AI Weekly Digest - Perplexity's Model Council, China's AI Blitz, and Gemini's 750M Users

Week in Review: February 10-15, 2026

If you wanted to understand where the AI industry is heading in a single week, this was the one to watch. On one side of the Pacific, Perplexity launched a feature that treats multiple frontier AI models as a deliberative council rather than competing oracles. Google pushed Gemini 3 Deep Think into territory that solves research problems humans have struggled with for a decade. And on the other side, China's tech giants turned the Lunar New Year into a full-blown AI arms race, flooding the market with new models, giving away money and luxury cars, and racing to prove that the DeepSeek moment was not a fluke but a permanent shift in global AI dynamics. Somewhere in between, 100 experts from 30 countries published a sobering assessment of where all of this is actually heading.

This was not a week of incremental progress. This was a week where the architecture of how we interact with AI changed, where the geographic center of AI innovation blurred further, and where the gap between what AI can do and what society is prepared for widened again.

The common thread running through every story this week is convergence. Convergence of models into multi-model systems. Convergence of Chinese and Western capabilities. Convergence of consumer adoption numbers that make AI the fastest-adopted technology category in human history. And convergence of safety concerns that the international community is finally trying to address with the rigor the moment demands.

---

1. Perplexity Model Council: The End of the Single-Model Era

Perplexity launched Model Council on February 5, and it took a week for the implications to fully register across the industry. The feature runs a single query across three frontier AI models simultaneously -- Claude Opus 4.6, GPT-5.2, and Gemini 3.0 among them -- then uses a synthesizer to reconcile their outputs into a single response that shows where the models agree and where they diverge.

This is not a gimmick. It is an architectural statement about the future of AI interaction. For the past three years, the AI industry has operated on the assumption that users choose a single model and interact with it as an oracle. Perplexity is arguing that the oracle model is fundamentally broken because every model has blind spots, biases, and failure modes that are invisible to the user unless they can see how other models respond to the same question.

The chair model in Perplexity's system does the heavy lifting. After the parallel responses are generated, it identifies areas of agreement, highlights points of disagreement, surfaces unique insights from each model, and evaluates the strength of evidence. The result is not a simple average of three outputs but a synthesized response that is meaningfully more reliable than any single model could produce alone.

The implications for enterprise adoption are significant. Organizations that have been reluctant to trust a single AI model for critical decisions now have a framework that mirrors human decision-making practices like peer review and committee deliberation. Multi-model consensus does not eliminate AI errors, but it makes them visible in ways that single-model interaction cannot.

Model Council is currently available for Perplexity Max subscribers on the web, with mobile support and potential Pro tier expansion coming. The pricing positions it as a premium research tool rather than a consumer feature, which is the right framing. This is infrastructure for serious inquiry, not a chatbot upgrade.

Why it matters: Perplexity is not building a better chatbot. It is building a verification architecture that treats disagreement between AI models as signal rather than noise. If this pattern catches on, it could fundamentally change how organizations evaluate AI-generated information.

---

2. Google Gemini 3 Deep Think: AI That Solves What Humans Cannot

Google released a major upgrade to Gemini 3 Deep Think on February 12, and the benchmark results are the kind that force you to recalibrate your assumptions about what AI can do. The updated model achieved an unprecedented 84.6 percent on ARC-AGI-2, verified by the ARC Prize Foundation. It scored 48.4 percent on Humanity's Last Exam without tools. It reached 3455 Elo on Codeforces and gold-medal level on the 2025 International Math Olympiad.

But the benchmarks are not the story. The story is that Gemini 3 Deep Think solved 18 previously unsolved research problems across mathematics, physics, and computer science. It disproved a mathematical conjecture from 2015 that human researchers had been unable to resolve for a decade. In early testing, it identified a subtle logical flaw in a technical mathematics paper that had passed through human peer review unnoticed.

Google is positioning Deep Think specifically for science and engineering applications. The model is available to AI Ultra subscribers through the Gemini app, and Google has opened an early access program for scientists, engineers, and enterprises through the Gemini API. The focus is deliberate. Deep Think is not meant to write better marketing copy. It is meant to accelerate scientific discovery.

The 18 solved research problems span domains from number theory to fluid dynamics to algorithmic complexity. Each represents a problem where human researchers had been stuck, not for days or weeks, but for years. The implication is not that AI is smarter than humans in some abstract sense but that AI reasoning approaches problems from angles that human intuition tends to overlook.

Why it matters: Gemini 3 Deep Think is the clearest evidence yet that AI reasoning models are crossing from "useful tools" to "research collaborators." When an AI system can disprove decade-old conjectures and catch errors that survive peer review, the relationship between AI and scientific research changes from augmentation to partnership.

