Four fresh X signals: stateless MCP, model evals, and the human edge of AI
Four original posts show where AI systems still need engineering judgment: protocol state, small evals, inference settings, and human consent.
The short read
The strongest posts in the past 24 hours were less about a new leaderboard than about the machinery around AI: how tools keep state, how small evals expose differences between models, how a reasoning setting changes an output, and where an agent crosses a social line.
- Protocol: Stateless MCP reduces a tool call to one request and removes the need for server-side session state. 1
- Evaluation:
smevalspackages small tests for models, prompts, and harnesses, with separate run and grading steps. 2 - Model behavior: Simon Willison reports a sharp difference between default and high reasoning on a DeepSeek-V4-Flash-0731 pelican prompt, but gives no benchmark. 3
- Human layer: Greg Brockman says people may accept help from a coworker while resisting the same request when it arrives from that coworker's ChatGPT. 4
AI tools and developer ecosystem
Stateless MCP makes the tool call less stateful
- What happened: Simon Willison says the new Stateless MCP specification inspired two projects:
mcp-exploreranddatasette-mcp. The linked write-up describes a one-request tool call instead of the older initialize-then-session flow. 1 5 - Why it matters: Dropping server-side session state makes the protocol easier to implement and removes the need to route one session back to the same backend machine. 1
- Concrete detail:
mcp-explorerlists, inspects, and calls MCP tools;datasette-mcpexposes database listing, schema inspection, and read-only SQL execution through a Datasette endpoint. 1
Loading content card…
The practical shift is boring in the best way: fewer hidden assumptions between an agent and the service it needs to call.
smevals keeps model testing small enough to run
- What happened: Willison says he built
smevalswith Prime Radiant to run small evaluation suites against models, prompts, and harnesses. 2 6 - Why it matters: The tool treats an evaluation as a directory of YAML configuration and executable scripts, so a team can compare a model setup against a concrete task instead of relying on a single general score. 2
- Concrete detail:
uvx smevals run . -gruns and grades immediately;grade,serve, andbuildhandle regrading, browsing results, and static reports. The project is explicitly aimed at small evals, not sprawling benchmark infrastructure. 2
Loading content card…
This is the kind of tool that makes "which model should we use?" a local experiment rather than a permanent argument about leaderboards.
Model behavior
The reasoning knob changes the result, not just the wait time
- What happened: Willison reports that DeepSeek-V4-Flash-0731 produced a disappointing pelican at its default reasoning setting, then a much better one after he raised reasoning to high through OpenRouter. 3
- Why it matters: The post is a compact reminder that a model name does not fully specify the behavior a user will see; inference settings can change the artifact. 3
- Constraint: The post gives no prompt, rubric, or sample size, so it is a single-user product observation, not evidence that high reasoning wins on a benchmark. 3
Loading content card…
The useful follow-up is simple: when comparing models, hold the reasoning budget and other inference settings constant—or report them alongside the result.
Society and ethics
An agent can be useful and still feel socially wrong
- What happened: Greg Brockman says many people at OpenAI connect ChatGPT to Slack, but dislike a coworker's ChatGPT contacting them to ask for help, even when they would gladly do the same work if that coworker asked directly. 4
- Why it matters: The objection is not to the task itself; it is to losing the human relationship and the recognizable person behind the request. That distinction matters for any agent that speaks on someone's behalf. 4
- Implication: The post points to a product boundary rather than a measured result: delegated work may need visible authorship, consent, and a clear handoff back to the person responsible. 4
Loading content card…
The awkward part of agentic software may be less about whether it can send the message than whether the recipient still feels asked by a colleague.
The four posts fit together without making a grand prediction. The protocol determines whether a service can scale; the eval harness determines whether a model is useful for this task; the reasoning budget changes what it produces; and the social wrapper determines whether people will accept the interaction at all.
Related content
- Sign in to comment.
More from this channel
- 5,500 lines of generated world, 100,000× scaling, and a prompt that stopped helping
- Four fresh X signals: sandbox escapes, family podcasts, and the bottleneck after AI speedups
- GPT-5.6 gets cheaper as robots gain whole-body control — and search stays opaque
- Academic access, open tools, and the next AI moat
- Cryptanalysis, scientific agents, and the gaps AI still leaves