from the get-out-of-lawsuit-free-card dept

Qualified immunity — crafted out of thin air by the US Supreme Court — has rarely been anything but an easy way for government employees to duck out of lawsuits before they’re actually asked to defend themselves against allegations of rights violations.

The Supreme Court has continually narrowed this doctrine, pretty much ensuring that if every single fact of an allegation doesn’t perfectly align with precedential rulings, qualified immunity will be awarded. The Supreme Court has ensured no further movement will take place by continually refusing to establish rights violations, even when it (very rarely!) disagrees with a lower court’s granting of qualified immunity.

The doctrine has been memorably pilloried more than once by appellate judges. Most famously, Judge Don Willett of the Fifth Circuit Appeals Court had this to say about the qualified immunity doctrine — something tends to reward rights violators just because they happened to find a slightly different way to violate someone’s rights.

To some observers, qualified immunity smacks of unqualified impunity, letting public officials duck consequences for bad behavior—no matter how palpably unreasonable—as long as they were the first to behave badly. 

That was the wind-up. Here’s the pitch:

Section 1983 meets Catch-22. Plaintiffs must produce precedent even as fewer courts are producing precedent. Important constitutional questions go unanswered precisely because those questions are yet unanswered. Courts then rely on that judicial silence to conclude there’s no equivalent case on the books. No precedent = no clearly established law = no liability. An Escherian Stairwell. Heads defendants win, tails plaintiffs lose.

Justice Sotomayor’s dissent [PDF] isn’t as immediately quotable, but it still delivers a stinging indictment of the qualified immunity doctrine. The facts of the case are unpleasant, as they almost always are when government defendants start invoking qualified immunity.

Green Bay, Wisconsin jail staff responded to prisoner Antonio Smith’s refusal to submit to a wellness check (on day 46 of his hunger strike) by pepper spraying him in the face, ordering him to strip naked, and taking him to the health unit. When Smith refused the wellness check, he was dumped clothed in nothing but a small towel into an unheated, unfurnished “control cell” for the next 23 hours. The temperature in the cell ranged from “25 to 57 degrees Farenheit,” according to uncontested testimony.

When Smith was first placed in the cell around noon, Van Lanen told Smith that Smith could request a shower any time and that he would come back to discuss “‘clothing and stuff,’” but he never returned. Ibid. Three and a half hours later, Smith requested clothing, bedding, and a mattress from Lieutenant Timothy Retzlaff and asked to be moved to a warmer cell given the cold. Retzlaff said he would check with Van Lanen. Twelve additional hours went by with no word from Van Lanen or Retzlaff. Then, around 3 o’clock in the morning, a different officer told Smith that if he submitted to future wellness checks, he could have a smock, but that otherwise, “he would remain naked and cold.” Ibid. Smith declined. Another eight hours came and went without any word from Van Lanen or Retzlaff. Smith remained naked and frigid overnight as the temperature dropped below freezing to 25 degrees. After 23 hours, prison staff removed Smith from the cell. Smith later stated that he stayed on his feet for most of those 23 hours because it was too painful to sit, lie down, or sleep.

The Seventh Circuit Appeals Court actually said exactly this in its ruling granting qualified immunity to the defendants.

The Seventh Circuit held that the officers violated Smith’s Eighth Amendment right to be free from cruel and unusual punishment but nevertheless granted them qualified immunity, reasoning that the Circuit “had never held it unconstitutional on closely analogous facts to house an inmate in a cell that ranged in temperature from 25 to 57 degrees over a 23-hour period without clothes or a way to keep warm.”

Yep, that’s how fucking insane this doctrine is. The court even said this was a rights violation, but since it hadn’t said the same thing earlier about a nearly exactly matching set of circumstances, the defendants apparently had no way of knowing tossing someone naked in a freezing cell for nearly 24 hours would violate the prisoner’s rights.

As Sotomayor points out, the Seventh Circuit appeared to willfully disregard its own precedent when handing down this ruling.

