5 topics covered
Signal Blocks Screenshots to Keep AI Agents From Reading Chats
What happened: Signal is restricting screenshots on Android and iOS to stop AI agents such as Gemini and Siri from reading users' messages and other data.
Key details:
- The screenshot block applies in the app on iOS 27 and Android, and is meant to prevent Siri, Gemini, and other Android AI assistants from reading on-screen content.
- Users can restore screenshot capability in Signal's app settings, but Signal warns this could "potentially expose your (and your contacts') Signal content to these AI agents, and break our privacy guarantees."
- Signal president Meredith Whittaker wrote that the remedy is "incomplete," since AI agents can still access app data through alternatives such as the device file system.
- Signal cannot implement these protections on macOS, Whittaker said, because Google, Apple, Microsoft, and others do not provide developers ways to restrict AI data access.
- Signal is calling on Apple, Google, and Microsoft for "simple technical tools and accurate documentation" and says tools are needed to stop third-party agents like Meta's Muse from using accessibility settings as a "skeleton key."
- Signal took a similar approach last year to block Microsoft's Recall feature.
Why it matters: Whittaker argues that OS-level AI agents without developer controls are an existential threat to application-layer privacy, and that the problem extends beyond Signal.
Practical takeaway: If you rely on Signal for sensitive chats, be aware that re-enabling screenshots in its settings lets on-device AI assistants read those messages.
Trump Threatens Tax on AI Data Centers Built Abroad
What happened: President Donald Trump promised a "tremendous tax" on US companies that build data centers outside the United States, even as states and cities restrict new data center projects at home.
Key details:
- Trump made the remarks at a press conference on Friday after his summit with China President Xi Jinping and Russia President Vladimir Putin.
- Quoted remarks: "We're not going to allow our companies to build data centers outside of our country." He also said, "if they do that, they can, but we're going to put a tremendous tax on them."
- Trump warned that if the US allowed data centers to be built abroad, "China will take them all."
- Trump did not share many details, and The Verge reports he appeared to be floating the proposal publicly for the first time.
- Trump was flanked by Vice President JD Vance and Secretary of State Marco Rubio.
- The Verge notes a new Democratic majority Congress will be sworn in within a couple of months, and that states are installing moratoriums and new restrictions on data center proposals.
Why it matters: The threat comes as data center buildout has become unpopular and local restrictions grow, which may push companies to look abroad; The Verge characterizes the threat as an attempt to make that option less attractive.
Practical takeaway: No formal policy or tax structure was announced, so watch for follow-up action from the administration or Congress before assuming any change in rules.
Hugging Face Harness Adapters: RL-Trained LoRAs for Agent Harnesses
What happened: Hugging Face published a recipe for training RL-based harness adapters, which are LoRA adapters trained for Qwen3.8-27B inside specific agent harnesses.
Key details:
- One LoRA adapter was trained per harness: Pi 1.5.2, OpenCode 2.2.7, and Hermes Agent v0.22.6.
- On Terminal-Bench 2.1 with the pi harness, the base Qwen3.8-27B scores 39% and uses on average 2.9K tokens per task; with a custom RL-trained LoRA, pass@1 reaches 55% at about 1.5K tokens.
- The dataset, harness-hub/agentic-coding-envs, holds 871 tasks from 35 repositories, with 171 designated as the qualified subset for GRPO training.
- Training used TRL's AsyncGRPOTrainer, with a reward of the verifier's correctness plus a 0.1 tool-efficiency bonus.
- The three LoRAs were merged into a single super-adapter that runs unchanged under pi, Hermes Agent, and OpenCode.
- Latent Space reports the super-adapter is a rank-96 adapter and raises SWE-Bench Pro hard from 45% to 61% in Hermes and from 47% to 67% in OpenCode.
Why it matters: The post argues that training a model inside the exact harness it will run in can improve both accuracy and token efficiency for that task mix.
Practical takeaway: The adapters, GGUFs, RL tasks, and training recipe are published in a Hugging Face collection if you want to reproduce the approach on your own harness and tasks.
Claimed Proof of the Riemann Hypothesis with Claude
What happened: Anthropic mathematician Levent Alpöge claims he has proved the Riemann hypothesis with the help of Claude, a result that has not yet been independently verified.
Key details:
- The proof was released as Lean code on Alpöge's website rather than as a full paper; Latent Space describes the Lean code as 1.3GB, roughly 250,000 page equivalent.
- Alpöge says the method extends to the Generalized Riemann Hypothesis and he expects, with more time, the Extended Riemann Hypothesis too.
- Elliot Glazer confirmed the formalized statement is GRH and puts the chance of a Lean bug well under 1%, per Latent Space; the Lean code was still compiling at the time.
- Anthropic has not commented on the announcement, per The Verge. In August, Anthropic had said it was working on the problem with Claude but that it did not expect the techniques would prove the hypothesis.
- Alpöge said he texted Anthropic CEO Dario Amodei about the solution "late Sunday night" and held off publication until after the company's IPO.
- Alessandro Masullo separately announced a roughly 940-page, human-written work-in-progress proof of RH, reported by Latent Space.
- The Clay Mathematics Institute requires a proposed solution to be published and withstand at least two years of scrutiny before it can be considered for the $1 million prize.
- Latent Space reports ANT stock fell about 10% on the day.
Why it matters: If verified, this would be another major AI-driven result in mathematics, following OpenAI's September Navier-Stokes claim, which is also unverified. Latent Space notes the AI-written proof is probably correct but hard for humans to read, while Masullo's proof is readable but may be technically incorrect.
Practical takeaway: Treat the claim as unverified until the Lean compilation completes and independent mathematicians review it.
Thinking Machines Releases Inkling 2 Mini Looped MoE Model
What happened: Thinking Machines Lab released Inkling 2 Mini, a small multimodal mixture-of-experts model built for local setups, with Hugging Face publishing its launch details.
Key details:
- The model has ~26B physical total parameters and 3B active parameters per token; it supports a 1M-token context and accepts image and text inputs.
- It uses a looped architecture: the 12 middle of its 24 decoder layers run three times, so it executes 48 layer visits per token.
- Each layer routes each token to 8 experts chosen from 128 (Inkling 2 Small uses 384).
- Shipped in mixed precision with NVFP4 and MXFP8 expert weights, plus a DSpark drafter for speculative decoding.
- Hugging Face says it leads 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL, and is state-of-the-art in its class on the Artificial Analysis Intelligence Index v5.
- Latent Space reports a 20.1 score on the AA Index and an Apache 2.0 license.
- Inference support includes transformers, llama.cpp (GGUF IQ4_NL and Q8_0 quants), MLX, and day-zero SGLang support; Latent Space reports day-zero vLLM support.
- A companion research checkpoint, TML-d24-varloop-research, lets users set the number of loop passes at inference.
Why it matters: The looped design aims to add effective capacity without adding physical layers, and the model is positioned as a local option for coding and knowledge work.
Practical takeaway: Developers with local hardware can run the GGUF, MLX, or transformers builds; Hugging Face notes more validation is needed beyond its vibe tests.