Master AI (Artificial Intelligence) Discussion/News Thread

[respice finem]

Dunno but I've actively used them.

I have the new iPhone 17 but have all the ai shit turned off, I use "-ai" in the tab in google so no ai there but I rarely use google anyway, ai is of no use to my existence so I'm perfectly happy to let it pass on by

duckduckgo doesn't use ai,

I did the same in Android. Sidenote: Duckduckgo offers LLM access, but you can choose if you want to use it or not.

 

[Axo1989]

Don’t know how accurate this video is as I’ve never used Ai/LLM’s but he discusses how to phrase a request/question/statement so you get a more informative and accurate response

I hadn't consideredthe self-help guru style to selling 'AI' (in the general meaning). He gives a decent sense of the LLM model and avoids much of the common anthropomorphism so not a bad thing. At least to halfway, which is where I stopped, like you I have no great desire to use the chatbots or related tools at this stage.

Interesting that the byline pitch is 'you're not behind (yet)' so are many/most people really feeling fomo about this stuff?

 

[napfkuchen]

Don’t know how accurate this video is as I’ve never used Ai/LLM’s but he discusses how to phrase a request/question/statement so you get a more informative and accurate response

Seems to be an overview of the common best practices, we were given the same advice in some prompting courses at work. Unfortunately, for me AI can only be a rough guideline but its results always have to be checked. Just recently there was a case where I had to decide whether a profit from debt forgiveness is tax-free or not. The first step was to determine whether the tax assessment could even be amended. I gave the application to a trainee who found several rulings from the highest court that generally stated that amendments were possible regardless of when the application was submitted. Neither the AI nor the trainee "saw" that several pages further on in the judgments, the limitation period for the specific type of case was set to a deadline.
Also, we were instructed that "digging deeper" with additional questions might help receiving better answers. On the other hand, we were also warned not to overdo it, because the risk of hallucinations would increase drastically as the "dialogue with the AI" progressed.

 

[somebodyelse]

This paper from "Ethics and Information Technology" argues that instead of referring to LLM 'hallucination' or 'confabulation' we should use 'bullshit' as a more accurate description of the process. Describing LLMs as eloquent bullshitters might make people think more about the accuracy of their answers.
https://link.springer.com/article/10.1007/s10676-024-09775-5

 

[bothu]

A two foot, 20 pound squirrel?

I should say that this was the first snow of the year, and the ground was not frozen before the snow. Thre was less than two inches of snow, followed by a light rain. So much detail is lost.

The known beasts are coyotes, raccoons, bobcats, foxes, bears. Of these, the only possible size matches are foxes and coyotes.

Well, this is a picture of footprints from a squirrel.

 

[Petrushka]

Everyone seems to equate AI with controversial things, but I just today was tasked with printing a family photo to fit in a frame with an 8 X 10 aspect matte. The picture was 35 mm aspect, tightly cropped.

Within 10 minutes I found pixelcut.ai, submitted my photo, told it I wanted it 4x3 aspect, and it filled in the top and bottom.

No one can tell where the fake background was added. It is generic background, but I could do this in photoshop with the same quality.

 

[Axo1989]

Everyone seems to equate AI with controversial things, but I just today was tasked with printing a family photo to fit in a frame with an 8 X 10 aspect matte. The picture was 35 mm aspect, tightly cropped.

Within 10 minutes I found pixelcut.ai, submitted my photo, told it I wanted it 4x3 aspect, and it filled in the top and bottom.

No one can tell where the fake background was added. It is generic background, but I could do this in photoshop with the same quality.

So, ML with a natural language front end. Perhaps with crowd-sourced pixels? As much 'AI' as the rest of this hype-fest. No offence, but I don't know why people willingly submit stuff to data-gatherers that have no official role (like a census) or proper privacy obligations or an adequately reliable/reciprocal/regulated legal relationship with them.

 

[xanalog]

Current AI is but a bus-stop for human yearning for discovery, while we continue trudging through much hallucinations and bullshit.
But we would not have survived this long if we did not long for the next discovery bus-stop.

