Los desarrollos en inteligencia artificial han sido rápidos y furiosos en las últimas semanas. Uno de los últimos, el “operador” de OpenAI, ha aparecido en los titulares de viajes con el apoyo de grandes socios de la industria, incluidos TripAdvisor, Uber, Hipcamp, Priceline y Booking.com.

La herramienta, que actualmente se encuentra en una vista previa de investigación y solo está disponible para los suscriptores profesionales, plantea preguntas sobre el futuro del marketing de viajes digitales. Después de todo, si la búsqueda se interrumpe, ¿eso también no cambiará las estrategias y el gasto para los vendedores de viajes?

“¿Todos entienden que la búsqueda de Google está fundamentalmente muerta en los próximos siete años? Que todos van a la IA? dijo Gary Vaynerchuck, presidente de Vaynerx y CEO de Vaynermedia, en un video publicado en LinkedIn.

Las marcas aliados con “operador”

Varios gigantes de viajes ya se están inclinando al éxito potencial del operador. Alinearse con una empresa que regularmente está en los titulares de su tecnología de vanguardia seguramente aumentará el perfil de una marca. Pero las empresas que trabajan con OpenAI para desarrollar el operador creen que hay un juego comercial a más largo plazo, aunque no está definido exactamente lo que podría verse. A corto plazo, mientras monitorean de cerca a la audiencia de OpenAI de unos 300 millones de usuarios semanales, las marcas están viendo la asociación como una forma de conocer a los viajeros donde están.

“Podría decir que la oportunidad de negocio va de la mano con las preferencias en evolución de los viajeros”, dijo Rahul Todkar, jefe de datos e IA en TripAdvisor. “Y así, en ese sentido, lo que diría es que siempre estamos buscando más fuentes de adquisición de tráfico”. TripAdvisor ha estado trabajando con Openai desde principios de 2023, con el primer producto un generador de itinerario que salió en julio de ese año. La compañía también se asoció recientemente con la perplejidad del motor AI-Answer para mejorar la planificación del viaje.

Priceline, que también se ha asociado con Openai antes para mejorar su propio chatbot de IA, también está trabajando con la compañía en el operador. Pero el director de productos Kevin Heery dijo que si bien sería especulativo en esta etapa decir que el mundo se está moviendo a una conexión de agente a agente, está prestando atención a las predicciones de la mayor adopción de agentes de IA en el comercio electrónico impulsado en parte por el crecimiento. interés del consumidor.

“Históricamente hemos tenido como objetivo anticipar las formas en que los consumidores querrán interactuar con nosotros y luego entregar esa experiencia, ya sea la web, las aplicaciones, el comercio social o las tecnologías de IA agente como el operador”, dijo Heery.

No está claro qué significará las asociaciones del operador de OpenAI desde una perspectiva financiera. Sin embargo, en algún momento, Todkar dijo que espera que evolucione un elemento de monetización.

“Si eso significa que nos convertimos en su fuente de referencia para esta información sin gastar necesariamente ningún medio de búsqueda o dólares de búsqueda en contra de eso, eso se vuelve bastante atractivo”, dijo Todkar, y agregó que es demasiado pronto para decir qué sucederá. “Si el tráfico que se presenta a través de esta característica de producto en particular es una intención más alta, entonces es probable que eso pueda convertirse mejor”.

El futuro de los sitios web

Internet, y los sitios web, la experiencia del usuario, etc., han sido construidas en gran medida para su uso por humanos.

Pero la interacción de agente a-website o agente a agente, como lo que el operador puede iniciar, podría requerir cambios importantes para los sitios web en el futuro.

Marina Petrova, CEO de Intentful, dijo que los sitios web deberán actualizarse gracias a los avances tecnológicos que permiten la interacción de agente a agente.

El contenido deberá centrarse en algo más que solo comercializar mensajes para priorizar información clara que los agentes autónomos puedan usar para realizar tareas, dijo Marina Petrova, CEO de Intentful, una compañía generativa de soluciones de IA.

“En SEO, el objetivo era clasificar lo más alto posible”, dijo. “En el proceso de decisión impulsado por el agente, que comienza con la búsqueda, hemos visto repetidamente en los primeros experimentos que el agente no solo va con el primer resultado, incluso si es el sitio web oficial de marca/negocio”.

La precisión se volverá primordial con los agentes que toman decisiones basadas en información veraz percibida en los sitios web. Pero Todkar no ve que la forma actual de sitios web desaparece.

“El papel de un sitio web se convierte en una fuente realmente auténtica y confiable de datos, ideas y recomendaciones”, dijo Todkar. “Eso no va a desaparecer, porque … tiene que haber una fuente de la que estos agentes consumirán información y eso se convierte en la próxima evolución de [the] sitio web.”

Heery dijo que si bien la reputación, la relevancia y el contenido de un sitio web seguirán siendo importantes, también puede convertirse en una prioridad para que los sitios web reconsideren cómo estructuran sus datos para el consumo de agente.

