The New Battle for E-Commerce, New Agent Stacks, and News from Europe
Yann LeCun founds a new company in Paris as an LLM alternative, and VC investments are (slowly) rising on the old continent. Plus: Agent stacks, e-commerce, Bytedance
The Battle for AI-Driven E-Commerce
The way consumers make purchasing decisions is increasingly shaped by AI. Amazon, Google, and OpenAI are competing to dominate the moment of product discovery and evaluation, even before a product lands in the shopping cart. Instead of opening numerous tabs and reading reviews, users ask a simple question to an AI assistant. Amazon is positioning itself as a smart shopping companion, Google wants to defend its role as the primary source of information, and OpenAI acts as a neutral layer above them. For consumers, this means more convenience but also a narrowing of choices. The trade-off is between speed and discovering new options, with most likely preferring simplicity. → Future Blueprint
Synthszr Take: A tectonic shift is happening here: value creation is moving from the transaction to the recommendation. Whoever controls the AI-powered pre-selection controls the market. Amazon has the transaction data, Google has the search history and inventory, OpenAI has the conversational layer. Visibility will no longer be achieved through organic search or ads, but through the ability to be interpreted by an AI as a relevant option. Finding the balance between advertising (=manipulation) and organic LLM results will be incredibly difficult for any newcomer—Google holds the best cards here.
Humans& launches with a $480M Seed Round
The AI startup Humans&, founded by former employees of Anthropic, xAI, and Google, has closed a $480 million seed funding round at a $4.48 billion valuation. Investors include Nvidia, Jeff Bezos, and various well-known VC firms. The company's philosophy is to use AI to enhance human collaboration rather than replace people. The team consists of prominent researchers and engineers from major AI labs. Humans& aims to use existing AI techniques to train models in new ways, such as enabling chatbots to request information from users and store it for later use. The goal is an AI that serves as a 'deeper connective tissue' for organizations and communities, requiring innovations in reinforcement learning, memory, and user understanding. → Techpresso
Synthszr Take: The funding for Humans& is a perfect example of the current AI supercycle: investors are not betting on an MVP, but on the density of talent from leading labs. It's about the assumption that these teams have an unfair advantage in recruiting and accessing the next generation of architectures. The 'human-centric' positioning is clever marketing to differentiate themselves from the pure scaling and efficiency rhetoric of the competition. The bottom line is: capital follows talent, not products.
The Agent Stack is Becoming Real Software
The development of AI agents is moving away from clever prompts and towards robust software architecture. Until now, the focus has been on improving the reasoning of models. However, the real problems in practice are persistence and coherence over long periods. Agents lose track, repeat tasks, or become unstable. Current research addresses these challenges as system problems. Topics like hierarchical caching of knowledge, memory as an executive control function, and the dynamic generation of tools at runtime are taking center stage. The infrastructure for training and operating these agents will be crucial. → 🔳 Turing Post
Synthszr Take: This is the transition from Phase 1 (magic) to Phase 2 (industrialization) of generative AI. The hype around 'prompt engineering' was a cargo cult for non-technicians. The real value creation and defensive moats are being built at the system level. Agent failures will soon resemble a 'memory leak' or 'state corruption' more than a wrong answer. The most sought-after experts will not be 'prompt engineers,' but systems thinkers who understand how to orchestrate resilient, long-running AI processes.
ByteDance Challenges Alibaba in the AI Cloud Market
ByteDance, the parent company of TikTok, is taking on Alibaba in the Chinese cloud and AI market. The company has significantly expanded its enterprise cloud business, Volcano Engine, leveraging its AI expertise and computing resources. With aggressive price cuts and an enlarged sales team, ByteDance is positioning itself as a serious competitor. The strategy focuses on custom AI agents for enterprise customers, based on the same proprietary models and data infrastructure that power TikTok and Douyin. IDC estimates that ByteDance already controls nearly 13% of the Chinese AI cloud market, placing it second behind Alibaba. → Techpresso
Synthszr Take: This is a classic 'full-stack' attack. ByteDance built one of the world's most advanced AI infrastructures for its internal needs (the TikTok algorithm). Now, they are externalizing this capability and commodifying the market from the outside. They are not just selling cloud storage, but an entire AI operating system that has been battle-tested in the toughest environment imaginable—the global attention market. For Alibaba and Tencent, this is a massive threat, because ByteDance isn't selling technology; it's selling a proven recipe for engagement.
