as simple as a screenshot and throwing it in Magic Patterns. Specialized tool closes a specific gap in an LLM. Is it perfect? No, it's not perfect. It's not designed for full app building, right? But does it give you a quick sketch sense? Is it designed for exactly what it does? Well, yeah, it does. Visily is another option there. It is a little bit cheaper than Magic Patterns. It focuses on rapid mockup creation rather than code generation. So, if you need the code components, don't go with Visily. If you just need the quick mockup, wireframing can be much faster with Visily. And so, again, like both are in the interface category. They do slightly different things. So, I want to lay them out as distinct. My goal here isn't to make these like competitors, but to actually help you understand how each tool is attacking a particular gap that a chatbot has.
All right, let's move on to the second one. Spreadsheet intelligence. What do we have? Shortcut AI is exploding. It's an early access. You may not be able to get it. It is definitely the best I've seen at tackling complex spreadsheet creation. And I want to underline the word creation. There are still some struggles with macros and existing spreadsheets, but if you want to create something and you are a Power Excel user, I am getting rave reviews on this one. Again, not me. People coming to me saying, "I'm trying Shortcut AI and it's incredible." And so, I suspect once this goes more widely public, there's a good chance it becomes the definitive answer for AI and Excel. The other solution, which is more widely available, is Numerous AI, which really focuses on embedding AI in your existing spreadsheet through custom functions. That's a different use case. It's supposed to help you add AI in useful ways to your current spreadsheets versus just creating new sheets. From a product strategy perspective, Shortcut is in the stronger position because they're inventing a solution to the entire spreadsheet problem I discussed rather than just trying to wrap AI in into your existing sheets. There's no way, as far as I can tell, for Numerous AI to create a brand new spreadsheet that is very complex from scratch and have it sort of handle the kind of complexity that Shortcut is bringing to the table, especially from a prompt. They just do different things. Again, it's not necessarily a competitor thing. They just do different things. And Shortcut is solving the bigger part of the spreadsheet intelligence problem.
Let's move on to another gap. Executing code. We wouldn't expect most LLMs to do this, but we do need solutions that include AI. I'm going to give two. I don't hear a ton about either of these, but I want to throw them out there and you can tell me what you think about them and which one you think is more useful. Uh the first one is e2b.dev. It starts at a free tier. It leverages AWS Firecracker. Um the the critical piece is that it's it's effortless to integrate. Like if you want to throw this up and make it easy to execute code, e2b.dev makes it easy to stand up a sandbox and try something. It's super quick. Daytona is not as cheap. I love that it's named Daytona, by the way. It's also a little bit, as you would expect, more established, right? It has ISO 27001, SOC 2, all that good stuff from a certification perspective. And it again makes it it makes it easy to execute code and ensures you won't damage production systems. And that is one of the biggest concerns that people have with vibe coding is that you're going to have the risk that it will damage production systems. So, the stakes here are real. And I think you're going to see a lot of traction in this space from startups like E2B.dev and Daytona. I'm curious if you guys have a strong opinion between the two.
Moving on to LLM observability. Really important to understand if you're running a lot of prompts through production grade AI, you have to understand how they're actually working. And I want to call out too, both of these are very established at this point. Uh, Helicone is very simply a clear visibility proxy that just sits across your stack and makes it really obvious where your chatbot logs are and how you can monitor them and enables you ultimately to track latency costs errors across more than 100 model providers in a single gateway. So far so good. I actually really like it. A lot of companies use it. Another one that is also strong is Langfuse. Uh you can have observability tracing evaluation frameworks again they have SOC 2 they have ISO 27001 uh you can track parent child relationships with execution tracing uh you can automate quality assessment in ways that Helicone doesn't always attempt to do so there's some differences between the two to dig into I think in a sense the observability piece is something that we have had a little bit more runway on so it's it's been more of an obvious problem for a while whereas I think the vibe coding execution ution piece and the sandbox piece is newer because vibe coding itself is only a few months old and so we're still figuring out where the winds are there.
