In a world where “bigger is better” has become the default maxim in AI, Samsung’s recent paper turns that narrative on its head. Their Tiny Recursive Model (TRM), with just 7 million parameters—orders of magnitude smaller than today’s sprawling foundation models—achieves state‑of‑the‑art results on some of the hardest reasoning benchmarks. It’s a provocative demonstration that smarter architecture, not brute force scaling, might be the next frontier.

The Scale Trap: Why Big Models Still Struggle with Reasoning

Over the past few years, the AI arms race has fixated on parameter counts. Models with hundreds of billions—and soon trillions—of parameters have become the norm, enabling fluent language generation, multimodal reasoning, and general-purpose capabilities. Yet, when it comes to multi‑step reasoning—solving puzzles, planning paths, logical deduction—these behemoths remain brittle. A single misstep early in generation can compound errors, leading to invalid conclusions.

To compensate, researchers introduced methods like chain-of-thought prompting, which encourages models to “think aloud” through intermediate steps. However, these methods come with costs: they increase computational load, require specialized prompting or training, and still don’t guarantee flawless logic.

Enter TRM—a model that targets reasoning directly with a recursive architecture built to self-correct, rather than relying on sheer scale or brute force.

The Tiny Recursive Model (TRM): A Minimalist with a Punch

The core insight behind TRM is deceptively simple: use recursion and self‑refinement to incrementally polish both the reasoning trace and the answer itself. The model receives the problem prompt, an initial guess at the answer, and a latent reasoning vector. It then cycles—up to 16 times—through a two-stage process: first, it updates the latent reasoning vector based on the prompt, current answer, and prior reasoning. Second, it uses the updated reasoning to propose an improved answer.

Rather than relying on fixed-point convergence theorems, TRM is trained by backpropagating through the full recursive process. Surprisingly, the researchers found that a shallow two‑layer network version of TRM outperformed a deeper four‑layer variant. Intuitively, restricting capacity may help avoid overfitting and force more generalizable reasoning patterns.

Blowing Benchmarks Out of the Water

The results are striking. On tasks where training data is sparse and reasoning precision is critical, TRM posts significant gains. On the Sudoku-Extreme benchmark, TRM hits 87.4 percent accuracy, compared to a baseline of around 56.5 percent using hierarchical reasoning models (HRMs) with more parameters and longer training. On Maze-Hard, which involves pathfinding in large 30×30 grids, TRM achieves 85.3 percent accuracy, significantly outperforming HRM’s 74.5 percent.

Most dramatically, on the Abstraction and Reasoning Corpus (ARC-AGI) benchmarks—designed to test fluid, general intelligence—TRM’s 7 million-parameter version achieves 44.6 percent on ARC-AGI-1 and 7.8 percent on ARC-AGI-2. These numbers not only beat HRMs with 27 million parameters but also surpass the performance of some of the largest commercial LLMs, such as Gemini 2.5 Pro, which scores around 4.9 percent on ARC-AGI-2.

These gains come without extravagant compute. TRM introduces an adaptive stopping mechanism (ACT) to decide when recursion is sufficient, reducing wasteful extra forward passes during training and inference.

Implications: Architectures Over Scale?

If TRM’s performance holds across broader benchmarks, this work could mark a pivotal shift in how we build AI.

Efficiency and sustainability become much more viable when you can achieve state-of-the-art results without expensive hardware or massive data centers. A 7 million-parameter model that outperforms giants in key reasoning tasks is a stark counterexample to the “bigger is always better” mindset.

Rather than forcing a gigantic general-purpose model to master every task, future systems might combine tiny, specialized reasoning modules with larger generative backbones. You might call a TRM-like module only when precise logic is needed.

ARC-AGI was created to test general fluid intelligence—the ability to solve new, abstract problems. That TRM does well here suggests that architectural cleverness may matter more than scale when it comes to true intelligence, not just pattern matching.

Caveats and Open Questions

TRM’s promise is compelling, but there are several caveats. The benchmarks used—Sudoku, Maze, ARC—are highly structured and well-defined. Real-world reasoning often involves ambiguity, commonsense, and incomplete information.