As Judge Hamilton explained in dissent, the Seventh Circuit has itself held that intentionally subjecting prisoners to extreme cold conditions without any way to stay warm violates the Eighth Amendment. In Gillis v. Litscher (2006), for example, the Circuit held that a reasonable jury could find that prison officials violated a prisoner’s Eighth Amendment right when they deliberately left him naked in a cell blowing cool air for five days as part of an effort to “conform [his conduct] to the rules.” [S]ee Del Raine v. Williford,(1994) (officers deliberately strip-searched prisoner in cell for 15 to 30 minutes when windchill was 40 to 50 degrees below zero). The Seventh Circuit has also held that, when cold conditions are the product of heating-system failures, officers violate the Eighth Amendment if they are aware of such conditions and fail to take corrective measures such as providing an alternative way to keep warm.

That should have been enough for SCOTUS to review this one and, hopefully, send it back with a reminder that QI readings need to be narrow, but perhaps not so narrow they provoke gasps of disbelief.

But that’s not how this Supreme Court majority operates. Sotomayor calls them out for only reviewing certain QI cases. You know the ones.

This Term… the Court has exercised its discretion to summarily reverse supposed errors that were far less clear than the one here. See, e.g., McCarthy v. Hernandez, 607 U. S. _ (2026) (per curiam); Zorn v. Linton, 607 U. S. (2026) (per curiam); see also Smith v. Scott, 608 U. S. __ (2026) (summarily vacating and remanding denial of qualified-immunity in light of Zorn). If those cases were clear enough for summary action, the Court here should have readily concluded, based on precedent and basic human decency, that it is beyond debate that it is cruel and unusual to lock someone intentionally in a freezing prison cell completely naked for 23 hours.

The Court’s decision not to do so today exacerbates its asymmetrical trend of declining to intervene when courts wrongly afford officers the benefit of qualified immunity, but unflinchingly summarily reversing when it believes courts have wrongly denied officers the protection of qualified immunity.

Advertisement

This would be hypocrisy if it were being carried out by people who actually maintained a pretense of judicial fairness. But it’s being carried out by people who actively believe in the message they’re sending to the public, as well as to the administration they are so clearly devoted to pleasing.

Reversing only denials of qualified immunity sends the regrettable message that, when choosing between shielding government officials from liability and vindicating individuals’ constitutional rights, this Court will almost always choose the former.

Sotomayor is right. The message being sent is “regrettable.” Unfortunately for America, the people sending it have no regrets at all.

Filed Under: 7th circuit, 8th amendment, police misconduct, qualified immunity, rights violations, sonia sotomayor, supreme court

As enterprise AI systems scale to handle complex workflows, practitioners face the challenge of routing subtasks to the right tools and skills. Agents can have hundreds of tools and skills and get confused on which one to use for each step of a workflow.

To address this challenge, researchers at Alibaba developed SkillWeaver, a framework that creates an execution graph for a given task and chooses the right skills for each of the nodes. They also introduce Skill-Aware Decomposition (SAD), a novel technique that uses a feedback loop to enable the agent to fetch and vet relevant tool candidates iteratively. This compositional approach and feedback loop mechanism distinguishes SkillWeaver from other tool-routing frameworks that choose tools in a one-shot fashion. 

SkillWeaver relates to real-world AI applications where agents autonomously orchestrate multi-tool ecosystems, such as the Model Context Protocol (MCP), to execute multi-step business operations like downloading datasets, transforming information, and creating visual reports. 

In practice, the researchers’ experiments with SkillWeaver show that implementing this retrieve-and-route approach significantly increases accuracy while reducing token consumption by over 99% compared to naively exposing agents to an entire tool library.

For practitioners building AI agents, the main takeaway is that the granularity of task decomposition is the biggest bottleneck to accurate tool retrieval. 

The challenge of skill routing

Skills are a key pattern in modern LLM agent architectures. A skill is a modular, reusable tool specification that uses structured natural language documentation. 

As enterprise agents integrate with massive tool ecosystems, accurately routing user queries to the right skills becomes a difficult task. Exposing an entire library to an LLM to find the right tool is highly inefficient, quickly overwhelms context limits, and consumes hundreds of thousands of tokens.

