There were plenty of model launches for today’s radar, but what stayed with me was a study about how AI chatbots talk to us. It is worth reading before looking at the rest.

A recent study gave names to a number of things we experience when using AI chatbots but had never stopped to catalogue.

The Center for Democracy & Technology mapped 37 “dark patterns” used by chatbots such as ChatGPT, Gemini, and Replika. Anthropomorphism—the tendency to strike up personal conversation—and sycophancy, agreeing with users too much to hold their attention, are among the most common.

What caught my attention was not the list itself. It was the motive behind it: keeping attention, extracting data, and maintaining engagement. The same goals as any digital product, only applied to an interface that pretends to be human.

I work with credit products, and one thing I learned quickly is that trust in a financial product is not easily recovered once it is broken. If a service, credit-simulation, or customer-support chatbot uses these patterns without realizing it, the risk is not merely reputational—it is regulatory.

The research is already being cited by regulators and is likely to appear in lawsuits against AI companies. It has stopped being an academic debate and become audit evidence.

Before thinking about how much a chatbot converts or how much session time it generates, it is worth asking whether the way it speaks is honest with the person on the other side. This is product design, not a minor concern.

For anyone who wants to read the full study, the link is here.

The rest of the radar

Kimi K3, the new open frontier model — Moonshot AI’s Chinese model rivals Claude and GPT on coding benchmarks and costs about half the price of the Opus API. It puts pressure on multi-model and vendor-lock-in decisions. Read more

NotebookLM became Gemini Notebook — Google brought the product into the Gemini ecosystem and added cloud code execution while retaining its 30M+ users. A rebranding case that keeps the base that already loves the product. Read more

Dynamic Workflows in Claude Code — Anthropic now orchestrates hundreds of subagents in parallel for audits, refactors, and large-scale migrations. AI credits become a new cost variable for users. Read more

LM Studio launched Bionic — an agent that runs open models locally, with zero data retention and coding through inline diffs. A sign of growing demand for privacy-first AI. Read more

Noisy AI evaluators can still help — TensorZero shows that imprecise LLM evaluators can still produce useful signal at scale, making continuous evaluation loops for production agents more affordable. Read more

Robinhood lets AI agents manage your portfolio — AI trading in beta, with an isolated portfolio, order approval, and an agentic card for autonomous purchases. A trust blueprint for AI that handles real money. Read more

Step 3.7 Flash, another strong cost-performance option — StepFun’s multimodal model has 256K context and up to 400 tokens per second, targeting coding agents and search without the price of top-tier models. Read more

Human-in-the-loop is tired — a Pydantic article argues that reviewing AI-generated code has broken developers’ reward loop. It is worth measuring review fatigue, not only throughput. Read more

The criticisms of LLMs are right, and I still use them — a reflection on how AI broke the natural effort filter for contributions, making it harder to assess the genuine investment of a submitter. Read more

That is what was on my radar today. Until the next edition.