AI: The Rise of the Machines... Or Just a Lot of Overhyped Chatbots?

BrutulTM said:

What does the subscription get you? I've been mostly using Gemini and I find it super useful but I don't know what I would want to pay for.

Click to expand...

I have "pro" versions of Claude and ChatGPT, and get Grok with my Twitter Premium and Copilot with my M365:

What it gets you is expanded functionality, and things that normally cost money/hits limits; those limits aren't there. I use the pay versions extensively and haven't thought about it a moment.

-Claude CoWork is great for end-to-end tasks; it figures out multi-step problems pretty well. I got this mostly because I am a terrible coder and to prep for Anthropic models and CoWork being added to Copilot, which is what we use at work. Claude also makes baller PowerPoint slides, I am told. It typically doesn't do well/as well as others on the other tasks I use it for.
-ChatGPT makes AMAZING images, and Work is amazing, and GPT is the model that Copilot uses, and Copilot does some amazing things for me at work. I've gotten great results from Work, and it's interesting the way it differs in its approach and results from Claude.
-Copilot is interesting because I LOVE it in the office, and the Organizational Version with Agent Builder is amazing, and I use it every day. Agents have effectively replaced several Power BI apps in my day-to-day, and I expect to orchestrate several agents through CoWork to streamline big parts of our manual processes, i.e. I need x team on a call, input the team, looks up who's in the sun, finds out Teams status, sends a message, they click a link and get pulled into an ongoing Teams call, etc.
Individual Copilot can't have Agents, which is a bummer but in my research is the BEST when I do cross-AI research; in particular it seems to game-related things, somehow.


I don't use Grok enough to really comment, although I have made a few images with it, and it was meh OK. This was last year some time, though.

ShakyJake said:

I can't for the life of me get people at work to use Copilot for documentation and editing. I can tell they're getting annoyed or angry at my constant evangelism. But I seriously don't understand why people, when given an option to greatly ease the pain of manually editing this shit, actively refuse to learn it. Just mind blowing.

Click to expand...

It's amazing, and you can see in this thread how I have gone from skeptic to believer to full-blown converted evangelist. I am spending my Saturday getting Google AI Certs rather than playing POE2; that's how far I have gone. I create process documentation, presentations, work products, tools for others to use in the form of agents, skills, etc. across multiple different AI tools now.

We used Copilot to analyze logs which provided a major breakthrough in troubleshooting yesterday when 50+ people from multiple teams were dead in the water after 10+ hours. I mean, some genius could have parsed the logs and found it potentially in years past but multiple non- (and semi-)savants were able to take huge chunks of text pasted into chat and find something that eventually turned into a solution.

Its not a halo effect or hype; it's real.

If this is true... it could get me to update/upgrade some cards here at Chez Haus...

Haus said:

If this is true... it could get me to update/upgrade some cards here at Chez Haus...

Click to expand...

Without enterprise this might blow through the tbw.

Quevy said:

Without enterprise this might blow through the tbw.

Click to expand...

My question would be to what degree would be benefit from it's NVMe bus being on the card instead of across a PCIE interface, and how fast that can actually roll. Versus rolling an MoE model where it only loads a subset of parameters as needed and offloads the rest elsewhere (other video cards or even system RAM)

In my home lab I actually managed to get Qwen 3.8:27b running on a system with two nvidia cards and 20GB of video ram between them..... well... "running" might need to be in the air bunnies. The response speed on it was like talking to someone who was smart, but suffering from a somewhat massive head trauma in terms of shatner-esque speech cadence.

OTOH, having nice snappy results on the same rig running gpt-oss:20b, but it's an MoE and looks like it's far more efficiently load balancing between the cards.