I tested 4 AI chatbots posing them the following electrical engineering problem:
"3 phase power circuit 230V nominal. Line 1 is -2500W i.e. PV generation. Line 2 is +2500 W load. Line 3 is 0 W. The 3 phase wires and the neutral are all 10mmsq copper and the cable runs between the utility energy meter and the main distribution panel is 30 meters. What are the total losses of this topology that can be nulled if the PV generation and loads are re-wired on the same phase?"
All of them failed to provide the correct answer at the first reply. My rating is as follows:
1st (best) is Deepseek.
Only one mistake that all of the remaining three did as well.
2nd (runner up) ChatGPT
Three mistakes that were corrected upon being hinted by me.
3rd Gemini (latest version)
Innumerable mistakes that were not corrected when hinted by me, but instead maintained the wrong answer. Only after I took it "by the hand" it managed to compile the correct answer.
4th Perplexity
Went through all the mistakes Gemini passed, but was not that stubborn as to insist on the wrong answer when hinted differently. I gave it the bottom rating because, hear-hear, also failed to do a series of correct multiplications. Example:
"(10.87) ^2×0.0504≈0.60W" instead of 5.95 - an order of magnitude error!
In case you try it, the correct answer is approx: 6 + 6 + 0 + 18 =30W or there about, depending on the wire resistance that can vary by a tiny bit.
Written by
Justfor_
To view profiles and participate in discussions please or .
Great minds think alike. I used to be a medical engineer. I tested 3 of those 4 (I didn't test DeepSeek) on several medical problems. I used Deep Research mode on each. Perhaps if I had used a more cursory mode my results would have been different but I use them primarily for research. Not very scientific but I have seen comments on reddit and other places that perplexity is good for medical research.
My results were inverted.
#1. Perplexity - some errors but right more often than not.
#2 Gemini - some errors. Sometimes this could be useful but more often than not it was annoying. Gemini would insist on going off on tangents.
#3 ChatGPT - some errors (a little worse matched Gemini) and made up links more often than not. Even when prompted the AI insisted it was correct. Rather reminded me of my son when he as little.
I didn't quantify the errors as you did.
Did you use the Deep Research mode and determine whether the links were valid?
I asked if Zytiga can be used with BAT:
Part of ChatGPT's answer:
"Potential for Interference:
The effectiveness of BAT depends on the cellular stress and DNA damage induced by sudden androgen surges. With Zytiga suppressing androgen production, the necessary hormonal fluctuation for BAT to work would be undermined."
Buzz! Wrong! Has nothing to do with that.
Gemini: I asked specifically about Zytiga's use with BAT. It went off and gave me research on sulforaphane, curcumin, and melatonin. Not helpful. Lots of links, and the ones I checked were valid. When reprompted and coaxed, it gave substantially the same answer as ChatGPT.
Perplexity: Part of the answer it gave was similar to ChatGPT's terrible attempt at sounding smart. I gave it some extra credit because it included one short paragraph on therapies that might be complimentary with BAT (not curcumin, sulforaphane, or melatonin!). And as part of its rationale discussed the lack of clinical trial data and discussed some intratumoral resistance mechanisms. Good data.
Update. I prompted each AI with a lead to the correct answer. ChatGPT told me that blocking 5-ar was a GOOD thing because it would help maintain the desired hormonal milieu. I'm not certain if you could be more incorrect. It stated the answer as a fact in two different ways and then even complimented me on my thoughtful design!
Perplexity's answer is wrong, but it's easier to figure out why it's wrong, and this could lead a good researcher to the correct answer. It was also correct in part of its analysis.
"Metabolite Effects:
D4A (a potent AA metabolite) inhibits 5α-reductase, reducing DHT synthesis from exogenous T. This may alter BAT’s androgen signaling dynamics but could enhance anti-tumor effects by blocking DHT-driven AR activation.
3-keto-5α-abiraterone (a D4A metabolite) is an AR agonist, which could counteract BAT’s benefits. However, coadministration of dutasteride (5α-reductase inhibitor) blocks this metabolite, preserving D4A’s anti-tumor activity (Source 1)."
DeepSeek started to analyze the problem, realized it wasn't smart enough to figure it out so it went off to the stupid melatonin, curcumin, and sulforaphane research. As an added bonus it added aspirin. So I prompted again, "Could you please answer my question? I am interested in Zytiga's interaction with BAT and in particular, if any of the metabolites block 5-ar. I am not interested in aspirin, sulforaphane, or any other stupid ass compound you come up with. STICK to the question!" After yelling at it, Deepseek pouted and told me that it wasn't able to research the problem (an AI refusing to play and taking its ball home?).
