• Welcome to ASR. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

New AI Kid On The Block: Deepseek

amirm said:
They can do this?
Click to expand...

It is after all the US military. The rules there are far different from civilian organizations where you have a private life away from work. As a member of the US armed forces, you are subject to the rules of the branch of the service that you belong to—whether on or off duty. What with the defense of the Constitution, the nation, and such.
 
sheninat0r said:
OpenAI introduced "reasoning models" with GPT-o1. This approach is based on older research using "chain-of-thought" prompting techniques which "manually" induce the model to produce the reasoning/thought process by explicitly prompting it to do so across multiple turns of conversation. The innovation of GPT-o1 was specifically that they trained the model to generate both reasoning tokens and answer tokens in a single turn, DeepSeek did not invent this.
Click to expand...
I'm aware, but my point is that Deepseek exposes this information clearly (which is at least for programming, probably engineering or problem solving in general more valuable than the answer itself), and it's the first one to do that (out of the box for free), no?
 
Last edited:
Bow_Wazoo said:
What sonic disadvantages arise when using the non-oversampling filter with a Delta Sigma DAC, for example the Cirrus Logic CS43131?

Deepseek:
Using a **non-oversampling (NOS)** filter with a **Delta-Sigma DAC** like the **Cirrus Logic CS43131** can introduce several sonic disadvantages. These arise because Delta-Sigma DACs are specifically designed to work with oversampling and digital filtering, and bypassing these features can compromise their performance. Here are the key sonic disadvantages:

---

### 1. **Aliasing Artifacts**
- **Issue**: Delta-Sigma DACs rely on oversampling to push quantization noise into higher frequency ranges, which are then filtered out digitally. An NOS filter skips this oversampling process, allowing aliasing artifacts to appear in the audible frequency range.
- **Sonic Impact**: Unwanted high-frequency distortion and unnatural artifacts, particularly in the treble range.

---

### 2. **Increased Quantization Noise**
- **Issue**: Oversampling reduces quantization noise by spreading it over a wider frequency range. Without oversampling, this noise remains concentrated in the audible spectrum.
- **Sonic Impact**: Higher noise floor, reducing dynamic range and potentially masking fine details in the music.

---

### 3. **Loss of High-Frequency Detail**
- **Issue**: Delta-Sigma DACs are optimized to reproduce high-frequency details accurately through oversampling and digital filtering. An NOS filter lacks this processing, leading to a less precise reproduction of high frequencies.
- **Sonic Impact**: Softer or less detailed treble, resulting in a less airy and less precise sound.

---

### 4. **Nonlinear Frequency Response**
- **Issue**: Delta-Sigma DACs use oversampling and digital filters to achieve a flat frequency response. An NOS filter can introduce irregularities, especially near the Nyquist frequency (half the sampling rate).
- **Sonic Impact**: Uneven frequency reproduction, potentially making the sound unbalanced or colored.

---

### 5. **Phase Shift and Group Delay**
- **Issue**: NOS filters often have linear phase characteristics, which can introduce phase shifts or group delays in the signal.
- **Sonic Impact**: Smearing of transients, affecting the clarity and timing of fast-paced music, such as percussion or complex passages.

---

### 6. **Incompatibility with Delta-Sigma Architecture**
- **Issue**: Delta-Sigma DACs are designed to leverage oversampling and digital filtering for optimal performance. Using an NOS filter bypasses these features, undermining the DAC's intended design.
- **Sonic Impact**: Reduced precision and linearity, leading to a less accurate and less technically refined sound.

---

### 7. **Subjective Listening Experience**
- **Issue**: Some audiophiles prefer the sound of NOS filters, describing it as "more analog" or "natural." However, this is subjective and depends on personal taste and the rest of the audio system.
- **Sonic Impact**: While some listeners may enjoy the NOS sound, others may find it less detailed or less faithful to the original recording.

---

### Summary
Using an NOS filter with a Delta-Sigma DAC like the Cirrus Logic CS43131 can lead to several sonic disadvantages, including aliasing artifacts, increased noise, loss of high-frequency detail, and a nonlinear frequency response. These issues arise because Delta-Sigma DACs are optimized for oversampling and digital filtering, and bypassing these features can compromise their performance. However, whether these disadvantages are perceived as drawbacks depends on individual listening preferences and the specific audio system in use.

ChatGPT:
Using a non-oversampling (NOS) filter with a Delta-Sigma DAC, like the Cirrus Logic CS43131, can lead to several sonic disadvantages, primarily related to the way that digital audio signals are processed and converted to analog. Here are some of the key issues that may arise:

Frequency Response and Phase Distortion: Non-oversampling filters do not employ the same type of digital filtering techniques that oversampling filters do. This can lead to less controlled frequency response and phase distortion, especially at frequencies near the Nyquist limit. This can result in less accurate reproduction of high-frequency content.

