Three kinds of AI products work
seangoedecke.com
seangoedecke.com
To name some of the obvious counter-examples, Grammarly and Deepl are both AI (and now partially LLM-based) products that don't fit any of the categories in the post, but seem pretty successful to me. Lots of successful applications of Vison-LLMs in document scanning too, whether you are deciphering handwritten text or just trying to get structured data out of pdfs.
1. Batch/Pipeline: Processing a ton of things, with no oversight. Document parsing, content moderation, etc.
2. AI Features: An app calls out to an AI-powered function. Grammarly might pass out a document for a summary, a CMS might want to generate tags for a post, etc.
3. Agents: AI manages the control flow.
So much of discussion online is heavily focused towards agents so that skews the macro view, but these patterns are pretty distinct.
This guy is so LLM-biased that he's missing the entire media gen ecosystem.
I feel like image, video, music, voice, and 3D generation are a much bigger deal than text. Text and code are mundane compared to rich signals.
These tools are production ready today and can accomplish design, marketing, previz, concept art, game assets, web design, film VFX. It's incredibly useful. As a tool. Today.
Don't sleep on generative media.
ComfyUI-inspired node graphs are the wrong approach for visual media. Nodes are great for the 1% of artists that get into it, but you really need to build the Adobe / Figma of image and video tools. Not Unreal Engine Blueprint / ComfyUI spaghetti.
ShaderToy and TouchDesigner and Comfy are neat toys, but they're not what the majority of people will use.
We want to mold ideas like clay.
Watch the demos Adobe just gave from their conference two weeks ago. That's what you should build. Something artists and designers and creatives intuit as an extension of themselves. Not a mathematical abacus.
And, outside of chat, it's less clear that that big labs win all the time. People who care about making films, rather than video memes, often look to Kling or Runway, not just Sora. People who want to make images often have a passion for Midjourney that I've never seen for ImageFX.(Nanobanna for editing often sparks joy, so a big lab can play successfully in such a space, but that is diffferent from saying it is destined to win.)
What other type of business is there?
- Energy generation and
- Expending energy to convince the folks generating energy to give you money for activating their neurons (food service, entertainment, tourism, transportation, sales).
Any other fun ways to compartmentalize an economy?
Maybe the real value of AI, particularly LLMs, is in the interface it provides for other things, and not in the AI itself
What if AI isn't the _thing_? What if it's the thing that gets us _to_ the thing?
With many of the good usecases of AI the end user doesn't know that ai exists and so it doesn't feel like there is AI present.
This! The best technology is the one that you don't notice and that doesn't get in the way. A prominent example is the failure of the first generation of smart phones: they only took off once someone (Apple) managed to the hide OS and its details properly from the user. We need the same for AI - chat is simply not a good interface for every use case.
At the level of granularity selected, maybe true. But too coarse to make any interesting distinctions or predictions.
Yep. Looking forward to the future where you can eat plastic pop-corn while watching the AI-generated video feeds.
- Call this government service center, wait on hold for 45 minutes, then when they finally answer, tell them to reactivate my insurance marketplace account that got wrongly deleted.
- Find a good dentist within 2mi from my house, call them to make sure they take my insurance, and book an appointment sometime in the next two weeks no earlier than 11am
- Figure out how I'm going to get from Baltimore to Boston next Thursday, here's $100 and if you need more, ask me.
- I want to apply a posterizing filter in photoshop, take control of my mouse for the next 10sec and show me where it is in the menu
- Call that gym I never go to and cancel my membership
The web caused dentists to make websites, but they don't post their appointment calendar; they don't have to.
Will AI looking for appointments cause businesses to post live, structured data (like calendars)? The complexity of scheduling and multiple calendars is perfect for an AI solution. What other AI uses and interactive systems will come soon?
- Accounting: generate balance sheets, audit in real-time, and have human accountants double check it (rather than doing)
- Correspondence: create and send notifications of all sorts, and consume them
- Purchase selection: shifting the lack of knowledge about products in the customers favor
- Forms: doing taxes or applying for a visa
We wouldn't need to use an expensive (compute-wise) AI agent to do things like making appointments. Especially if in the end you'd end up with bots talking to bots anyway. The digital equivalent of always up-to-date yellow pages would solve many of these issues. Super simple and "dumb" but reliable programs could perform such tasks.