---

3. China's AI Model Blitz: RynnBrain, Seedance 2.0, and Kling 3.0

The week's most consequential story may not be any single announcement but the collective weight of what China's AI ecosystem produced in the span of five days. Alibaba released RynnBrain for robotics. ByteDance launched Seedance 2.0 for video generation. Kuaishou shipped Kling 3.0 with native 4K and multi-shot storyboarding. And behind all three, a dozen smaller companies released models that would have been headline news in isolation.

Alibaba RynnBrain: Open-Source Robotics AI

Alibaba's DAMO Academy released RynnBrain on February 10, an open-source robotics foundation model that set new records across 16 embodied AI benchmarks, outperforming Google's Gemini Robotics-ER 1.5 and NVIDIA's Cosmos-Reason2. The model ships in multiple configurations: 2-billion and 8-billion parameter dense versions and a 30-billion parameter mixture-of-experts variant, all available on Hugging Face.

RynnBrain is trained on Alibaba's Qwen3-VL vision-language model and can map objects, predict trajectories, and navigate cluttered environments like kitchens and factory assembly lines. The architecture combines vision-language understanding with fine-grained motor control, enabling robots to interpret natural language instructions and execute complex physical tasks.

ByteDance Seedance 2.0: Video Generation Goes Cinematic

ByteDance unveiled Seedance 2.0 on February 12, and it immediately drew comparisons to the original DeepSeek moment. The model generates multi-shot cinematic videos up to 20 seconds long with synchronized audio, integrating four input modalities -- text, image, video, and audio -- to produce narratives with consistent characters, emotion-driven expressions, and phoneme-level lip-sync in over 8 languages.

The technical leap is substantial. Seedance 2.0 delivers up to 2K resolution at 30 percent faster generation speed than its predecessor. The model incorporates physics-aware training that produces videos where gravity works, fabrics drape correctly, fluids behave realistically, and object interactions look believable.

Hollywood has already responded. The Motion Picture Association issued a statement demanding ByteDance "immediately cease its infringing activity," claiming massive-scale unauthorized use of copyrighted works within a single day of launch.

Kuaishou Kling 3.0: The Creator Tool

Kuaishou launched Kling 3.0 on February 5, including Video 3.0, Video 3.0 Omni, Image 3.0, and Image 3.0 Omni. The headline features include native 4K ultra-high-definition output, multi-shot storyboarding where users can specify duration, shot size, perspective, and camera movements for each segment, and native audio generation across multiple languages and dialects. Since launching in June 2024, Kling AI now serves over 60 million creators and has produced more than 600 million videos.

Why it matters: China is not catching up. In robotics AI, video generation, and creator tools, Chinese companies are setting the benchmarks that Western competitors must chase. The DeepSeek moment was not an anomaly. It was the beginning of a structural shift in global AI capability distribution.

---

4. The Lunar New Year AI War: $700 Million in Cash and Cars

The Year of the Horse kicked off on February 15, and China's tech giants turned the holiday into the most expensive AI marketing campaign in history. ByteDance handed out 100,000 prizes during the country's main holiday TV gala, including luxury cars, along with digital red envelopes worth up to CNY 8,888 (about $1,280). Baidu allocated CNY 500 million (about $72 million) to promote its Ernie chatbot. Alibaba pledged $432 million in subsidies. Tencent matched with its own campaign. Total spending across the major players approached $700 million.

The strategy is straightforward: acquire users during the highest-engagement period of the Chinese calendar, then convert them into habitual AI app users over the following months. ByteDance's Doubao chatbot leads with 155 million weekly active users. Alibaba's Qwen topped the Apple App Store free-download chart in China after launching its promotions.

But the campaign drew a sharp response from Beijing. China's top market regulator summoned the leading tech companies to demand an end to what it called "involutionary" competition -- a term describing destructive, zero-sum rivalry where companies burn capital without creating sustainable value. The regulator warned that the giveaway war risked destabilizing the AI market before it had time to mature.

The tension between market growth and regulatory concern captures a contradiction at the heart of China's AI strategy. Beijing wants Chinese AI companies to lead globally. But it also wants orderly markets and does not want AI competition to become a subsidy war that destroys margins before the companies can generate sustainable revenue.

Why it matters: The scale of spending signals that Chinese tech companies view consumer AI adoption as a winner-take-most market where early user acquisition determines long-term dominance. The regulatory pushback signals that the government is worried the race is producing heat without light.

---

5. Google Gemini Crosses 750 Million Monthly Active Users

Google disclosed during its Q4 2025 earnings that the Gemini app has surpassed 750 million monthly active users, adding 100 million users in a single quarter. The growth was driven substantially by the launch of Gemini 3, which Google described as the fastest-adopted model in the company's history. Gemini models now process over 10 billion tokens per minute via direct API use.