Let's move on to story delivery. Another gap that we called out I want to call out. This one again is maybe not quite as widely accessible as it could be. I believe it's in public beta now, but I do worry that they're going to get overwhelmed. Chronicle is out. So, Chronicle enables really, really high quality storytelling like pixel perfect components, built-in interactivity and motion. And the idea is that you want to get to the sort of massive consultant army that is always building PowerPoints and that is struggling to use AI to do so effectively. And so, you want to look for a workflow that is keyboard first, that enables presentation creation in 8 or 10 minutes versus hours. And you want to be in a position where you can deliver on that promise in a way that is near perfect out of the gate as long as you know what you want to say from a story perspective. You'll notice I am not mentioning Gamma here and that is because Gamma has been able to evolve but has not gotten to the level of professional quality where I or anyone else who presents to a serious CEO would really want to use that tool. It just hasn't been able to master the combined uh storytelling arc in visual and text. Storydoc is an option. It's a little bit more mature. It really is designed to create elements that ChatGPT can't conceptualize. It does not it does not fit neatly into the PowerPoint bucket in the same way that Chronicle does. And so I think in a sense part of what Chronicle is looking to do is to become the new PowerPoint with more dynamic features that PowerPoint just can't do. Uh and it's designed to key off the fact that we really like slides and we've had slides in the workplace for 40 years. So in that sense, I think Chronicle is better positioned for high stakes presentations where excellence matters, especially design excellence. And Storydoc is really handy if you just need to put together a quick somewhat visual doc. Maybe sales teams can use it, marketing content, that kind of thing.
All right, let's go to voice intake. So, we talked about the fact that ChatGPT, you know, just summarizes notes. There's a lot of note takingaking out there. I I use Granola. Granola is actually not what I'm going to talk about here. I want to talk about Nata and WhisperFlow. So, Nata is an extremely accurate, high quality audio transcriber, and it can process hour-long recordings in just like 5 minutes. Like it's it's very efficient at process and recording. It handles uh 58 different transcription languages, a bunch of them. And if you're just trying to get to meeting notes and transcribe them really effectively, not as great for that. It's definitely going to be, as most of these purpose-built tools are, better than your standard chatbot for the experience. WhisperFlow has a different approach. So, Nata is just obsessed with transcriptions, right? WhisperFlow is more like we think voice is the new interface and we're going to enable systemwide dictation. So you're going to be able to use WhisperFlow in all of your existing apps, which some people really like. Like they want to move to voice as the interface because we can talk faster than we can type. And so WhisperFlow gives them 3 or 4x on their traditional typing speed in a wide range of apps. And I think it's sub-second latency. I've tried it. It's not always sub-second latency, but it's it's quite fast. And it supports a hundred some languages with automatic detection. Again, I think it's really interesting to see in these examples how these products are solving different pieces of the problem. WhisperFlow really conceptualizes voice as an interface and so they're looking to plug into your existing apps whereas Nata conceptualizes voice as something that needs accuracy to transcribe and so they're just obsessed with that and it's just a very clean point solution. Your mileage is going to vary as is your problem set. You have to think about where you really care about the workflow speed up.
And as we wrap this up, that's where I want to leave you. I want you to think about your biggest time sink. I want you, if you're in a team, to think about your biggest shared pain point. Really, what you need is to get clear on that and then go back in and look at tools that make sense. I've laid out 12 tools here that I think are useful for some structural gaps that come up in LLMs. Your biggest time sink, your biggest pain point may or may not be one of those six issues that I identified with AI chatbots like ChatGPT. You may have a different one. That's okay. The point is this video should challenge you to think about where you are overindexing on time spent in a chatbot or time spent working around a chatbot flow and ask yourself, is there a point solution for AI that could solve this that I just haven't taken the time to invest in? And so if it would save you 10 hours weekly, it's worth finding out if there's an AI tool that can do it. And there are lots of stories across the 12 tools I've described that are in that category because you can imagine if you're using Shortcut and you can create a bunch of Excel sheets and that's your living, it's going to save you a lot of time. Similar way with Nata, if you're just trying to transcribe stuff, it's going to save you a ton of time. And so my challenge to you is to not regard the 100,000 tool universe of AI tools as this blank uh sea of tools that are just impossible to parse. There are useful tools in there. And the way to fish them out is understanding your own pain points. That's what really matters. That's what distinguishes people that can add tools strategically to their stack that fix what ChatGPT can't do versus people who are just rolling their eyes and saying it's too much. I can't do it. So there you go. Do you know your own pain points?