Model & Agent Releases
Hydrogen-1.0 research model. Deep Principle released Hydrogen-1.0, a scientific problem-solving model built on GLM-5.3. It focuses on assisting with scientific questions, derivations, and open-ended research problems through text interaction.
Anthropic's FlashProve. Anthropic released FlashProve, an open-source zkSNARK system that produces cryptographic proofs that a response came from a specific model. It can verify identity without revealing weights and proves about 110 tokens/second on a 400B MoE model with one H100.
Siri AI third-party actions. Apple is rolling out Siri AI access to third-party apps in iOS 27.3, allowing the assistant to use personal context, on-screen awareness, and new app skills. It also introduced always-on Audio Intelligence in watchOS 27.3, including Siri Recap and Live Rewind.
GPT-6 Astra designs hardware. OpenAI says GPT-6 Astra designed a device for talking to the Codex coding agent from scratch in under a week, handling enclosure, PCB layout, firmware, and UI. A newer internal model scores 0.875 on BenchCAD 2.0 versus Astra's 0.594, though Astra needed several revision rounds.
Anthropic Message Threads. Anthropic launched Message Threads in public beta, transmitting only changes since the last message instead of resending full conversation history. This cuts time to first token by about 35% for requests between 500K and 1M tokens.
Tools & Frameworks
LangSmith BYOC GA. LangSmith Bring Your Own Cloud is now generally available on AWS and in beta on Azure and Google Cloud. Customers keep traces, evals, datasets, prompts, and deployments in their own cloud account while LangChain manages the control plane.
Optimizing browser agents. LangChain details how it optimized Browser Use for deep agents, using Stagehand tools built for agents and iterating with tracing and evals. They note agentic browser requests jumped 4.7x in September.
Voice agent STT. Soniox released STT v6, a speech-to-text API that reads the agent's own text input and uses it to improve recognition and filter echo. Pricing is $0.12/hour streaming and $0.10/hour async, with no model size disclosed.
Research Highlights
Reading and editing weights. Multiple teams replicated the 'weight-space programming' idea: LLMs like GPT-6 Astra, Qwen 3.8 27B, DeepSeek V4.1 Flash, and Claude Opus 5.5 can interpret raw model weights and even edit a few neurons to fix behaviors. Replications include changing translation behavior in Gemma 4 and fixing Qwen's reasoning loops.
- Like GPT-6 Astra, Qwen 3.8 27B can interpret model weights(opens in a new tab)
- I tried MIT's "weight-space programming" on my own use case: Claude rewrote 8 neurons of Gemma 4 12B to change how it translates one Italian tense(opens in a new tab)
- DeepSeek V4.1 Flash 1017 can also do Weight-Space Programming(opens in a new tab)
- What Qwen is thinking when it gets stuck saying "Wait"(opens in a new tab)
- Qwen 3.8 27B on my laptop can interpret model weights, too(opens in a new tab)
LLM research expiration. A Stanford study tested 50 widely cited claims about LLMs against 20 recent models and found 44% have already expired. Claims about what models can't do expired fastest, with only 16% still holding after two years.
Long-context state tracking. Hugging Face introduced StateBench, a benchmark for testing how models track state changes about entities over long documents. It distinguishes simple latest-update tasks from harder chains of changes, revealing gaps in current models.
Agentic Applications
Agents for drug discovery. Google DeepMind showed Antigravity with Cheminformatics Skills accelerating drug discovery workflows. An agent searched thousands of compounds and proposed novel macrocycles with synthetic routes in an hour, down from days.
AI agents for SETI. SETI's Allen Telescope Array now uses AI agents on Nvidia Holoscan to steer the telescope and analyze signals in real time, processing up to 2.4 Tbps. The team detected a pulsar in the Crab Nebula in real time as a first result.
Open-source robot harness. Tsinghua and UC Berkeley researchers built uni-robot-agent, an open-source harness letting Qwen 3.8 Flash Next control real robots without fine-tuning. It matched closed-model performance on 19 manipulation tasks at $1.09 per task versus GPT-6 Astra's $21.29.
Industry & Business
Palantir enterprise slowdown. Palantir's Q3 revenue grew 96% YoY to $2.31B, but U.S. commercial revenue growth showed signs of slowing, leaving investors lukewarm. The company raised full-year guidance but retail traders sold slightly.
Fal doubles valuation. Inference startup Fal raised $1.3B led by Bessemer at a $16B valuation, double its April figure. It surpassed $800M annualized revenue, though inference companies have thin margins due to compute costs.
Anthropic compute margins. Anthropic's revenue reached $17B in the September quarter, but compute costs rose to 58% of revenue from 48% the prior quarter. The worsening margin is a concern as it approaches IPO.
Revolut cuts coding costs. British fintech Revolut slashed AI coding costs 10x by dropping Claude and moving to self-hosted models, according to a Reddit post. The company now runs its own models for code generation.
Policy & Society
Google AI in schools. Google will disable AI tools for K-12 students by default, but schools that opt out also lose access to Google Search entirely. The company says Search and AI are deeply integrated, forcing administrators to choose all or nothing.
Apple 18+ AI companions. Apple is requiring dating apps and AI companions with sexual or romantic content to carry 18+ ratings starting November 16. The move targets apps like Character.AI and Replika often used by teens.
Nexus Mods AI rules. Nexus Mods will require mod authors to declare whether AI was used in their work and to explain how for AI-assisted mods. The site says it is pro-human creativity but won't ban AI outright.
That's everything for this day – about a 5-minute read.