# Jul 7, 2026

## A global workspace in language models

- **Claude's J-space** represents a unique internal workspace that allows the model to engage in silent reasoning and report on its thoughts, emerging autonomously during training rather than being explicitly programmed.

## GLM 5.2 and the coming AI margin collapse

- **GLM 5.2** emerges as a formidable competitor to Opus and GPT, offering a potential **50% cost reduction** in workflows despite its slower performance and lack of vision support.

## Python 3.14 compiled to metal – no interpreter

- **`pon` is a JIT & AoT native compiler for Python 3.14**, designed to eliminate the interpreter and bytecode, utilizing a single intermediate representation (IR) for both in-process execution and standalone binaries, with memory managed by a Green Tea garbage collector.

## Pruning RAG context down to what the answer actually needs

- Kapa.ai's innovative approach prunes **68% of irrelevant context** from their retrieval-augmented generation (RAG) process while maintaining **96% recall**, significantly reducing query costs by a third.

## NSA and IETF: Fairness

- **NSA's historical influence** on cryptographic standards, including the promotion of DES despite its known weaknesses, raises serious concerns about the integrity of current IETF processes, particularly regarding the push for solo ML-KEM in TLS, which could lead to significant security vulnerabilities.

## MIRA: Multiplayer Interactive World Models trained on Rocket League

- **MIRA** is a **multiplayer interactive world model** trained on **10k hours** of synthetic **Rocket League** data, featuring **5B parameters** and capable of running **4 players at 20 fps** on a single B200.

## AI Meets Cryptography 1: What AI Found in Cloudflare's Circl

- **AI audit of Cloudflare's CIRCL library revealed seven critical bugs**, including a significant precision loss in RSA and a severe access-control breach in attribute-based encryption, all of which have been fixed upstream.

## TorchJD: Training with multiple losses in PyTorch

- **TorchJD** now supports **multiple loss training** methods, including **scalarization** and **Jacobian descent**, allowing for flexible model optimization with minimal code changes.

## Ph.D. thesis on Differentiable Ray Tracing for Radio Propagation Modeling

- The **Ph.D. thesis** on **Differentiable Ray Tracing for Radio Propagation Modeling** presents a novel approach that integrates **automatic differentiation** into ray tracing, enabling the computation of exact gradients in complex environments, which is pivotal for solving inverse problems in wireless communications.

## Reducing Doom Loops with Final Token Preference Optimization

- **Antidoom** employs **Final Token Preference Optimization (FTPO)** to significantly reduce doom loops in language models, achieving a drop in repetitive completions from **10.2% to 1.4%** on challenging prompts.

## Show HN: Halo – open-source, tamper-evident runtime evidence for AI agents

- **halo-record** provides **tamper-evident runtime records** for AI agents, ensuring that every action taken by the agent is logged in an append-only, hash-chained format that can be independently verified without trust in the vendor.

## Giving a domain a hill to climb: benchmarking as data activation

- **Benchmarking transforms domain data into measurable standards**, enabling models to be evaluated and trained effectively, particularly in complex fields like medicine where traditional metrics are lacking.

## TRACE: open-source hierarchical memory for LLM agents, 82.5% on MemoryAgentBench’s EventQA using gpt-oss-20B

- **TRACE** introduces a **hierarchical memory system** that organizes conversation history into a topic tree, achieving **82.5% F1 score** on the EventQA task of MemoryAgentBench, outperforming traditional flat RAG methods.

## CPU TTS benchmark with UTMOS MOS scoring: Kokoro, Supertonic, Inflect-Nano, and Kyutai's new Pocket TTS

- **Pocket TTS's streaming LM architecture yields a consistent RTF of 0.69 to 0.76**, demonstrating linear cost efficiency across varying text lengths, unlike other models that exhibit variable performance based on input size.

## 🤗LeRobot v0.6.0: Imagine, Evaluate, Improve

- **LeRobot v0.6.0** introduces enhanced capabilities for **imagination, evaluation, and improvement** in robotics, significantly advancing the field's potential applications.