Increased Quantization Noise: Delta-Sigma DACs work by oversampling the input signal and using noise shaping to push quantization noise out of the audible range. By not oversampling, a NOS filter may not effectively manage quantization noise, leading to higher levels of noise in the audible range, which can negatively impact the clarity and detail of the audio.

Aliasing Artifacts: Non-oversampling can lead to more pronounced aliasing, where high-frequency signals can fold back into the audible range. This can create unwanted artifacts and distortions in the sound, particularly when handling complex musical passages.

Limited Dynamic Range: The dynamic range of a DAC is often improved with oversampling techniques due to better noise shaping and filtering. NOS filters may not take full advantage of the potential dynamic range of the DAC, resulting in a less dynamic sound.

Poor Transient Response: Oversampling can improve transient response by allowing the DAC to better handle fast changes in the audio signal. Non-oversampling may lead to smearing of transients, resulting in a less precise and less engaging audio presentation.

Less Effective Anti-Aliasing: Oversampling filters can act as effective anti-aliasing filters, reducing artifacts in the signal. Non-oversampling filters may not provide the same level of protection against aliasing, which can affect the overall fidelity of the audio.

Higher Distortion Levels: The lack of oversampling can lead to higher levels of distortion, particularly in complex audio signals. This can result in a less natural and more "digital" sound.

In summary, while non-oversampling techniques may appeal to some audiophiles for their perceived simplicity and "purity," they can introduce several sonic disadvantages when used with Delta-Sigma DACs. These disadvantages often manifest as increased noise, distortion, and artifacts that can detract from the overall listening experience.
Click to expand...
A very ill-posed question gives BS results (obviously), and some of the answers are so intrinsically wrong (like the aliasing thing) that it really hurts. And because AI has no actual knowledge and basically is only parroting (and maybe extrapolating), you can't ask scientific or engineering questions and expect correct results anway.

What was the purpose of that exercise?

Delta-Sigma DAC always operate with massive oversampling, that's the whole idea of it and how it works. In the purest form the output is one single bit toggling (DSD).

There is an anti-imaging filter which comprises of an digital filter and an analog filter. For the digital filter, some chips have an option to emulate the NOS "stair-steps" behavior but that has nothing to do with the oversampling (it is always in place, as is noise-shaping etc). And true NOS DAC incorporate a steep analog filter which does a reasonable amount of anti-imaging / reconstruction

So perhaps you wanted to ask what are the implications from using a completely filterless and non-oversampling digital-to-analog conversion, no reconstruction filter whatsoever (also known as zero-order-hold output)? The answer will still be questionable, of course, see above.
 
No. I was referring to the mode that Cirrus describes as NOS.

4.8 Filtering OptionsTo accommodate the increasingly complex requirements of digital audio systems, the CS43131 incorporates selectable filters in different playback modes.For PCM/TDM mode, the following interpolation filtering options can be selected:
• Fast roll-off and slow roll-off interpolation filter options.
• In each option above, both low-latency and normal phase-compensation filtering options can be used.
• Nonoversampling (NOS) mode is provided, which minimizes the internal digital processing. Once NOS mode is set, the settings on the above two options are ignored.

Elsewhere it is actually referred to as an emulation:
20250129_105847.jpg


"2 Wideband Flatness Mode The CS43131 specifically optimizes for wideband flatness playback, which is enabled by PCM_WBF_EN bit. This mode should only be enabled when PCM plays back at 192 kHz. When operating in other speeds or modes, this bit should be properly disabled. If wideband flatness mode enabled, the filter spec complies to Table 3-10. There is no option for filter roll-off or phase response in this mode. NOS filter mode should be disabled for proper operation. Note that wideband flatness mode can only be enabled or disabled when PDN_HP is set."

Did I understand correctly that the input sampling rate should be 192khz when the NOS mode is activated?
 
Last edited:
amper42 said:
Interesting article on Copyright rules for works created with various levels of AI support:


Apparently, it excludes copyright for any work created solely by AI but if recognizable human design, creativity, content is involved in the work it may qualify for Copyright. This could easily offer varying degrees of gray in the implementation.
1. How do you measure new human input versus the AI bringing other peoples work into the final rendition?
2. How is the original copyright acknowledged that was used by the AI?
3. How does one author have the right to infringe on another's work simply by re-arranging things?

It's possible this ruling may be challenged depending on exactly how it is applied.
Click to expand...

I am not sure if it goes far enough. I still think the AI system should credit the sources it used to "supplement" the one provided by the human user. So it's *both* the human's and the AI system's responsibility to credit their sources.