Es probable que los sitios web tengan cuatro audiencias objetivo en el futuro, dijo Petrova: los humanos, sus agentes de inteligencia artificial, asistentes de IA del sitio web y un SEO o rastreador de optimización de motores generativos.

Y Mario Gavira, vicepresidente de Growth and Brand en Kiwi.com, quien recientemente escribió sobre por qué cree que los agentes de IA podrían sobrealimentar las agencias de viajes en línea, dijeron que los sitios web podrían volverse menos enfocados transaccionalmente.

“Si en el futuro, los sitios web se convierten en el punto final para el tráfico humano y agente, actual [user experience] Es posible que las prácticas necesiten ser revisadas para encontrar un equilibrio entre optimizar el flujo para dos casos de uso muy diferentes ”, dijo.

Los intercambios de agente a agente serán impulsados ​​por las conexiones API, que serán más escalables y más confiables que las interacciones tradicionales del sitio web, dijo Gavira.

¿Cómo podría cambiar el marketing?

Las estrategias de marketing tendrán que mantenerse flexibles. De la misma manera, las marcas tienen que considerar los cambios con el aumento del marketing de influencia y los cambios posteriores como una posible prohibición de Tiktok, deberán considerar la IA agente en el futuro. Es probable que el gasto también evolucione.

“Los fundamentos no cambian, así es como activamos los que están evolucionando”, dijo Petrova. “Muchas empresas y organizaciones a nivel mundial se han basado en gran medida en el marketing de rendimiento debido a la naturaleza inmediata de la” validación “. Pero esta no es una estrategia saludable a largo plazo. … Debido a estos nuevos dos segmentos de audiencia [such as AI agents or AI assistants]toda la estrategia de marketing requiere una actualización holística, no solo un sitio web “.

Petrova cree que habrá un mayor enfoque en la construcción de marcas, incluida la limpieza de canales de medios propios para hacer que las marcas sean más descubiertas.

Heery acordó que la marca podría convertirse aún más en una prioridad dada que un agente tiene menos probabilidades de hacer clic en un anuncio, además de que los clientes pueden estar instruyendo a los agentes que usen sus sitios web preferidos.

Y Gavira dijo que cree que los dólares de marketing pueden cambiar de costo por clic a una búsqueda costo por agénica.

¿Qué tan rápido cambiará el panorama de búsqueda y marketing?

Hay mucho ruido alrededor del operador. Pero lo rápido que se implementará está en el aire, y solo el tiempo dirá si los consumidores querrán entregar el proceso de reserva completo a un agente de IA.

A gran escala, VaynerChuck cree que una transición a la búsqueda orientada a AI llevará un tiempo.

“Pero que no haya confusión: esto es como cuando vi a Web Crawler en 1994, estaba claro que sería un problema para las páginas amarillas, las páginas amarillas comenzaron a sentir realmente los efectos de esto en 2001/2/2/2/2/2/2/2/2/2/ 3, así que casi 10 años después ”, dijo Vaynerchuck en el video. “Las cosas se mueven más rápido ahora y hay más adopción de nuevas herramientas, aplicaciones y tecnología en general”.

Petrova cree que la adopción viene rápidamente. Y ella cree que ahora es el momento para que las marcas comiencen a posicionarse para un gran cambio.

“Los primeros motores tendrán una ventaja significativa en la navegación de este cambio de manera efectiva”, dijo. “Creo que estamos viendo 12-24 meses para una adopción más amplia de agentes, pero todas las cosas de búsqueda generativa ya están en vivo y la proporción del descubrimiento de IA está creciendo”.

Gavira no está convencido.

“Incluso si las capacidades de IA están progresando a un ritmo exponencial, los humanos estamos comenzando a aprender qué hacer con él”, dijo. “Google, con sus productos profundamente entretejidos en nuestra vida cotidiana, podría tener un impacto mucho mayor con ‘marine’ o cualquiera que sea el nombre final de su agente de IA”.

Sin embargo,, como con todo, el interés humano es importante, según Heery.

“El comportamiento y las preferencias del consumidor dictarán el ritmo”, dijo.

Hi, this is Tony! Happy Chinese New Year! Welcome to this issue of Recode China AI (for the week of January 27, 2025), your go-to newsletter for the latest AI news and research in China.

This week, DeepSeek came under scrutiny for potentially using generated outputs from OpenAI’s platforms to train its models, which violates OpenAI’s terms of service. This practice, known as Knowledge Distillation (KD), refers to transferring advanced capabilities from leading large language models (LLMs) to relatively less powerful ones.

In this post, I will break down what KD is, share my thoughts on whether DeepSeek distilled GPT models, and, if so, explore what this means.

Meanwhile, on January 29 (Beijing Time), the first day of the Chinese New Year, e-commerce giant Alibaba unveiled its latest foundation model, Qwen2.5-Max, which claims surpassing DeepSeek-V3 in multiple benchmarks. Are we entering an age where multiple Chinese AI labs can churn out frontier models?

