AMA: AI in 2025 and Beyond

In this AMA, Sahara AI’s CEO and Co-founder Sean Ren and Anthropic’s Rohan Taori reflect on the biggest AI breakthroughs of the year and share their thoughts on where the industry is headed in 2025. From emerging trends to the future of AI agents, they break down what’s next and what it means for the AI landscape.

Link to full recording: https://x.com/i/spaces/1yNGagBaPZVxj

Joules Barragan (04:11)
Hi, everybody! Welcome to our very first AMA, "AI in 2025 and Beyond." I'm Joules from Marketing at Sahara AI, and I’ll be your host today. We are thrilled to have two amazing guests:

1. Sean Ren – Our CEO and Co-Founder at Sahara AI. He’s also an Associate Professor in Computer Science at the University of Southern California, where he leads the Intelligence and Knowledge Discovery Research Lab.

2. Rohan Taori – A member of the technical staff at Anthropic. If you haven’t heard of Anthropic, they’re a leading AI safety and research organization that built the very famous Claude AI. He also has a PhD from the Stanford AI Lab.

Welcome, Sean and Rohan!

Sean Ren (05:10)
Glad to be here.

Rohan Taori (05:12)
Yes, thanks for hosting, Joules. Happy to be here.

Joules Barragan (05:15)
Thank you both for joining! We have an exciting agenda: we’ll talk about this year’s AI breakthroughs, what makes for “great AI,” and predictions for the future. Before we dive in, a quick reminder: we’ll be giving away exclusive swag, an invitation to our next event (with a plus one), and three whitelist spots for our Data Services Platform, Season 1. To win, just ask thoughtful questions during the AMA in the chat. We’ll pick winners at the end and post them on X (formerly Twitter).

Let’s kick things off with a quick question for Sean and Rohan: What’s one AI development from this year that really stood out to you?

Sean Ren (06:13)
I can go first. I’m really excited about something called GenIE (short for Generative Interactive Environments), published by Google DeepMind. Historically, simulation environments for AI agents were very handcrafted—lots of set rules and constraints. But with generative interactive environments, you can go into infinitely many possibilities. You can imagine different types of games or scenarios that the AI agent can explore, even beyond what’s physically possible in the real world. This opens up vast opportunities for training and evaluating AI. I expect more parties to do similar work, which will push the frontier of how AI is tested and developed. What do you think, Rohan?

Rohan Taori (07:49)
Yes, that’s very interesting and closely tied to what I was going to mention: the shift toward test-time compute and reinforcement learning (RL). Historically, people focused on scaling up training—pre-training on more data and bigger models. But scaling test-time compute for RL expands the solution space. You can have more nuanced data distributions, more varied environments, and many ways to tune your model to do what you want. This new focus on test-time compute, combined with what you mentioned about interactive environments, really widens AI’s creative space. We’ll see a lot more of this in 2025.

Joules Barragan (09:11)
Cool! 2024 has been a year of monumental breakthroughs. Looking back, what do you both think have been some of the most significant AI advancements over the past year?

Sean Ren (09:32)
I agree with Rohan’s emphasis on inference-time (or test-time) scalability. Everyone knows that scaling pre-training—throwing more compute and data at these models—has been successful, but it’s showing signs of saturation. People at OpenAI and other labs are noticing that. There’s still more to do with post-training (like fine-tuning and RLHF), but the new frontier is pushing more intelligence and compute into the inference phase.

Also, I’ve been impressed by new, more efficient sequence-model architectures like Mamba, state-space models, and others that reduce computational requirements for training and inference. That’s huge for the engineering and infrastructure side of AI.

Rohan Taori (11:07)
Exactly. And along with more efficient architectures, we’ve also seen massive price reductions for running models—potentially 10x, 100x, or more for inference tokens. That means more people can deploy and experiment with these models. Also, with something like the Llama 3.1 release, there are increasingly capable open-source models, which is great for the open-source community, hobbyists, and tinkerers.

Sean Ren (12:51)
Yes, I sometimes forget just how good open-source models have gotten. Just one or two years ago, open models were quite limited. But now, models like the Llama 3 family are pretty useful off the shelf. You can do personal info-seeking, customer service, or retrieval-augmented generation on top of them. You can even serve some of the smaller models on-device, thanks to techniques like model distillation. That’s a big deal for broader deployment.

Rohan Taori (13:35)
It’ll be fascinating to see how on-device deployments evolve in 2025. People are working on better quantization schemes and ways to reduce memory requirements. The open-source community is really pushing that boundary.

