The premise behind this vibe-coded website is excellent. It would have been great if a bit more effort had been invested in the visualization side and, if instead of using "AI-rendered" images, you'd have fetched scientifically accurate representations from papers and open repositories.
Sorry, but I really wouldn't trust a website that uses Gemini's image models for generating scientific diagrams. The website itself is also vibe-coded. That's not an issue by itself, but there are lots of layout issues visible, see https://i.imgur.com/SidB6pI.png
A bit offtopic: If anyone wants to check whether a specific image is generated by Google models without using Gemini, go to https://images.google.com/, upload the image, click the "About this image" section. It'll say "Made with Google AI" if it was generated with their models.
AI really loves that purple-blue style with rounded corners. I asked chatGPT to make a few sites to see how it worked and any time I said "make the site look nicer", it did that. I wonder why.
At least put a big warning in your introduction box that highlights that those data might be invalid because you are not an expert in this topic. Let the use acknowledge this before continuing.
This is so riddled with inaccuracies that I can spot them immediately despite not being a phage biologist. For example, the PhiX image has a DNA with about 20 base pairs - wildly not to scale. M13 is also wildly scaled, and it clearly has a double stranded DNA which is labeled as single stranded.
What the hell is this amino acid view? This is not how genes work at all. This is biology 101 and it's completely wrong. Why did you buy a domain name to share disinformation that you don't even understand?
None of this is displayed in a way that would be useful to working biologists, and I don't see how this could be used as a teaching tool even if all the errors were corrected. This simply doesn't provide any insight into how phages work. Looking at a raw sequence is pointless (also that color scheme is incredibly garish) - you need annotations! The 3D structures don't have their domains labeled and you can't connect sequence features to structural elements.
Why wouldn't you just use all of the existing tools that already do all of this correctly? Look, I don't mean to gate keep, and it's great that you learned something (assuming you didn't vibe code this), but this is a lot of effort that could have been avoided if you had had a single conversation with a biologist of any background, or asked an LLM to critique your idea, or made a single reddit post asking if this would be useful.
Edit: This may come across as super harsh - but really, I love the enthusiasm and I hope you keep pursuing this. But the right place for this passion at this point in your life is a classroom or some kind of structured course.
Yeah we're all quickly figuring out that LLMs shift the engineering work from computer science to bullshit detection. You basically have to become that guy on the Internet who's always trying to prove you wrong when working with Claude Code. Otherwise you're going to build yourself a false reality and get skewered if you try to share it. I mean I've done it myself, because we're so used to blindly trusting the things other people built, that we forget we're the ones building it. Nothing in life is free.
LLMs may be enabling, but OP explicitly stated "I wanted to dig deeper into the subject, but not by reading a boring textbook", which, of course, would have eliminated the issues in this tool, or at least made it clear to them that they needed to dive deeper before publishing. I feel like there might be some analogy with blaming cars for drunk driving - by definition, not possible without cars, but you can drive responsibly if you choose to.
That's how Plato taught Aristotle. I'd much rather have a dialog than read a textbook. You just have to think critically and fact check. You can't just trust whatever the robot says because it interpolates knowledge.
Sure, but there apparently wasn't enough of a dialog either. With a textbook, you're confronted with facts and explanations that you didn't ask about, or even knew to ask about. Don't get me wrong - my original recommendation to take an interactive course is still the best option in my mind, as simplifications made for the benefit of the learner often lead to apparent contradictions that an instructor can clarify. But at some point you do just need the set of raw facts to be able to work with these systems.
Only genuine socialization does that, since you can flip through a textbook. It's the main reason why people go to universities. To hear all the answers they didn't think to ask. LLMs actually do that a little bit. It's both a wonderful and terrifying prospect, since there's so much risk with that kind of feature to introduce bias, but it's great when it works.
I find this to be a bizarre sentiment. It’s an artifact that exists. The chances of me making this by hand are 0. This would be a full time job for 3 years to research and build this. For 10 people. And it would have to charge a ton of money to access in that case.
This is ridiculous. If you are capable of editing the libraries when necessary, and you don't ask someone else to do it for you, then yes. If you can't do it yourself and need to ask someone to do it for you, then no.
Asking for something out of your head, whether you're requesting a custom sandwich at a drive-thru restaurant or you're the project manager of a software company, is not the same as making it yourself. If it were, then you wouldn't need to drive through the restaurant or be a mid-level managerial parasite to get what you envisioned.