What’s new: Speculation about KD in DeepSeek’s models first surfaced in December 2024 when users noticed that the DeepSeek chatbot, powered by DeepSeek-V3, sometimes identified itself as ChatGPT. Such discussions continued to spread after DeepSeek-R1 was released in January and became available on the DeepSeek chatbot.

This week, amid the DeepSeek frenzy, Bloomberg reported that Microsoft security researchers discovered individuals suspected to be linked to DeepSeek extracting large amounts of data using OpenAI’s API in the fall. Microsoft notified OpenAI, which subsequently blocked their access.

Later, the Financial Times reported that OpenAI had found evidence suggesting that DeepSeek may have used GPT-generated outputs for model training. OpenAI’s public statement echoed concerns raised by David Sacks, an AI and crypto advisor in the Trump administration, who stated that “it is possible” intellectual property theft had occurred.

What is Knowledge Distillation? KD is a technique where smaller or less advanced AI models learn from more powerful proprietary models like GPT-4 or Gemini. This method is akin to an experienced teacher guiding a student.

According to the paper A Survey on Knowledge Distillation of Large Language Models, KD serves three main functions:

• Enhancing model capabilities: Smaller models improve their contextual understanding, task specialization, and alignment with human intent by learning from more advanced models.

• Compression of LLMs: KD reduces the model size while maintaining performance to make them more efficient for low-latency deployment.

• Self-improvement: Open-source models can iteratively refine themselves using their own distilled knowledge.

Traditional KD techniques rely on logits – the raw, unnormalized output scores (often transformed into a soft probability distribution) from the teacher model – to train the student model. However, this approach requires the teacher model to be white-boxed, meaning its internal architecture, parameters, and outputs must be fully accessible.

Since advanced LLMs like GPT-4 and Gemini are proprietary, black-boxed models that do not provide access to logits, alternative distillation methods have been developed. One common method is supervised fine-tuning (SFT), where the student model is trained directly on the generated outputs of the teacher model.

Distilling knowledge from proprietary models into open-source LLMs is a common practice. For instance, Stanford University’s Alpaca LLM was a 7B LlaMA model trained on a 52K-example dataset, which was generated by prompting GPT-3.5 to produce high-quality, instruction-following responses.

Beyond simple answer generation, Chain-of-Thought (CoT), a step-by-step thought process, can also be distilled. Microsoft’s Orca and Orca 2 leveraged GPT-4 to generate multi-step reasoning chains for complex questions, and then fine-tuned smaller models to mimic this thought process.

Another method is using GPT-4 as a judge to score model responses. The student model is then trained to prefer high-scoring responses to improve their alignment and response quality.

A complementary technique closely intertwined with KD is Data Augmentation (DA), which is a process of generating synthetic training data from a small amount of existing data to improve model performance.

With this understanding of KD, let’s examine what DeepSeek’s papers reveal.

What did DeepSeek papers say? The DeepSeek-R1 paper provided detailed breakdown of its training methodologies. The DeepSeek-R1-Zero was trained exclusively through RL without fine-tuning on datasets, meaning it did not rely on ground-truth data to guide its outputs. In contrast, DeepSeek-R1 was first trained using a cold-start dataset derived from DeepSeek-R1-Zero’s outputs, refined by human annotators, before training through RL.

Another LLM, DeepSeek-V3, was trained on 14.8 trillion tokens. Given the widespread presence of AI-generated content on the internet, it is likely that the dataset contained GPT-generated texts. After pre-training, the model was fine-tuned using 1.5 million examples across multiple domains. For reasoning tasks, training data was collected using an internal DeepSeek-R1 model built on DeepSeek-V2.5, while non-reasoning data was curated using DeepSeek-V2.5 directly.

While DeepSeek did not mention any inclusion of GPT-generated outputs for training, a recent paper, Distillation Quantification for Large Language Models, suggests otherwise. It found that DeepSeek-V3 and Qwen-Max-0919 demonstrated higher distillation levels, aligning closely with GPT-4o, whereas models like Claude-3.5 Sonnet and ByteDance’s Doubao exhibited lower distillation levels. The study looked at how models handle identity-related information, i.e. what model are you, and whether they do so consistently. It also measured how similar the outputs of models were compared to GPT-4o.

If Bloomberg’s report is accurate – given that the data pulling was discovered in the fall of 2024 – individuals linked to DeepSeek may have extracted outputs from OpenAI’s o1-preview API, which was released on September 12, 2024. However, o1 intentionally concealed its CoT process. At best, DeepSeek could have accessed final model outputs, but not the underlying thought process.

My guess is that this extracted data may have been used to develop an early, inside iteration of R1, by fine-tuning DeepSeek’s older base models, such as DeepSeek-V2.5. It later played a role in training DeepSeek-V3, the base model for the official R1. While DeepSeek successfully improved its model’s reasoning ability through pure RL, it seems unlikely that its initial experiments on reasoning models were conducted without high-quality, reference data.