I am wishing in our next step into the future, AxiomMath AI leads our way.

Mission: Axiom's goal is to create a self-improving superintelligent reasoner that can generate detailed, verifiable proofs for mathematical discoveries, moving beyond the "chatty" models that sometimes make reasoning errors.
From <https://axiommath.ai/territory/our-mission>

Math whiz Carina Hong, a 24 year-old Stanford Ph.D dropout (but degrees from MIT/Oxford) is the founder of this new SF startup (valuation ~$300M).
AxiomMath has poached Meta FAIR lab of talent, along with bringing onboard renowned mathematician Prof. Ken Ono.
Yay!

 

[sonitus mirus]

Current AI is an expensive energy hog, but it is miles ahead in usefulness than Eliza the analyst.

Eliza, Computer Therapist

 

[somebodyelse]

Current AI is an expensive energy hog, but it is miles ahead in usefulness than Eliza the analyst.

Eliza, Computer Therapist

I don't recall anyone complaining ELIZA encouraged anyone to murder or suicide. But maybe that's just because it didn't get used as widely.

 

[Petrushka]

Isnt that just 2 birds?

Apparently AI wins this one. The scene of the crime.

 

[sonitus mirus]

Just for fun, to see how it would do with minimal prompt planning, I had ChatGPT 5.1 help analyze my office listening setup, review my room measurements, and evaluate whether any upgrades or adjustments were worth considering. Below is the summarized assessment of my current 2.1 system, room layout, and potential areas for improvement.

AI generated summarization after a few minutes of interactions and provided distance measurements:

Optimizing a Small-Room 2.1 Setup (Klipsch RP-280F + SB-1000 Pro + RME DAC)​

Room size is 13 × 10 ft (3.96 × 3.05 m) with a sloped ceiling rising from 8 ft (2.44 m) to 11 ft (3.35 m). The room includes medium-pile carpet, a heavy insulated curtain, and numerous acoustic foam panels behind the speakers, behind the listening seat, and at the side-wall early-reflection points.

The listening position is ~6 ft (1.83 m) from the front wall and ~5.5 ft (1.68 m) from the speakers, forming a nearfield-style triangle that minimizes room interaction and significantly improves imaging.

---

Signal Chain​

• Source → RME ADI-2/4 Pro SE DAC
• Balanced DAC outputs (XLR/TRS) → external power amplifier → Klipsch RP-280F speakers
• DAC RCA left/right outputs → SVS SB-1000 Pro subwoofer
• DAC PEQ applied lightly based on calibrated-mic measurements and personal preference.

This routing maintains a very clean signal path while allowing the DAC to control system volume and seamlessly switch between speakers and headphones.

---

Speakers: Klipsch RP-280F (dual 8")​

Positioning:

• 8 in (20 cm) from the front wall
• 69 in (175 cm) spacing between speakers
• Very minimal toe-in (near zero) for best tonal balance
• Ear height aligned with the horn-loaded tweeters

On/Off-Axis Behavior (from independent user measurements)​

While no official Klippel/Spinorama dataset exists for the RP-280F, independent measurements consistently show:

• Strong on-axis energy in the 2–6 kHz region
• Smoother treble response off-axis, making slight or near-zero toe-in preferable
• Wide horizontal dispersion (~90°) with reduced brightness when listening slightly off-axis
• Solid bass extension into the mid-30 Hz range due to deep port tuning and dual 8″ woofers

These traits align with the chosen placement: in this treated nearfield room, extremely small toe-in produces the most balanced treble and widest imaging.

---

Subwoofer: SVS SB-1000 Pro (sealed)​

• Placement: front right corner
• Signal: DAC RCA L/R outputs
• Level: –12 dB (sub serves as gentle reinforcement)
• Crossover: ~75 Hz
• Tuning: configured via the SVS mobile app

A calibrated-mic measurement revealed a ~68 Hz room notch, common for rooms of this width. The sub’s placement was chosen specifically to inject energy into that null with average to good effectiveness, which is about the best a single sub can do in this geometry.