The number puts Gemini in striking distance of ChatGPT's estimated 810 million monthly active users, though both trail Meta AI's reported 1 billion. The competitive picture is more nuanced than raw user counts suggest. Gemini benefits from deep integration with Google's existing product ecosystem -- Search, Android, Workspace, Chrome -- giving it distribution advantages that standalone AI apps cannot match. ChatGPT, by contrast, has built its user base almost entirely through direct product appeal.

The 10 billion tokens per minute through API usage is perhaps the more telling metric. It reflects not just consumer chatbot usage but enterprise and developer adoption at scale. When organizations build applications on top of Gemini's API, they create switching costs that consumer users do not.

Why it matters: AI assistant adoption is approaching the scale of social media platforms. At 750 million monthly users, Gemini has more active users than Twitter/X ever achieved at its peak. The AI assistant market is not a niche technology category anymore. It is a mass consumer market with implications for every industry that touches information work.

---

6. International AI Safety Report 2026: A Global Reckoning

The second International AI Safety Report was published on February 3, authored by over 100 AI experts from more than 30 countries under the leadership of Turing Award winner Yoshua Bengio. The report provides the most comprehensive science-based assessment to date of general-purpose AI capabilities, risks, and the state of safety practices across the industry.

The findings are simultaneously reassuring and alarming. On the capability side, the report documents that general-purpose AI can now converse fluently in numerous languages, generate code, create realistic images and videos, and solve graduate-level mathematics and science problems. Reasoning models have become more common, with improved performance in biology, chemistry, and protein design.

On the risk side, the report notes that models remain unreliable on multi-step tasks, continue to produce hallucinations, and struggle with physical world reasoning. More critically, the report found that while 12 companies published or updated frontier AI safety frameworks in 2025, there is no unified approach to risk management across the industry. Each company defines risk differently, measures it differently, and mitigates it differently.

The report's most striking finding is about adoption velocity. More than 700 million people now use leading AI systems every week, making the rate of adoption faster than the personal computer, smartphones, or social media. The gap between how many people are using AI and how well we understand its risks is widening, not narrowing.

Why it matters: The International AI Safety Report is the closest thing the global community has to a consensus scientific assessment of AI risks. Its finding that adoption is outpacing safety understanding should concern everyone building, deploying, or depending on AI systems.

---

7. The DeepSeek Aftermath: China's Low-Cost Model Offensive

One year after DeepSeek shocked the industry with models that matched Western performance at a fraction of the cost, the aftermath is clear: every major Chinese AI lab has adopted the playbook. According to a RAND report cited this week, Chinese AI models now cost roughly one-sixth to one-fourth of comparable American systems. The price advantage is not a temporary promotional strategy. It reflects fundamental differences in development efficiency, labor costs, and willingness to accept lower margins.

The Spring Festival timing is not coincidental. Chinese companies timed their model releases to coincide with the highest-attention period in the Chinese calendar, maximizing visibility and user acquisition. DeepSeek is preparing its next-generation V4 model. Alibaba is readying Qwen 3.5 with improved math reasoning and coding performance. ByteDance is upgrading Doubao. And behind the major players, companies like Zhipu AI, Moonshot, and MiniMax are releasing models that would have been competitive with Western frontier systems just 18 months ago.

The competitive dynamic has shifted from "can China match Western AI" to "can Western companies justify premium pricing when Chinese alternatives are one- fourth the cost." For enterprise customers evaluating AI infrastructure, the cost differential is becoming impossible to ignore, particularly for applications where the marginal capability difference between a $100 and a $22 model does not justify the premium.

Why it matters: The DeepSeek price war has permanently altered AI market economics. The question is no longer whether Chinese models can compete on capability but whether Western models can justify four to six times higher costs for incrementally better performance.

---

8. By the Numbers: This Week's Key Metrics

---

Analysis: The Multi-Polar AI World

The defining characteristic of this week is not any single announcement but the emergence of a genuinely multi-polar AI landscape. For the past three years, the narrative has been relatively simple: US companies lead, Chinese companies chase, and everyone else watches. That narrative is now obsolete.

Consider what happened in a single week. A US company (Perplexity) launched multi-model consensus using models from three different providers. A US company (Google) pushed reasoning capabilities past what human researchers could achieve. Chinese companies (Alibaba, ByteDance, Kuaishou) released models that set new benchmarks in robotics, video generation, and creator tools. An international coalition of 100 experts published a safety assessment that no single country could have produced alone. And a Canadian institution (Vector Institute) prepared a conference that will bring all of these threads together.

This is not a US-versus-China story anymore. It is a story about the emergence of specialized excellence across geographies, modalities, and applications. The US leads in reasoning and multi-model architectures. China leads in open-source model economics and consumer AI adoption. Europe leads in regulatory frameworks and safety science. Each node in this network has strengths the others lack.

