Very informative video. Thank you. I am in Canada and had a client with a German Pension. I needed to translate it into English. We scanned it as a first step. I found that Claude worked best, because it could use a fixed width font in the artifacts window to replicate the columns.
Ok I did this by hand and noticed my number came in a lot lower than these at $148,644. Found the difference. The Deepthink and ChaptGPT models are missing a BIG deduction. The Qualified Business Income (QBI) deduction. I understand there is the 50% W-2 or 2.5% UBIA phaseout that kicks in at $380,000 taxable income for MFJ. However, in this scenario the QBI deduction would be worth probably about $84,868 in deduction. With such a large deduction the AI should have asked clarifying questions otherwise at the 35% tax rate they're leaving about $29,700 on the table. Also should be noted I used 2024 tax brackets.
Thank you Hector, this was fantastic test. Good to know that our clients will as confused as us till further notice 🤣 Also, i personally ranked Deepseek and Claude over gpts heck even perplexity over gpt.
Love the comparison, Hector. Can you also include your assessment of how far each of these is to your best answer (i.e. accuracy measure). Also, it would be good to understand the difference in cost for each of the models.
I would have suggested that you would have figured out the correct answer. Not sure why waiting 10 minutes for an answer would be a bother. I find most clients would not mind. But accuracy way more important than speed? Not can you figure it out. I am in Canada, and specialize in Canadian Taxes. Thank you. Always appreciate your videos.
@@RichardJessup-k8f thats right. But keep in mind that accuracy will be easily tweaked as the prompting gets better. I was simulating a prompt from a non-expert
@@HectorGarciaCPA Hey totally fair, appreciate all the videos. Gave you a thumbs up now that i understand the context it was made for. Thank you and love the training videos you make.
@@HectorGarciaCPA However, as the tax law updates consistently, I wouldn't use it for good. As a tax pro, we should have the ability to adapt any law changes and apply as needed ,that differentiate between the human brain and others.
I think he means the accounting profession is about to be decimated if not obliterated by these AI learning models. Ditto design engineering. About to get a lot less costly for the end customer of these high end services.
@@HectorGarciaCPAI think he means the accounting profession is about to be decimated if not obliterated by these AI learning models. Ditto design engineering. About to get a lot less costly for the end customer of these high end services.
@HectorGarciaCPA fellow beancounter here. A wrapper is a pre-configured app/front-end that uses a frontier model such as open AI or Gemini as it's "brain". Pre configured directions/instructions on its purpose and output with structured sources to draw on, such as state and federal tax libraries or case precedents. This improves the quality of the response.
@@HectorGarciaCPA I'm confused then, it is my understanding R1 requires LLMs to be taught before it runs, faster yes, but still requires the subscription based data?
To run **DeepSeek R1**, the minimum chipset requirements depend on the model size you're using. Here are the general requirements: - **For the 7 billion parameter model**: At least an **NVIDIA RTX 3060** with **12GB of VRAM**. - **For the 33 billion parameter model**: At least an **NVIDIA RTX 4080** with **16GB of VRAM**.
LOL. DeepSeek claimed? It is proven, you can install and run DeepSeek on your personal computer. As you are not an engineer, you are ill to say or advice others. I have both an accounting degree and computer science which makes me more qualified to challenge your claim. You can even build your own cloud as cloud is nothing but virtual servers and data servers. You can buy them with less than $10,000 to set them in your home.
![Lp. Сердце Вселенной #60 РОЖДЕНИЕ ЛОЛОЛОШКИ [Финал] • Майнкрафт](http://i.ytimg.com/vi/YoR0pAV9FVQ/mqdefault.jpg)
I did a another video focused on accuracy that shows ChatGPT very accurate: ua-cam.com/video/nggdCHhjgGo/v-deo.html
Very informative video. Thank you.
I am in Canada and had a client with a German Pension. I needed to translate it into English. We scanned it as a first step. I found that Claude worked best, because it could use a fixed width font in the artifacts window to replicate the columns.
I like Claude for that stuff
Ok I did this by hand and noticed my number came in a lot lower than these at $148,644. Found the difference. The Deepthink and ChaptGPT models are missing a BIG deduction. The Qualified Business Income (QBI) deduction. I understand there is the 50% W-2 or 2.5% UBIA phaseout that kicks in at $380,000 taxable income for MFJ. However, in this scenario the QBI deduction would be worth probably about $84,868 in deduction. With such a large deduction the AI should have asked clarifying questions otherwise at the 35% tax rate they're leaving about $29,700 on the table. Also should be noted I used 2024 tax brackets.
