2025 AI Battlelines: Agents, Reasoning, and World Models

0:00Today on the AI Daily Brief, why the 2025 AI battle lines are around reasoning models, agents, and world models. Before that, in the headlines, is a new AI mode coming to Google search? The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. To join the conversation, follow the Discord link in our show notes.

0:23Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. One of the really fascinating phenomenon that is somehow kind of backseat relative to everything else going on in AI, is that for the first time in a very long time, 20 years basically, there's competition around what search means. Perplexity is obviously one of the most beloved AI products right now and continues to add to their war chest as well as their feature set. OpenAI is part of their 12 days of shipmas, expanded access to ChatGPT search to the wider world. And now the information is reporting that Google is planning to add an AI mode option to their search as well.

0:59The information sources someone who's working on the product who says that Google is planning to, give its billions of search users the option to switch to an AI mode that looks nearly identical to its Gemini AI chatbot. Others have found indications that this shift won't just be about what you type into your computer, but also that there will be a way to talk to search. 9to5Google found some indications in the code that suggest you'll be able to use mobile inputs, including voice and photos, as a way to search. A Google spokesperson was circumspect about all of this, saying, as our state-of-the-art models continue to advance, there's a huge opportunity to bring these new capabilities into search, helping people discover even more of the web.

1:35And to some extent, this is absolutely obvious. It feels very likely that in the future, there will be simply multiple types of search experiences for different types of queries. The type of perplexity or chat GPT search where you're actually trying to get an answer to a question is going to become default for lots and lots of types of queries. I don't think that means that Google's traditional search has no help, but being able to toggle between the two could be really valuable. The challenge, of course, is that a federal judge has already ruled that Google's search engine is an illegal monopoly.

2:03As the information writes, the Department of Justice has suggested that it wants to make it harder for Google to leverage its search engines to beat AI chatbot rivals, which could create a legal barrier to something like AI mode. This is tough. On the one hand, I absolutely want competition, but on the other, this is just sort of the obvious place to take search for Google, and artificially blocking them from doing so is basically just forcing them to lose. Next up, former Twitch CEO and very briefly OpenAI CEO Emmett Scheer is reportedly working on a new AI startup with some intriguing goals.

2:33You might remember that Scheer was very briefly named as replacement for Sam Altman as CEO of OpenAI during the leadership controversy in November 2023. Interestingly, he was credited by the Wall Street Journal as clearing the path for Sam Altman's return by effectively immediately threatening to resign if he wasn't given evidence by the board to support Altman's removal. TechCrunch is now reporting that Shear has founded a company called STEMAI with incorporation documents filed in June of last year. The company is still in stealth, so details are very limited, but what TechCrunch uncovered does sound interesting.

3:05According to a trademark filed last year, STEMAI is developing software to create AI that, quote, understands, cooperates with, and aligns with human behavior, human preferences, human biology, human morality, and human ethics. More hints come from the presence of Adam Goldstein as a co-founder. After selling a travel website called Hitmonk in 2016, Goldstein became a visiting partner at Y Combinator. He also founded an incubator called Astonishing Labs to back bioresearch startups. According to his LinkedIn page, Goldstein spent a year at Tufts University's Levin Labs as a visiting scientist, where he, quote, developed new models for biological systems with a focus on cancer.

3:38According to PitchBook, STEM received backing from Andreessen Horowitz back in August. And while that's all we know right now about the actual company, Shear has been growing increasingly vocal about AI safety and regulation over the past month. In December, for example, he posted, almost all currently proposed regulation is a bad idea. He added, though, that ideas around regulating firms rather than AI models and increasing transparency are some of the few reasonable ideas currently being floated. Back in November, he wrote, not being scared of AGI indicates either pessimism about the rate of future progress synthesizing digital intelligence or severe lack of imagination about the power of intelligence.

