News

Will the HomeView be Apple’s first smart home success?

Apple will announce a smart home hub called HomeView on Tuesday, according to Bloomberg’s Mark Gurman. He describes a Siri-powered device with a display and an A18 chip. The hub is designed to control other smart home devices.

Countertop and wall models anchor a wider device push

Gurman reports that HomeView will come in two models. One is for countertops, and the other is for wall mounting. Updated Apple TV and HomePod mini devices are also scheduled for unveiling, according to his report. He says Apple plans to announce a partnership with LG. That partnership would create a smart lock, thermostat and doorbell. (TechCrunch)

thinkidiot take: The reported two-model lineup gives HomeView a place on either a countertop or a wall. I would choose the mounting option before deciding where to put a screen for home controls. The planned LG lock, thermostat and doorbell make the surrounding devices as interesting to me as the hub. My judgement is that those controls deserve more attention than the A18 chip.

Silicon Valley is falling in line with Trump’s attempt to rename AI

President Donald Trump announced an executive order requiring government agencies to use the term “super intelligence.” Elon Musk promised to carry that wording into his own company’s name. He said SpaceXAI would become SpaceXSI.

The new label reaches research funding and safety language

Marc Benioff announced that Salesforce’s AIForce would become SIForce. Nvidia used the phrase “super intelligence research” in its announced $1 billion commitment to scientific advancement. Satya Nadella repeatedly used “super intelligence” in a post calling for an AI “emergency brake.” Most of the broader population is still using AI, according to The Verge. (The Verge)

thinkidiot take: Government agencies face a required vocabulary change under Trump’s announced order. I would keep calling the technology AI in my own setup, as most people still do. Musk and Benioff have attached the new wording to promised name changes, while Nadella uses it to discuss an emergency brake. I find that alignment with a political naming choice a poor reason to change technical language.

Can the AI industry persuade data center opponents by getting rid of NDAs?

Amazon said it will stop using NDAs in data center negotiations with local governments. The change follows Microsoft’s lead and removes Amazon’s use of those agreements from the talks. Amazon announced the decision in a longer blog post arguing that data centers benefit communities.

Short moratoriums meet construction timelines measured in years

On TechCrunch’s Equity podcast, Rebecca Bellan questioned the lasting effect of data center moratoriums lasting one or two years when data centers can take two to five years to come online. Sean O’Kane described how some negotiations disclose the project type but withhold the identity of the facility’s user or occupant. (TechCrunch)

thinkidiot take: Amazon is dropping NDAs from its data center negotiations with local governments. I would still ask who will occupy the facility, given O’Kane’s account of talks that disclose only the project type. Knowing that a proposal concerns a data center does not answer that question. I judge the change by whether residents get that answer, rather than by Amazon’s argument that these facilities benefit communities.

AI agent teams waste massive tokens for barely measurable quality gains, research finds

Vals AI tested GPT-6 Sol and Claude Opus 5.5 working alone and in teams on Vibe Code Bench at medium and maximum reasoning effort. Teams cost 1.8 to 5.1 times as much. Only one of four comparisons showed a statistically significant improvement: GPT-6 Sol at medium reasoning scored 7.3 points higher.

Faster completion is a separate payoff from better answers

Anthropic’s Opus 5.5 tests showed only slight score gains after 24 hours when teams grew from 10 to 100 agents. Fable 5.1 gained more on Lean theorem proving above 10 agents but stayed below Opus 5.5 across all tests. Its knowledge base score fell when the team grew from 30 to 100 agents. OpenAI researcher Noam Brown said four agents solved tasks twice as fast at twice the cost. He reported slightly lower efficiency at 16 agents. Brown said web research and math parallelize well, while novel writing does not. (The Decoder)

thinkidiot take: Teams cost up to 5.1 times as much in the Vals AI tests, with a statistically significant gain in only one of four comparisons. I would start with a solo agent and require a measured benefit before paying for a team. Brown’s result gives me a concrete reason to spend more when completion time matters: four agents finished twice as fast at twice the cost. Agent teams belong behind a deliberate spending decision, not in my default setup.

A diary of running local AI on one Mac Studio

The Verge’s author installed the open-source Hermes Agent on an M5 Ultra Mac Studio with 256GB of unified memory. They selected Qwen 3.8 Flash Next, a 125-billion-parameter model approximately 105GB in size. The setup let the author explore local AI after concerns about giving personal data to cloud services had held them back.

A phone connection turns the local setup into a daily routine

Hermes is a self-hosted AI agent desktop app for macOS, Windows and Linux. It is free to use with local LLMs. The author made Hermes and Qwen controllable from a phone through a Telegram bot. They also set up a daily morning briefing. The article describes an imminent RTX Spark Windows machine lineup aimed at agentic AI, with up to 128GB of RAM. (The Verge)

thinkidiot take: The author ran an approximately 105GB model on a Mac Studio with 256GB of unified memory. I would start by testing the morning briefing, because it gives the setup a concrete daily job. Hermes being free with local models does not make that model’s memory footprint any smaller. For my own setup, a useful recurring task is a stronger reason to run local AI than the chance to load a large model.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-10-12.

ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents

Building a program starts before anyone writes code: someone must work out what it should do and how to organize it. ProgramGym gives coding agents practice at this by asking them to recreate programs they can investigate. It pairs a way to make checkable practice tasks with a workflow that helps agents carry a project through analysis, coding and review.