TRM’s recursion depth is fixed and bounded; some problems might require more flexible or unbounded reasoning chains. It also remains to be seen how TRM-style modules integrate with large language models and whether similar strategies scale to multimodal or open-ended tasks.

Conclusion
Samsung’s Tiny Recursive Model points toward a bold alternative to the current scaling regime: leaner, smarter architectures that recursively self-correct rather than relying on mind-boggling parameter counts. If this approach generalizes, we may be witnessing the dawn of an AI paradigm where efficiency and elegance outstrip brute force.

In one corner stands TikTok, the reigning king of short-form entertainment and social engagement. In the other—barely a week old—emerges Sora 2, OpenAI’s audacious experiment in blending generative AI with social media. The two platforms share a superficial resemblance: vertical video, endless scroll, algorithmic feeds. But beneath the surface, they diverge dramatically. Comparing them is like contrasting a stadium concert with an improvisational theater performance. This piece explores how these platforms differ in purpose, audience, appeal, and potential—while examining whether Sora 2 is a passing novelty or the start of a creative revolution.

The Platforms at a Glance: Legacy vs. Disruption

TikTok is already a household name, with over 875 million global downloads in 2024 alone and more than 1.5 billion monthly active users worldwide. It has cemented its position as a cultural and commercial powerhouse. Users flock to it not just to consume content, but to engage in creative expression, trends, and community. TikTok’s algorithmic feed—known as the “For You” page—serves as a launchpad for virality, social discovery, and even political discourse. It offers a toolkit for creators, including monetization options, live streaming, and e-commerce integration, reinforcing its role as a full-spectrum media ecosystem.

Sora 2, by contrast, is the newest contender on the scene. Built around OpenAI’s powerful text-to-video model, it enables users to generate short, AI-crafted videos by entering prompts or remixing existing ones. Unlike TikTok, where the content is user-recorded and often tied to real life, Sora 2 is more speculative—a kind of dream-machine for visual storytelling. Although it is still in invite-only stages in many regions, the app surged to the top of iPhone app store charts shortly after its release. This suggests that curiosity, if not yet loyalty, is already high.

What Users Can Do—and What They Actually Want

TikTok thrives on personal performance and cultural participation. Users film their own videos—ranging from dance routines and lip-syncs to comedy sketches and DIY tutorials. These clips are then shared, remixed, or commented upon, creating a dynamic social loop. Engagement is driven by recognition and interaction: creators build loyal followings, often turning their digital personas into careers. The app is optimized for viral success, with ordinary users able to reach millions overnight. It’s a space where authenticity, relatability, and personal flair are often more valued than polished production.

Sora 2, on the other hand, shifts the focus from “what I can do” to “what I can imagine.” Instead of uploading filmed footage, users generate video snippets through textual prompts, often resulting in surreal, stylized, or entirely fictional outputs. There’s a significant emphasis on remix culture—users can take someone else’s AI-generated video, tweak it, and publish their own version. Some are even creating mashups involving real or fictional figures, sometimes controversially featuring celebrities or historical personalities. The app includes mechanisms for managing consent and attribution, but the social norms are still forming.

While TikTok encourages real-time creativity based on lived experience, Sora 2 promotes imaginative storytelling unbound by reality. Its users are more like directors or prompt-engineers than performers.

What Makes Them Attractive

TikTok appeals because of its familiarity. Its content is rooted in real life, its trends reflect popular culture, and its social loops—likes, comments, shares—create a sense of community. Viewers recognize the people behind the videos, connect with their stories, and return to see what they’ll post next. There’s also the powerful allure of virality; the platform has made stars out of previously unknown teenagers and sparked music hits and fashion movements across the globe.

Sora 2’s charm lies in novelty and surprise. The unpredictability of AI-generated content—imagine a reimagined New York skyline filled with cats or a synthetic Tupac rapping Shakespeare—can be mesmerizing. Its strength is in speculative creativity, turning dreams, jokes, and “what if” questions into videos. For now, it’s more of a curiosity cabinet than a social space. But that might change if users begin to build persistent identities or recurring themes within their AI-generated content.

TikTok rewards authenticity and performance, while Sora 2 celebrates imagination and synthesis. Both are creative, but they differ in what kind of creativity they prioritize.