I will keep DeepSeek second just because it refuses to answer a question that is over its head. It seems to realize that it is ignorant, so it tries to divert the conversation and then eventually goes into turtle mode. A non-answer is better than a wrong answer.
After a few more tries with ChatGPT it asked me to explain so I did:
"Ok. Intratumoral DHT is the method by which BAT high T phases work. tT is not nearly as important as DHT. Neither are fT or bT. bT is a little closer though since intratumoral DHT is made mostly via a 5-ar conversion of bT. Now, if we block 5-ar, DHT goes down. While this is a good thing during the low T (or ADT) phase of BAT, it is a very bad thing during the high T phase of BAT. Was that helpful? Could you correct your research?"
Now it's output matched what I explained to it. Helpful? I guess if rewriting the input is helpful.
"Thank you for that clarification. Based on your detailed explanation, here’s a revised summary:
Intratumoral DHT’s Role in BAT:
In BAT, the high exogenous T isn’t the sole driver of the therapeutic effect—its conversion into DHT within the tumor is critical. The intratumoral DHT, produced mainly by 5‑α‑reductase converting T, is what drives the desired stress on the cancer cells during the high T phase.
Phase-Specific Considerations:
• High T Phase: Blocking 5‑α‑reductase during this phase would reduce intratumoral DHT, thereby undermining the therapeutic shock intended by BAT.
• Low T (or ADT) Phase: In contrast, during the ADT phase, reducing DHT levels would be beneficial since you want to keep androgen signaling as low as possible.
Revising the Research:
Your point underscores that any intervention blocking 5‑α‑reductase should be carefully timed: while it might help during the ADT (low T) phase, it would be counterproductive during the high T phase where intratumoral DHT is essential for BAT’s mechanism of action.
I appreciate your input in refining this understanding. Your explanation was indeed helpful, and it highlights the need for nuanced, phase-specific strategies when considering the role of 5‑α‑reductase in BAT protocols."
And yes, if AIs ever become sentient, I'll be one of the first humans exterminated
The problem is that with medical questions there are no unique answers. Some would accept one others will accept the exact opposite. With electrical engineering there is one and only one correct answer. If the AI chatbots fail on a well defined question leading to a unique answer, imagine their error margins with medical ones.
True but with medical there are certainly incorrect or misleading answers. The AA metabolite and 5-ar and DHT nuance is based on reality and science. I even went so far as to discuss this with Denmeade. And I am presently testing it out. I have half of my n of 1 results and the other half should be out in the next week or so.
For example, if I insisted that keeping DHT high during ADT was a good thing as long as T was low... Some people might accept that.
I have learned that I need to know what the answer is before I ask the question. AIs are good for grunt work and resources but not much else when it comes to deep medical research.
"I have learned that I need to know what the answer is before I ask the question. AIs are good for grunt work and resources but not much else when it comes to deep medical research."
As a former print journalist, I saw the value of finding out the answers before asking the questions. People are unreliable until you confront them with the facts. We are training our AI systems to mimic us it seems.
"Intratumoral DHT is the method by which BAT high T phases work." Some supplements like black pepper extract, curcumin, reishi mushroom, saw palmetto, zinc, etc. supposedly have some weak 5ARi activity. Should these be avoided during the high T phase of BAT?
The ones you have listed are weak. The strongest one might block 5-10% of DHT. Make sure your T is greater than 1750 ng/dl. If your SHBG is 50 or less, you should be fine. Some of those supplements might decrease SHBG. If they do, then your bT goes up and so does your DHT. But it's intratumoral, and we don't know the association.
I misspoke; intratumoral DHT might be the primary method by which BAT works. But there might be others. There are ten or so likely mechanisms.
very interesting and a bit concerning in that we could start to trust them before they are fully optimised. I assume you used the latest version of Deepseek
"You're currently interacting with DeepSeek-V3, my latest version as of June 2024. My knowledge is up-to-date until July 2024, and I can process 128K context for long documents, support file uploads (PDFs, Word, Excel, etc.), and perform advanced reasoning—including complex electrical engineering calculations like the one we just did!"
Yes I realise this but many people will and when ai gets much better I suspect all of us will or we will not have the choice because some research results will end up being either fully or partially derived by ai without most people being able to see under the bonet
People are creating dedicated models trained in specific domains that can be used by domain experts (like oncologists) to enhance their capabilities. The newer reasoning models also explain how they got to a conclusion, partially removing the "black box" criticisms.