Another possibility is that a certain amount of GPT-generated outputs were included – intentionally or unintentionally – in DeepSeek-V3’s 14.8T pre-training dataset or its 1.5 million SFT examples.

Please note this remains speculative and lacks definitive evidence.

Why it matters: As mentioned earlier, open-source LLMs distilling from proprietary models is common practice, particularly for startups and university labs with limited budgets for data collection and cleaning. In the past, OpenAI rarely raised complaints about such activities. But DeepSeek is an exception – it presents strong competition to OpenAI in both consumer and enterprise markets.

DeepSeek-V3 and DeepSeek-R1 have demonstrated performance comparable to Anthropic’s Claude-3.5 Sonnet and OpenAI’s o1 models, respectively, but at a fraction of the training and inference costs. As of this writing, DeepSeek remains the most popular free app on the iOS App Store. U.S. enterprises are rushing to integrate DeepSeek into their applications. In response, OpenAI released o3-mini on Friday, a lightweight version of its most powerful reasoning model, o3.

OpenAI has clear legal grounds to warn DeepSeek, as its terms of service explicitly prohibit using OpenAI’s model outputs to train other AI models. Some companies have distanced themselves from the practice to avoid any potential legal consequences. For example, ByteDance emphasized in its Doubao 1.5 release that “in all model training processes, we did not use any data generated by other models, ensuring the independence and reliability of our data sources.”

But the ethical debate surrounding KD is complex for open-source LLMs, which are widely credited with driving AI innovation. OpenAI’s claims against DeepSeek have sparked backlash, with critics questioning why OpenAI itself is allowed to train on unauthorized web data, while simultaneously raising concerns about DeepSeek’s use of generated outputs.

The distillation controversy doesn’t seem to slow down DeepSeek’s strong industry adoption. Microsoft, Dell, Nvidia, and Amazon recently announced support for DeepSeek models, allowing their enterprise customers to deploy and fine-tune DeepSeek R1. Citing a question from Nikkei Asia’s Yifan Yu:

Could anyone please explain why Microsoft would put DeepSeek R1 in its Azure AI Foundry for enterprise customers if they believed the AI model was involved in IP violations or other unethical behavior?

From a tech standpoint, there are indeed risks associated with KD, known as distillation tax. Over-reliance on proprietary model outputs can lead to data homogenization, thus reducing response diversity. If a model depends too heavily on KD, it is unlikely to surpass the teacher model. In DeepSeek’s case, even if some GPT/o1-generated outputs were included in its training data, they are not the sole reason for its strong model performance.

As DeepSeek aims to pursue AGI and become a leading AI lab, stricter adherence to data ethics will be crucial.

What’s new: On January 29, the first day of Chinese New Year, Alibaba rushed to unveil its latest and most powerful LLM, Qwen2.5-Max, positioning it as a direct competitor to DeepSeek-V3 and other leading LLMs.

The model is available via Alibaba Cloud’s API and Qwen Chat but is not open-sourced.

How it works: Qwen2.5-Max adopts a Mixture-of-Experts (MoE) architecture, which aligns with the broader trend among China’s top LLMs, such as DeepSeek-V3 and MiniMax-o1. Alibaba has two other MoE models, Qwen2.5-Turbo and Qwen2.5-Plus, which perform competitively against GPT-4o-mini and GPT-4o, respectively.

The model was pretrained on 20 trillion tokens of diverse data—surpassing the 18 trillion tokens used for Qwen2.5. However, despite its massive training, Qwen2.5-Max supports only a 32K-token context window.

Post-training involved curated SFT and a multistage RL approach, combining offline Direct Preference Optimization (DPO) and online Gradient-based Reinforcement Preference Optimization (GRPO).

As a result, Qwen2.5-Max achieved a 89.4% accuracy on Arena-Hard, surpassing GPT-4o and Claude 3.5 Sonnet, but lagged behind both models on MMLU-Pro. It also outperformed DeepSeek V3 in benchmarks such as LiveBench, LiveCodeBench, and GPQA-Diamond.

For API pricing, Qwen2.5-Max costs $1.60 per million input tokens and $6.40 per million output tokens — cheaper than GPT-4o and Claude 3.5 Sonnet, but still more expensive than DeepSeek-V3, which charges just $0.27 per million input tokens and $1.10 per million output tokens.

Why it matters: Over the past year, Alibaba’s Qwen series and DeepSeek have emerged as the two leading LLM brands in China, driven by rapid iteration and an aggressive open-source strategy. The Qwen series has been adopted among Chinese developers due to its various model sizes catering to diverse use cases.

However, DeepSeek’s meteoric rise has clearly put pressure on Alibaba, forcing it to accelerate its own releases. Two days before the unveil of Qwen2.5-Max, Alibaba open-sourced its latest multimodal model, Qwen2.5-VL, which can understand videos and control PCs and smartphones. The release includes both base and instruction-tuned models in three sizes: 3B, 7B, and 72B. Qwen2.5-VL-72B outperforms OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash in document analysis, video understanding, and agent tasks.