• Problem: There are few training setups where agents must discover a program's behavior, choose a design and have their work checked automatically. Existing workflows also struggle with lengthy projects that require studying a working program before writing its replacement.
• New idea: ProgramGym joins a task builder called Program-Env with an agent workflow called Program-Agent. Program-Env turns code repositories into practice tasks, using feedback to repair isolated execution setups and tests that check behavior without inspecting the implementation. Program-Agent separates investigation, coding and review into distinct roles. It also manages code revisions to check whether programs compile and help recover from failures.
• Simple example: Think of recreating a kitchen appliance by trying its controls and observing what happens. The agent must work out the behavior, build a replacement and check that the same inputs produce the expected results.
• Evidence: The pipeline produced 1,648 training environments covering C, C++, Go and Rust. On ProgramBench, fine-tuning Qwen3.8-27B with training runs from these environments raised performance from 34.87% to 42.84%. Program-Agent brought the base model to 55.71% and the fine-tuned model to 56.25%, exceeding some flagship models.
• Limitation: The abstract reports results on ProgramBench but does not establish whether the gains carry over to software projects beyond program reproduction.
• Why it matters: This gives coding agents a way to practice discovering what to build as well as writing the code.
• Paper: ProgramGym: Scaling Training Environments for Long-Horizon

Here the World in Stereo: Learning Dynamic Spatial Correspondence for Immersive Joint Video-Audio Generation

A moving object should sound as though it is moving too. StereoBind aims to make that happen when a model creates video and audio together. It uses the paths of visible sound sources to guide where their sounds seem to come from.

• Problem: Current models can match sounds to video content and get their timing right, but often overlook stereo sound. Virtual reality and interactive games also need a sound's perceived location to follow the object producing it.
• New idea: StereoBind uses motion tracks, records of where visible sound sources move, to guide stereo audio, which uses left and right sound channels. One component connects each visual source to its sound. Another represents source positions, while a third represents relative movement. The researchers also built StereoWorld-29K, a video and stereo audio dataset with matching motion tracks, and StereoWorldBench, a test of whether sound locations agree with visual movement.
• Simple example: Imagine a car crossing the screen from left to right. Its engine should seem to travel with it, rather than stay in one place while the picture moves.
• Evidence: The abstract reports substantially better agreement between visual movement and stereo sound location than existing models, while maintaining overall video and audio quality. It gives no numerical results.
• Limitation: The abstract gives no numerical scores or details about the comparison models, so it does not show how large or consistent the improvement is.
• Why it matters: Sound that follows visible movement matters for making generated scenes feel convincing in virtual reality and games.
• Paper: Here the World in Stereo: Learning Dynamic Spatial

SoloQ: Calibration-Free Quantization for Diffusion Language Models

Some language models produce text by repeatedly refining a sequence instead of writing straight through from beginning to end. That repeated work can demand substantial computing power and memory. SoloQ shrinks the numerical representations these models use, aiming to make them cheaper to run without first tuning the compression on sample data.

• Problem: Diffusion language models repeatedly refine text through costly calculations, and versions that work in blocks also store intermediate information that consumes memory. Existing methods for reducing numerical precision rely on sample data for calibration, even though the models' internal values change across refinement steps and patterns of hidden text.
• New idea: SoloQ uses quantization, which represents numbers with fewer bits, without calibration data, the samples normally used to choose how that conversion works. It rotates and normalizes model weights, the learned parameters, and activations, the values computed during processing, to give them predictable numerical distributions. A small scale correction helps support this conversion. For models that refine text in blocks, it also compresses the KV cache, stored intermediate information used in later calculations, when a block is committed, leaving the block still being refined untouched.
• Simple example: Think of rounding measurements before storing them in a smaller notebook. Putting them on a predictable scale first helps choose a rounding scheme without inspecting a sample of entries.
• Evidence: Across LLaDA, Dream, Fast-dLLM v2 and Nemotron-Labs-Diffusion, SoloQ retained accuracy with 4-bit quantization and outperformed calibration-based methods on knowledge and reasoning benchmarks. Using the hardware-supported NVFP4 number format, it reduced peak memory by up to a factor of 2.61 and sped up end-to-end inference by up to a factor of 2.24.
• Limitation: The abstract gives no individual accuracy scores or hardware details, so it does not establish how broadly the maximum memory savings and speed gains apply.
• Why it matters: SoloQ offers a way to run diffusion language models with less memory and computation without preparing calibration data.
• Paper: SoloQ: Calibration-Free Quantization for Diffusion Language

Trending AI Repositories

Ranking source: GitHub Trending.

lyogavin/airllm

AirLLM reduces the memory needed for large language model inference. It does this without quantization, distillation or pruning.

• What it is: The README says AirLLM can run a 70-billion-parameter model on a single 4GB GPU. It includes a quickstart, configuration guidance and a macOS section.
• What it does: AirLLM 70B inference with single 4GB GPU
• Who it helps: AirLLM helps people who have limited GPU memory run large language models. They can run inference without quantizing, distilling or pruning the model.
• Limitation: The supplied README excerpt gives no inference speed measurements.
• Repository: lyogavin/airllm