Should we have AI assisted thread summaries?

[neptronix]

Saw someone mention that they are developing AI assisted thread summaries for our forum software on the xenforo forum.

I think it would be pretty useful to have a TLDR of all threads for new people or casual ES enjoyers or someone who wants to dip their toe into a thread instead of deep dive.

I would be interested in deploying this only if the accuracy was very high and would be willing to tinker with it until it got there. I have a lot of experience running AI models on local hardware by now.

Here is a preview from the developer of this.

Let me know what you think of this feature.

 

[E-HP]

Saw someone mention that they are developing AI assisted thread summaries for our forum software on the xenforo forum.

I think it would be pretty useful to have a TLDR of all threads for new people or casual ES enjoyers or someone who wants to dip their toe into a thread instead of deep dive.

I would be interested in deploying this only if the accuracy was very high and would be willing to tinker with it until it got there. I have a lot of experience running AI models on local hardware by now.

Here is a preview from the developer of this.

Let me know what you think of this feature.

View attachment 374377

I have no familiarity with how these tools work, except for seeing things like how Google or other sites summarize things. Especially how Ai determines what information is valid vs bad or misleading. If someone creates a thread, and introduces bad information, how does Ai handle that, vs other good information that may be in the responses?

If you search Google on ebike stuff, it may draw a conclusion in the summary that's totally false, and clicking on the footnotes frequently bring you back to ES, with information taken completely out of context or associating it with info in another article to draw a really misleading conclusion.
If you search how to make an ebike faster, several articles may be referenced, but may be total garbage like this. Even the referenced article looks like it may have been Ai generated.:

"The motor provides power for the electric bike. Therefore, the wattage of the motor can directly determine the speed of the e-bike. If the above operations still can not meet your demand for speed, then the most direct way is to upgrade the motor and the controller that matches it. Of course, a 1000W motor will be much faster than a 750W one. You can replace the 500W motor with a 750W, 1000W, or even bigger one. But in addition to replacing the motor, there is one more thing that you must do. You need to replace the controller that matches the motor and battery to prevent the controller from overloading."

If the user quoted the above in their thread, how does the Ai summarization handle "Of course, a 1000W motor will be much faster than a 750W one." vs other valid information in the thread? If the Ai summary is wrong or misleading, but the user treats it as good, it seems like it could lead to some bad outcomes.

Anyway, I don't have enough knowledge about how mature the tech is so I have no real input, except wanting to understand more. On the other hand, if the tech can perform some sort of indexing functions, rather than summarizing the technical content, on some of the long 100+ page threads, so that the user can locate the information they are looking for, that could be a huge help. I don't know if that take more Ai brainpower or less, but it would be a nice to have.

 

[neptronix]

I'm in complete agreement.
I would not accept low accuracy results and i see how using an AI to make summaries without having the capabilities to do a good job is essentially a harm done to the forum instead of a benefit.

What about the inverse situation where we could figure out how to make it do a good job >= 95% of the time?
And also we put big warnings that the computer might be gaslighting you

 

[Epithemeus]

What would the cost and energy consumption look like? I think the value provided would be low (but not zero) so the cost (both monetary and environmental) would also have to be correspondingly low.

 

[harrisonpatm]

My opinion is that it wouldn't be very much value added, for the level of work that would need to go into setting it up

 

[neptronix]

What would the cost and energy consumption look like? I think the value provided would be low (but not zero) so the cost (both monetary and environmental) would also have to be correspondingly low.

Beyond awful to be honest, anyone playing with this technology is a pioneer still. I'm investing about $10k on hardware and it will need a dedicated 120v circuit to operate.

On the other hand, we can pre-generate the summaries maybe once weekly and only update threads that are active every ~7 days, which would reduce the summary generation time to maybe 1% of what it would be if you were just constantly refreshing the entire site.

I'd guess the cost is 30kw-hr of energy per month, which is about a gallon of gas.

About 150,000 different people visit ES in a 28 day period. This is a pretty good scale to for the amount of extra co2 we must burn using this technology while it is barely past an infantile state of development.



As far as the trajectory of open source AI software, it seems to be getting 10% better every 6 months per the amount of resources it consumes. hardware is moving slower. I think we will burn half a gallon of gas per month of electricity to provide this functionality next year.