Let’s be honest: dropping LLMs on the first day of Chinese New Year is a crime against naps and dumplings. Still, kudos to the Qwen team — they pulled off an impressive launch under pressure.

Looking ahead, as progress on next-gen frontier models like GPT-5 is slowing further than anticipated, it’s increasingly likely that Chinese AI labs will continue to narrow the performance gap with U.S. leaders — while maintaining a significantly lower price point.

Oscars are going through a ton of drama, trade of Luka Dončić and Anthony Davis trade has everyone in the NBA shocked and….

Before you go further, Attention is like money–you can waste it or invest it. If you invest it in my paid newsletter, you’ll get dividends for years.

Join Paid Newsletter

• Google opens up Meridian Marketing Mix Model, helping marketers measure across channels. (AR)

• Meta AI will use its ‘memory’ to provide better content recommendations on IG & FB. (CT)

• OpenAI releases o3-mini, a free and their first small reasoning model. (read)

• LinkedIn lawsuit over use of customer data for AI models is dismissed. (NB)

• Senators revive bill to ban kids under 13 from social media. (read)

• DeepSeek App holds top global spot in downloads, led by India.

• Pinterests shares a guide to the top Pinterest marketing moments for 2025.

• F.C.C. Chair orders investigation into NPR and PBS sponsorships. (read)

• Billionaire Ambani’s Reliance brings Shein back to India after 2020 app ban. (AK)

• Majority of US teens have lost trust in Big Tech. (read)

• Nielsen pushes new Big Data + Panel TV ratings product in the US, flags end of panel-only ratings. (read)

• Rhode to Retail – The Brand’s entry to Sephora delayed by partnership politics and more. (read)

• Amazon launches new tool for AI-based TV audience building. (SP)

• Is internet-induced brain rot a myth? (read)

• Google at the big game: YouTube AdBlitz, 50 small business stories and more.

• Your next apartment rental may also be a marketing campaign.

In partnership with Supermetrics

We’ve all been there: the whirlwind of seasonal campaigns leaves you with a mountain of disconnected data and not enough time to make sense of it all. Think in-store sales reports sitting in your POS, loyalty program data stuck in your CRM, and online ad metrics scattered across dashboards. The result?

• Attribution headaches: Was that sales spike driven by Facebook ads or your holiday email campaign?

• Missed opportunities: Without a unified view, real-time optimization is almost impossible.

• Budget black holes: How do you know where your dollars worked hardest?

One retailer I spoke to struggled to connect offline and online data during their Christmas campaign. In-store purchases couldn’t be linked back to their digital ads, leaving them in the dark about what truly drove sales.

The fix? A more integrated approach to their reporting. By consolidating their offline and online data streams, they gained clarity on their customer journey, optimized their campaign mid-season, and more.

And yes, tools that let you blend structured data from different sources can help here. Something like Custom Data Import might be worth exploring for those tackling similar challenges—it’s all about bridging gaps, not just filing reports.

• Creative Directors: Glenn Martens goes to Maison Margiela and Kim Jones steps down at Dior.

• JBL is going mega-viral on TikTok, thanks to Fetty Wap’s Again, viral track and memes. (TR)

• Paramount brings live shoppable ads to the NFL playoffs. (CL)

• WhatsApp says journalists and civil society members were targets of Israeli spyware. (AK)

• Bluesky adds a video tab to user profiles. (AX)

  • The platform has reached 30 million users.

  • Bluesky’s science takeover: 70% of Nature poll respondents use platform.

• OpenAI, Microsoft, Trump admin claim DeepSeek trained AI off stolen data. (TG)

• UPS shares tank 14% after weak guidance, plan to slash Amazon deliveries by more than half. (NS)

• White House says it is opening briefing room to “Podcasters, Social Media Influencers and Content Creators”. (DK)

• M&S targets family shoppers in renewed value push.

• Ipsos acquires Melbourne’s Where*to* Research to bolster public sector expertise. (read)

• Inside a network of AI-generated newsletters targeting “small town America”.

• The World’s Most Admired Companies: How They Stack Up. (List)

• Amazon secretly tracked consumers through their cellphones, lawsuit alleges.

• Co-op preloads app on Samsung O2 Telefónica phones. (view)

• AI research team claims to reproduce DeepSeek core technologies for $30 — relatively small R1-Zero model has remarkable problem-solving abilities. (TR)

• Louis Vuitton Is revved up about its Formula 1 Partnership. (Read)

• Dentsu and Adobe speed up client AI content production with Adobe GenStudio Dentsu+. (RT)

• LinkedIn’s ongoing efforts to woo video creators paid off, according to new figures. (TF)

• Fragmentation comes to search advertising as marketers grapple with shifting search behavior. (read)

• NFL doubles down on creator spotlights with a Kai Cenat vs IShowSpeed Super Bowl pre-game. (WD)

• McDonald’s teams up with Basketball Champion Angel Reese for first-ever female athlete meal. (DX)

• Crackdown on tariff exemption snares U.S. E-Commerce retailers. (WT)

• Sky News to overhaul newsroom around paid-for content.