Challenges and Ethical Dimensions

TikTok is no stranger to controversy, facing criticism for data privacy, content moderation, mental health effects, and algorithmic addiction. However, its scale and longevity have allowed some of these concerns to be addressed through policy changes, public scrutiny, and user familiarity with its risks.

Sora 2 enters even murkier territory. Its very premise—generating video with AI—raises questions about ownership, ethics, and representation. Users have already begun creating deepfakes and fictionalized portrayals of real people, including public figures, without clear legal boundaries. OpenAI has implemented visible watermarks and consent tools, and has promised to enforce policies around impersonation and misinformation. But the speed at which users are pushing the platform’s limits suggests enforcement will be a constant challenge.

There’s also the problem of saturation. While TikTok’s content is grounded in endless human variation, Sora 2’s AI-generated clips may start to feel repetitive once the novelty wears off. If every video is a remix of the same surreal themes, users might disengage, especially without emotional or social anchors.

Another issue is demographic imbalance. Reports indicate that Sora 2’s public feed is currently dominated by teenage boys, with very little female participation. This skew could hinder its appeal and slow its evolution into a truly inclusive social platform.

Metrics, Momentum, and Uncertainty

TikTok’s dominance is clear. It commands over a billion active users and enjoys institutional scale, established monetization paths, and a wide-ranging creator economy. Sora 2 is still in its infancy. While it’s impossible to know how many active users it currently has, early signals show enormous interest. Its app store debut was explosive, and public discussion is already likening it to the “ChatGPT moment” for video.

OpenAI is positioning Sora 2 not just as a creative toy, but as a foundational platform for generative media. Some observers believe it could redefine what social media looks like in the age of synthetic content. Others are skeptical, viewing it as another hype-driven AI experiment that could implode once the novelty fades and the moderation issues pile up.

A Bubble or the Beginning?

Sora 2 has enormous potential, but it walks a tightrope. On one side, it could empower a new generation of storytellers, lowering the barrier to visual creativity and spawning new genres of content. It could even integrate into broader ecosystems—ChatGPT, plugins, or creative suites—making it a key node in the generative web.

On the other hand, the risks are substantial. If the platform fails to build strong social bonds, offers little creator monetization, or becomes overrun with ethically fraught content, it could fade quickly. It might remain a powerful tool—but not a lasting social platform.

TikTok’s strength is in its deep entrenchment in culture. It mirrors life, amplifies identity, and thrives on community. Sora 2 is more like a lucid dream: stunning to watch, fascinating to interact with, but not yet grounded in sustained, emotional or social relevance.

Final Thoughts: Two Different Realities

TikTok is about showing the world who you are. Sora 2 is about showing the world what you can imagine. One reflects life; the other reshapes it. One builds community through shared experience; the other through shared creativity.

It’s too early to declare a winner—and maybe that’s the wrong frame. Sora 2 doesn’t need to replace TikTok. If anything, it might redefine what the next phase of digital creativity looks like: more automated, more collaborative, more surreal. Whether it becomes a new cultural mainstay or fades into the long list of tech novelties will depend not just on its technology, but on whether it can foster real, meaningful connections in a world increasingly full of synthetic voices.

In the rapidly evolving world of generative AI, text-to-video has become the new frontier. The release of OpenAI’s Sora 2 and Google DeepMind’s Veo 3 has ignited fresh debate over which model currently leads the charge. Both promise cinematic-quality video from text prompts, yet their strengths—and limitations—reveal very different approaches to solving the same problem. So, which one is truly pushing the envelope in AI-generated video? Let’s take a closer look.

The Shape of a New Medium

Sora 2 and Veo 3 aren’t just iterative updates; they represent a leap forward in AI’s ability to understand, simulate, and visualize the physical world. Veo 3, unveiled as part of Google’s Gemini ecosystem, emphasizes realism, cinematic polish, and high-fidelity audio. Sora 2, OpenAI’s successor to its original Sora model, doubles down on deep physics simulation, coherence across time, and intelligent prompt understanding.

Both models target similar creative workflows—commercials, short films, visual storytelling—but their design choices show stark contrasts in how they get there.