We should still judge people's skills on the merits of what they can create themselves. I've never had much interest in people who just ordered things up from their underlings, and claimed they were the creators. That's stolen merit, if you ask me. Show me someone who can create a piece of code from scratch, and then we can evaluate how well they leveraged an AI or a team of people to do the next piece of code for them. Not every piece of code has to be written from scratch, but it needs to be understood. You seemed to assert that this project would've taken many people many years or it wouldn't otherwise have been written. That's not true. It could be written in a week by a competent human. The fact that it wasn't could be a clue that it doesn't add value to begin with, which in turn implies that its output may be unreliable (even if it were useful).
- i pressed "Amino Acids", and nothing updated below the toolbar. can't figure out what it does
- the "Tools" buttons looks like a segmented picker, but both seem to actually initiate a modal presentation
this tool seems interesting, but it would be worth polishing some of these ui quirks because my first impression was that it seems a bit broken (or confused me)!
but seems like a cool project otherwise, love people building and sharing explainers as they learn stuff!
It looks nice but very "LLMsy". I liked that it specifically annotates "Lytic" vs "Lysogenic" categorization for phages. Important in terms of virome analysis.
https://pmc.ncbi.nlm.nih.gov/articles/PMC11097211/
Very good idea and very interesting. Great use of AI. The next step in this project, if you were funded, would be to have human experts double check the accuracy of the information and images.
Since this isn't funded, you could have a frontier model evaluate all the information for inaccuracies and fix them it can't find any. Then have another frontier model do the same. Then go back to the first model and see if it finds any inaccuracies.
Keep at it. You are doing a great job, but there is more work to be done before you can productionize it.
What used to be a life time project that would inspire awe and respect and make OP an instant hire for most managers is now a fun 1-week stunt that makes you go “cool, how many tokens?” Of course, the result is cool and maybe even useful (I wouldn’t dare say _correct_, being ignorant of the topic), but I cannot help but think that this would have been tremendously better if done the old (proper?) way.
Also, I suspect that OP would have learned so much more on the topic.
Yeah but the choice isn't "do I spend two weeks on this, or do I spend a lifetime?". It's "I have two weeks, do I spend them making the whole thing with AI, or 5% of it without?".
It's absolutely no issue to build all the technical and design stuff with AI, but science facts must be science facts and not just AI gibberish presented like facts.
What's the benefit of presenting not trustworthy data?
It must be me getting old, but definitely 5% of it without, if your goal is to learn something. 5% of thinking about a hard problem is more valuable than 100% of letting it being done by someone else.
Especially when the value of the end result is close to zero, since it is now so easy to do and replicate.
I'm pretty much on the no-AI side (learning, art, decision making, etc.), but this is the kind of thing I can appreciate. I suppose OP didn't want to learn more about coding this kind of visualization, but rather learn from the visualization. Any tool that can help with that is acceptable, whether it produces code or not. That the tool produced code has the advantage that OP can share this with us. I only hope it doesn't contain fundamental errors, because that would make this project a negative contribution.
Sure, I don’t discard the contribution altogether, but I am dubious that it is possible to properly draw the line between what to vibe code and what to do it “by hand” to make sure you get the benefits of building.
Errors are just nonsense that shows up in the console until you spend more tokens to make them go bye bye, right? I think you're talking more about the idea of true and false information, which is such a human bias. Will it really matter to anyone in 5 years whether this accurately depicts phages? By then AI will have solved everything. /s
Research into phages is paramount, as they represent one of our best hopes for combating the rapidly increasing problem of antibiotic resistance (largely driven by the overuse of antibiotics, even including last-resort antibiotics, in industrial animal agriculture so industrial farms can place more animals per sq/m without them dying from lack of space and cut-off body parts so they take even less space).
I really resonate with your goal of creating a more intuitive tool than a boring textbook. Being able to visually see how complex genetic code translates into physical geometry (3D structures) would be incredibly helpful for students. Thank you for sharing such a wonderful educational tool!
The premise behind this vibe-coded website is excellent. It would have been great if a bit more effort had been invested in the visualization side and, if instead of using "AI-rendered" images, you'd have fetched scientifically accurate representations from papers and open repositories.
Sorry, but I really wouldn't trust a website that uses Gemini's image models for generating scientific diagrams. The website itself is also vibe-coded. That's not an issue by itself, but there are lots of layout issues visible, see https://i.imgur.com/SidB6pI.png
A bit offtopic: If anyone wants to check whether a specific image is generated by Google models without using Gemini, go to https://images.google.com/, upload the image, click the "About this image" section. It'll say "Made with Google AI" if it was generated with their models.