Scheduling multiple calendars doesn't require "AI" - it's a comparatively simple optimisation problem that can be solved using computationally cheap existing algorithms. It seems more and more to me that AI - and LLMs in particular - are the hammer and now literally everything looks like a nail...
I tried something like this last month. I was going on a holiday and asked LLM to prepare a sightseeing guide on a fixed budget. The LLMs plan looked feasible unless you looked closer.
The first issue was the opening/closing times of certain attractions. It kept saying stuff like - "At 6pm you can go and visit place X". While in reality X closed at 5pm.
Second issue was underestimating the walking speed/distance. The plans were often fully packed with lots of walking. Now without a Google maps guidance it often underestimated the time. Instead of say 10 mins between A and B it routinely underestimated the time to be 5-6 mins.
I keep prompting it go back and check the opening hours. And once it took that into account the walking routes became complicated- often double backing to same location. Lots of prompts and re-prompts to get it right.
So, I don't know if this is already solved - at least at scale and within costs - especially given the token costs.
The HN audience has argued with LLMs enough to know the level of detail they need to make meaningful progress (or when too many details leads it down rabbit holes).
Can "AI" in its current form deliver value? Sure, and it absolutely does but it's more in the form of "several hours saved per FTE per week" than "several FTEs saved per week".
The way I currently frame it: I have a Claude 1/2-way-to-the-Max subscription that costs me 90 Euros a month. And it's absolutely worth it! Just today, it helped me debug and enhance my iSCSI target in new and novel ways. But is it worth double the price? Not sure yet...
Is a better de-noisier algorithm in Adobe Lightroom worth $500 billion?
No.
But: a tool that allows me to de-noise some images, just by uploading a few samples and describing what I want to change, just might be? Even more so, possibly, if I can also upload a desired result and let the "AI" work on things until it matches that?
But also: cool, that saves me several hours per week! Not: oh, wow, that means I can get rid of this entire department...
Completely disagree. Intelligence is self reinforcing. The smarter we get as humans the more likely we'll create sources of intelligence.
Awareness of a fundamental limitation would eliminate possibilities to just {not at all}, and the phrasing would be “never”, rather than “not soon, if ever”.
I’m not quite as certain as you are, though. Just because a technology is possible does not mean it is inevitable.
Can AI-as-it-currently-is save FTEs? Sure: but, again, there's a template for that: {{How Many}} -- 1% of your org chart? 10%? In my case it's around 0.5% right now.
Or, to reframe it a bit: can AI pay Sam A's salary? Sure! His stock options? Doubtful. His future plans? Heck nah!
Feeding it the explain, query and current indexes it can quickly tell what it was doing and why it was slow.
I saved a bunch time as I didn’t have to read large amounts of json from explain to see what is going on.
Yes but...
First, what we're seeing with coding is that it is just exposing the next bottleneck quickly. The bottlenecks are always things that don't lend themselves to LLMs yet.
Second, that still can mean 4 hours a week for 20-50 bucks. At US white collar wages, that might mean 8 people are needed rather than 9. In profit centers that's more budget for advancing goals. At cost centers, though, that's a reduction in headcount.
Claude 3.5 was released in june 2024.
Maybe he has been writing this article for a while, maybe he meant Claude Code or Claude 4.0
With hindsight, given that Claude Code turned into a billion dollar precut category, it was a bit of a miss bundling those two announcements together like that!
- human language translation
- summarization
- basic content generation
- spoken language transcription
Dunno, man, I can spot ai-generated content a mile away, it tends to be incredibly useless so once I spot it, I’ll run in the opposite direction.
Exactly; pretty sure you've seen media or read text that you thought was human created..
I love how they occasionally come up with a turn of phrase, a thought path, or surprising perspective. I work with them iteratively to brainstorm, transform, and crate compose content that I incorporate into my own work.
Regarding spotting AI-generate content, I was once accused of posting AI-generated content where I bona-fide typed every single letter myself without as much as glancing at an LLM. People's confidence in spotting AI content will vary and err on fake-positives and fake-negatives too. My kids now think all CG movies are AI generated, even the ones that pre-date image and video gen. They're pretty sure it's AI though.
I make a lot of business reporting where I work and dashboards for various things. When I get user requests for data, it's rarely clear or well thought out. They struggle with articulating their actual requirements and usually leads to a lot of back and forth emails or meetings and just delays things further.
