XenForo AI roadmap?
- Thread starter back
- Start date
smallwheels
Well-known member
Currently nothing and also not for the forseeable future. They already have more than a challenge to bring out the announced and massively overdue releases 2.4 and 3.0 and there are a gazillion of feature wishes open. So nothing to expect.
Maybe, maybe not. As you say: It is a hype. Probably there are already toilet seats with AI on the market. The question is: What problem would you like to solve? Is it worth being solved? Is it feasible to solve? Can AI be part of a solution and if so: Is it worth using it or are there other/better/easier ways?
Implementing AI just to have AI in the product would help with marketing, for anything else it would just be a waste of time and therefor even damaging the product as the time to implement AI would be missing from implementing other things.
Other than that there have been countless discussions about possible uses and integrations AI in XF in this forum - you will find them with the forum search. Pretty much any aspect has been discussed from various points of view and there are also add ons available that bring AI to XenForo for those who want it.
ebikes.com
Active member
AI service is still very expensive and extremely power hungry. It also barely works at this point.
Commercial AI services are rationed, so taking advantage of their highly subsidized starter plan won't work on any kind of scale.
Another issue that is not solved - if you have a commercial AI provider and you get thrown into a tier higher than a basic plan, you're getting charged per token and that can get quite expensive. Some bot running a DDOS on that website could bankrupt the operator by abusing the AI features. So now you have to design a protection system too.
I think a lot of forum operators would not like the kind of bills they could rack up - this is probably best integrated via plugins, since it's not going to be an affordable feature for a while.
I can build self hosted AI services right now. To get adequate intelligence means i need about $10k of hardware. This hardware can serve 2 people simultaneously. It also draws 1.5kw when being used, so we also have another $100-$200/mo power bill on top of that. And that isn't enough power to do anything in realtime as users aren't willing to wait >20 seconds for a response.
The result is basically AI service that's 1/4 the cost of major providers. But your upfront cost is huge.
In short, this is so expensive at the moment that it's not financially feasible for a majority of people to use it. If i was the Xenforo team, i'd think about integrating it if/when it becomes affordable to use AI service.
Commercial AI services are rationed, so taking advantage of their highly subsidized starter plan won't work on any kind of scale.
Another issue that is not solved - if you have a commercial AI provider and you get thrown into a tier higher than a basic plan, you're getting charged per token and that can get quite expensive. Some bot running a DDOS on that website could bankrupt the operator by abusing the AI features. So now you have to design a protection system too.
I think a lot of forum operators would not like the kind of bills they could rack up - this is probably best integrated via plugins, since it's not going to be an affordable feature for a while.
I can build self hosted AI services right now. To get adequate intelligence means i need about $10k of hardware. This hardware can serve 2 people simultaneously. It also draws 1.5kw when being used, so we also have another $100-$200/mo power bill on top of that. And that isn't enough power to do anything in realtime as users aren't willing to wait >20 seconds for a response.
The result is basically AI service that's 1/4 the cost of major providers. But your upfront cost is huge.
In short, this is so expensive at the moment that it's not financially feasible for a majority of people to use it. If i was the Xenforo team, i'd think about integrating it if/when it becomes affordable to use AI service.
ebikes.com
Active member
You could do a hell of a job with this with AI, the problem is the cost of today's hardware.
Digital Doctor
Well-known member
the question with AI seems to be the cost.
If people want to improve xenforo search they should use google to do it for them.
DEMO:
Home
Our main goal is to provide easy-to-understand help to pool owners and to give them the tools to care their pools easily and stress-free.
www.troublefreepool.com
ebikes.com
Active member
I'm building an AI box as we speak.
It's got 3 Nivida 5090's and a 1500w power supply and costs about $10k to build.
This will serve a mid level AI coding assistant for a total of 3 people concurrently.
Extrapolate that cost up to 10's-100's of simultaneous requests for a forum and your cost is gonna get insane, you'll also need a commercial facility with multiple 240v circuits to serve that many people.
It's got 3 Nivida 5090's and a 1500w power supply and costs about $10k to build.
This will serve a mid level AI coding assistant for a total of 3 people concurrently.
Extrapolate that cost up to 10's-100's of simultaneous requests for a forum and your cost is gonna get insane, you'll also need a commercial facility with multiple 240v circuits to serve that many people.
ebikes.com
Active member
the question with AI seems to be the cost.
If people want to improve xenforo search they should use google to do it for them.
DEMO:
Home
Our main goal is to provide easy-to-understand help to pool owners and to give them the tools to care their pools easily and stress-free.
No kidding, that's the route i went, the xenforo search engine is hugely memory intensive and worse than phpbb, so i made google search the default option whenever possible
Digital Doctor
Well-known member
Cool !
What coding is it going to do ?
How hard would it be for me to make my own local AI scribe ?
I use this and it is pretty good.
Heidi Health - Best AI Medical Scribe for All Clinicians
Heidi Health is the best AI medical scribe that transcribes visits & generates notes to save time & focus on care. Trusted in 50+ countries – try for free today!
www.heidihealth.com
RobinHood
Well-known member
Does it not make more sense to use cloud services for this for lower cost of entry and considering how fast the landscape is changing?
ebikes.com
Active member
It's hard to say.