>> what I ended up with has taken a sickening number of tokens to generate
I was a bit confused by this as to whether it related only to the graphics or to the UI as well.
AI really loves that purple-blue style with rounded corners. I asked chatGPT to make a few sites to see how it worked and any time I said "make the site look nicer", it did that. I wonder why.
At least put a big warning in your introduction box that highlights that those data might be invalid because you are not an expert in this topic. Let the use acknowledge this before continuing.
This is so riddled with inaccuracies that I can spot them immediately despite not being a phage biologist. For example, the PhiX image has a DNA with about 20 base pairs - wildly not to scale. M13 is also wildly scaled, and it clearly has a double stranded DNA which is labeled as single stranded.
What the hell is this amino acid view? This is not how genes work at all. This is biology 101 and it's completely wrong. Why did you buy a domain name to share disinformation that you don't even understand?
None of this is displayed in a way that would be useful to working biologists, and I don't see how this could be used as a teaching tool even if all the errors were corrected. This simply doesn't provide any insight into how phages work. Looking at a raw sequence is pointless (also that color scheme is incredibly garish) - you need annotations! The 3D structures don't have their domains labeled and you can't connect sequence features to structural elements.
Why wouldn't you just use all of the existing tools that already do all of this correctly? Look, I don't mean to gate keep, and it's great that you learned something (assuming you didn't vibe code this), but this is a lot of effort that could have been avoided if you had had a single conversation with a biologist of any background, or asked an LLM to critique your idea, or made a single reddit post asking if this would be useful.
Edit: This may come across as super harsh - but really, I love the enthusiasm and I hope you keep pursuing this. But the right place for this passion at this point in your life is a classroom or some kind of structured course.
Yeah we're all quickly figuring out that LLMs shift the engineering work from computer science to bullshit detection. You basically have to become that guy on the Internet who's always trying to prove you wrong when working with Claude Code. Otherwise you're going to build yourself a false reality and get skewered if you try to share it. I mean I've done it myself, because we're so used to blindly trusting the things other people built, that we forget we're the ones building it. Nothing in life is free.
LLMs may be enabling, but OP explicitly stated "I wanted to dig deeper into the subject, but not by reading a boring textbook", which, of course, would have eliminated the issues in this tool, or at least made it clear to them that they needed to dive deeper before publishing. I feel like there might be some analogy with blaming cars for drunk driving - by definition, not possible without cars, but you can drive responsibly if you choose to.
That's how Plato taught Aristotle. I'd much rather have a dialog than read a textbook. You just have to think critically and fact check. You can't just trust whatever the robot says because it interpolates knowledge.
Sure, but there apparently wasn't enough of a dialog either. With a textbook, you're confronted with facts and explanations that you didn't ask about, or even knew to ask about. Don't get me wrong - my original recommendation to take an interactive course is still the best option in my mind, as simplifications made for the benefit of the learner often lead to apparent contradictions that an instructor can clarify. But at some point you do just need the set of raw facts to be able to work with these systems.
Only genuine socialization does that, since you can flip through a textbook. It's the main reason why people go to universities. To hear all the answers they didn't think to ask. LLMs actually do that a little bit. It's both a wonderful and terrifying prospect, since there's so much risk with that kind of feature to introduce bias, but it's great when it works.
It's rather nice looking, but it would have been so much nicer if you'd done it yourself.
I find this to be a bizarre sentiment. It’s an artifact that exists. The chances of me making this by hand are 0. This would be a full time job for 3 years to research and build this. For 10 people. And it would have to charge a ton of money to access in that case.
> This would be a full time job for 3 years to research and build this. For 10 people.
No it wouldn't.
He did do it himself.
If you meant "without AI", then it would have never been done in the first place, so you can choose your preference there.
If he paid the same amount of money to an offshore worker to build it, would you say he'd done it himself?
Yes? A product designer does make a product, even though they don't write any of the code.
Perhaps we have a different definition of doing something yourself.
If you write an application, but use libraries, did you do it yourself?
This is ridiculous. If you are capable of editing the libraries when necessary, and you don't ask someone else to do it for you, then yes. If you can't do it yourself and need to ask someone to do it for you, then no.
