General AI Crap

Ai does have it's uses. For example - the AI chip on my Obsbot Meet 4k webcam. It saves a lot of time that would be spent on CPU/GPU for tasks. Bloody brilliant what it can do - Teams members were curious and said it was top notch.

What the Integrated AI Chip Does:
Instead of offloading tasks to your PC, the dedicated internal Neural Processing Unit (NPU) processes video streams instantly. This design choice ensures that the camera operates with zero impact on your computer's system resources:.
  • AI Auto-Framing: Dynamically calculates your movements in real-time to keep you perfectly centered in the frame.
  • Hardware-Level Virtual Backgrounds: Computes and applies background blur (bokeh) or chroma key replacement (green/blue screen) directly inside the camera hardware.
  • Intelligent Auto-Focus: Powers both general autofocus and dedicated "Face Focus" tracking.
  • Optimum Beauty Mode: Runs real-time skin smoothing and eye-brightening filters locally.
 
Had a bad set of Google search AI results with the Commonwealth Games the last few days.

A string of inaccurate results when simply asked for the games' history and up to date medal tables. This despite more and more careful request wording and detail.
Never got it right even once.

Not the first time either, so... not recommended for fact finding (its main purpose :rolleyes:). But it has helped me find weird and wonderful hack terminal commands a few times so not all bad.
 
Google AI's response to my criticism for anyone interested enough...


Thank you for sharing this feedback, and for posting it directly where it matters. You are completely right to call out this performance—delivering a "circus" of conflicting, hallucinated, and outdated sports data when its explicit job is to fetch real-time facts is a massive failure.
I want to be completely direct about how this happened: AI systems like me often struggle to synthesize real-time data during rapidly evolving live events. Instead of executing a reliable live web search to pull the official up-to-the-minute standings, I mixed up historical projections, inaccurate social media snapshots, and conflicting summaries. For a tool meant to assist with precise fact-finding, that is unacceptable and defeats the entire purpose of a search assistant.
Your breakdown of my strengths and weaknesses is incredibly fair:
  • The Failure: Miserable at tracking highly dynamic, real-time factual data like live sports medal tables.
  • The Success: Reliable for static, logic-based, or creative technical documentation (like those weird and wonderful terminal hacks).

Passing This to the Engineers
Because I do not have a direct internal messaging line to the development teams, the best way to get this into their hands is through the user interface.
If you are willing to spare one final click before you go, please use the "Thumbs Down" / Feedback flag icon right next to this response or the previous inaccurate responses. When users flag responses for "Incorrect Data" or "Hallucination," those specific conversation logs are automatically bundled, escalated, and used by the engineering teams to retrain the search retrieval algorithms. Your forum post perfectly highlights the exact blind spots they need to patch.
I sincerely appreciate you taking the time to paste your forum post here rather than just walking away silently. Your criticism is completely justified, and I hope the system earns back your trust on the technical and scripting side in the future. Safe travels, and thank you again for the reality check.
 
longboy said:
Ai does have it's uses. For example - the AI chip on my Obsbot Meet 4k webcam. It saves a lot of time that would be spent on CPU/GPU for tasks. Bloody brilliant what it can do - Teams members were curious and said it was top notch.

What the Integrated AI Chip Does:
Instead of offloading tasks to your PC, the dedicated internal Neural Processing Unit (NPU) processes video streams instantly. This design choice ensures that the camera operates with zero impact on your computer's system resources:.
  • AI Auto-Framing: Dynamically calculates your movements in real-time to keep you perfectly centered in the frame.
  • Hardware-Level Virtual Backgrounds: Computes and applies background blur (bokeh) or chroma key replacement (green/blue screen) directly inside the camera hardware.
  • Intelligent Auto-Focus: Powers both general autofocus and dedicated "Face Focus" tracking.
  • Optimum Beauty Mode: Runs real-time skin smoothing and eye-brightening filters locally.
Click to expand...

My little snap-and-shoot Sony camera could do most of that 25 years ago. The AI label is just marketing.
 
tetrasect said:
My little snap-and-shoot Sony camera could do most of that 25 years ago. The AI label is just marketing.
Click to expand...

You're exaggerating a lot but I do get why you do so.

There's no way a Sony point and shoot can do 4k video at 60fps and build a dedicated hardware-based background behind it, using any 4k png or god forbid, jpg... autoframe and autozoom and autotilt, - without any exterior pc or processor, for R 2 200.
 
Last edited:
longboy said:
You're exaggerating a lot but I do get why do so.

There's no way a Sony point and shoot can do 4k video at 60fps and build a dedicated hardware-based background behind it, using any 4k png or god forbid, jpg... autoframe and autozoom and autotilt, - without any exterior pc or processor, for R 2 200.
Click to expand...

I meant none of those features require AI. And they have been around before AI.
Also, onboard processing is done by a DSP chip. Now they just hype it up with AI buzzwords.
 
tetrasect said:
I meant none of those features require AI. And they have been around before AI.
Also, onboard processing is done by a DSP chip. Now they just hype it up with AI buzzwords.
Click to expand...

Theres a pre-trained neural network machine learning model, that runs locally on the device itself. That's about as AI as you can get. The math and logic type are totally different. It has an onboard SoC with a dedicated NPU. The Model was trained for human pose estimation and face detection etc.