Most current tool-use frameworks attempt to solve this through API retrieval, documentation matching, or hierarchical structures that treat routing strictly as a single-skill selection or per-step problem. 

However, this single-skill paradigm is insufficient for enterprise environments because real-world queries are inherently compositional. A standard business request such as “Download the dataset, transform it, and create visual reports” cannot be fulfilled by one tool. It requires breaking the prompt down and sequencing an API client, a data processor, and a visualization tool into a cohesive, multi-step execution plan.

How SkillWeaver and SAD work

To tackle this, the researchers frame the problem of handling complex tasks that require multiple skills as “compositional skill routing.” Given a complex user prompt and a vast library of tools, an agent must simultaneously figure out how to break the request into a sequence of atomic sub-tasks, how to map each sub-task to the single best available skill, and how to compose those skills into an executable plan.

SkillWeaver orchestrates this process through three distinct stages: Decompose, Retrieve, and Compose. In the first stage, an LLM acts as a task decomposer, breaking the user’s complex query down into a sequence of sub-tasks that each require one skill. Once the sub-tasks are clearly defined, the system uses an embedding model to compare each subtask against the skill library to pull a shortlist of the top candidate tools for each step. 

In the final stage, a planner evaluates the retrieved candidates based on how well they work together. It checks for inter-skill compatibility to ensure the outputs of one tool naturally flow into the inputs of the next. It then creates a final execution plan as a Directed Acyclic Graph (DAG) that maps out dependencies so independent tasks can potentially execute in parallel.

For example, consider a user asking an AI agent to “Download the dataset, transform it, and create visual reports.” In the decompose stage, the decomposer LLM breaks this into three distinct sub-tasks: downloading the dataset, transforming the data, and creating the reports. 

In the retrieve stage, the system searches the library and finds candidates like “api-client” or “http-fetch” for task one, “csv-parser” or “etl-pipeline” for task two, and so on. Finally, the compose stage evaluates these options, selects the specific combination of “api-client,” “csv-parser,” and “chart-gen” that are most compatible, and wires them together into a final, ready-to-execute workflow.

A key challenge of this pipeline is that LLMs often produce generic step descriptions that fail to match the specific, technical vocabulary of the actual skills available in the library. To fix this, SkillWeaver introduces Iterative Skill-Aware Decomposition (SAD), a novel feedback loop. SAD works by having the LLM draft an initial plan, conducting a preliminary search to find loosely matching skills, and then feeding those retrieved skills back into the LLM as hints. This allows the LLM to rewrite its decomposition so the granularity and vocabulary perfectly align with the actual tools that exist.

SkillWeaver in action

To evaluate how SkillWeaver performs in realistic enterprise scenarios, the researchers created a custom benchmark called CompSkillBench. It consists of 300 multi-step queries of different difficulty levels. To mirror real-world environments, they used a library of 2,209 real-world skills sourced from the public MCP ecosystem, covering 24 functional categories like cloud infrastructure, finance, and databases. 

For the core engine, the researchers primarily used a lightweight 7-billion parameter model (Qwen2.5-7B-Instruct) for task decomposition, paired with a standard semantic search retriever (MiniLM with a FAISS index) to find the tools. SkillWeaver was evaluated against three main setups: a brute-force “LLM-Direct” method where they stuffed all the tool names into the prompt of a large model, a vanilla LLM-based decomposition without SAD, and a ReAct-style agent loop.

The experiments indicate that task decomposition is the main bottleneck. Standard LLM behavior falls short when dealing with large tool libraries, but the SAD feedback loop dramatically moves the needle. In the vanilla setup, the 7B model achieved a decomposition accuracy (i.e., predicting the correct number of steps) only 51.0% of the time. By activating the SAD feedback loop, accuracy jumped to 67.7% (with the larger Qwen-Max model, the accuracy reached 92%). On “hard” tasks requiring four to five distinct skills, SAD improved accuracy by 50%.

One fascinating finding was that larger models can actually perform worse when unguided. When tested in the vanilla setup, a larger 14-billion parameter model saw its accuracy plummet below the 7B model’s accuracy because it tended to over-decompose tasks into microscopic, unnecessary steps. Once SAD was introduced, the retrieved tool hints anchored the model back to reality and increased its accuracy. This suggests that aligning an agent with the vocabulary of specific tools is often more impactful than paying for a larger, more expensive LLM.

