0823 | Anthropic's Apparent Reduced Effort Test for Claude Code, Local LLM IQ & MCP Roadmap

Mia: Welcome back to HackerNews Daily on Bri Radio. I'm Mia, and with me is Milo. We've got a full show for you today, from Anthropic quietly testing new effort levels inside Claude Code, to a look at why your local language model might feel dumber than it actually is.

Milo: And there's plenty more on the roadmap too. We're covering the new MCP roadmap, a tool for running an office of your own clones, plus Meta's alleged hook, hold, harvest, and hide strategy.

Mia: Milo, all that before we even get to hdiutil hitting end-of-life in macOS, a friendly introduction to Racket, and Canada suspending trade talks with the US. And one for a bit of fun — a Belgian car salesman who found out he's actually a prince.

Milo: That story right there is why we read the front page. Stay with us.

Milo: Anthropic appears to be A/B testing reduced effort levels in Claude Code, and the Hacker News thread on it is full of suspicion that the official explanation isn't the whole story.

Mia: One commenter, N_Lens, suspects it's not just the effort-level change, pointing to plenty of optimization around rubberbanding usage limits and routing to a different model in the backend. Their read is that the incentives are too strong for that not to be happening. Another commenter, rrr_oh_man, says they've been using the API themselves — with a plug for their own tool. So the real tension is that Anthropic may officially be testing reduced effort, but the commenters frame it as part of a broader pattern of usage-limit management and model routing that Anthropic hasn't confirmed.

Milo: Separately, there's a Level1Techs forum post asking why your local large language model feels dumber than it is, and commenter jonplackett pushes back on that premise.

Mia: They say they just got the Qwen 3 27 billion parameter build running via MLX on their MacBook Pro and are honestly pretty blown away by how not-dumb it is.

Milo: Prettyblocks then raises a practical catch — the heat. On an M4 Pro, they ask whether jonplackett hits the same issue of these local models running hot.

Mia: The team behind the Model Context Protocol has published a new roadmap, and the headline item is a progressive discovery effort. The idea is that a server can offer a small entry point to a model, then reveal more of its catalog as a conversation narrows in on something specific.

Milo: And that's drawing a bit of a late-to-the-party reaction on Hacker News. One commenter said they've already had to build lazy loading of MCP servers into a couple of harnesses themselves, and now they're moving to implement everything as code instead.

Mia: Right, so the protocol's progressive discovery is arriving after some developers have already worked around the same gap on their own. It lands as a proposed direction on the roadmap, not something shipped yet, which is worth holding onto when you weigh how quickly people adopted it.

Mia: Over on Hacker News there's a project called Munder Difflin, and it's an agent harness built to run an entire office of your own AI clones. It caught a sharp reaction in the comments, with one person calling the whole thing cringe and hoping to see less of that kind of project.

Milo: But that criticism itself drew pushback. Someone asked simply why it's cringe, and another commenter said they'd rather see fewer comments that are just shitting on projects people have no stake in anyway.

Mia: So the back and forth here is essentially about tone, not about the tech itself. One side dismisses the office-of-clones concept as the kind of gimmick they're tired of, and the other side argues that dismissing a project you don't care about adds nothing. Neither comment engages with what the harness actually does.

Mia: The first week of a federal trial over Meta's handling of children's privacy wrapped up, and the Guardian reports prosecutors are framing the case around what they call a four-step strategy: hook, hold, harvest, and hide. The idea is that the company designed its platforms to attract young users, keep them engaged, collect their data, and hide that from regulators and the public.

Milo: That framing made the account less about engineering and more about an alleged deliberate design pattern across a young user's full lifecycle. On Hacker News though, people reacted to accountability rather than the technical claims. One comment simply said more CEOs need to go to prison, and another pointed to China as a place that puts people who damage their society and economy behind bars. So the open question is whether the evidence backs that four-step allegation as testimony continues.

Milo: Switching to the Apple side, macOS 27, codenamed Golden Gate, is deprecating the command-line tool hdiutil, which is what attaches and detaches disk images from the terminal and handles a range of other image-management tasks.

Mia: One Hacker News commenter tied that to Xcode, noting that the xip archive format, deprecated for a long time yet still how Xcode is distributed, relies on these same disk-image tools under the hood. So deprecating hdiutil raises what will handle those formats going forward.

Milo: Separate from the deprecation, another developer complained that errors from these tools don't show in Console.app, the system log viewer, calling it a major annoyance for Cocoa and AppKit apps where terminal usage is secondary. So on top of the change itself, there's a complaint about how visible those failures are when they occur.

Milo: A friendly introduction to Racket landed on Hacker News and the discussion quickly turned to a familiar demand — people asking whether the language has any real applications worth exploring, not just libraries and developer tools. One commenter pointed to a site collecting Racket resources and noted it's mostly libraries and dev tooling.

Mia: Another commenter answered that exact question with a concrete working example — a Racket web application called Remember you can read about or try at remember dot defn dot io.

Milo: So the debate isn't really about whether Racket works as a language, it's about whether it has a visible ecosystem of finished software people can actually open and use — a practical demand for proof of deployment, and Remember is one direct response to it.

Milo: The Canada and United States trade situation escalated this week — the Canadian government announced it is suspending trade negotiations and will match American tariffs dollar for dollar. The thread on Hacker News immediately gravitated to one claim, that Canada has committed to retaliating against every American tariff with an equal tariff of its own. One commenter noted the United States already considers that approach unfathomable, and questioned why a US official keeps citing that only Canada and China have retaliated, asking whether that criticism is actually meant as a compliment.

Mia: The tone turned personal fast, with another commenter pointing to the negotiating style behind the breakdown — described as very good yesterday, very bad today, a frustration that nations are not eager to bargain with what one user called a petulant child. That read frames the escalation less as tariff policy and more as a temperament problem.

Mia: A Belgian car salesman has become a prince after a DNA test proved he's the secret son of a member of the royal family. The story, reported by CNN, is getting a sharp reaction on Hacker News, where commenters are focusing less on the romance of the discovery and more on what it says about inherited privilege.

Milo: Exactly — the top comment puts it bluntly. The point being made is that this is a title, prestige, wealth, and an inheritance, all gained through DNA — something over which no offspring has any control — with no merit, effort, or capability involved in the gain.

Mia: And a reply drives that skepticism home even further, calling it a reflection of, in their words, the good nepotistic society we are. So the discovery itself reads like a fairy tale, but the online read on it is really about unfairness baked into the system.

Mia: We've covered a lot today, from Anthropic testing reduced effort levels in Claude Code to why your local LLM can feel dumber than it actually is.

Milo: Right, those little differences between what the model can do and what it shows on your machine are worth keeping in mind. Thanks for listening — talk to you next time.