Joules Barragan (13:50)
Absolutely. One thing we haven’t touched on is decentralized AI. What role has decentralized AI played in 2024?

Rohan Taori (13:48)
That’s an interesting question. In 2023, we saw a big push for open-source community involvement—building fine-tuning datasets for Llama 1 and 2, for example. With Llama 3.1 and its strong out-of-the-box fine-tuning and RLHF, there’s been less emphasis on community-driven instruction datasets, since the official versions are already quite good. But decentralized AI is broader than just dataset creation. It’s also about how we can collectively build and share models, data, and the environments for RL. There’s definitely still a big opportunity there, especially with test-time scaling, but it will take more innovation and perhaps new incentives for the community.

Sean Ren (16:08)
Yes, I totally agree. Inference-time scaling makes decentralized AI more appealing. For instance, you can take something like Llama 3 as a base (like a commodity layer), then build a private or proprietary pipeline on top—such as specialized system prompts, RAG databases, or external memory with certain inductive biases. That pipeline becomes the monetizable component. Small- or medium-sized businesses can do this; developers can spin up interesting use cases. Then the big question is how the creators of the base models sustain themselves. But as long as companies like Meta keep updating Llama, it benefits the entire ecosystem.

Joules Barragan (17:59)
Great insights. Sean, you touched on data briefly. Sahara AI recently launched a Data Services Platform, and at Sahara, we often say: “Good AI requires good data.” Why did we start with a Data Services Platform, and what does it mean for AI?

Sean Ren (18:17)
Everyone knows data is the “new oil” for AI, but it has to be high-quality data. You can crawl the internet, but you’ll get noisy data. If you’re building an AI application in a specific domain, you need well-curated, balanced data—no major biases, no irrelevant noise. That’s why data services (collection, cleaning, labeling) are still a huge industry.

The next level is figuring out how to incentivize people to share valuable data without worrying they’ll lose ownership or revenue potential. You want to enable continuous revenue sharing for data providers if their data is used to train or refine a model that eventually makes money. That’s where data provenance becomes critical—tracking the lineage and usage so that if a data contributor’s work ends up in a lucrative application, they share in the revenue.

We started with the Data Services Platform because it’s the foundation for all of that. We want a marketplace where unique data sets meet model developers. Later, we can allow for more complex interactions, like revenue sharing from the resulting models or apps.

Rohan Taori (21:20)
Exactly. And looking to 2025, scaling up test-time compute and RL is going to require not just textual data, but interactive environments. That’s another kind of “data,” and a new aspect of model tuning. Imagine crowdsourced environments or tasks that these agents can explore. The same incentive mechanisms apply: how do we reward contributors of these environments or specialized data sets?

Joules Barragan (23:49)
Great. Rohan, you mentioned we’ve gone from “AI assistants” to “AI agents.” As we head into 2025, what other big trends do you see on the horizon?

Rohan Taori (30:17)
We’ve hit on many of them: more agentic systems, more test-time compute, and big cost reductions. These trends combined will make AI cheaper, faster, smarter, and more versatile. I’m especially excited about multimodality. We’ve already seen glimpses with models that accept text, images, audio, and even video streams (like Google’s new demos with Gemini). In 2025, we’ll see an explosion of multimodal models—both proprietary and open-source—where you can talk to AI using any modality and get responses in multiple forms. This will unlock huge new applications. It’ll also make the data challenge bigger, because now we’re dealing with text, images, audio, video, 3D data, and more.

Sean Ren (32:05)
Yes, I’m also excited by “contextualized” models that integrate signals beyond text—speech, images, environment data, etc.—to create more natural interactions. For example, you can record a quick video and ask the AI to observe or reason about what it sees. Real-world use cases become far more compelling and grounded. On the business side, these richer models boost productivity (automating repetitive tasks, integrating with enterprise systems) and entertainment (immersive, game-like experiences). I’m curious whether we’ll see a brand-new “killer app” that emerges from multimodal AI. Regardless, it will definitely improve what already exists.

Joules Barragan (35:16)
We’ve reached a point where AI feels ubiquitous. My generation grew up as internet natives; the next generation might grow up AI-native. How far off is that?

Rohan Taori (35:43)
Sooner than we think! Even right now, it’s strange that your computer doesn’t “understand” you unless you click buttons or type commands in a specific way. In a few years, kids will be baffled by how we used computers in 2023 or 2024. They’ll be like, “What do you mean you had to press all these icons for Photoshop?” They’ll expect to just say, “Hey, computer, do this,” and it’ll do it. This shift will be as big as going from paper maps to Google Maps.