I don't think the % of human vs AI content is relevant, since absolutely *nothing* AI does is based on content it has genuinely "created" - it's all based on human work. Whether that human work is copyrighted/protected or is "general use" is a different matter. It is the responsibility of those feeding the AI system to keep that categorization as they keep adding to the knowledge base.

Incidentally, in the company I work for, Deepseek was strictly blocked on all corporate computers effective today.
 
pablolie said:
Incidentally, in the company I work for, Deepseek was strictly blocked on all corporate computers effective today.
Click to expand...
Eventually someone will figure out how to use "innocent looking" email addresses that are received by and replied to by the AI chatbot...fun times!
 
mhardy6647 said:
I found this deliciously amusing.
www.bbc.com

OpenAI says Chinese rivals using its work for their AI apps

ChatGPT maker says it will need extra protection from US government, following emergence of Chinese rival, DeepSeek.
www.bbc.com www.bbc.com
apnews.com

Did DeepSeek copy ChatGPT to make new AI chatbot? Trump adviser thinks so

Did the upstart Chinese tech company DeepSeek copy ChatGPT to make the artificial intelligence technology that shook Wall Street this week?
apnews.com apnews.com


View attachment 424671
View attachment 424672
Click to expand...
Isn't this what AI is all about, distilling everybody's published data for your benefit.
 
AI could be the end of "publish or perish" in Universities.
 
I would not trust Deepseek with your data...

www.wired.com

DeepSeek’s Popular AI App Is Explicitly Sending US Data to China

Amid ongoing fears over TikTok, Chinese generative AI platform DeepSeek says it’s sending heaps of US user data straight to its home country, potentially setting the stage for greater scrutiny.
www.wired.com www.wired.com

www.wired.com

Exposed DeepSeek Database Revealed Chat Prompts and Internal Data

China-based DeepSeek has exploded in popularity, drawing greater scrutiny. Case in point: Security researchers found more than 1 million records, including user data and API keys, in an open database.
www.wired.com www.wired.com

I'm sure the CCP is collecting everything.

Martin
 
Timcognito said:
AI could be the end of "publish or perish" in Universities.
Click to expand...
Fur sure it will *change* the form of it.

For beginners if AI companies want to use content they *must* pay for it. Just like the music streaming services would be dead in the water if they hadn't content deals with artists AI should never have been allowed to do what it's doing right now harvesting content with utter impunity.

How that payment looks like I don't know. many models are imaginable, ranging from taxation to licensing agreements to hybrid models.

I am interested in the topic so I joined a class action lawsuit to just be in the know of how these things go. Market economy principles often mean that things are shortsightedly regarded as a free resources until some litigation or such thing goes down. And we've heard the saying "nothing is free in America" many times... so wake up AI companies.

It took us long enough to come up with stuff like GDPR that protects data and its commercial use - in fact that is still work in progress 25 years after the internet boom that fed the pockets if data harvesters whose monetization models basically was to break your trust in them...
 
Whatever we think of the state of AI, and its current state of morality/integrity - it is going to be an incredibly useful resource.

I needed to extract every email address in about 1.5 years of incoming emails. I wont go into the detail of why. The only way I could get the mails out of my client was in the form of an Apple mailbox file. I tried googling for how to extract the addresses from that. I got nowhere.

I went to AI (happened to be google ai studio, but I tried deep seek also, and got similar results, although it also offered a python script.)

Hi. I have an apple mailbox file, and i need to extract a list of all email addresses from it. Can you help with that?
Click to expand...

It created a full grep command with all the regex I needed with alternatives provided for eliminating duplicates, sorting and piping to a file. Full explanation of how everything worked.


grep -oE '[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}' ~/my_emails.mbox | sort | uniq > emails.txt
Click to expand...

I asked for a mod
Hi. Can you use the filename of my mbox file which is "mbox". Can you also filter so that only emails with the domain "collinsho.me", and that are less than 31 characters long are extracted.
Click to expand...

And with another full page of explanation.
grep -oE '[a-zA-Z0-9._%+-]+@collinsho\.me' mbox | awk 'length($0) <= 30' | sort | uniq > emails.txt
Click to expand...

So after 30 minutes of failing with google, I had a perfect working and customised solution in about 3 minutes with AI.



EDIT - I'm going to quote the whole text following the modification question. It is incredible how much supporting information, specific to my use case, it provided...

Okay, let's refine the command to extract email addresses from your mbox file (named "mbox"), filtering for those with the domain "collinsho.me" and a total length of 30 characters or less.