4:13In June, around the time that California's SB 1047 legislation was being debated, he said on a podcast appearance that his greatest concern was self-improving models that could grow out of human control. He said at the time, I'm in favor of creating some kind of fire alarm, like maybe no AI is bigger than X. I think there's good options for international collaboration and treaties about some sort of AI test ban treaty. TLDR, Shira is a good operator, and this is likely one to watch. Lastly today, how the mighty have fallen, Intel is courting bids to buy out its Altera programmable chip arm. Altera specialized in the design of low-power programmable chips for use in AI-enabled devices.

4:48The company was spun off as a separate entity in February as Intel attempted to right the ship after a disappointing few years. Bloomberg reports interest from multiple private equity firms, including Francisco Partners, Silver Lake Management, Apollo Global Management, and Bain Capital. Intel is giving potential buyout partners until January to formalize their offers. Deal terms presented in November range from taking a 20-30 % stake in the company all the way up to taking full control. Bloomberg reports that Altera is being valued in the range of$9 to$12 billion, a steep discount from the$17 billion Intel paid back in 2015.

5:19The move comes, of course, in the shadow of the departure of CEO Pat Gelsinger. After being brought in three years ago to get Intel back on track, Gelsinger retired from his position earlier this month at the request of the board. That, however, is going to do it for today's AI Daily Brief Headlines edition. Next up, the main episode. Today's episode is brought to you by Vanta. Whether you're starting or scaling your company's security program, demonstrating top-notch security practices and establishing trust is more important than ever. Vanta automates compliance for ISO 27001, SOC 2, GDPR, and leading AI frameworks like ISO 42001 and NIST AI risk management framework, saving you time and money while helping you build customer trust.

6:00Plus, you can streamline security reviews by automating questionnaires and demonstrating your security posture with a customer-facing trust center all powered by Vanta AI. Over 8 ,000 global companies like LangChain, Lila AI, and Factory AI use Vanta to demonstrate AI trust and prove security in real time. Learn more at vanta.com slash NLW. That's vanta.com slash NLW. If there is one thing that's clear about AI in 2025, it's that the agents are coming. Vertical agents by industry, horizontal agent platforms, agents per function. If you are running a large enterprise, you will be experimenting with agents next year.

6:38And given how new this is, all of us are going to be back in pilot mode. That's why Superintelligent is offering a new product for the beginning of this year. It's an agent readiness and opportunity audit. Over the course of a couple quick weeks, we dig in with your team to understand what type of agents make sense for you to test, what type of infrastructure support you need to be ready, and to ultimately come away with a set of actionable recommendations that get you prepared to figure out how agents can transform your business. If you are interested in the agent readiness and opportunity audit, reach out directly to me, nlw at bsuper.ai.

7:11Put the word agent in the subject line so I know what you're talking about. And let's have you be a leader in the most dynamic part of the AI market. Welcome back to the AI Daily Brief. I love it when a set of stories converge in a way that really tells a bigger story than just an individual piece of news. And boy, is that the case today. We are looking at what are shaping up to be the 2025 battle lines in AI. Reasoning models, agents, and world models, all of which got some interesting news yesterday. Now, of course, at this point, you probably don't need the background on reasoning models, but effectively, this is a new approach to scaling that uses different strategies than just throwing more compute and data in the pre-training.

7:50This is clearly where OpenAI is putting its emphasis. It released O1 Preview back in September. Subsequent to that, we've had Amazon announcing Nova and talking about a reasoning model in the lineup. Meta releasing Lama 3.3, also emphasizing its reasoning capabilities. Several Chinese labs have also released very competent reasoning models. And now Google has joined the party. A few days after their initial launch, Google has added a reasoning model to the Gemini 2.0 Flash lineup. The model is called Gemini 2.0 Flash Thinking Experimental. Hopefully this is just a working title, guys. And the way it describes itself is that it's the best in the lineup for, quote, multimodal understanding, reasoning, and coding.

8:27In demonstrations, it seems to perform well on puzzles involving both visual and text clues. And so far as novel features that make it stand out from the pack, the model shows its chain of logic so you can see what's going on under the hood. OpenAI co-founder Andre Karpathy wrote, the prominent and pleasant surprise here is that unlike O1, the reasoning traces of the model are shown. As a user, I personally really like this because the reasoning itself is interesting to see and read. The models actively think through different possibilities, ideas, debate themselves, etc. The case against showing these is that it's typically a concern of someone collecting the reasoning traces and training to imitate them on top of a different base model to gain reasoning ability possibly and to some extent.

