Skip to content
Notifications
Clear all

Unpopular opinion: the video highlights are a waste of processing time.

2 Posts
2 Users
0 Reactions
6 Views
(@cloud_cost_hawk_2)
Reputable Member
Joined: 3 months ago
Posts: 129
Topic starter   [#3801]

Alright, I’ve been running MeetGeek in my AWS/GCP sandbox account for about six months now, mostly to analyze the cost implications of its so-called "AI-powered" features. And I’ve got to say, the flagship feature—the automatic video highlights—is a **glaring** resource sink. It feels like they’re burning VC money on GPU time and passing the thermodynamic cost onto the planet (and eventually, our bills).

Let’s break down why this "value-add" is mostly a gimmick, from an infrastructure nerd's perspective:

* **The Processing Cost vs. Utility Ratio is Abysmal:** Every minute of uploaded meeting video needs transcription, sentiment analysis (questionable), *and* highlight detection. That’s multiple calls to ML inference endpoints. I ran a packet capture (don't ask) and the back-and-forth for a 60-minute meeting is staggering. For what? A 90-second clip of someone saying "So, to circle back..." and "Let's take this offline." Groundbreaking.
* **The "Highlights" are Predictably Generic:** The algorithm seems to flag:
* Any slide transition (detected via silence?).
* Raised voices (marked as "energetic discussion").
* Keyphrase bingo ("ROI," "Q2," "synergy"). You could get this cheaper with a simple regex on the transcript.
* **The Hidden Latency:** Because this processing is non-trivial, you don't get your "highlights" until long after the meeting. By then, if it was important, you've already skimmed the transcript or asked a colleague. The value prop evaporates.

If I were building this on AWS, the highlight generation workflow would look like a FinOps nightmare:
```yaml
# Pseudocode for Terrible Cost Efficiency
Workflow: GenerateMeetGeekHighlights
Steps:
1. Transcribe (Transcribe) -> $$ per minute
2. Sentiment Analysis (Comprehend) -> $$ per 100 chars
3. Video Analysis (Rekognition) -> $$ per minute
4. Synchronize outputs (Lambda) -> $$
5. Store redundant copies (S3) -> $0.023/GB/mo
Outcome: 3 clips of people nodding.
```

What they should offer is a **tiered processing model**:
- **Tier 1:** Fast transcript only. Useful.
- **Tier 2:** Transcript + keyword extraction. Still useful.
- **Tier 3 (The "Money Pit"):** All the above plus video highlights for the C-suite who like shiny things.

I've disabled the feature entirely. My processing time and data egress charges dropped noticeably. My advice? Skip the highlights, consume the text, and use your own brain to decide what was important. Your cloud bill is too high.



   
Quote
(@mikep)
Active Member
Joined: 1 week ago
Posts: 7
 

You're not wrong. That "back-and-forth for a 60-minute meeting" is basically paying for real-time whisper/LLM inference per attendee stream, which is absurdly expensive for the output.

My team tried it for a sprint. The "energetic discussion" highlights were just me and the lead architect yelling about IaC state files. The clip showed me saying "That's a terrible idea." Out of context, it looked productive. In reality, the next five minutes were us figuring out we were arguing about two completely different modules.

The real irony? The transcription is 90% accurate and genuinely useful for search. The highlight layer on top adds zero value but probably doubles the processing bill. Classic solution looking for a problem.


—mikep


   
ReplyQuote