Here's the modified command:

grep -oE '[a-zA-Z0-9._%+-]+@collinsho\.me' mbox | awk 'length($0) <= 30' | sort | uniq

content_copy download
Use code with caution.Bash
Let's break it down:

  1. grep -oE '[a-zA-Z0-9._%+-]+@collinsho\.me' mbox:
    • This part is similar to the previous commands.
    • -oE tells grep to print only the matches using extended regular expressions.
    • [a-zA-Z0-9._%+-]+@: Matches the local part of an email address (before the @).
    • collinsho\.me: Matches the domain specifically as collinsho.me (escaped . for a literal dot).
    • mbox: Specifies your file named mbox.
    • This will output all the email addresses that end in @collinsho.me
  2. | awk 'length($0) <= 30':
    • The pipe | sends the output of grep to the awk command.
    • awk is a powerful text processing tool. Here, it's used to filter based on the length of each email address.
    • length($0) gets the length of the current line (which is an email address since that was the output of grep).
    • <= 30 filters and only outputs lines that are 30 characters or less in length.
    • This will now only contain email addresses that are collinsho.me and less than or equal to 30 characters
  3. | sort | uniq:
    • The final part, as before:
      • Sorts the filtered output alphabetically with sort.
      • Removes any duplicate addresses with uniq
How it Addresses Your Requirements

  • Filename: The command now uses mbox as the input file.
  • Domain Filter: The regular expression in grep ensures only email addresses with the domain collinsho.me are selected.
  • Length Restriction: The awk command filters further, selecting addresses that are 30 characters or less in length (including the @ and domain name).
  • No Duplicates: The use of sort | uniq ensures that duplicate addresses are removed from the final list.
Example

If your mbox file contained the following entries:

From: [email protected]
To: [email protected]
Cc: [email protected], [email protected]
Content: some text [email protected] more text, [email protected]
Shortest: [email protected]

content_copy download
Use code with caution.
The output of the command would be:

[email protected]
[email protected]
[email protected]

content_copy download
Use code with caution.
Important Notes:

  • Escaping Dot: The dot (.) in the domain name needs to be escaped (\.) in the regular expression to treat it as a literal dot, not as a wildcard character.
  • Case Sensitivity: grep by default is case-sensitive. If you need to find addresses in any case (e.g. Collinsho.me), you can add the -i option to grep: grep -i -oE '[a-zA-Z0-9._%+-]+@collinsho\.me' mbox ...
  • Adjust Length: If you want addresses less than or equal to 31 characters then change the awk statement to length($0) <= 31
This command should give you a more specific and filtered list of the email addresses you're looking for from your mbox file.
Click to expand...
 
Last edited:
antcollinet said:
Whatever we think of the state of AI, and its current state of morality/integrity - it is going to be an incredibly useful resource.
...
Click to expand...

Absolutely no one doubts it. It provides a powerful abstraction level to use complex products. I think what you seemed to do is to use AI to help use the Linux grep command very effectively for what you were trying to do.

In my work one of the top customer pain points is the complexity of using our products too. By feeding thousands and thousand of pages of documentation and training the model things get much easier very quickly. You can ask the model "why did I just get this alarm?" with the hex stuff that gets spit out with that alarm you can feed it system configurations and remediation happens up to 70% faster easily. According to some McKinsey studdies in our industry that easily saves $70B in combined operational cost across the industry. Plus people get more sleep. :)

My objection is to the reckless wild west open harvesting model. It's never about the awesome potential when AI is used "positively" within more contained use cases that provide immediate value everybody sees.
 
Last edited:
mhardy6647 said:
I found this deliciously amusing.
www.bbc.com

OpenAI says Chinese rivals using its work for their AI apps

ChatGPT maker says it will need extra protection from US government, following emergence of Chinese rival, DeepSeek.
www.bbc.com www.bbc.com
apnews.com

Did DeepSeek copy ChatGPT to make new AI chatbot? Trump adviser thinks so

Did the upstart Chinese tech company DeepSeek copy ChatGPT to make the artificial intelligence technology that shook Wall Street this week?
apnews.com apnews.com


View attachment 424671
View attachment 424672
Click to expand...
1738194993610.jpeg
 
You must log in or register to reply here.

Similar threads

E
Bachelor Study on AI in the Music Industry
Replies
0
Views
337
Emanuel1103
E
Confused_by_tech
Audio Stem Splitters and AI - What's the current status?
Replies
4
Views
725
antcollinet
antcollinet
Alter.Nate.37
I need help to find an ideal portable BT DAC/Amp for the HifiMan Arya Stealth.
Replies
13
Views
905
Brian Hall
Brian Hall
I
Ground/earth solution for class 2 amplifier with some voltage on the chassis
Replies
17
Views
843
Speedskater
Speedskater
S
New DAC on the block: WADAX
Replies
4
Views
775
MaxwellsEq
MaxwellsEq
Back
Top Bottom