Visual Realism and Cinematic Quality

On first impression, both Sora 2 and Veo 3 impress with sharp resolution, consistent lighting, and smooth transitions. Veo 3, in particular, demonstrates a clear edge in cinematic effects: seamless camera movement, depth-of-field rendering, and visually stunning transitions that mimic professional film work. Veo’s ability to replicate human-directed cinematography stands out.

Sora 2, by contrast, leans harder into realistic physics and object behavior. Where Veo 3 dazzles with filmic beauty, Sora 2 seems more intent on ensuring that what happens on screen makes sense. Vehicles move with believable momentum, liquids splash and flow realistically, and characters interact with their environment in ways that respect gravity and friction. This physics-aware realism may not always be as visually glossy as Veo 3, but it adds a layer of believability that matters for narrative coherence.

Temporal Coherence and Scene Continuity

A major weakness of early video generators was temporal inconsistency: objects morphing frame-to-frame, faces flickering, or scene geometry drifting. Sora 2 makes significant strides in solving this. Across 10-second (and sometimes longer) videos, objects remain stable, actions continue naturally, and the scene retains structural integrity.

Veo 3 also shows improvement here, but with caveats. While its short clips (typically 4–8 seconds) hold together well, subtle issues can emerge in complex motion sequences or rapid cuts. In side-by-side prompts involving a person dancing through a rainstorm or a dog running through a forest, Sora 2 often preserves object integrity and movement more effectively over time.

However, Veo 3’s strength in lighting and composition can sometimes make its videos appear more polished—even when inconsistencies are present.

Audio Integration and Lip Sync

Here’s where Veo 3 pulls ahead decisively. Veo 3 not only generates realistic visuals but also supports synchronized audio, including ambient noise, sound effects, and even lip-synced speech. This makes it uniquely suited for use cases like video ads, dialogue scenes, and social media content that require full audiovisual immersion.

Sora 2 has made progress in audio generation, but lip-sync remains rudimentary in current versions. While OpenAI has demonstrated Sora’s ability to match ambient sounds to visuals (like footsteps or weather effects), it has not yet caught up to Veo in producing realistic spoken dialogue.

For creators working in multimedia formats, Veo 3’s audio capabilities are a game-changer.

Prompt Control and Creative Flexibility

Controllability—how much influence users have over the generated output—is key to unlocking creative potential. Veo 3 offers a relatively straightforward prompting system, often yielding high-quality results with minimal fine-tuning. However, it sometimes sacrifices precision for polish; complex multi-step prompts or shot-specific instructions can be hard to achieve.

Sora 2, in contrast, supports a more nuanced form of instruction. It appears better at following detailed, layered prompts involving camera angles, character action, and scene transitions. This makes it especially appealing to storytellers or developers who want fine-grained control over the output.

If you’re crafting a multi-part scene with shifting perspectives and nuanced interactions, Sora 2 often delivers a more controllable, logically grounded result.

Limitations and Access

Despite their power, both models remain gated behind layers of access control. Veo 3 is currently integrated into Google’s suite of tools and remains limited to selected creators, while Sora 2 is available through invite-only access via OpenAI’s platform.

Sora 2 also enforces stricter prompt filtering—especially around violence, celebrities, and copyrighted characters—making it less permissive in some creative contexts. Veo 3, while still governed by safety policies, appears slightly more lenient in some edge cases, though this can change with updates.

Both models are also computationally intensive, and neither is fully accessible via open API or commercial licensing at scale yet.

Final Verdict: Different Strengths, Different Futures

If you’re choosing between Sora 2 and Veo 3, the best answer may not be “which is better?” but “which is better for you?”

• Choose Veo 3 if your priority is audiovisual polish, cinematic beauty, and natural soundscapes. It’s ideal for creators looking to generate short, eye-catching content with minimal post-processing.
• Choose Sora 2 if your work demands physical realism, temporal stability, or precise narrative control. It’s a better fit for complex scenes, storytelling, and simulation-heavy tasks.

Both are leading the charge into a future where the boundary between imagination and reality blurs further with every frame. As the models continue to evolve, the true winners will be the creators who learn to harness their distinct strengths.