• KFC is taking “open 24/7” to a new level by removing the doors from its restaurants. (AP)

• American public vs Politics: 3 in 5 (60%) U.S. adults said that DEI initiatives are either “very” or “somewhat” important to the success of most businesses, and this share was even higher (67%) among those who are currently employed full-time. (TC)

• System 1 and ITV name UK’s best ads of 2024. (watch)

• McDonald’s U.S. Marketing Chief steps down. (read)

• Copyright law doesn’t cover art made solely by AI, according to new decision.

• Macy’s ends program that offered tuition-free college degrees to employees.

• How hopelesscore became even more hopeless. (Read)

• MrBeast and the CEO of Roblox are trying to buy TikTok for over $20 Billion.

• TikTok partners with EE BAFTA Film Awards 2025.

• Trump says Microsoft eyeing TikTok bid to keep app in the US. (AR)

• TikTok launches Behind the Breakthrough content campaign. (read)

• TikTok tells advertisers they are not breaking the law by running ads. (DW)

• Pence says Trump should uphold TikTok Ban, allow US Steel Deal.

• TikTok testing a new IG-like buttons layout in Edits section.

• TikTok launches 2025 Marketing Calendar.

• TikTok to invest $3.8 bln in Thailand data hosting project, investment board says.

• TikTok limits access to its beta app.

• TikTok introduced new resources in recognition of Holocaust Remembrance Day. (read)

• How TikTok is preparing for the 2025 Federal Election in Germany.

• TikTok celebrates Data Privacy and Protection Day.

• TikTok’s traffic bounces back despite being pulled off app stores, fears of shutdown. (read)

• TikTok Studio app gets Inspiration and trending tabs added to app’s homepage. (DT)

• TikTok joins Tech Firms establishing a presence in Zurich.

• TikTok star Khaby Lame appointed UNICEF Goodwill Ambassador. (read)

• Perplexity AI revises Tiktok merger proposal that could give the U.S. government a 50% stake.

• Instagram updates insights with ‘View rate’ for reels and other old but official releases.

• Threads adds another 20M monthly users since December, reaching 320M.

• Instagram improves translation feature in the DMs. (AM)

• Instagram adds Transparency Labels to Business Chats.

• Threads adds a ‘media’ tab and the ability to tag people in photos. (TF)

• Instagram is finally testing a Reels pause feature that works easily. (AT)

• Threads continues to improve search function.

• Instagram shares notes on Stories Ranking.

• Threads experimenting with spoiler tags and post templates. (ST)

• Twitch adds built-in exporting to Instagram.

• Leaked Instagram deals reveal Meta is offering TikTok creators as much as $300,000 to post. Read the contract terms.

• Instagram publishes new carousel about post dimensions.

• Meta’s Threads is getting a feature you’d have thought it already had. (Tagged)

• Mark updates audience targeting options for Advantage+ catalog ads. (FB)

• Meta to pay $25m to settle Trump lawsuit over ban. (CT)

• Meta’s Safety Advisory Council says It’s prioritizing Politics over Safety.

• Meta considers testing Chinese AI Models in its Generative AI tools for advertisers.

• Meta’s pullback on fact-checking puts brand safety back in the spotlight.

• Reflecting on Meta’s $8 billion investment in Privacy.

• Facebook testing Partnership ads, your partners can run ads, add your testimonials and access the audience.

• Meta in talks to reincorporate in Texas or another state, exit Delaware, WSJ reports. (read)

• ICYMI: Meta posts sharply higher Q4 profit, revenue, topping Wall Street’s expectations.

• Meta blocked news in Canada. Ads for scams are taking its place.

• WhatsApp testing ability to create and share events in individual chats.

• WhatsApp for iPhone warning: Don’t use view once feature without updating. (TC)

• Meta sues EU antitrust watchdog over $831.7 Million marketplace fine, Bloomberg says.

• Mark Zuckerberg teases a 2025 return to ‘OG Facebook’.

• Meta AI launches in Middle East, extends support for Arabic language.

• Meta says end of fact-checking hasn’t impacted ad spend.

• Meta’s Reality Labs posts $5 billion loss in fourth quarter.

• X reveals new partnership with Visa for digital payments. (DS)

• Why Didn’t Netflix check Karla Sofía Gascón’s Twitter History?

• X expands lawsuit over advertiser ‘boycott’ to include Lego, Nestlé, Pinterest, and others.

• Amazon raises its ad spending on X/Twitter, in major reversal. (ST)

• X’s Grok 3 seemingly went live for some users. (TC)

• X is working on a new feature called ‘X CHAT’. (TG)

• xAI is working on Grok Pro.

• YouTube’s new Discord-like Communities are rolling out to more creators. (FT)

• YouTube testing cost-per-hour masthead. (SL)

• Nickelodeon takes on CoComelon with first original animated series on YouTube.