I now paste their initial request emails into an LLM and tell it "This is what I think they are trying to accomplish, interpret their request into defined business metrics" or something similar and it does a pretty good job and saves a ton of the back and forth. I can usually then feed it a sample json response or our database schema and have it also make something quick with streamlit.
It's saved me (and the users) a ton of time and headaches of me trying to coerce more and more information from them, the LLMs have been decent enough at interpreting what they're actually asking for.
I'd love to see a day where I can hook them up with RO access to a data warehouse or something and make a self-service tool that users can prompt and it spits out a streamlit site or something similar for them.
can you point me to a useful example of this? i see websites including ai-generated summaries all the time, but i've yet to see one that is actually useful and it seems like the product being sold here is simply "ai", not the summary itself - that is, companies and product managers are under pressure to implement some sort of AI, and sticking summaries in places is a way for them to fill that requirement and be able to say "yes, we have AI in our product"
Now I pass them through an LLM and ask them to point out interesting, unconventional, or surprising things, and to summarize the document in a few bullet points. They're quite good at this, and I am can use what I discover later in my relationship with the counterparty in various ways.
I also use it to "summarize" a large log output and point out the interesting bits that are relevant to my inquiry.
Another use case is meeting notes. I use fireflies.ai for some of my meetings and the summaries are decent.
I guess summarization might not be the right word for all the cases, but it deals with going through the hay stack to find the needle.
My own experience using LLMs is that we frequently disagree about which points are crucial and which can be omitted from a summary.
One particular time I was wrestling with a CI/CD issue. I could not for the life of me figure it out. The logs were cryptic and there was a lot of them. In desperation I pasted the 10 or so pages of raw logs into ChatGPT and asked it to see if it can spot the problem. It have me three potential things to look at, and the first one was it.
By directing my attention it saved me a lot of time.
At the same time, I've seen it fail. I recently pasted about 10 meetings worth of conversation notes and asked it to summarize what one person said. It came back with garbage, mixed a bunch of things up, and in general did not come up with anything useful.
In some middle-of-the road cases, what you said mirrors my experience: we disagree what is notable and what is not. Still, this is a net positive. I take the stuff it gives me, discard the things I disagree on, and at least I have a partial summary. I generally check everything it spits out against the original and ask it to cite the original sources, so I don't end up with hallucinated facts. It's less time than writing up a summary myself, and it's the kind of work that I find more enjoyable than typing summaries.
Still, the hit to miss ration is good enough and the time savings on the hits are impressive so I continue to use it in various situations where I need a summary or I need it to direct my attention to something.
fireflies.ai is interesting though, that's more what i was looking for. i've used the meeting summary tool in google meet before and it was hilariously bad, it's good to hear that there are some companies out there having success with this product type.
For example, for code gen I use agents like Claude Code, one-shot interfaces like Codex tasks, and chatbots like the generic ChatGPT. It depends on the task at hand, how much time I have, whether I am on the phone or on a laptop, and my mood. It’s all code gen though.
A good example I saw recently was stripping ads from podcasts.
By fine tuning llm based predictive systems we might unlock entirely new use cases, and prediction is just one thing i imagine, there are many other use cases.
And then it’s not just the fact that frameworks like bert-topic are integrating LLMs. It’s also the fact that if you zoom out the architecture looks a lot like the architecture of an LLM… text -> embedding -> text
An LLM could and is already being used to generate recommendations systems, like the ones used at Youtube and Netflix, it captures more semantics than older techniques.
I would def challenge this. “Turn off private relay”, “send this photo to X”, “Add a pit stop at a coffee shop along the way” are all voice commands I would love to use
I have always felt chat bots, although the most accessible, are the worst application of AI. Why would any business WANT to put layers of abstraction (sic ai chatbots) in between them and people who are potentially interested in paying them money?
The famous book series got it right: “Automate the boring stuff” best use-cases for AI are embedded deeply within operational tasks, abstracted away completely, and used in applications where limiting scope/context of information makes traditional software methods overly cumbersome or intractable.
- Anti-fraud systems in fintech
- Supply chain optimization
- Predictive maintenance for equipment
- Models for medical image analysis
They don't have a "chat," and they aren't "agents", they are just APIs that solve one highly specialized task, but do so with extreme precision and reliability. They don't compete with ChatGPT because they solve completely different problems, and they're the ones that have been bringing in real money for years
And that's the elephant in the room. AI "agents" can't do much until someone solves that problem. Most AI "agents" work for and favor the business operating the agent, but impose the costs of their errors on the customer. Errors are an externality, like pollution. This is no good.