Yann LeCun on His Bet Against LLMs
In an interview, Yann LeCun explains the vision behind his new Paris-based company, Advanced Machine Intelligence (AMI). He positions it as a European alternative to the dominant US and Chinese AI firms and advocates for an open-source approach. LeCun criticizes the industry's current focus on Large Language Models (LLMs) as a dead end. He is convinced that LLMs alone cannot achieve human-like intelligence because they lack a true understanding of the real world. Instead, AMI is focusing on 'world models' and his self-developed JEPA architecture, which learns abstract representations of the world from observations (like videos) to create a foundation for planning and action. → Techmeme
Synthszr Take: LeCun is playing a different game here. While Silicon Valley is digitizing language and thus the cumulative intellectual capital of the Gutenberg galaxy and the web, LeCun in Europe wants to model the physics of the real world. This is a fundamentally contrary but potentially far more valuable bet. Agents based solely on LLMs will always be brittle because they cannot predict the consequences of their actions. An agent with an internalized world model can. LeCun is not trying to build the next ChatGPT, but the foundation for truly autonomous systems. LeCun's 'pivot' is the intellectual equivalent of an earthquake in the AI establishment and an attack on the pure scaling hypothesis in the Valley.
Adaptive RAG Systems Optimize Efficiency
Enterprise systems for Retrieval-Augmented Generation (RAG) are evolving into query-adaptive architectures. These systems dynamically classify queries based on their complexity. Simple, factual questions are answered through a fast, single-hop retrieval process, while more complex queries requiring logical reasoning are routed to a multi-hop synthesis process. Leading implementations report latency reductions of 30-40% and an accuracy increase of 8%. At the same time, the cost of LLM calls can be reduced by up to 50%. Success depends on robust complexity detection and continuous feedback loops. → TLDR Data
Synthszr Take: This is the invisible but crucial next stage of AI industrialization. Away from monolithic 'one-size-fits-all' models, towards intelligent, resource-efficient pipelines. Adaptive RAG is essentially a load balancer for cognitive load. It's the realization that not every question needs a sledgehammer, i.e., a huge model. This kind of optimization in the 'engine room' of AI is what makes enterprise applications profitable and scalable. The moat is not in the model, but in the orchestration.
AI Drives European Venture Capital
In 2025, venture capital investments in European startups rose slightly by 9% to $58 billion. For the first time, AI was the leading sector for investment with around $17.5 billion, a significant increase from $10 billion the previous year. This trend shows a shift towards deep tech, including data centers, wearables, defense, and quantum computing. Despite the growth, Europe has not experienced the same AI-driven boom as North America, where investments increased by 46%. The United Kingdom remains the leading country, although its share of the total capital slightly decreased, while countries like France, Germany, and Switzerland gained ground. → Techmeme
Synthszr Take: The numbers show that Europe is waking up, but still riding in the slipstream of the US. The investments are flowing, but the question remains: are the next platform companies being created here, or just highly-funded users of US technology? The focus on deep tech is promising, as it plays to European strengths in research and engineering. The litmus test will be whether they can succeed in creating market-dominant products and ecosystems from this scientific excellence, rather than just another generation of highly innovative but ultimately insignificant niche players.
Anthropic and OpenAI: Revenues Rise, Margins Fall
OpenAI's CFO Sarah Friar stated in a blog post that the company's annualized revenue has grown at the same rate as its consumption of computing resources. The compute demand tripled from 0.2 gigawatts in 2023 to 1.9 gigawatts in 2025, while revenue increased from $2 billion to over $20 billion in the same period. OpenAI is currently closing more deals with cloud providers and chip manufacturers to meet the growing demand. → The Information AM
Anthropic has lowered its gross margin forecast despite a sharp increase in revenue. The company recorded an annualized revenue of over $9 billion in 2025, more than double the previous year. At the same time, the expected gross margin was reduced to 40%. This adjustment reflects the high costs associated with training and operating large AI models. → Martin Peers
Synthszr Take: Elegantly phrased messages to reassure investors who are nervously watching the exploding compute costs. The statement 'revenue grows like compute' obscures the crucial question: how are the marginal costs per query developing? As long as every interaction burns money, growth is just pre-funded scaling. Currently, no scaling dividend is visible: not good news for investors. Either they find a way to make the models radically more efficient, or they have to specialize in high-margin enterprise workflows that justify these costs. The pure sale of API access is a race to zero margin.
The Consolidation of the LIDAR Market
The LIDAR market, which once had over 200 companies, is consolidating to a few remaining players. According to Omer Keilaf, CEO of Innoviz, the market is approaching an endgame with one or two dominant Western suppliers. A key driver is the increasing focus on Level 4 autonomy (robotaxis, trucks), sparked by Waymo's success. At the same time, Level 3 programs, which had long stagnated, are being accelerated again. The technical difference between the autonomy levels is crucial: while L2 systems have a driver as a fallback, L3 and especially L4 require nearly 100% availability and robustness of the sensors, even under adverse conditions like dirt. → Austin Lyons from Chipstrat
Synthszr Take: This is the classic development of a technology market following the Gartner Hype Cycle. The 'inflated expectations' phase with hundreds of suppliers is over. Now comes the 'trough of disillusionment' and consolidation, where only companies that can deliver industrial-grade, reliable technology for specific use cases survive. It's no longer about demos, but about long-term supply contracts with OEMs. The market is rationalizing, and the winners are those who not only have the best technology but also the most robust supply chain and the strongest customer relationships.