• You can now tag products in YouTube posts.

• YouTube publishes new video explaining how algorithm works in 2025.

• Snapchat now allows users to promote their snaps.

• Google’s new ‘Ask For Me‘ AI tool calls businesses to get your questions answered.

• Google offers tagging support appointments for advertisers. (SJ)

• Google updates Demand Gen campaigns with additional placement options and new experiences.

• Google expands AI overviews in AI circle to Search. (ST)

• Google to migrate all reCAPTCHA services to cloud platform. (SL)

• Google tests AI-generated backgrounds for Shopping Ads.

• Google confirms Alt text is not primarily an SEO decision.

• Google reverses stance on Performance Max campaign controls.

• Google testing new-look search results in EU. (SL)

• Google expands Site Reputation Abuse enforcement to German websites.

• Google loses bid to dismiss US states’ lawsuit over digital ads. (ST)

• OpenAI in talks to raise funding that would value AI startup at up to $340 billion. (SJ)

• Sensitive DeepSeek data exposed to web, Israeli cyber firm says. (MG)

• Google quietly announces its next flagship AI model. (TC)

• Books written without AI can now receive new ‘Human Authored’ certification. (XD)

• you.com & Perplexity AI deploys USA-hosted DeepSeek AI model. (SR)

• ChatGPT’s mobile users are 85% male, report says. (TC)

• AI startup Perplexity sued for alleged trademark infringement. (TV)

• OpenAI launches ChatGPT Gov. (NT)

• Copilot’s powerful new ‘Think Deeper‘ feature is free for all users – how it works.

• OpenAI used this subreddit to test AI persuasion. (TC)

• The US Copyright Office’s new report on AI. (read)

• Hugging Face makes it easier for devs to run AI models on third-party clouds.

• LinkedIn cofounder launches Manas AI with $24.6M for AI drug discovery. (VB)

• Block’s new open-source AI agent ‘goose’ lets you change direction mid-air.

• Anthropic’s CEO says DeepSeek shows US export rules are working. (TC)

• AI tools used for child sexual abuse images targeted in Home Office crackdown.

• Alibaba releases AI model it says surpasses DeepSeek.

• Microsoft makes OpenAI’s o1 reasoning model free for all Copilot users. (SL)

• Microsoft stock moves lower after Q2 earnings report, cloud revenue miss.

• LinkedIn removes accounts of AI ‘Co-Workers’ looking for jobs. (read)

• Microsoft Performance Max testing LinkedIn targeting, measurement tools.

• LinkedIn updates ‘Top Voice’ badge requirements.

• Microsoft tests ‘scareware blocker’ for Edge that uses computer vision to detect scams.

• LinkedIn expands Top-of-Feed News Banner Test.

• Microsoft’s AI business is booming — Xbox, not so much. (BA)

• Introducing new Surface Copilot+ PCs for Business.

• How Amazon Prime Video made itself an essential pick for brand media plans.

• Elementor rolls out WordPress AI site planner. (SJ)

• CHEQ acquires Deduce, expanding Go-To-Market security platform with identity graph for Human and AI-Generated fraud prevention. (RT)

• Walmart Canada to invest more than $4 bln to expand stores, supply chain. (BN)

• Yahoo testing new AI Search features.

• The MIT Grad whose prediction market can reveal what Netflix show will be #1.

• Marketing automation vendor Act-On acquired for $53 million. (TM)

• Brands are set to win the Super Bowl, Fox’s ad sales chief explains why. (AK)

• Deloitte acquires SimplrOps, a leading SaaS tech company. (JG)

• Rakim is helping Notes startup ensure artists get paid in full.

• USIM partners with IRIS.TV to enhance CTV campaign precision and engagement.

• Fuji TV leaders resign amid sexual misconduct case. (NR)

• EU to make Temu, Shein and Amazon liable for ‘unsafe’ goods, FT reports.

• Prophet integrates Influencer discovery feature within ‘Taylor’, its AI-powered journalist discovers and content creation product. (NG)

• Consulting IQ and Mastercard easy savings specials come together to empower micro, small and medium-sized businesses with AI-driven solutions.

• Marketing automation vendor Act-On acquired for $53 million. (MP)

• 3 Feb: Elmo’s Birthday / National Golden Retriever Day / National Carrot Cake Day / National Sickie Day / St. Blaise’s Day / American Painters Day / National Wedding Ring Day.

• 4 Feb: International Day of Human Fraternity / Transit Equity Day / World Cancer Day / Torture Abolition Day / National Sweater Day / National Hemp Day / Facebook’s Birthday / Farmer’s Day.

• 5 Feb: National Weatherperson’s Day / World Nutella Day / World Read Aloud Day /Global School Play Day / Western Monarch Day / National Signing Day / National Girls & Women in Sports Day.

• 6 Feb: Pay a Compliment Day / National Chopsticks Day / National Chopsticks Day / World May Thai Day / New York Fashion Week begins / Optimist Day / Waitangi Day.