Thanks!... there you go... There are two big questions here, which is wether the two K-1's and Schedule C would qualify for QBI
Can you provide the prompt
Thank you Hector, this was fantastic test. Good to know that our clients will as confused as us till further notice 🤣
Also, i personally ranked Deepseek and Claude over gpts heck even perplexity over gpt.
Perplexity is one of my faves
Love the comparison, Hector. Can you also include your assessment of how far each of these is to your best answer (i.e. accuracy measure). Also, it would be good to understand the difference in cost for each of the models.
Ill be doing videos on accuracy later on. For now I wanted to address speed
I would have suggested that you would have figured out the correct answer. Not sure why waiting 10 minutes for an answer would be a bother. I find most clients would not mind. But accuracy way more important than speed? Not can you figure it out. I am in Canada, and specialize in Canadian Taxes. Thank you. Always appreciate your videos.
10mins vs 30 seconds? The problem is client expectations
so who's the most accurate according to the your calculation?
They were technically all OFF, as none of them considered QBI, but that can easily be trained.
@@HectorGarciaCPA thank you.
You need to click the deepsek search button to get the current values
I did click on that
Ask it for the justification for an 80 p/e for INTU?
@@jamescassidy6717 he he, hard to justify
It appears useless if you did not figure out the answer and accuracy. Speed is nothing if none returned the near correct results. ?
@@RichardJessup-k8f thats right. But keep in mind that accuracy will be easily tweaked as the prompting gets better. I was simulating a prompt from a non-expert
@@HectorGarciaCPA Hey totally fair, appreciate all the videos. Gave you a thumbs up now that i understand the context it was made for. Thank you and love the training videos you make.
I wonder if you could upload all the applicable IRS publications to notebook LLM and see how it could do.
In Theory yes, thats what TaxGPT and CPApilot is for
@@HectorGarciaCPA However, as the tax law updates consistently, I wouldn't use it for good. As a tax pro, we should have the ability to adapt any law changes and apply as needed ,that differentiate between the human brain and others.
thanks for sharing
You're welcome!
what is the correct amount?
none of those were 100%... but ChatGPT o1 and DeepThink R1 were closest
No, when you compare a paid model to a open source model already not apple apple comparison.
@@pizizhangsg1319 at $20/m is practically free.. so its going to be a similar comparison
U have to conclude the video with which AI has utmost accuracy
@@vignesh1326 none were accurate! Some were close
For context these are zero shot questions with no guidance or input sources. Wrappers are done.
Wrappers?
I think he means the accounting profession is about to be decimated if not obliterated by these AI learning models. Ditto design engineering. About to get a lot less costly for the end customer of these high end services.
@@HectorGarciaCPAI think he means the accounting profession is about to be decimated if not obliterated by these AI learning models. Ditto design engineering. About to get a lot less costly for the end customer of these high end services.
@HectorGarciaCPA fellow beancounter here.
A wrapper is a pre-configured app/front-end that uses a frontier model such as open AI or Gemini as it's "brain". Pre configured directions/instructions on its purpose and output with structured sources to draw on, such as state and federal tax libraries or case precedents. This improves the quality of the response.
@ and you think with Deepseek those wrappers are toast?
Running locally vs Subscription model. R1 still needs the large language model.
I did not run it locally, I used all the webbased tools in this video
@@HectorGarciaCPA I'm confused then, it is my understanding R1 requires LLMs to be taught before it runs, faster yes, but still requires the subscription based data?
R1 runs on your home based chipset, Groq is a specifically designed chipset for LLM?
To run **DeepSeek R1**, the minimum chipset requirements depend on the model size you're using. Here are the general requirements:
- **For the 7 billion parameter model**: At least an **NVIDIA RTX 3060** with **12GB of VRAM**.
- **For the 33 billion parameter model**: At least an **NVIDIA RTX 4080** with **16GB of VRAM**.
RTX 3060 for $300 or A100 for $9,000
LOL. DeepSeek claimed? It is proven, you can install and run DeepSeek on your personal computer. As you are not an engineer, you are ill to say or advice others. I have both an accounting degree and computer science which makes me more qualified to challenge your claim. You can even build your own cloud as cloud is nothing but virtual servers and data servers. You can buy them with less than $10,000 to set them in your home.
What’s your UA-cam channel, so I can learn more from you?