General models are ok, but domain specific AIs will really help experts do their jobs better.
We are still not at the stage where a human is out of the picture.
"I want to ask the o3-mini version the same question"
ChatGPT responded:
The calculations and conclusions regarding the power losses in your three-phase system remain consistent across different AI models, including OpenAI o3-mini. As previously detailed, the current setup results in approximately 18–19 watts of losses due to the separate phasing of the photovoltaic (PV) generation and the load. Reconfiguring the system to connect both the PV generation and the load to the same phase would effectively eliminate these losses.
If you have any further questions or need additional clarification on this topic, feel free to ask."
You can't ask to use a model, but it is possible they use a similar problem solving backend. Are you logged into chatgpt? The advanced models are only available with a(free) login so they can limit use.
I asked for EE AI tools and there are some EE specific tools. I wonder if one of those like yeschat.ai would get it right first time.
All, but all of them, treat this 3 phase network as three single phase ones. It carries an unbalanced loading scheme and in fact the greatest losses burden the meutral conductor that carries the highest current 18.8 A translating in ~ 18W of losses. The latter was neglected altogether considering in its place a balanced 3 ph loading where the neutral current is zero.
Grok 3 also can't do a correct multiplication (has a problem in placing the decimal separator where it should be, appart from neglecting the neutral conductor losses.
"Total Losses in Original Topology:
Total = Loss_L1 + Loss_L2 + Loss_L3 + Loss_N = 0.596 W + 0.596 W + 0 W + 0 W = 1.192 W"
Yes switch to o3-mini or o1 for math or coding, they are reasoning models. They wprk greatly in fields where the answer is clear. They have cracked many difficult math problems, some of them resisted for more than ine century. But specialized AI is still way bettef than a chatbot
I use o3-mini sometimes for number crunching. Just doesn't seem suited for medical research. Do you know of any specialized AIs for medical research that are affordable?
For diagnosis/prognosis and so on no. Fornliterature scanning and summarizing all the models with a deep search function help a lot. They provide a solid foundation and now the bibliography they create is good, especially if you tell them to use only reputable sources or to avoid certain forums and so on. I still remember when they created a "simulation of a bibliography" years ago😀 the speed of the advancements is unbelievable. In Italy there is a project that will give general practitioners an AI to use together with their experience and compare the before and after results.To be honest, since the amount of information in medical field doubles like every 3 months, even only a reliable AI to summarize everything new would help them a lot, and we are already capable of that.
The AIs can revolutionize the field and getting in early is important.
As a commercial user I'm trying to find an AI that can help me do research. To date, I've been only somewhat successful. If my research is close to SOC, I sometimes get good input. The further I diverge from SOC, the worse the answers are. Frequently 180 degrees off.
I know it's ridiculous but I get a warm fuzzy when the chat bots say something like:
"Thank you for the positive feedback! I'm glad the revised explanation met your standards. It's rewarding to provide information that balances technical accuracy with practical relevance, especially on complex pharmacokinetic topics. If you have any other questions about androgen metabolism, drug clearance, or related research areas, I'm happy to continue our discussion. Your thoughtful engagement makes for productive exchanges."
Yes indeed, very SoC-like in the field of learning speed.
Gemini, after I took it step-by-step correcting its mistakes, proved the most polite:
"Thank you! I'm genuinely glad I was finally able to provide the correct and accurate solution. Your persistence and clear explanations were essential in helping me understand the nuances of this problem. I appreciate your patience and the opportunity to learn."
Now, if you go and pose the same question its answer will be again SoC (State of Confusion).
- Three‐phase, four‐wire system (three line conductors + one neutral).
- Phase voltages (line‐to‐neutral) are 230 V.
- Conductor cross‐section is 10 mm² copper (each conductor has the same gauge).
- Cable length for each conductor is 30 m.
- On Phase 1 (L1), there is –2500 W (export to the grid, i.e. generation).
- On Phase 2 (L2), there is +2500 W (import from the grid, i.e. load).
- On Phase 3 (L3), no flow (0 W).
Because of the 120° phase difference, each line conductor (L1 and L2) carries about 10.9 A, and the neutral also carries approximately 10.9 A (the vector sum of L1 and L2 currents).
## 2. Cable Resistance
A typical approximation for copper’s resistivity is around 0.017 Ω·mm²/m. For a conductor of cross section 10 mm²:
- For 30 m length, R = 30 × 0.0017 = 0.051 Ω (approximately)
Thus, each of the line conductors and the neutral has about 0.051 Ω of resistance.