The implications for organizations trying to build AI strategies are profound. Betting on a single model, a single provider, or a single geography is increasingly risky. Perplexity's Model Council is a response to this reality at the model level. Organizations need equivalent strategies at the vendor, infrastructure, and policy levels.

The multi-polar dynamic also changes the safety calculus. The International AI Safety Report found no unified approach to risk management across even the twelve companies that published safety frameworks. In a multi-polar world where models from different countries, companies, and architectures are combined in systems like Model Council, the safety question becomes exponentially more complex. Whose safety framework governs a response synthesized from Claude, GPT-5.2, and Gemini 3.0? The answer, right now, is nobody's.

---

The Week Ahead: February 16-22, 2026

Vector Institute Remarkable Conference (Feb 19-20)

The Vector Institute's third annual Remarkable conference in Toronto brings together AI researchers, industry leaders, and policymakers. Given the density of this week's developments, the conference will likely serve as the first major forum where the implications of multi-model AI, Chinese model advances, and the AI Safety Report are debated in a structured academic setting. Watch for announcements about Canadian AI research priorities and potential international collaboration frameworks.

Chinese Model Releases Continue

The Spring Festival model release window extends through the end of February. DeepSeek's V4 model is expected to drop during this period, along with Alibaba's Qwen 3.5 series and ByteDance's Doubao upgrades. Each release will pressure Western pricing and capability benchmarks. The question for Western companies is not whether to respond but how quickly they can adjust their strategies.

Gemini 3 Deep Think API Access

Google's early access program for Deep Think through the Gemini API is expected to begin onboarding enterprise and research users this week. The initial use cases will likely focus on pharmaceutical research, materials science, and mathematical problem-solving. Watch for early reports from research teams about whether Deep Think's benchmark performance translates to real-world discovery acceleration.

Seedance 2.0 Copyright Battles

The Motion Picture Association's demand that ByteDance cease Seedance 2.0 operations will likely escalate this week. Expect legal filings, potential geographic access restrictions, and broader industry debate about whether AI video generation constitutes transformative fair use or systematic copyright infringement. The outcome will set precedent for every AI video model, not just Seedance.

Perplexity Model Council Expansion

Perplexity is expected to announce expansion plans for Model Council, potentially including additional models, mobile support, and Pro tier availability. The key metric to watch is whether multi-model consensus drives measurably higher user retention and subscription conversion compared to single-model queries.

Beijing Regulatory Follow-Up

Following the regulator's summons of tech companies over the Lunar New Year giveaway war, expect specific guidance or restrictions on AI subsidy spending. Beijing's response will signal whether the government views the current competitive dynamic as healthy market development or destructive involution that needs intervention.

---

Final Thought

There is a moment in every technology wave where the question shifts from "will this work?" to "how do we manage what it has become?" This was the week AI crossed that threshold on a global scale. Seven hundred fifty million people use Gemini. Seven hundred million use AI weekly across all platforms. Seven hundred million dollars were spent in a single week to acquire more users in China alone.

Perplexity's Model Council is the intellectual answer to the complexity this creates: when you cannot trust any single AI system, build a system of systems that makes disagreement visible. The International AI Safety Report is the institutional answer: when no single country can govern AI, build a coalition of 100 experts from 30 nations to at least describe the problem.

But the operational answer -- how organizations, governments, and individuals actually navigate a world where multiple AI systems from multiple countries with multiple safety frameworks are simultaneously available, increasingly capable, and radically inexpensive -- that answer does not exist yet. This week made it clear that we need it soon.

The Lunar New Year AI War is not just a marketing spectacle. It is a signal that AI adoption in the world's most populous country is being accelerated by hundreds of millions of dollars in direct subsidies. The Chinese model blitz is not just a technology story. It is an economics story about whether premium pricing can survive when comparable capabilities are available at one-sixth the cost. And the safety report is not just an academic exercise. It is a warning that the 700 million weekly users of AI systems are running ahead of every governance framework that exists.

Next week's Vector Institute conference, the continued Spring Festival model releases, and the Seedance copyright battles will test whether the industry can keep pace with its own momentum. History suggests that the answer is usually no. But this industry has a way of proving history wrong -- sometimes for better, sometimes for worse.

---

Related Coverage

For deeper analysis of how China's AI ecosystem evolved to this point, see our coverage of DeepSeek R1 disrupting the reasoning AI market.

Our analysis of Big Tech spending $650 billion on AI infrastructure provides context for the compute economics driving both Western and Chinese model development.

For background on the enterprise AI adoption dynamics that these consumer milestones reflect, see our coverage of the AI pragmatism shift in 2026.