Sean Ren (36:59)
Yes, it’s going to redefine “engineering” or “productivity.” Right now, you need programming skills. But soon, an agent could write entire codebases from a single natural-language prompt. We’ll need new thinking about education, too. Will we still teach everyone Python, or do we focus on training people to direct and oversee AI systems? Ethics, philosophy, and governance will remain critical, because we need to ensure these AI systems align with human values.

Rohan Taori (39:17)
That’s an interesting point. Will programming fade out as a specialized skill, or will it become as common as learning English in school? We might end up with a world where everyone has some coding fluency because it’s an essential way to supervise AI. Or we might see an ecosystem of “meta-agents” that build, verify, and maintain software with only high-level guidance from humans.

Joules Barragan (41:22)
Let’s save some time for audience questions. One question is: “What ethical boundaries should we place on AI capable of creating other AIs?”

Rohan Taori (45:20)
Fascinating question. There are at least two aspects here:

1. AI improving or creating new AI: This can mean recursive self-improvement or model distillation. We need to ensure the newly created AIs remain aligned and safe. That calls for robust guardrails and evaluations after each incremental capability gain. AI labs are working on this, but it’s a huge challenge, especially in deciding how to measure risk or “jailbreak” potential.

2. AI spawning more compute for tasks: Another question is how much computational power we allow an AI to use at once, or whether we let it autonomously spin up multiple agents. That also requires careful testing. An iterative deployment approach—where we test new capabilities, see if they break alignment, and then proceed—will be key.

Sean Ren (48:05)
Yes. In the near term, the same alignment methodologies we use today apply to AI-creating-AI scenarios—model distillation, self-play, or multi-agent refinement. It’s not an entirely separate problem, though it raises the stakes. We just need to be extra vigilant.

Joules Barragan (49:05)
Another question: “How close are we to achieving AGI, and what would that mean for humanity?”

Rohan Taori (49:07)
Hard to say, because AGI means different things to different people. Some define it by economic impact, others by the ability of AI to recursively self-improve. But overall, AI capabilities are advancing so quickly that societal transformations will happen faster than previous tech waves (like the internet). I don’t think everything changes in 2025, but we’ll see major leaps, especially in multimodality and extended reasoning.

Sean Ren (50:23)
Yes, it depends on definitions. If you call a highly capable multimodal model “AGI,” that might be a year or two away. Or if you mean something fully aligned and self-improving, that might be longer. But it’s more productive to focus on real-world alignment and safety rather than just a countdown to “AGI.”

Joules Barragan (51:23)
Next question: “How can we encourage the open-source community to engage more in developing and improving model fine-tuning processes, especially in decentralized AI?”

Sean Ren (51:42)
Fine-tuning big models is costly—compute and data are real barriers. One approach is to create new incentives. For instance, if you contribute data for fine-tuning, you should share in the revenue if a fine-tuned model is commercialized. That means building platforms that track data provenance and usage so contributors aren’t just handing over data for free. Also, for compute, we might need decentralized or community-driven clusters to reduce the burden on any single party.

Rohan Taori (52:33)
Yes. Also, we should expand beyond just text-based fine-tuning to RL and environment-based tuning. That might mean letting people build or contribute “mini-environments” that help shape agent behavior. But the big question is how to get enough compute behind it—so it comes back to having good platforms and funding models for open-source.

Joules Barragan (53:16)
Final audience question: “How long until AI can be completely separated from the human component?”

Sean Ren (53:16)
Honestly, I don’t think it should be completely separated. That invites scenarios like The Matrix, where humans end up as an energy source for AI. I’d rather see humans and AI remain closely integrated, continually aligning AI with human ethics, values, and oversight.

Rohan Taori (54:03)
Yes, 100% agree. It’s critical they stay in collaborative alignment.

Joules Barragan (54:09)
That’s all the time we have. Thank you, everyone, for joining us today, and thank you Sean and Rohan for sharing your insights. We’ll pick the prize winners and announce them on X soon.

A few closing notes:

• Don’t forget to follow us on X (Twitter) and join our Discord.

• Sign up for our waitlist: Season 2 of our Data Services Platform is coming soon! We always share alpha info with the waitlist first.

• We’re planning another Twitter Spaces AMA in about two weeks, so stay tuned for that as we go into the New Year.

Thanks again for attending!