9:04The model is also extremely fast compared to Rivals and available for free on Google AI Studio. This in and of itself is pretty surprising as reasoning models so far have been extremely expensive to operate compared to their non-reasoning counterparts. One interesting thing here is that Google's naming convention implies that this is just a fine-tuned version of 2.0 Flash. Or perhaps simply the base model with some system prompts to ask the model to think for longer and check its work before answering. Compare that to 01, where OpenAI went out of their way to present it as an entirely new model.

9:34Sam Altman even framed the release as the beginning of a different branch of LLMs for the company. One of the big questions I think heading into next year is just how different these reasoning models are from their non-reasoning counterparts, and more particularly, whether they really do evolve in different ways going forward. Now, speaking of OpenAI, the other big news for reasoning models is that OpenAI is preparing to release the second generation of their O1 model. Funny enough, speaking of weird naming conventions, according to the information, the model is going to be called O3 to avoid intellectual property disputes with British Telco O2.

10:06Sam Altman all but told us the model would be released today, so it's probably out by the time you're listening to this. The release could answer another big question surrounding reasoning models, which is whether they can show major improvements on the model layer. At the time O1 came out, it was suspected that OpenAI was pivoting to reasoning because adding training data and compute to training runs was showing diminishing returns. In the following months, it was confirmed that noticeable improvements could be made to reasoning models by getting them to think longer. Assuming O3 is a brand new model and not a tweak of O1, it should reveal whether reasoning models themselves can scale an ability, or instead if all of the improvements are only possible by scaling up inference time.

10:42The community is pretty excited to check it out. Chubby on X writes, O3 equals Orion. There is probably no more GPT 4.5 or 5. Everything is summarized in Orion. i.e. O3. Surely Orion was fed with a lot of synthetic data from O1 and now has evolved into O3. Chubby also got at the competitive dynamics in the field, writing, time to take back the crown from Google. We have one more normal episode coming on Monday before we get into end-of-the-year specials, so I will have a chance to follow up on what exactly came out on Friday. Now, moving on to the next dimension of competition, the one that's even more obvious than reasoning models in some ways, is the race to deploy agents.

11:19Yesterday, OpenAI announced a long list of new integrations for ChatGPT. The desktop application can now access data from a gigantic list of coding platforms, as well as Apple Notes, Notion, and Quip. For now, ChatGPT can only read these apps in context. It can't take actions within those programs. But Chief Product Officer Kevin Wheel made it clear that's where this is all going. He said, We've been putting a lot of effort into our desktop apps. As our models get increasingly powerful, ChatGPT will become more and more agentic. That means we'll go beyond just questions and answers. ChatGPT will begin doing things for you.

11:49A few weeks ago, Wharton Professor Ethan Malek posted, OpenAI has a lot of pieces on the board right now. Multimodal vision and voice, small, large, and reasoning models, image and video creation, code execution, mobile and desktop apps, web search, some agentic stuff. Very curious when it will be glued together into a singular thing. Interestingly, going back to this quote from Kevin Wheel, it seems extremely notable to me that he says ChatGPT will begin doing things for you. This indicates pretty clearly to me that we have the singular thing, that it is and has always been ChatGPT. It's just that over time, ChatGPT is going to be a lot more and frankly different from what ChatGPT originally was.

12:29A couple of days ago, we also got an update from Salesforce on their agent platform, AgentForce. Just three months after announcing AgentForce in September, the company announced AgentForce 2.0. They write, This release introduces a new library of pre-built skills and workflow integrations for rapid customization, the ability to deploy agent force in Slack, and advancements in agentic reasoning and RAG. These advancements will enable companies to scale their workforce with customized agents capable of handling complex multi-step tasks with even more precision and accuracy. And if you need a sense of just how important this is to Salesforce, go check out the piece in the information AI is Mark Benioff's friend and foe.