For example, Alibaba's Qwen just dropped a small model that runs on a 24gb gpu and is known for being astoundingly fast and efficient. This tiny model competes with big commercial quality and runs on a tiny system. It's about 10% better than it's predecessor for the same resources. We see leaps like this every 6 months now.

 

[neptronix]

My opinion is that it wouldn't be very much value added, for the level of work that would need to go into setting it up

I'm happy to do the work on my own time as an experimental application. I'm self interested in learning ways to utilize the technology.

Upfront:
I am very willing to kill the idea if it turns out bad.
All data will be processed firstparty and is absolutely not for sale.

I think AI has been a joke until now and the technology is just starting to get good.
At our development shop we use it to good effect for a constrained set of uses. it has been so impactful that it's having to change our business model because we occasionally get 5 hours of work done in 15 minutes.

I think the technology is eventually going places and if we can utilize it in a way that provides tangible benefit, it could be a slam dunk for improving information accessibility by providing a way to glance at a 200 pager.

 

[neptronix]

Here's another wild thought.

You can finetune AI on a huge dataset and we have the largest dataset on our topic.
Finetuning AI produces much more accurate results and lower resource consumption.
When you see these dumb google summaries of our website, this is the product of not finetuning and really thinking about our content holistically.

It is feasible to achieve better results than what google is providing. I just don't know how much better.
You have to run the experiment in the computer lab to find out.

The only problem is that finetuning requires a buttload of resources which may or may not be out of our reach financially. But some organizations exist that aid open source organizations and i think we could qualify as one.

I pass the mic to the community on the topic now, i'll read feedback tomorrow.

 

[scianiac]

I think the real killer application here for AI is AI assisted search. Being able to read the posts directly and gain the knowledge that way means you get high accuracy the problem is being able to find them when the search tools often fail unless the thing is very specific. We just have so many technical and non-technical names and ways about talking about things it can be very hard to find what you are looking for without plugging in a dozen different names for something until you get lucky.

 

[neptronix]

I also think AI search would be a slam dunk but it could get really energy intensive to operate since every query is made to order and you don't have much possibility of caching at our scale since there's so much permutations of how a question can be asked in natural language.

Google does a good job of this but i think deepseek R1 is actually better.
But to get a good answer can take up to a minute out of deepseek R1.
Energy consumption wise it's less like burning a gallon a month and instead doing a cross country trip in a Hummer

But the software + hardware could someday open up the possibility.

 

[BlueSwordM]

Saw someone mention that they are developing AI assisted thread summaries for our forum software on the xenforo forum.

I think it would be pretty useful to have a TLDR of all threads for new people or casual ES enjoyers or someone who wants to dip their toe into a thread instead of deep dive.

I would be interested in deploying this only if the accuracy was very high and would be willing to tinker with it until it got there. I have a lot of experience running AI models on local hardware by now.

Here is a preview from the developer of this.

Let me know what you think of this feature.

View attachment 374377

Try something like Qwen3-30B-A3B-2507-Instruct to see how it performs.

Being a a 30B-A3B MOE, it should run very fast and very efficiently; it's also a super good LLM, not even considering the size.

 

[A-DamW]

I'm happy to do the work on my own time as an experimental application. I'm self interested in learning ways to utilize the technology.

Upfront:
I am very willing to kill the idea if it turns out bad.
All data will be processed firstparty and is absolutely not for sale.

I think AI has been a joke until now and the technology is just starting to get good.
At our development shop we use it to good effect for a constrained set of uses. it has been so impactful that it's having to change our business model because we occasionally get 5 hours of work done in 15 minutes.

I think the technology is eventually going places and if we can utilize it in a way that provides tangible benefit, it could be a slam dunk for improving information accessibility by providing a way to glance at a 200 pager.

I'll chime in and say go for it, based on what you said here Nep'

I am very willing to kill the idea if it turns out bad.

Nuff said.

 

[Epithemeus]

Unfortunately the type of forum user most likely to benefit from this is also the least likely to give input in this thread.

Glad you are well learned on the costs involved, not surprised I suppose I'm getting the vibe that you've been wanting to explore and play with AI and have just been looking for the right application. Sounds worth the cost to me then. I look forward to seeing what an individual can do with the tech rather than a big company.