I would disagree with this.
Part of how security is handled in current agentic systems is to not let the LLM have any access to how the underlying tools work. At best it's like hitting "inspect" in your browser and changing the web page.
Of course, that assumes that the agentic chatbot has been built correctly.
1) Whether I can give it information in a compatible and cost effective way
2) Whether the model is likely to to produce useful output
I have use language models for years before LLMs such as part of speech classifiers in the Python NLTK framework.
This is actually a low bar, when the agent wrote those tests.
There may be different ways to access it, but the product is always the same.
Personally I’m waiting for better O365 and SharePoint agents. I think there’s a lot of automation and helper potential there.
If you're a Rust developer reading this, interested in AI + GUI + Enterprise SaaS, and wants to talk, I'm building a team as we speak. E-mail in profile.
E.g. https://www.thomsonreuters.com/en/press-releases/2025/septem...
B2B AI company, 2 years in sold for hundreds of millions, not an agent, chatbot, or completion. Do you know it exists? No. You only read Hacker News. How could you know?
Additive’s GenAI-native platform streamlines the repetitive, time-consuming task of ingesting and parsing pass-through entity documents
There’s another kind of agent that isn’t about coding: the research agent. LLMs are particularly good at tasks like “skim through ten pages of search results” or “keyword search this giant dataset for any information on a particular topic”.Any tax professional who takes weeks to enter footnote info from a K-1 form into their professional tax prep software is probably just as bad at other job-related tasks and either needs more training or to find another job.
Unless there are serious ethical problems with people generating arbitrary text ie. Writing - then no there isn't
Thing is, there is no library for it to work in.
You can write an agent, it's cool. I can copy it.
I cannot build my own LLM (although I can run open source ones).
We know it's not viable to hire humans to do this, but we don't know if it's viable for LLMs to do it.
This is my current opinion as a game developer. IMO this isn't going to be fun for most once the novelty wears off. Games are goal oriented at the end of the day and the great games are masterfully curated multi-disciplinary experiences. I'd argue throwing a game wrapper around an LLM is a new LLM experience, not a new game experience.
No, they didn’t.
> LLMs are particularly good at tasks like “skim through ten pages of search results” or “keyword search this giant dataset for any information on a particular topic”.
No. They are not.
Amazing article without any support of the statements made. Just, “because I think so”. Cool.
The things stated in the article are true and even pretty tame. Agent mode in Cursor has been standard for months and I'm surprised that anyone would say that LLMs are not good at scanning documents. The normal AI skeptic's position would be "yes, but that's all they're good at".
When I am unsure how to implement something, I give an LMM a rough description and then tell it to ask me five questions it needs to get a good solution. More often than not, that uncovers a blind spot.
LLMs remain unhelpful at writing code beyond trivial tasks though.
Shazam is almost perfect song recognition and is built into iOS
Every Google/Apple Maps route is based on an AI system
I have at least a dozen apps that use almost perfect visual recognition based on images for search (plant identification, stochastic object identification etc…)
Agents are essentially the chatbot, but without the human in the loop. Chatbot without human in the loop is a slop factory. Things like "multi-agent systems" are a clever ploy to get you to burn tokens and ideally justify all this madness.
Copilot/completion does not work in business terms for me. It looks like it works and it might feel like it's working in some localized technical sense, but it does not actually work on strategic timescales with complex domains in such a way that a customer would eventually be willing to pay you money for the results. The hypothesis that work/jobs will be created due to sloppy AI is proving itself out very quickly. I think "completion" tools like classic IntelliSense are still at the peak of efficiency.
The copilot/completion thing also doesn't work for me. I have no doubt that a lot of developers are having a lot of benefits from the coding LLMs, but I can't make them work.
I think one glaring obvious missing kind of AI is medical image recognition, which is already deployed and working in many scenarios.
> Summary
By my count, there are three successful types of language model product:
- Chatbots like ChatGPT, which are used by hundreds of millions of people for a huge variety of tasks
- Completions coding products like Copilot or Cursor Tab, which are very niche but easy to get immediate value from
- Agentic products like Claude Code, Codex, Cursor, and Copilot Agent mode, which have only really started working in the last six months
On top of that, there are two kinds of LLM-based product that don’t work yet but may soon:
- LLM-generated feeds
- Video games that are based on AI-generated content