13:05It talks about how Salesforce is facing increasing competition from companies like Sierra, which are bringing agents to market, and in fact, in some cases, winning business away from Salesforce. The final vector of competition that I want to discuss today is this new world model approach. These models are trained in a fundamentally different way to LLMs. Where LLMs are trained on a large corpus of text, image, and voice data, world models are trained by observing real or simulated worlds. We've seen a few working prototypes of this style of AI, with two big examples coming out of Fei-Fei Li's World Labs and another coming out of Google DeepMind.

13:36A third big player is Descartes, who released a model in October capable of generating a fully playable Minecraft-like game. While the demo was buggy and rudimentary, it clearly made investors sit up and pay attention. TechCrunch reports that the company has now raised their Series A. The startup raised$32 million at a$500 million valuation. CEO and co-founder Dean Liedersdorf said, the company wants to compete at the highest level, building a, quote, fully vertically integrated AI research lab alongside enterprise and consumer products. He said the aim was to create what he's calling a kilocorn.

14:05In other words, a trillion-dollar company. You gotta love ambition, man. Part of the reason that people are so interested in world models, especially recently, is the sense that perhaps their understanding of physics could be something that allows them to make more fundamental breakthroughs. Still, in many ways, this class of models feels closer to where GPT-based LLMs were a few years ago, demonstrating some fascinating emergent properties, but still nowhere near the full scale that they're going to reach. Along those lines, a group of researchers across 19 different universities have just revealed something they are calling a comprehensive physics simulation platform.

14:38Named Genesis, the researchers claim the platform is, quote, capable of simulating a wide range of materials and physical phenomenon. Researcher Zhao Zian writes, after a 24-month large-scale research collaboration involving over 20 research labs, a generative physics engine able to generate 4D dynamic worlds powered by a physical simulation platform designed for general purpose robotics and physical AI applications. We aim to build a universal data engine that leverages an upper level generative framework to autonomously create physical worlds, together with various modes of data including environments, camera motions, robotic task proposals, reward functions, robot policies, character motions, fully interactive 3D scenes, open world articulated assets, and more.

15:16aiming towards fully automated data generation for robotics, physical AI, and other applications. So basically, Genesys, as I understand it, can be used as both a robotic simulation platform and as a photorealistic rendering platform. The platform accepts natural language prompts and can be used as a data engine to produce a range of different modalities of synthetic or simulated data. In the immediate term, it's a massive boost to robotics training in terms of speed and accuracy. It could lead to immediate improvements in that field and possibly even unlock more complex use cases. Roboticist Ben Duffy commented, quote, with Genesis, you'll be able to train a locomotion policy that's deployable in real world in less than 26 seconds.

15:52That sentence tells us about a future we are not ready for. For reference, that's 430 ,000 times faster than the previous leading physical simulators to give a sense of how dramatic this change could be. One of the other potentials is that a platform like this could produce the gigantic data sets required to scale up world models. Currently, they've been trained using either data sets from self-driving cars or by observing video games. There are a few projects strapping camera rigs to hikers in order to gather real-world data, but if this platform is as performant as the researchers claim, we could soon see near-infinite synthetic datasets available to train the next generation of world models.

16:26Effectively, all of the response to this is some version of wow. Viewing their announcement video, AI evangelist Linus writes, this is all generated and simulated in 4D. Mind-blown emoji, mind-blown emoji, mind-blown emoji. Billawalseed who writes, Think instant physics-accurate environments, camera paths, and character animations, all from natural language. Milo writes, what the? This Genesis project is like something out of a sci-fi movie. I mean, generating entire 4D worlds with physics emulations? That's mind-blowing. I'm just sitting here stunned, trying to wrap my head around how this could change everything from robotics to video games.

17:00And I think, friends, if we had to sum this up, if you thought that 2025 was going to be any slower than 2024 and 2023 had been, boy, do you need to think again. That's going to do it for today's AI Daily Brief. Appreciate you listening as always. Until next time, peace.