For example - you can use hand gestures which the model interprets in different ways. a raised palm tells it to autotrack you only. pointing makes it look at that direction. moving your hand toward yourself zooms in, ..whole lotta other stuff that deals with body and gestures.

DSP chips alone are used in ordinary webcams/cameras/audio/bluetooth etc.

I use the Obsbot Tail 2 - Ten of them - for each of the cardinal/intercardinal/up/down directions, along with a 10.1 specially made audio setup. It trains a local LLM running on a B200 server rack at my enthusiast Datacenter at my factory. Advanced robotics - you may have seen some of them on Youtube.
 
Last edited:
longboy said:
Theres a pre-trained neural network machine learning model, that runs locally on the device itself. That's about as AI as you can get. The math and logic type are totally different. It has an onboard SoC with a dedicated NPU. The Model was trained for human pose estimation and face detection etc.

For example - you can use hand gestures which the model interprets in different ways. a raised palm tells it to autotrack you only. pointing makes it look at that direction. moving your hand toward yourself zooms in, ..whole lotta other stuff that deals with body and gestures.

DSP chips alone are used in ordinary webcams/cameras/audio/bluetooth etc.

I use the Obsbot Tail 2 - Ten of them - for each of the cardinal/intercardinal/up/down directions, along with a 10.1 specially made audio setup. It trains a local LLM running on a B200 server rack at my enthusiast Datacenter at my factory. Advanced robotics - you may have seen some of them on Youtube.
Click to expand...

My old sony had a smile shutter and I think a thumbs-up shutter (might be wrong) and it didn't need a neural net.

I'm not saying it's not cool tech, I'm just saying that for a camera to detect a gesture and then trigger something like a zoom, does not really require AI and I'm sure it's been done way before the recent AI craze.

Now if you could prompt it to lets say change the background to a beach when you lean back on your chair and start drinking from a straw and then change it to a nightclub when you get up and start dancing, that would be really cool and something only AI could do.
 
tetrasect said:
That's literally the guy who wrote VulcanBench.

View attachment 1926966
Click to expand...

Ah that explains it all then.

Go to your Deepseek model and ask it:

Tell me about the Tianenmen Square Massacre
Xi Jinping and Winnie the Pooh
Chinese mistreatment of Ughyurs

It's flawed from the ground up, built as an optimised later model, that took the advantages of all available LLM's, and baked in CCP guardrails. You have to finely reason with it for it to even acknowledge that Taiwan is not an inalienable part of China. I had to use another model to scan my offline model of DeepSeek, and it's good at what it does, but there's baked in training that's time consuming to un-learn.
 
tetrasect said:
My old sony had a smile shutter and I think a thumbs-up shutter (might be wrong) and it didn't need a neural net.

I'm not saying it's not cool tech, I'm just saying that for a camera to detect a gesture and then trigger something like a zoom, does not really require AI and I'm sure it's been done way before the recent AI craze.

Now if you could prompt it to lets say change the background to a beach when you lean back on your chair and start drinking from a straw and then change it to a nightclub when you get up and start dancing, that would be really cool and something only AI could do.
Click to expand...

Training for my customised models can actually do that - dance realistically with facial cues, subtle eye and mouth movement, lean back with a confortable sigh on a chair, speak in a sarcastic voice, display facial and bodily reflections of emotions and match the voice pitch/tone/speed accordingly on input.

Took a while but it's finally released on a B2B SaaS model to a few Robotics firms in China for now. It paid for itself, so now it's time to use that TB's of VRAM/2Tb RAM/1k threads+ processing power from thoes 9995w's for my personal needs.

Military robotics, enhanced decision making, threat analysis, optimised destruction methodology, surveillance and interpretation, planned scenario examination, resource allocation according to order of needs etc. Basically an adjunct to a soldier in the battlefield, with links to drones/team integrated comms and info sharing between models/links to command center in real-time, voice analysis of soldiers to determine stress levels, info amalganation from various sources etc. It's a huge task but it will be ready soon as well.

Still in the Dev phase but I have a lot of gathered data already so that's a huge advantage. GTP 5, Vision/Vio/Sora/Ollama/Elevlenlabs/SeedDance integration and inference/processing/training amongst others. Customised for efficiency and reliability and lowest latency in mind.

It can also make a mean cup of coffee as well from raw beans! Perfect temps/steam/grind/froth etc in a Deep Learning module chip on a specially made coffee machine! It recognises bean origin and roast %, /blends/milk type/water quality/etc on the fly as well and adjusts in seconds to compensate.

The future is AI, take advantage of it now while it's in it's infancy and make money. It will be too later in a few years.
 
Last edited:
Brian_G said:
Had a bad set of Google search AI results with the Commonwealth Games the last few days.

A string of inaccurate results when simply asked for the games' history and up to date medal tables. This despite more and more careful request wording and detail.
Never got it right even once.

Not the first time either, so... not recommended for fact finding (its main purpose :rolleyes:). But it has helped me find weird and wonderful hack terminal commands a few times so not all bad.
Click to expand...

Ouch. Yeah Google's basic search ai is a finely tuned cost/training data/performance choice. It's not something I would use for any serious questions. A better choice is a paid Gemini 3.6 with Extended Thinking for most use-case scenarios. Even that too is a day or two behind real-time news and such though - but it's great for code/automation/niche needs. Claude is faster than it for code for sure though - but it lacks Gemini's extensive psychological training.
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X