 

[offGridDownUnder]

If we didn't have the AI approach to consider, what would we be doing?

We would be putting more effort into the Knowledgebase, for instance, focusing on the cultural changes needed to get that into focus as the way to find knowledge.

AI leads to stupid people. As the entry-level jobs are picked off, what then is the path for people to gain the experience to become mid- or top-level performers?

AI leads to incorrect results. So who has the knowledge to discriminate what is incorrect, and how did they develop this knowledge? How are newbies going to do the hard yards needed for a person to develop those skills?

With an AI inquiry path, how and when does the not-knowledgeable person express what they are doing such that anyone else can notice that the AI is feeding them crap? They interact with the AI - they don't ask the forum. This leads to silent failures when people come to your AI site and interact with it, spend money and time and fail, and decide your site isn't worth bothering with - all without anyone knowing they've done this. And then they tell others that this site is crap.

The value of this forum is the interaction with people. The cost of developing the ability and tools to think with is not somehow short-cut by the use of an AI. The AI is of benefit to the people who have already done the hard yards and already developed the reasoning skills - it sounds like a "solution" in search of a problem. Do it at home yourself if you're so fascinated - I think it's a dis-service here.

And the environmental costs clinch the matter.

 

[neptronix]

Hm, i'm not sure if a paragraph summary on the first page of a thread would lead to such nasty effects. I think you are imagining a chatbot?

AI obviously has bad reputation as there's been a lot of overpromise/under deliver and super misguided misapplication of the technology. And also some kinds of usage can actually lead to brainrot or reliance.

If we were to implement AI search at any point, my requirement would be that it cites sources accurately and leads people into conversations rather than preventing them from occurring. For sure i don't want robot interaction replacing human interaction.

Right now there is a substantial problem with the web ecosystem because google summaries finally got good lately. It's dropped web visits to pretty much everyone. But conversely it's made looking things up on the web a lot faster for many, so it's improved the usefulness of the search engine.

I think if we fine tuned an AI model we could provide a better AI summary than google could and it's possible that the output is useful most of the time.

I'd love to have a great knowledgebase but i think in the meantime, i'm looking to make ES content easier to dive into and digest by any means possible.

If most people don't like the idea, even if well executed, then i'll nix this one in service of the above goal.

 

[Papa]

Plug the following (without the quotes) into Mr G'.... or your A-eye of choice.

"site:endless-sphere.com How do i program my bafang controller for more speed?"
or...
"site:endless-sphere.com Who can help me fix my dead ebike?"

 

[neptronix]

Bang up job on google's part.

Here's the response from Deepseek R1 ( either the first or second best open source model you can run )


To program your Bafang controller for more speed, follow these steps cautiously. Modifying speed settings may void warranties, violate local e-bike laws, and strain your motor/battery. Proceed at your own risk.

---

Tools Needed​

1. USB Programming Cable:
   • Bafang-specific cable (e.g., Bafang USB Programming Cable).
2. Software:
   • Bafang Configuration Tool (Windows) or BESST Tool (requires dealer login).
   • Download: SDZ Bafang Config Tool (community version).

---

Step-by-Step Guide​

1. Connect Hardware:
   • Turn off your bike.
   • Plug the USB cable into the motor's display port (near handlebars) and your computer.
2. Install/Open Software:
   • Install the Bafang Config Tool.
   • Run the software and select your motor model (e.g., BBS02/BBSHD).
3. Read Current Settings:
   • Click "Read from device" to load your current configuration.
4. Adjust Speed Settings:
   • Speed Limit (km/h): Increase from default (e.g., 25 km/h → 45 km/h).
   • Wheel Diameter: Set accurately (e.g., 26", 700C). Avoid under-reporting to "trick" the system.
   • Current Limits:
     • Increase Basic Current (e.g., 18A → 25A) for more torque.
     • Caution: Exceeding 30A on BBS02/BBSHD risks overheating.
5. Write Settings to Controller:
   • Click "Write to device" to save changes.
   • Restart the bike and test.