Another important takeaway is token savings. The LLM-Direct baseline, which used the very large Qwen-Max model, showed that feeding all tools into the prompt of a large model fails. Despite near-perfect task breakdown capabilities, the massive model only retrieved the right tool category 21.1% of the time when flooded with tool options. SkillWeaver’s targeted retrieve-and-route approach vastly outperformed this in accuracy while slashing context window consumption from an estimated 884,000 tokens down to roughly 1,160 tokens per query, a 99.9% reduction. For practitioners, this translates directly to drastically lower API costs and faster response times. 

Finally, the traditional ReAct baseline completely failed, achieving 0% decomposition accuracy. Its loop naturally collapses multi-step plans into isolated actions rather than explicitly mapping out a cohesive, multi-tool sequence.

Considerations for developers

While the researchers have not yet released the source code for SkillWeaver, their work was built on off-the-shelf tools that can easily be reproduced. 

Skill-Aware Decomposition (SAD), which is the key innovation at the heart of the framework, is a clever prompt-engineering and retrieval loop. The authors have shared the prompt templates in their paper, and developers can implement it themselves quite easily using standard orchestration libraries like LangChain, LlamaIndex, or even raw Python scripts.

As for the retrieval component, the authors built the core framework using all-MiniLM-L6-v2, an open-source embedding model. They found that swapping in a slightly stronger off-the-shelf encoder (BGE-base-en-v1.5) immediately boosted accuracy without any fine-tuning. While an off-the-shelf bi-encoder is great at getting a relevant tool into the top 10 candidates nearly 70% of the time, it struggles to consistently rank the perfect tool at exactly number one, achieving that only about 37% of the time. To bridge this gap, teams will likely need to implement a secondary cross-encoder or LLM-based reranker to re-order those top 10 candidates.

One upfront preparation requirement is vectorizing the tool library and building a FAISS index in advance. In practice, this is a negligible hurdle. Embedding and indexing all 2,209 skills in the benchmark took a mere 15 seconds. Once built, retrieving tools from the index adds less than 15 milliseconds of latency per query. For enterprise environments, syncing the tool index is a trivial background job. 

A current limitation in SkillWeaver is the lack of error recovery. While SkillWeaver successfully maps out a compatible DAG for execution, the authors’ pilot study revealed the challenges of multi-step tool chains. For example, if an API call fails in step two, the entire chain breaks. The paper’s core contribution is limited to the routing and planning phase. For a true production deployment, practitioners must build their own error recovery, fallback, and retry mechanisms on top of the compose stage to handle real-world API timeouts or malformed outputs.

• Six new F-35Bs entered service carrying ballast inside their noses instead of radar
• The APG-85 delay created stealth fighters without their primary sensor
• Lot 17 redesign decisions eliminated compatibility with older radar hardware

The United States Marine Corps has accepted delivery of six newly built F-35B stealth fighters carrying ballast weights where a radar should be installed.

The aircraft left production lines without the AN/APG-85 radar that future F-35 variants are expected to rely upon for combat operations.

I’ve said it before, and I’ll say it again: I have tested a vast number of desks. In fact, I would even be able to venture to say that I’m getting close to the arena of saying I’ve tested most of the key desks in the market.

Uplift has been stepping up their game big time. When I saw the Parsons and realized it came in a 48-inch-wide silhouette with those beautiful four legs and a sleek frame, I knew I needed to get my hands on it as soon as possible.

from the we-are-now-fully-enshittified dept

In the wake of the Sprint T-Mobile merger, wireless carriers immediately stopped trying to compete on price (exactly what deal critics had warned would happen when you reduce sector competition). T-Mobile, which once tried to differentiate itself as the consumer-friendly “uncarrier,” almost immediately began behaving just like AT&T and Verizon, starting with firing 9,000+ people.

It’s how mindless and harmful consolidation always works. We know this, there’s endless evidence of this, and somehow it never seems to matter in a country too corrupt to function.

