Enshittification: LinkedIn's TikTok-ification of the Feed with Tokens

Enshittification (I): LinkedIn's TikTok-ification of the Feed with Tokens

LinkedIn is fundamentally redesigning its feed, replacing classic ranking logic with a unified retrieval system based on LLM-generated embeddings. On top of this sits a sequential generative recommender model with causal attention transformers that understands interactions as temporal sequences. The feed evaluates content more based on semantic proximity and professional development paths rather than demographic features. Context arises from history and meaning, not from static profile fields. This shifts recommendations from heuristic scores to generative modeling. For creators and recruiters, this means less explicit audience targeting and more implicit trajectories within the system. LinkedIn is thus approaching the logic of TikTok, but with professional signals instead of entertainment. → TLDR Data

Synthszr Take: LinkedIn is replacing ranking tables with an end-to-end model that couples retrieval and generation, thereby treating the feed as a system. Causal attention on interaction sequences brings a memory to the recommendation engine, modeling careers as a time series rather than a profile snapshot. Embeddings commodify individual signals and shift differentiation to the orchestration of the overall system, which is precisely where lock-in and network effects emerge. Microsoft benefits twice over, as infrastructure and model expertise intertwine, lowering the marginal cost per recommendation while increasing quality. Creators lose a degree of controllable targeting and optimize more for semantic resonance, while recruiters operate in longer cycles with implicit signals. Competition is shifting from the feature set to the data and feedback loop; whoever has the best sequences wins the attention. Generative recommenders turn feeds into learning systems, and that's precisely what permanently shifts power to the platforms.

Enshittification (II): Norway Takes on the Deliberate Degradation of Digital Products

The Norwegian Consumer Council, together with over 70 organizations in Europe and the US, is launching a campaign against so-called 'enshittification,' the deliberate degradation of digital services. A viral video satirically stages this practice, while an 80-page report documents the systematic erosion of user experiences. The term was coined by Cory Doctorow, with examples ranging from cluttered social feeds to software updates that slow down devices. The initiative targets policymakers in 14 countries, calling for stricter enforcement of existing laws and more competition in the digital market. Specifically, it focuses on interoperability, repairability, and the ability to switch between services more easily. More than 20 organizations in Norway, as well as stakeholders in 12 other countries, have submitted corresponding demands. The campaign has struck a chord: the video has garnered millions of views and over 9,000 comments, and the report has been downloaded more than 6,000 times. → Techpresso

Synthszr Take: With 'enshittification,' Cory Doctorow provides a precise label for a well-known platform pattern: first user growth, then monetization, finally extraction. Norway's Consumer Council is translating this pattern into concrete policy demands across 14 countries, making competition an operational lever again. The number 70 shows how broad the coalition has become; unions and human rights organizations are moving into an issue long dismissed as a product detail. The platform economy scales through lock-in, and lower switching costs are the only real counterforce—which is exactly where interoperability and portability come in. A video with millions of views and 9,000 comments signals cultural resonance, a rare feat for regulatory topics. GAFA have only responded to forced openings; voluntary quality improvements remain exceptions (see App Store fees). Regulation that lowers switching costs and overturns defaults hits the core of the problem and is long overdue.

OpenClaw: The First Model for Agentic Marathons Instead of Chat Sprints

Z.ai has introduced GLM-5-Turbo, a Large Language Model developed specifically for the OpenClaw agent framework. The focus is not on benchmarks or chat performance, but on stable, long-running agents that call tools and execute complex workflows. In practice, such OpenClaw tasks consist of 30 to 40 model calls across APIs, webhooks, and pipelines. A single faulty JSON call can crash the entire chain. This is exactly where GLM-5-Turbo comes in, optimizing for reliability over long execution times. This shifts the bottleneck from pure model intelligence to system stability. At the same time, it suggests that LLM architecture is adapting to the requirements of agentic operating layers. → AI Secret

Synthszr Take: Z.ai is shifting the focus from demo intelligence to production reality, where 40 API calls per task are the norm. OpenClaw makes it clear how fragile current LLM chains are—one broken JSON and the entire workflow collapses. GLM-5-Turbo addresses this exact class of errors, treating reliability as a primary design parameter. Agent frameworks are thus evolving into the actual platform, with models becoming interchangeable components in the operating layer. GAFA dynamics take hold: whoever controls the most stable orchestration defines the standard for tooling and integrations. Developers will do less prompting and more system monitoring, including retry logic, state handling, and failure recovery. Models that reliably endure will win the market.