---

Key Settings for Speed​

Parameter	Default Value	Safe Max	Risk
Speed Limit	25 km/h (15 mph)	45–50 km/h	Motor overheat, legal issues
Basic Current	15–18A	25A (BBSHD: 30A)	Battery/controller failure
Keep Current %	50–70%	80–90%	Reduced battery life
Wheel Diameter	Accurate size	—	Set correctly for speed accuracy

---

Important Notes​

• Legal Limits: Most regions cap e-bikes at 25 km/h (15.5 mph). Exceeding this may classify it as a moped/motorcycle.
• Safety Risks:
  • Overheating motor windings or MOSFETs.
  • Premature battery drain or failure.
• Display Tweaks: Some Bafang displays (e.g., DPC-18) allow temporary speed limit overrides via hidden menus (check your manual).
• Cadence Matters: Higher speeds require fast pedaling to maintain motor assistance.

---

Alternative Solutions​

• Voltage Upgrade: Swap to a higher-voltage battery (e.g., 52V instead of 48V) for a ~10% speed boost.
• Gearing Changes: Install a larger front chainring (e.g., 46T–52T) for higher top speed.

Disclaimer: Modifications may damage components or violate regulations. Consult local laws and a professional if unsure.

For advanced tuning, see the Endless Sphere Bafang Forum.




... they must have trained this on our website
I think this answer is pretty good and i like that it links to ES, despite misnaming us.


Here's the same question on Qwen3-235B-A22B-2507 ( half size version of the 1st or second best open source model right now )

Programming Your Bafang Controller for More Speed​

There are several methods to adjust your Bafang controller for higher speed, but it's important to understand the legal and safety implications first.

Methods to Increase Speed​

Using Display Menu Settings​

Some Bafang controllers allow speed adjustments through the display menu. You can set the maximum speed to 50km/h and adjust the speed percentage to achieve your desired maximum assist speed. For some models, you can simply stop and press specific buttons on the handlebar display unit to set speed limits.

Using Programming Cable and Software​

For more precise control, you'll need:

• A programming cable with USB connector on one end and green Higo connector
• Bafang Config Tool software

The programming process involves:

1. Connecting to your drive system
2. Finding your COM port
3. Opening the Bafang Config Tool
4. Adjusting specific parameters

When using the config tool, you can modify settings such as current limits and assist levels (e.g., Level 0: 0, Level 1: 10, Level 2: higher values).

Important Considerations​

• Legal Limits: Many regions have legal speed limits for e-bikes (commonly 32km/h or 20mph). Exceeding these may make your e-bike illegal for road use.
• Risks Involved: There are significant risks when modifying your controller, including potential damage to components, reduced safety, and voided warranties.
• Model Specific: The exact process varies depending on your specific Bafang model (BBS01, BBS02, M620, etc.).

Before proceeding, I recommend:

1. Identifying your exact Bafang controller model
2. Researching the legal speed limits in your area
3. Understanding the risks of modifying your e-bike's programming
4. Considering whether the potential benefits outweigh the risks and possible warranty voidance

If you decide to proceed, the Bafang central motor programming guide provides detailed instructions on configuring your motor while protecting it from possible breakdowns.



Again the first cited source is endless sphere.
Interesting because i've been kicking alibaba off this website whenever i see them because of abusive scraping practices. I guess they picked up just enough data!
This answer is less impressive but this AI model has 1/3rd the data packed into it as deepseek.


Asking it a more difficult question:



I didn't expect this smaller AI model to do anything but hallucinate things but it did okayish.

How'd deepseek do?



Luke "Liveforphysics" Workman is a legendary figure in the electric vehicle (EV) community, particularly on the Endless Sphere forum. Here's a comprehensive overview of his contributions, identity, and impact based on available sources:

1. Identity and Persona​

• Username Origin: His handle "Liveforphysics" reflects his obsessive passion for applied physics and extreme engineering, particularly in high-performance EVs29.
• Real Name: Luke Workman, a self-taught engineer known for hands-on experimentation (e.g., surgically implanting a rare-earth magnet in his finger to "feel" electromagnetic fields)211.
• Personality: Described as a "modern-day Nicola Tesla" – brilliant, unconventional, and relentlessly innovative, with a focus on solving real-world problems211.

2. Technical Contributions on Endless Sphere​

• Forum Activity: A prolific contributor to Endless Sphere, especially in high-power ebike/motorcycle builds and battery technology threads2911.
• Deathbike Saga: Documented his creation of the "Deathbike," a DIY electric motorcycle that reached 80+ mph and famously outpaced a Tesla Model S P85 in a drag race (110+ mph)911.
• Controller/Motor Expertise: Shared advanced controller programming techniques (e.g., tweaking Bafang Ultra Max controllers for safety at high power) and advocated for robust motor systems like the Agni axial flux motor29.