• 7 Feb: Rose Day / National Ballet Day / National Bubble Gum Day / Number Day / National Periodic Table Day / National Woman’s Heart Day.

• 8 Feb: Opera Day / National Boy Scout Day / Global Movie Day / National Iowa Day / International Snowmobile Ride Day / Prešeren Day / UFC 312.

• 9 Feb: Super Bowl LIX / Chocolate Day / National Pizza Day / World Marriage Day / Frevo Day / Super Chicken Wing Day / National Toothache Day / Autism Sunday.

• 42% said the stress of working in New Business or on pitches had made them think of changing job. An alarming 33% have considered leaving the industry because of ‘the stress of working in New Business or on pitches’. (Adworld)

• As President Trump advances his agenda through executive orders, 74% of Americans, including 73% of Republicans, believe that President Trump should “always follow the Constitution, even if means he sometimes can’t get things done.” (read)

• Americans to consume 1.47 billion chicken wings watching the Kansas City Chiefs and Philadelphia Eagles battle for the Lombardi Trophy. (read)

• Tripadvisor’s research shares mentions of skincare treatments are up 87% in Tripadvisor forums and beauty bookings are up 400% + searches for bookings containing words like ‘astrology,’ ‘tarot’ and ‘full moon’ have shot up over 500% in the last two years. (read)

• January’s fall in U.S. consumer confidence was led by consumers under 55 years old. By income group, the sharpest decline in confidence was seen in households earning over $125K, while consumers at the bottom of the income range reported the strongest gains. (MR)

• Google AI Overviews Found In 74% Of Problem-Solving Queries. Telecommunications had the highest percentage of keywords linked to AI Overviews at 56%, while Beauty and Cosmetics had the lowest at 14%. About 33.3% of non-brand searches show an AI Overview, while only 19.6% of brand searches do. (read)

• Only 63% of shoppers said they ‘never’ failed to scan an item when using self-checkouts. A third (32.5%) also admitted to having weighed loose items incorrectly, while 38% had put through an incorrect loose item. (UK)

• Digital advertising expenditure increased by 10.8% in 2024 reaching €1,011.1 million, representing over 65.5% of total advertising revenue. In Northern Ireland, the print market fell by approximately 11% in 2024, recording a spend of £6.3 million (€7.45 million), a decline of 8.7%. (Outlook)

• One in three (36%) digital experiences are actively failing people, creating a divide between what brands promise and what they deliver online. (DX Index)

• D1A’s must-read study “Life Actually, A No Bullshit” shares 29% of their respondents fall into the bucket we’re calling Neo-Traditionalists, 33% of the respondents, has existed for as long as we’ve talked about generations. This moderate, thoughtful, slightly cautious person takes a pragmatic approach to life. 38%, takes what some might call a modern approach to life, Internet-Age explorer.

• The median wealth of older US millennials, between the ages of 36 and 45, was 37% above expectations. The wealth of younger millennials and older Gen Zers, or those aged 26 to 35, exceeded expectations by 39%.

• The individuals who maintain relationships with luxury brands for early access (69%) also insist those brands demonstrate environmental leadership (77%). And while 73% value being the first to own a new release, an equal percentage insist that their luxury benefits local communities. (Luxury Consumer Study)

• The Handmaid’s Tale for Gen Z – How BookTok made a dystopian novel from the ’90s into an indie best seller.

• Over the course of the concert dates in Toronto (Nov 14-24), spending at restaurants near the arena in Toronto increased by 31% compared to what spending levels would be if the event never occurred, while spending at hotels in Toronto rose by 45%. (Taylor Swift)

• You are what your feeds make you: A study of user aggressive behavior on Twitter

• More than 8-in-10 (81%) California voters support the legislation, and 63% want the state to expand and improve its recycling infrastructure to find ways to give a new life to plastics. That compares to only 27% who prefer the state eliminate single-use plastic and stop manufacturing new plastics.

• According to a survey conducted in December 2024 on actions to reduce waste among Generation Z in Indonesia, 50 percent of female respondents stated they brought their own food containers to reduce waste. In comparison, 42 percent of male respondents also did the same to reduce waste.

• More than 50% of respondents cite cost as the most influential factor in their purchasing decisions, while around 30% prioritize quality. Sustainability is a primary consideration for only a small fraction of travelers, ranging from 7% to 11%, even within the most environmentally conscious segments. (Travel)

• 67% of gen z say they like the idea of escaping to a different reality using technology (Reality Shift, 1), reflecting desire for agency and optimism in a chaotic world + All generations, and especially gen z, are embracing of solo travel for self-discovery (Destination Solitude, 22) and contrasting trends such as Social Saunas (82) and Agrihoods (47). (Future 100)

• Nearly half of Gen Z men (45%) believe that ‘we have gone so far in promoting women’s equality that we are discriminating against men’ and 44% think women’s equal rights have gone far enough. (Gen-Z Trends & Truth)