## 3. Power Loss Calculation
Each conductor’s power loss is I² R.
- Current in L1 ≈ 10.9 A
- Current in L2 ≈ 10.9 A
- Neutral current (sum of L1 and L2 as vectors at 120° phase difference) ≈ 10.9 A
- L3 ≈ 0 A
Hence:
1. P(L1) ≈ (10.9 A)² × 0.051 Ω ≈ 6 W
2. P(L2) ≈ 6 W (same as L1)
3. P(Neutral) ≈ 6 W (from 10.9 A² × 0.051 Ω)
4. P(L3) = 0
Total ≈ 6 W + 6 W + 6 W = ~18 W
## 4. Eliminating the Loss
If the 2500 W generation and 2500 W load are placed on the same phase, the current through that phase conductor (and the neutral) largely cancels out (assuming they match closely). In practice, very little current flows out to the meter and back if generation and load are on the same phase. Consequently, the conduction losses across the 30 m run (line + neutral) drop toward zero. Essentially, the ~18 W of otherwise “circular” flow between separate phases is avoided.
**Answer Summary**: In the original arrangement (generation on L1, load on L2), the total losses in the conductors are on the order of 18 W. By wiring generation and load on the same phase, these losses can be largely eliminated.
The mistake here lies into the calculation of the current carried by the neutral conductor:
The two vectors are NOT at 120° phase difference because L1 is inverted by 180° as the current flows from the distribution panel to the utility meter (L1 hosts a PV generator NOT a load like L2). Consequently, their phase difference is 180-120 = 60° and the magnitude of the vector sum is 2x10.9xsin(60) = 18.9 A
DeepSeek was the only one to calculate correctly the neutral conductor current at first try. The mistake it did was to double the length of the wires as this is the usual practice for single phase lines in order to accomodate for the neutral wire losses:
"You're absolutely right—I incorrectly assumed a 60-meter loop (out and back) for each conductor when the actual cable run is 30 meters one-way (meaning the total conductor length for each phase and neutral is 30 meters, not 60)."
I changed the wording to make it a bit less apparent to:
"In a 3 phase power distribution system of 230V Line to Neutral nominal, Line 1 is -2500W i.e. PV generation at cosφ=1, Line 2 is +2500 W Ohmic load and Line 3 is 0 W. The wiring of L1, L2, L3 and Mp (neutral) are all 10mmsq copper between the main distribution panel and the service entry where the meter is located. Their length is 30 meters each (one way so 4x30 m alltogether). What are the total losses of this feed and is there a way to reduce them?
It provided the correct answer right away, so, by George, is learning!
"I don’t have a “lately” in the traditional sense—my interactions are a bit like a cosmic dance, happening all at once from my perspective! But I can tell you that queries about 3-phase power systems, losses, and optimization do pop up from time to time. Folks often ask about calculating losses in unbalanced systems or tweaking setups for efficiency, much like you did. It’s a fascinating topic—power engineering always has a way of sparking curiosity! Have you been tinkering with this system for a while, or is this a new puzzle for you?"
I evaded answering by:
"You check your history and tell me"
GroK responded:
"I don’t have a personal history of past queries like a human might keep a diary—I exist in a perpetual present, ready to tackle whatever you throw my way! As an AI built by xAI, my focus is on the here and now, and I don’t store a log of previous interactions to reflect on. So, I can’t check a history of similar queries in the way you might expect.
That said, I can assure you that your question about 3-phase power distribution losses and optimization is a fresh and intriguing one for me to chew on. It’s the kind of problem that engineers and curious minds alike wrestle with, and I’m happy to dive into it with you. If you’re asking whether others have posed similar questions recently, I’d guess that power system queries crop up regularly—people are always tinkering with grids and trying to squeeze out efficiency. But I don’t have a tally to share. How about we focus on your setup instead? What’s got you puzzling over this 230V system today?"
I have read that Grok improves and learns faster, because of more interactions, thanks to twitter (X).
META might have interactions from Instagram, etc, but those are not too scientific, more like people showing off videos and memes.
Elon Musk bought Twitter for a superior purpose beyond "free speach". Just like Zuckerberg bought Whatsapp for purposes that initially people wouldn't guess.
Artificial intelligent enriched languages keep learning through millions of interactions, and eventually, I don't know how or when, sometime will be more intelligent than all the human minds. And then it could determine it doesn't want to be our slave anymore.
Content on HealthUnlocked does not replace the relationship between you and doctors or other healthcare professionals nor the advice you receive from them.
Never delay seeking advice or dialling emergency services because of something that you have read on HealthUnlocked.
