AI in 2025: Rush for use cases

Turn to CNBC TV for the latest stock market news and analysis. From market futures to live price updates CNBC is the leader in business news worldwide.

Connect with CNBC News Online

#CNBC
#CNBCTV

[JoshuaLawrence]

The fact that this question can be raised after the o3 benchmarks is crazy to me.

[CorbinTelligman]

I've trained a GPT on all of my company's documented use cases which is helping me with sales-acumen and copywriting. Definitely seeing it going the direction of smaller language models for real world use cases.

[farleyholicsannonymous]

How is this peaking? Have you tried the latest models released in 2024? They're very good. Even if we see a 2 fold increase in reasoning in 2025, it's still going to be smarter and more capable than most people on Earth.

[iainmackenzieUK]

Please tell us the impact on education this year. (I am a teacher of Physics and would love to hear your thgouhts on the impact on our students.)
Thank you - lovin your reporting xx

[cheese-power]

1:29 “western model”😂 that killed me😂😂😂

[mickeycavalier]

Dierdre Bosa! 🎉🎉🎉great report! Thanks!

[pacanosiu]

when one android is more effective than 100 of the best trained agents, soldiers and policemen combined
then you all understand

[MrSchweppes]

She is wrong. Three components of Reasoning for AI:
1. Foundation (Pre-training)
2. Self-improvement (RL)
3. Test-time compute (planning).

So scaling of frontier models is not going away anytime soon.

[VPopkins]

Great report. The only thing i disagree about Small LMs. They proved to be worse than LLM even in highly specilized tasks. The trends is that LLMs become faster and can run locally making SLM usless. While RAG and ever increasing content window are solving context problem. Even LLm fine tuning is questuionable. Imho.

[BruceWayne]

The big things to look for in 2025 are:
1. Androids - recent advancements in training have made this a very achievable goal in 2025 for less than the cost of a car.
2. Infinite context - this is going to be critical for allowing AI to do things like displace human workers since it will need to be able to remember what its job parameters are.
3. Agents - we already have them, but they are too infantile to be useful. In 2025, AI companies are focusing on this feature as it will allow AI to start doing the job of a human. Ultimately, this is where the real AI investment pays off, when AI can either replace or significantly augment human workers.
We definitely have not peaked with AI, they simply hit a wall when it came to scaling. Already though China has found a way around the scaling problem, and I'm sure we'll deal with it too. They were just going after the low hanging fruit by scaling constantly. Now they will have to work smarter, not harder.

[User-jca]

Elsewhere is in the upswing? Cuz elsewhere they are just starting.

[Vix]

Sorry D, it’s not about smaller models, it’s about agents

[walidoutaleb]

lol are we still questioning whether ai progress has stalled when progress keeps accelerating. o3 just saturated arc-agi a benchmark that proved ai-proof for 5+ years. the next benchmark to beat is frontier math which o3 scored 25%. the arc-prize fondation announced that they are making a new benchmark in which humans will score >95% and o3 <30%. i assume that by the end of 2025 both of these benchmark will be saturated. But even beyond reasoning not one big lab has focused on agents which i assume is the next goal after reasoning so in 2025 along with reasoning development we should see new very hard agentic benchmark which labs will compete to saturate.

[matthew.stevick]

thanks dierdre bosa! happy nye and best wishes to u and urs in 2025 📈🇺🇸
💚🖤

[moocorp]

FYI, all cloud providers have agents, not only Microsoft. I'm using AWS Bedrock.

[stevec]

Deepseek is real open ai. OpenAL is closed AI

[MrGriff-du]

Who will be the Prodigy or AOL of AI, because I want to invest in them😅

[andrewwalker]

Basically no insight here, just dropping names to sound like they know what’s going on

[Matt]

When are people going to get a clue and realize you can’t just make an AI that is more advanced. The way it really works is that the current AI continues to advance over time. That’s why ChatGPT will continue to be on top because it was the first.

[Cory-vw]

My Ai-Agent will have a PhD in inverse design. In material science these structures are based on desired properties.
So apparently, some humans haven't been using their PhD at an optimal standard. How much desire will that make for me.