With commercial services, you lose a lot of privacy and confidentiality. I'm not okay with that because i work on projects i have to sign a NDA on.
Commercial services also cost 5x-15x more than running an open source model.
Your middle ground solution is to go with a service like fireworks.ai or deepinfra and rent an open source model.
Hardware will remain expensive until software improves because the next generation in 2026-2027 won't be dramatically better ( we're at the end of moore's law )
But it's still going to end up much more expensive than the server bill, especially if users have access to send queries to an AI.
ebikes.com
Active member
A little more expensive than the services i mentioned, way harder to integrate, and doesn't carry the cutting edge chinese AI models, which are doing an amazing job of reducing the hardware requirements, so you're stuck with more expensive models over there.
ebikes.com
Active member
Here's an overview of recent open source models i prepared for a software development client.
GLM 4.5 air, for 2 user simultaneously runs on 2 Nvidia 5090's ( $5000 ) and matches in quality some middle level commercial AI
You can get it as cheap as $1.1 per million tokens
Claude Opus for 2 users simultaneously needs about $40k of hardware to run.
You're going to pay $90 per million tokens.
You want to step down on Claude to their low range model, haiku.. which is worse than GLM 4.5?
That's about $4.8 per million tokens.. still 4.5 times the price
What's a million tokens? i'd equate it to about 10000 chatbot requests.
Huge difference in price, small difference in quality of output in reality. You really want some AI service that offers these Chinese models if you want to make adoption anywhere close to affordable. Even still, you want to put some controls in against bots abusing it.
GLM 4.5 air, for 2 user simultaneously runs on 2 Nvidia 5090's ( $5000 ) and matches in quality some middle level commercial AI
You can get it as cheap as $1.1 per million tokens
Claude Opus for 2 users simultaneously needs about $40k of hardware to run.
You're going to pay $90 per million tokens.
You want to step down on Claude to their low range model, haiku.. which is worse than GLM 4.5?
That's about $4.8 per million tokens.. still 4.5 times the price
What's a million tokens? i'd equate it to about 10000 chatbot requests.
Huge difference in price, small difference in quality of output in reality. You really want some AI service that offers these Chinese models if you want to make adoption anywhere close to affordable. Even still, you want to put some controls in against bots abusing it.
tonmo
Well-known member
I've implemented AI on my forum through my exclusive partnership with Xpertloop, using their API, which is using open.ai / GPT5.
[edit: Importantly, we had Xpertloop ingest my forum's content, so that it can provide responses prioritized with my forum's content, and if not enough context is provided, it falls back on ChatGPT (other models can be used as the backup, e.g. Claude, etc). I can set up multiple agents for different topics, but for now I just have one general agent to set tone, tenor, rules and formats, etc.]
I am thinking about packaging the solution as an addon, and offering onboarding and ongoing support.
Personally, I think it's awesome, but it's slow to take with our general membership, which is fine with me for the time-being. I'm basically the sole user, but it comes in handy for dead/dying threads or addressing topics that are wrapping up, but inconclusive. Octobot adds a lot of value in these scenarios, which are not uncommon. I'll be monitoring heavily however, to avoid abuse or garbage content. Limiting to Supporters makes a lot of sense in this regard.
Here's how it is positioned with members:
[edit: Importantly, we had Xpertloop ingest my forum's content, so that it can provide responses prioritized with my forum's content, and if not enough context is provided, it falls back on ChatGPT (other models can be used as the backup, e.g. Claude, etc). I can set up multiple agents for different topics, but for now I just have one general agent to set tone, tenor, rules and formats, etc.]
I am thinking about packaging the solution as an addon, and offering onboarding and ongoing support.
Personally, I think it's awesome, but it's slow to take with our general membership, which is fine with me for the time-being. I'm basically the sole user, but it comes in handy for dead/dying threads or addressing topics that are wrapping up, but inconclusive. Octobot adds a lot of value in these scenarios, which are not uncommon. I'll be monitoring heavily however, to avoid abuse or garbage content. Limiting to Supporters makes a lot of sense in this regard.
Here's how it is positioned with members:
Last edited:
Digital Doctor
Well-known member
Google is much better than any xenforo search - AINEC
goyo
Active member
I'm sure it helps in certain scenarios.. and I hope we're referring to the same implementation
I ran it on a forum (≈5M messages, 600K+ users), and for single-word searches (that was 90%+ of searches), it was inferior to ES/XES - surfaced inactive threads, reshuffled results between pages (often empty next page), white background on dark styles, etc.
Users didn't like it, so it's gone. This was 2-3 years ago, and not much has changed.
- Permissions - does not honor per-user visibility
- Language handling - Hungarian morphology/stemming and custom synonyms worked far better with ES analyzers. CSE treated words with and without accents as the same (can have totally different meanings)
- Freshness - crawl delays mean new/edited posts obviously lag vs XES
- Relevance control: very limited control of weights vs XES
- Theming/UX: embedded UI constraints
- Pagination stability: re-ranking between pages, "next page", or a page number can be empty or reshuffled
- Privacy/ads/quotas: logging lives with Google; ads unless you pay; quotas on larger forums; and changing terms.
Similar threads
- Question