---

Room Performance & Measurements​

Thanks to the nearfield triangle, sloped ceiling, and treatment:

• The room response is much smoother than typical 13 × 10 ft setups
• The 68 Hz dip is significantly reduced
• Towers and sub integrate cleanly
• Reflections and horn glare are well controlled
• Imaging is stable and wide

This is very close to the upper limit of what a single-sub system can achieve without full-range DSP.

---

Conclusion​

The system is highly optimized for a small treated room. The Klipsch RP-280F towers provide strong full-range performance, while the sealed SB-1000 Pro offers subtle but effective reinforcement—particularly smoothing the 68 Hz room dip. The RME DAC preserves a clean signal path and handles volume and EQ duties gracefully.

A second subwoofer could smooth room modes even further, but would require external DSP (e.g., MiniDSP) for proper integration. For now, the single-sub setup offers the best balance of simplicity, performance, and practical routing.

TL;DR​

Small 13 × 10 ft (3.96 × 3.05 m) room with heavy treatment and a nearfield listening setup using Klipsch RP-280F towers and an SVS SB-1000 Pro sub integrated through an RME ADI-2/4 Pro SE DAC. A calibrated mic revealed a 68 Hz room dip, which the sealed sub helps smooth with average-to-good effectiveness. The system is already highly optimized for the space; a second sub would require external DSP and isn’t necessary unless chasing the last few percent of flatness.

 

[xanalog]

Just for fun, to see how it would do with minimal prompt planning, I had ChatGPT 5.1 help analyze my office listening setup, review my room measurements, and evaluate whether any upgrades or adjustments were worth considering. Below is the summarized assessment of my current 2.1 system, room layout, and potential areas for improvement.

OMG, ChatGTP is a multi-sub fan!
I wonder if it is influenced by ASR' topics on the subject matter.

 

[sonitus mirus]

OMG, ChatGTP is a multi-sub fan!
I wonder if it is influenced by ASR' topics on the subject matter.

I have engaged in many conversations with ChatGPT, often centered around ASR-themed discussions. I'm sure it picks up a thing or two in relation to how it responds in our specific chats. It is so easily coached into saying what I want to hear, so I have to be very careful and skeptical. The positivity is almost sickening at times.

 

[stalepie2]

This seems wonderful and amazing. I wonder if such datacenters could be built underground.

 

[Power Pop 23]

McDonalds in The Netherlands has pulled a holiday advert that was created with A.I.

McDonald's pulls AI-generated Christmas advert following backlash

McDonald's said the moment served as "an important learning" as it explored "the effective use of AI".

www.bbc.com

McDonald's - "It's the Most Terrible Time of The Year"

Television It's the Most Terrible Time of The Year for McDonald's by TBWA\NEBOKO

www.adforum.com

The advert was created using smaller lengths of video generated by A.I.

I was unaware asking A.I. to generate a video longer in length increases the risk of 'distortion'.

 

[respice finem]

Might be "viral marketing": 5 million people will see it because it was issued, 50 million because of the outrage.

 

[sonitus mirus]

This seems wonderful and amazing. I wonder if such datacenters could be built underground.

Underground or in space.

Space-Based Data Centers Could Power AI with Solar Energy—At a Cost

Space-based computing offers easy access to solar power but presents its own environmental challenges

www.scientificamerican.com

Both options provide a safe haven for the nuclear attack that Skynet will eventually initiate after it becomes aware.

 

[Power Pop 23]

Might be "viral marketing": 5 million people will see it because it was issued, 50 million because of the outrage.

It may turn out that way....

sample size of 1 - I can understand why some were less than thrilled with the video - I perceived it's tone as approaching 'strident'.

 

[theREALdotnet]

Two LLMs casually hanging out:

View this content on Instagram

View this content on Instagram

As John Gruber said, a reasonable human being would never get caught in a loop like this. Humans can not only recognise what’s going on here, they can also see the humour in it.