3. Professional Work & Innovations​

• Zero Motorcycles (2009–?): As a battery engineer, he revolutionized their packs, tripling range in his first year and quadrupling it later. He also integrated Agni motors into prototypes911.
• Luna Cycle Collaboration: Consulted on potted battery packs (Luna Fusion), making them waterproof/shockproof. Pioneered alternatives to spot-welding for stronger pack construction2.
• Battery Recycling Startup: Co-founded a company focused on ethical lithium battery recycling, addressing a critical gap in the EV ecosystem4.

4. Signature Projects and Philosophies​

• Performance as Spectacle: Believed high-speed EVs could inspire adoption by "being a spectacle" – e.g., his 80 mph ebike raced against gas bikes at Willow Springs9.
• Extreme DIY Culture: Built Franken-bikes with salvaged parts (e.g., a Zero DSR with custom charging ports and disabled safety interlocks)2.
• Open Knowledge Sharing: Advocated for open-source principles, echoing Endless Sphere's ethos. His posts demystified high-amperage systems and thermal management211.

5. Legacy and Industry Impact​

• Influence on EV Culture: Inspired a generation of builders, including professionals like Stéphane Melançon (battery expert) who credits Luke for his career path11.
• Speed Records: Held unofficial ebike speed records; his Deathbike remains a benchmark for DIY performance911.
• Environmental Mission: Viewed EVs as tools to "wean society off fossil fuels," driving his work in batteries and recycling24.

6. Controversies and Risks​

• Safety Boundaries: Known for disabling safety features on prototypes (e.g., throttle interlocks), leading to crashes and stripped gears2.
• Skepticism Toward Projects: His battery recycling venture faced initial doubts about feasibility, though supporters emphasized his track record4.

For deeper technical discussions or build logs, explore his Endless Sphere threads (forum link)39. His work exemplifies how DIY ingenuity can propel industry-wide advances in sustainable transportation .





Damn this is actually quite good, i'm not sure about the rare earth magnet in finger part though.

Deepseek R1 is really great at synthesizing a response from a bunch of web links in a way that's coherent, and all these numbers in the response are weblink references. I like that it cites it's sources.

What about generating a short summary?



..okay a little long, but the question was vague. Imagine this as a thread summary. Would you say the AI did a good job?
I read a gazillion liveforphysics posts back in the day and i'd say, this is generally on point and pretty good.

The only question is how well it would handle a single thread. That's the part we don't know.
Luckily, some Xenforo developer is currently writing the software that could do the thread summary part and we don't have to engineer this to find out, we just gotta buy their plugin to run the experiment.

 

[BlueSwordM]

View attachment 374487

Bang up job on google's part.

Here's the response from Deepseek R1 ( either the first or second best open source model you can run )


To program your Bafang controller for more speed, follow these steps cautiously. Modifying speed settings may void warranties, violate local e-bike laws, and strain your motor/battery. Proceed at your own risk.

---

Tools Needed​

1. USB Programming Cable:
   • Bafang-specific cable (e.g., Bafang USB Programming Cable).
2. Software:
   • Bafang Configuration Tool (Windows) or BESST Tool (requires dealer login).
   • Download: SDZ Bafang Config Tool (community version).

---

Step-by-Step Guide​

1. Connect Hardware:
   • Turn off your bike.
   • Plug the USB cable into the motor's display port (near handlebars) and your computer.
2. Install/Open Software:
   • Install the Bafang Config Tool.
   • Run the software and select your motor model (e.g., BBS02/BBSHD).
3. Read Current Settings:
   • Click "Read from device" to load your current configuration.
4. Adjust Speed Settings:
   • Speed Limit (km/h): Increase from default (e.g., 25 km/h → 45 km/h).
   • Wheel Diameter: Set accurately (e.g., 26", 700C). Avoid under-reporting to "trick" the system.
   • Current Limits:
     • Increase Basic Current (e.g., 18A → 25A) for more torque.
     • Caution: Exceeding 30A on BBS02/BBSHD risks overheating.
5. Write Settings to Controller:
   • Click "Write to device" to save changes.
   • Restart the bike and test.