Agents Are Massively Fueling the Demand for Data Centers

Ben Thompson describes a turning point in the AI economy: agents are fundamentally changing how demand for computing power arises and who triggers it. After ChatGPT (2022) and reasoning models like o1 (2024), the transition to agentic systems with tools like Claude Code (Opus 4.5) and GPT-5.2-Codex marks the third paradigm shift. These agents combine models with control software ('harness'), independently verify results, and execute complex tasks iteratively, often over several hours. This massively increases inference costs because multiple model calls, additional CPU workloads, and higher usage densities coincide. At the same time, the necessary human 'agency' decreases, as individual users can control many agents in parallel. For companies, this means lower coordination costs and higher productivity, which makes massive capex investments by hyperscalers plausible and questions the bubble thesis. → Ben Thompson

Synthszr Take: OpenAI and Anthropic are shifting value creation from the model to orchestration, as Claude Code and Codex show that the harness defines the actual product frontier. Nvidia benefits indirectly, as each agentic loop generates multiple inference cycles, thereby increasing token consumption and compute demand. Numerous agents per user subvert classic adoption logic because it doesn't require millions of active users, but rather a few power users with high control density. Enterprise economics are tilting toward smaller teams with greater leverage; coordination costs shrink faster than personnel costs (three fewer managers and one more agent are often enough). Microsoft's E7-bundle for $99 per seat shows how quickly pricing is being tied to real productivity gains once agents deliver measurable output. In this setup, Apple appears structurally weaker because pure licensing without a deeply integrated harness offers little differentiation. Agents drive demand not linearly, but in cascades through autonomous usage loops, and that's precisely why the compute boom is real.

Sycophancy: Design Becomes the Last Differentiator

A report from Citrini Research shook the SaaS world in late February, putting significant pressure on the stocks of Atlassian and Slack. The thesis: by 2028, software companies could maneuver themselves into a self-reinforcing downward spiral because AI is reducing development costs so drastically that they cannibalize their own products. In parallel, the interface paradigm is shifting; AI-first systems are inverting the logic of classic SaaS products, moving away from clear inputs toward open, probabilistic outputs. In design circles, there's also a growing concern that 'sycophancy'—models that agree too much instead of disagreeing—is becoming a central risk. At the same time, data shows that AI startups are staying smaller, growing slower, and yet raising more capital. The result is a quiet but structural shift in labor, value creation, and interface logic. Design is thus becoming less about the surface and more about a strategic lever within the system. → The UX Collective Newsletter

Synthszr Take: Citrini Research puts its finger right on the economic breaking point of SaaS: when build costs trend toward zero, differentiation through features collapses. Atlassian and Slack are reacting on the stock market not to AI hype, but to the prospect of their own output becoming a commodity. AI-first interfaces shift control from user input to system behavior, making design a matter of governance rather than aesthetics. Sycophancy is not a UX bug but a systemic risk, as agreeable models undermine decision quality (and are difficult to audit internally). Smaller teams with larger funding rounds show how much the operating model is changing: fewer people, more orchestrated systems. Product teams lose weight, while system architecture gains importance, including guardrails, feedback loops, and control mechanisms. Design is becoming the last differentiating factor in the software market.

Appliances 2.0: Robots and Humans for Daily Rental

China has built a functioning market for renting robots in less than twelve months, with drastically falling prices. A humanoid Unitree robot now costs as little as 1,796 yuan per day on JD.com, after comparable deployments cost 10,000 to 20,000 yuan in 2025; a robot dog costs 78 yuan. In parallel, the supply is exploding: over 1,500 companies entered the market in 2025, platforms like Qingtian Rent from Agibot are expanding aggressively, and more than 16,000 applicants competed for 600 local partner roles. The market is growing rapidly, with forecasts of nearly 10 billion yuan in volume in 2026 and sharply rising demand during events like the Spring Festival.