Asking for something out of your head, whether you're requesting a custom sandwich at a drive-thru restaurant or you're the project manager of a software company, is not the same as making it yourself. If it were, then you wouldn't need to drive through the restaurant or be a mid-level managerial parasite to get what you envisioned.
We should still judge people's skills on the merits of what they can create themselves. I've never had much interest in people who just ordered things up from their underlings, and claimed they were the creators. That's stolen merit, if you ask me. Show me someone who can create a piece of code from scratch, and then we can evaluate how well they leveraged an AI or a team of people to do the next piece of code for them. Not every piece of code has to be written from scratch, but it needs to be understood. You seemed to assert that this project would've taken many people many years or it wouldn't otherwise have been written. That's not true. It could be written in a week by a competent human. The fact that it wasn't could be a clue that it doesn't add value to begin with, which in turn implies that its output may be unreliable (even if it were useful).
some quick feedback on the user interface:
- i pressed "Amino Acids", and nothing updated below the toolbar. can't figure out what it does
- the "Tools" buttons looks like a segmented picker, but both seem to actually initiate a modal presentation
this tool seems interesting, but it would be worth polishing some of these ui quirks because my first impression was that it seems a bit broken (or confused me)!
but seems like a cool project otherwise, love people building and sharing explainers as they learn stuff!
It looks nice but very "LLMsy". I liked that it specifically annotates "Lytic" vs "Lysogenic" categorization for phages. Important in terms of virome analysis. https://pmc.ncbi.nlm.nih.gov/articles/PMC11097211/
Very good idea and very interesting. Great use of AI. The next step in this project, if you were funded, would be to have human experts double check the accuracy of the information and images.
Since this isn't funded, you could have a frontier model evaluate all the information for inaccuracies and fix them it can't find any. Then have another frontier model do the same. Then go back to the first model and see if it finds any inaccuracies.
Keep at it. You are doing a great job, but there is more work to be done before you can productionize it.
What used to be a life time project that would inspire awe and respect and make OP an instant hire for most managers is now a fun 1-week stunt that makes you go “cool, how many tokens?” Of course, the result is cool and maybe even useful (I wouldn’t dare say _correct_, being ignorant of the topic), but I cannot help but think that this would have been tremendously better if done the old (proper?) way.
Also, I suspect that OP would have learned so much more on the topic.
Yeah but the choice isn't "do I spend two weeks on this, or do I spend a lifetime?". It's "I have two weeks, do I spend them making the whole thing with AI, or 5% of it without?".
If it's inaccurate it's worthless anyhow.
It's absolutely no issue to build all the technical and design stuff with AI, but science facts must be science facts and not just AI gibberish presented like facts.
What's the benefit of presenting not trustworthy data?
That's an entirely different argument, though. Let's not move the goalposts.
It must be me getting old, but definitely 5% of it without, if your goal is to learn something. 5% of thinking about a hard problem is more valuable than 100% of letting it being done by someone else.
Especially when the value of the end result is close to zero, since it is now so easy to do and replicate.
Eh, my goal is to make something, personally.
I'm pretty much on the no-AI side (learning, art, decision making, etc.), but this is the kind of thing I can appreciate. I suppose OP didn't want to learn more about coding this kind of visualization, but rather learn from the visualization. Any tool that can help with that is acceptable, whether it produces code or not. That the tool produced code has the advantage that OP can share this with us. I only hope it doesn't contain fundamental errors, because that would make this project a negative contribution.
Sure, I don’t discard the contribution altogether, but I am dubious that it is possible to properly draw the line between what to vibe code and what to do it “by hand” to make sure you get the benefits of building.
Errors are just nonsense that shows up in the console until you spend more tokens to make them go bye bye, right? I think you're talking more about the idea of true and false information, which is such a human bias. Will it really matter to anyone in 5 years whether this accurately depicts phages? By then AI will have solved everything. /s
Look at the commit history. I’ve been working on this essentially every single day for over a month.
Research into phages is paramount, as they represent one of our best hopes for combating the rapidly increasing problem of antibiotic resistance (largely driven by the overuse of antibiotics, even including last-resort antibiotics, in industrial animal agriculture so industrial farms can place more animals per sq/m without them dying from lack of space and cut-off body parts so they take even less space).
I really resonate with your goal of creating a more intuitive tool than a boring textbook. Being able to visually see how complex genetic code translates into physical geometry (3D structures) would be incredibly helpful for students. Thank you for sharing such a wonderful educational tool!