---

Key Settings for Speed​

Parameter	Default Value	Safe Max	Risk
Speed Limit	25 km/h (15 mph)	45–50 km/h	Motor overheat, legal issues
Basic Current	15–18A	25A (BBSHD: 30A)	Battery/controller failure
Keep Current %	50–70%	80–90%	Reduced battery life
Wheel Diameter	Accurate size	—	Set correctly for speed accuracy

---

Important Notes​

• Legal Limits: Most regions cap e-bikes at 25 km/h (15.5 mph). Exceeding this may classify it as a moped/motorcycle.
• Safety Risks:
  • Overheating motor windings or MOSFETs.
  • Premature battery drain or failure.
• Display Tweaks: Some Bafang displays (e.g., DPC-18) allow temporary speed limit overrides via hidden menus (check your manual).
• Cadence Matters: Higher speeds require fast pedaling to maintain motor assistance.

---

Alternative Solutions​

• Voltage Upgrade: Swap to a higher-voltage battery (e.g., 52V instead of 48V) for a ~10% speed boost.
• Gearing Changes: Install a larger front chainring (e.g., 46T–52T) for higher top speed.

Disclaimer: Modifications may damage components or violate regulations. Consult local laws and a professional if unsure.

For advanced tuning, see the Endless Sphere Bafang Forum.




... they must have trained this on our website
I think this answer is pretty good and i like that it links to ES, despite misnaming us.


Here's the same question on Qwen3-235B-A22B-2507 ( half size version of the 1st or second best open source model right now )

Programming Your Bafang Controller for More Speed​

There are several methods to adjust your Bafang controller for higher speed, but it's important to understand the legal and safety implications first.

Methods to Increase Speed​

Using Display Menu Settings​

Some Bafang controllers allow speed adjustments through the display menu. You can set the maximum speed to 50km/h and adjust the speed percentage to achieve your desired maximum assist speed. For some models, you can simply stop and press specific buttons on the handlebar display unit to set speed limits.

Using Programming Cable and Software​

For more precise control, you'll need:

• A programming cable with USB connector on one end and green Higo connector
• Bafang Config Tool software

The programming process involves:

1. Connecting to your drive system
2. Finding your COM port
3. Opening the Bafang Config Tool
4. Adjusting specific parameters

When using the config tool, you can modify settings such as current limits and assist levels (e.g., Level 0: 0, Level 1: 10, Level 2: higher values).

Important Considerations​

• Legal Limits: Many regions have legal speed limits for e-bikes (commonly 32km/h or 20mph). Exceeding these may make your e-bike illegal for road use.
• Risks Involved: There are significant risks when modifying your controller, including potential damage to components, reduced safety, and voided warranties.
• Model Specific: The exact process varies depending on your specific Bafang model (BBS01, BBS02, M620, etc.).

Before proceeding, I recommend:

1. Identifying your exact Bafang controller model
2. Researching the legal speed limits in your area
3. Understanding the risks of modifying your e-bike's programming
4. Considering whether the potential benefits outweigh the risks and possible warranty voidance

If you decide to proceed, the Bafang central motor programming guide provides detailed instructions on configuring your motor while protecting it from possible breakdowns.



Again the first cited source is endless sphere.
Interesting because i've been kicking alibaba off this website whenever i see them because of abusive scraping practices. I guess they picked up just enough data!
This answer is less impressive but this AI model has 1/3rd the data packed into it as deepseek.


Asking it a more difficult question:

View attachment 374488

I didn't expect this smaller AI model to do anything but hallucinate things but it did okayish.

How'd deepseek do?



Luke "Liveforphysics" Workman is a legendary figure in the electric vehicle (EV) community, particularly on the Endless Sphere forum. Here's a comprehensive overview of his contributions, identity, and impact based on available sources:

1. Identity and Persona​

• Username Origin: His handle "Liveforphysics" reflects his obsessive passion for applied physics and extreme engineering, particularly in high-performance EVs29.
• Real Name: Luke Workman, a self-taught engineer known for hands-on experimentation (e.g., surgically implanting a rare-earth magnet in his finger to "feel" electromagnetic fields)211.
• Personality: Described as a "modern-day Nicola Tesla" – brilliant, unconventional, and relentlessly innovative, with a focus on solving real-world problems211.