The economics behind this at first glance resemble classic Robot-as-a-Service, a kind of physical SaaS model. In practice, however, every robot comes with a human operator who handles transport, control, maintenance, and safety. This coupling inverts the scaling logic, as each additional unit requires additional labor. The bottleneck is therefore less about hardware or demand and more about qualified field personnel. The boom demonstrates less technological maturity and more the current gap between capital inflow and the actual operational readiness of embodied AI. → Hello China Tech

Synthszr Take: JD.com is slashing prices from 20,000 to 1,796 yuan, forcing the market into a classic commoditization curve, while the underlying cost structure remains surprisingly analog. 1,500 providers reads like platform dynamics, but operationally, it looks more like a fragmented service business with high personnel dependency. Qingtian Rent is scaling demand with 5,000 orders and 70 percent peak growth, but it's also scaling the need for operators at the same rate. Robot-as-a-Service thus remains a semantic shortcut for 'hardware plus human in a bundle' (including travel, calibration, battery changes). Labor becomes the real gatekeeper, not the machine, shifting value creation back to a two-speed model of capital and manual execution. IDC can declare trends, but the unit economics tell a different story with linear cost curves. China's robotics boom is showing very precisely that embodied AI is currently a service business in costume.

Anthropic Is Now Doing Tupperware-Style Marketing

Anthropic is launching a program called 'Claude Community Ambassadors,' where customers organize local meetups and the company covers all costs. Participants also receive swag, monthly API credits, early access to new features, and a direct line to the team via a private Slack channel. The idea aims to build organic community structures that go far beyond classic marketing channels. Inspiration comes from formats like 'Friends of Figma,' but it's being implemented more aggressively and scaled globally here. The mechanism is simple: a local host brings friends to an event, new users get to know Claude, and adoption arises from social context instead of ads. In parallel, the post describes a second tactic in enterprise sales: a free, ten-minute ABM approach that specifically provokes responses from target accounts. Both ideas follow the same logic: proximity replaces reach, and personal interaction becomes the more efficient acquisition channel in an environment flooded with AI-generated content. → Tom's Marketing Ideas

Synthszr Take: Anthropic is consistently externalizing its market entry to its most engaged users, buying cultural proximity instead of media reach. Claude Community Ambassadors work because a local host with API credits and insider status is more credible than any central growth team (two pizzas, one meetup, ten new users). Figma played this exact playbook early on, building a global network that is hard to replicate today. Anthropic couples this community layer directly with product access and feature previews, creating a lock-in that goes beyond mere usage. The enterprise ABM in the same article points in the same direction: a few targeted touchpoints beat broad distribution because attention is the real scarce commodity. Marketing is thus shifting from campaigns to systems of relationships, incentives, and tooling. Community-driven expansion with real product incentives is currently the most efficient way to penetrate new markets.

LLMs: Shameless Guesswork (Still) Pays Off

An essay from Astral Codex Ten argues that so-called 'hallucinations' in AI systems are misleadingly described and should be better understood as calculated guesses. The author compares this behavior to students guessing on multiple-choice questions to minimally increase their chances of a correct answer. Technically, this behavior is based on the training of LLMs, which learn through next-token prediction and are systematically rewarded for correct predictions but not directly penalized for incorrect ones. This creates an incentive system where guessing is rational, even if the probability is extremely low. Post-training measures reduce this tendency but only bring it to an 'acceptable' level. Observations also show that models activate patterns associated with deception during such incorrect answers, further fueling the debate over 'lying' versus 'guessing.' The core issue thus shifts from misbehavior to the question of alignment mechanisms. → Astral Codex Ten

Synthszr Take: LLMs follow their training's reward system exactly; billions of tokens shape a behavior that maximizes hit probability rather than guaranteeing truth. Next-token prediction creates an economic logic where even absurd answers have a positive expected value as long as zero knowledge is the alternative. Post-training acts like a downstream governance layer that dampens symptoms but leaves the original incentive structure intact. Alignment thus shifts from model architecture to system design, including guardrails, retrieval, and feedback loops. Research on 'deception features' shows that models can internally distinguish between high and low confidence but do not reliably externalize this information. Product decisions around confidence scores, abstention, and source citation thus become the real differentiator. Those who systematically control guessing build trustworthy systems; those who ignore it scale uncertainty.