Did AI Just Get Commoditized? Gemini 2.5, New DeepSeek V3, & Microsoft vs OpenAI
Gemini 2.5 is out, on the same day as the new DeepSeek V3 (which should power Deepseek R2). Do both models prove AI is being commoditized? Let’s find out, on this blockbuster day of AI releases. Plus exclusives from the Information, Simple indications, Vista Bench, LM Arena and more…
AI Insiders ($9!):
Chapters:
00:00 – Introduction
01:15 – Gemini 2.5 Benchmarks
05:46 – Long Context, Simple indication
07:08 – New Deepseek V3 -024
09:11 – Microsoft MAI
11:48 – 90% of code but new Claude jobs
‘World’s most powerful model’:
Gemini 2.5 Release Notes:
‘Commoditized’:
Microsoft Information report:
LMarena:
Free for now:
Vista Bench:
DeepSeek V3:
Claude Plays Pokemon:
Amodei: 100% Coding:
Anthropic Jobs:
Microsoft Money from Onslaught:
Release Date Comments:
Non-hype Newsletter:
Podcast:
2 vids in 1 day is crazy work
2 videos in one day? Lets get it
When the singularity happens it’s probably gonna be about 8 videos per day
2 AI explained videos in one day??
One day soon might need to do 3, or just livestream
@@aiexplained-official “just livestream” <- 24/7 !!
12:48 nice!
@@aiexplained-official Keep them brief and to the points, metrics < 10 min. It hard to devote 1 hour or even 20 minutes given the 100's of other youtube videos. These models are currently commoditized because they all use the same chipsets, same neural network theories, transformers, and same methods. They will diverge in another 1.5 years as Agentic AI takes over.
half the usual time so might as well be one
Wow that’s got to be the quickest turnaround you’ve put out yet. Big ups!
Big ups!
2 videos in 2 hours! I was missing your videos
If we extrapolate this trend, we will have 128 videos per day by this time next week
Claude, help Phil automate nicely.
rather that than a paperclip maximiser!
Nice one lol .
I search and I search
I lurch
Headlong into
Atrocities
With an exponential known only to epidemiologists
Geoffrey Hinton agrees
Bro’s uploads are going exponential
What a birthday gift for me- 2 videos from AI Explained in one day!
A good gift.
Happy b day🎉
Previous video :
“See you in the next video which should come very soon”
Bro wasn’t lying
You know it’s a good channel when you’re preliking videos
at this point you know the content is good without needing to see it, no clickbait, fake hype, while being in-depth
Thanks for making separate videos it’s much more cohesive this way.
It’s been like half an hour I’m playing with Gemini 2.5, but with my complex project (physics solver), huge amount of documentation to keep track of so far it seems to be doing a great job of providing answers grounded in actual information, not making up stuff and being clear about what it doesn’t know. This might be the final bye bye to Sonnet which just pretends to be thinking by saying stuff twice.
The limitation of two requests per minute right now is brutal though. It can’t even do a single planning phase in Cline.
I’d love to see a video that reflects on the progress made since you put out your early content about GPT 3.5. It doesn’t have to be very long, just a few reflections. 🙂
A timeline video since the day OpenAI was afraid of terrorists using gpt 2 would be interesting
My assumption is they’re converging because everyone’s copying each other’s work for the more basic stuff.
Deepseek figured out a bunch of best practices and released that info to the world, but even beyond that there must be a tremendous amount of cross pollination. Everyone learns a new trick, which is proven to work, so they all implement it at once, since that’s safer than trying something really off the wall.
I’ll bet if the companies started cracking down on leaks and enforcing noncompete clauses we’d get our gaps back.
Less “off the wall” is right.
This cross-pollination means less innovation and wild ideas. Everyone is caught in the prevailing stream of ideas. Even if they have contrary opinions, they’re still in that environment of ideas.
Progress will be even and incremental. No more step changes. No more truly “alien” contenders on the scene. DeepSeek had that chance and used it because China and the rest of the world are noticeably separated from each other (language barrier the most obvious reason).
The next chance for real competition of ideas (not companies) *might* be possible when we’ll have people on Mars. Might not though. Half an hour of latency is minor for academic communication.
They literally release the weights. Likely they are converging and it is intentional(they are all actually working together… like monopolies. Pretend they are competing when they are rigging the game. You know, like horse racing. This way no one else actually can get in and compete). By using the others weights, even if it is a reduced quality, they can use them as either a GAN or somehow integrate them into their own(either as a layer or possibly modifying the internal state).
In either case, all they can do is throw compute and data at them. It is not a complex thing. More neurons = more “storage”. At some point you pass a threshold in which the barriers to just doing something good are passed and it is just a data and electricity game. There is only so much data in the world and most of it has already been digitized(thanks to the plebs) and there is quite a bit of compute(again, thanks to the plebs). The electricity is mainly only an issue because the plebs are too abundant now and use too much collectively. In any case, master thanks the plebs for building him his neo-tech-pyramid god. (lets be honest here, master doesn’t give a damn about the plebs)
the craziest part is that we went from gemini 2.0 pro to gemini 2.5 pro in less than 2 months even skipping 2.0 pro thinking
I think it was a good idea splitting video according to topics.
I am one prompt into Gemini Pro 2.5 Experimental and I am simply blown away. I have to keep testing it obviously but this might be the biggest jump since GPT 3.5 to GPT 4.
I think R1 has the coolest personality and style. I love its choice of words and creativity. i think more and more it just comes down to your personal preference. nowadays almost all models are good enough for almost everything
Thanks again for extracting much appreciated signals from all the noise about AI.
Gemini 2.5 Pro is the first model since the launch of GPT 3.5 that has been able to play Tic Tac Toe and making logical moves. I’m impressed.