2. Technical Contributions on Endless Sphere​

• Forum Activity: A prolific contributor to Endless Sphere, especially in high-power ebike/motorcycle builds and battery technology threads2911.
• Deathbike Saga: Documented his creation of the "Deathbike," a DIY electric motorcycle that reached 80+ mph and famously outpaced a Tesla Model S P85 in a drag race (110+ mph)911.
• Controller/Motor Expertise: Shared advanced controller programming techniques (e.g., tweaking Bafang Ultra Max controllers for safety at high power) and advocated for robust motor systems like the Agni axial flux motor29.

3. Professional Work & Innovations​

• Zero Motorcycles (2009–?): As a battery engineer, he revolutionized their packs, tripling range in his first year and quadrupling it later. He also integrated Agni motors into prototypes911.
• Luna Cycle Collaboration: Consulted on potted battery packs (Luna Fusion), making them waterproof/shockproof. Pioneered alternatives to spot-welding for stronger pack construction2.
• Battery Recycling Startup: Co-founded a company focused on ethical lithium battery recycling, addressing a critical gap in the EV ecosystem4.

4. Signature Projects and Philosophies​

• Performance as Spectacle: Believed high-speed EVs could inspire adoption by "being a spectacle" – e.g., his 80 mph ebike raced against gas bikes at Willow Springs9.
• Extreme DIY Culture: Built Franken-bikes with salvaged parts (e.g., a Zero DSR with custom charging ports and disabled safety interlocks)2.
• Open Knowledge Sharing: Advocated for open-source principles, echoing Endless Sphere's ethos. His posts demystified high-amperage systems and thermal management211.

5. Legacy and Industry Impact​

• Influence on EV Culture: Inspired a generation of builders, including professionals like Stéphane Melançon (battery expert) who credits Luke for his career path11.
• Speed Records: Held unofficial ebike speed records; his Deathbike remains a benchmark for DIY performance911.
• Environmental Mission: Viewed EVs as tools to "wean society off fossil fuels," driving his work in batteries and recycling24.

6. Controversies and Risks​

• Safety Boundaries: Known for disabling safety features on prototypes (e.g., throttle interlocks), leading to crashes and stripped gears2.
• Skepticism Toward Projects: His battery recycling venture faced initial doubts about feasibility, though supporters emphasized his track record4.

For deeper technical discussions or build logs, explore his Endless Sphere threads (forum link)39. His work exemplifies how DIY ingenuity can propel industry-wide advances in sustainable transportation .





Damn this is actually quite good, i'm not sure about the rare earth magnet in finger part though.

Deepseek R1 is really great at synthesizing a response from a bunch of web links in a way that's coherent, and all these numbers in the response are weblink references. I like that it cites it's sources.

What about generating a short summary?

View attachment 374489

..okay a little long, but the question was vague. Imagine this as a thread summary. Would you say the AI did a good job?
I read a gazillion liveforphysics posts back in the day and i'd say, this is generally on point and pretty good.

The only question is how well it would handle a single thread. That's the part we don't know.
Luckily, some Xenforo developer is currently writing the software that could do the thread summary part and we don't have to engineer this to find out, we just gotta buy their plugin to run the experiment.

If running locally, I'd recommend using llama.cpp/vLLM.

That way, you can much more easily control for context, system prompts, LLM probability tuning, etc.

Furthermore, I think there should be some local grounding using RAG to tell it some basic knowledge about lithium-ion cells and electric motors in general to prevent it to make too many mistakes

 

[neptronix]

We like lmstudio at the shop. Will soon have a multi GPU array and vLLM is the way to go for that because it has nice concurrency.

Running RAG on our knowledgebase ( if we could get a decent amount of information in there ) could produce some pretty awesome results i think.

 

[offGridDownUnder]

Do the membership terms permit you to use content through an AI?

 

[neptronix]

Kind of, what we don't want is it slop masquerading as communication. It's okay to use it for assistive or translation purposes.

 

[neptronix]

Maybe this would be a better use for AI:

What if you could ask some 200 page thread a question in english and get a coherent response that pointed you to the different pages containing what you're interested in?

it would be like our in-thread search, but on steroids.