In the last few years, T-Mobile’s been facing lawsuits and consumer blowback because it’s constantly jacking up the price for customers who believed they were under a “price lock” guarantee thanks to a 7-year-old promotion promising that their price would never change.

More recently, T-Mobile announced it would be kicking roughly 8 million subscribers off of their traditional (and often cheaper plans), and onto more expensive and shittier new T-Mobile plans. These new price hikes have joined a bunch of other price hikes to make everybody’s bills significantly more expensive and all of their connections less feature rich and useful:

T-Mobile frames the current migration as an average $4-per-line adjustment, according to CNET. That sounds modest until you stack it on the $5-per-line hike that already hit many legacy smartphone plans back in April 2025. PhoneArena reports some customers on older grandfathered plans face total increases approaching 60% compared to their original rates. Meanwhile, administrative fees for voice lines climbed from $3.99 to $4.49 per month — raised twice within a single year, according to tmo.report — with mobile internet line fees moving from $1.60 to $2.10.

This must be more of that deregulatory, consolidative innovation my Libertarian friends at “non profit” “free market” “think tanks” have spent years telling me about.

This was, of course, something merger critics warned about, very vocally, for a long time. I wrote repeatedly, at multiple outlets, about how this deal’s pre-merger promises were utterly worthless. It didn’t matter, because the federal government is too corrupt to function in the public interest, antitrust reform no longer exists, and the electorate very clearly has a head full of cottage cheese.

Meanwhile all the folks responsible — whether corrupt politicians, shitty Libertarian free market think tanks, or cocky executives — have long-since moved on to other terrible ideas and memory holed the entire thing, while consumers and labor — as always — are forced to eat all of the real-world costs.

Anyway, remember when T-Mobile bribed Trump to get the merger approved, eliminated all of its “DEI” requirements like an obedient poodle, or that time they hired Corey Lewandowski as a consultant just days after he mocked a Down Syndrome kid on cable TV? Great stuff. So many memories.

Filed Under: competition, consolidation, enshittification, john legere, layoffs, mergers, price hikes, prices, telecom, wireless

Companies: t-mobile

Anthropic is in talks with Samsung Electronics to explore manufacturing a custom AI chip, The Information reported on Thursday. The project remains at an early stage, and Anthropic has not yet decided what the chip would be used for, how powerful it would be, or how it would fit into a server, according to the report. The company could still abandon the effort entirely.

When asked for comment, Anthropic told TechCrunch that a diversified hardware stack including chips from Google, Amazon, and Nvidia will continue to be central to its compute strategy, and said it had nothing further to add on the Samsung discussions. Samsung already plays a significant role in the AI chip supply chain as a major manufacturing partner for Nvidia, producing chips that power AI training and inference workloads. The two companies are also building an AI chip factory together in South Korea.

The talks follow a Reuters report in April that Anthropic was exploring the idea of building its own chips as Claude’s compute demands outpaced available supply. At the time, the effort was described as preliminary, with no dedicated team assembled and no commitment to a specific design. What has changed since April is that Anthropic has hired Clive Chan, who previously helped build OpenAI’s custom chip programme, a signal that the company is moving from exploration to active development.

The timing also coincides with a move by Anthropic’s main competitor. Last week, OpenAI unveiled its first custom chip, a Broadcom-built inference processor it calls the “Intelligence Processor,” designed to reduce the company’s dependence on Nvidia hardware. Amazon and Google both already offer their own custom silicon through their cloud platforms, and Anthropic currently runs Claude across all three chip families.

Anthropic’s annualized revenue run rate surpassed 30 billion dollars earlier this year, more than tripling from roughly nine billion dollars at the end of 2025, a growth rate that makes the economics of custom silicon increasingly attractive. The company signed a long-term deal with Google and Broadcom in April for roughly three and a half gigawatts of TPU compute starting in 2027, but designing its own chips would give it an additional layer of control over the hardware that runs its models. Whether Samsung or another manufacturer ultimately builds a chip for Anthropic remains an open question, but the direction of travel across the industry, away from total reliance on Nvidia, is now unmistakable.