This could be much more useful than an AI summary of a thread.
This could improve participation instead of reduce it by quickly getting people very quickly where they want.

deepseek r1 does the citation thing fairly well and it's the only AI i really like for research. I can quickly check the AI's work so i can be sure it's accurate. Whereas with other systems, the quality is assurance is: 'trust me, bro'

Prompt: please explain the endless sphere website in a long paragraph. Cite as many sources as possible.



The only problem with this idea is that we could absolutely not afford it. I think it would be the best use of AI but.... yeah, we need the cost of hardware and electricity to go WAY down.

Right now the limit of our AI power is a dedicated 120v residential circuit at my house. That's enough juice to generate summaries. but for this AI search needs a small data center worth of power.

maybe we are some years out from implementing the technology in a way that:
- respects everyone's privacy ( hard, you have to run your own infrastructure )
- has very high accuracy ( top level tech is close to meeting this requirement )
- increases forum usage instead of reduces it ( intentional design is necessary )
- somehow fits into our budget ( pray to Jensen Huang 4x/day to increase the chance of a hardware miracle )


Thinking this as i see steveo come back on the forum and say this on a KT thread:

I will dig through the 200 pgs of posts over next few days, as its been quite a while since i played with these.
-Steveo

 

[BlueSwordM]

Maybe this would be a better use for AI:

What if you could ask some 200 page thread a question in english and get a coherent response that pointed you to the different pages containing what you're interested in?

it would be like our in-thread search, but on steroids.

This could be much more useful than an AI summary of a thread.
This could improve participation instead of reduce it by quickly getting people very quickly where they want.

deepseek r1 does the citation thing fairly well and it's the only AI i really like for research. I can quickly check the AI's work so i can be sure it's accurate. Whereas with other systems, the quality is assurance is: 'trust me, bro'

Prompt: please explain the endless sphere website in a long paragraph. Cite as many sources as possible.
View attachment 374836


The only problem with this idea is that we could absolutely not afford it. I think it would be the best use of AI but.... yeah, we need the cost of hardware and electricity to go WAY down.

Right now the limit of our AI power is a dedicated 120v residential circuit at my house. That's enough juice to generate summaries. but for this AI search needs a small data center worth of power.

maybe we are some years out from implementing the technology in a way that:
- respects everyone's privacy ( hard, you have to run your own infrastructure )
- has very high accuracy ( top level tech is close to meeting this requirement )
- increases forum usage instead of reduces it ( intentional design is necessary )
- somehow fits into our budget ( pray to Jensen Huang 4x/day to increase the chance of a hardware miracle )


Thinking this as i see steveo come back on the forum and say this on a KT thread:

If you want something that's quite a bit better than Qwen3-30B-2507, you could try GLM 4.5-Air.

105B MOE with 12B active parameters that should run quite well quantized to 6-bit on Strix Halo/Apple MX Studios.

 

[neptronix]

cc @kauaicycler because i'm a nvidia enjoyer and can't run this big of a model.

Personally i'm waiting for Qwen3-32B-2507 because i was really impressed with the previous 14b model for the speed/accuracy when programming, it could do some relatively hard stuff!

Yeah i know the 30b-a13 or whatever is super fast but not as intelligent as the 'dense models'. I definitely prefer quality of output over raw speed.

Seems like AMD is also making an aggressive push on lowering the cost of hardware btw:
https://videocardz.com/newz/amd-pre...with-navi-48-xtw-gpu-and-32gb-memory-on-board
Too bad i just bought a 5090

 

[BlueSwordM]

cc @kauaicycler because i'm a nvidia enjoyer and can't run this big of a model.

Personally i'm waiting for Qwen3-32B-2507 because i was really impressed with the previous 14b model for the speed/accuracy when programming, it could do some relatively hard stuff!

Yeah i know the 30b-a13 or whatever is super fast but not as intelligent as the 'dense models'. I definitely prefer quality of output over raw speed.

Seems like AMD is also making an aggressive push on lowering the cost of hardware btw:
https://videocardz.com/newz/amd-pre...with-navi-48-xtw-gpu-and-32gb-memory-on-board
Too bad i just bought a 5090

5090

RIP bank account.
Should have gotten a 48GB RTX 4090 from Taobao smh.